CN117930815A

CN117930815A - Wind turbine generator remote fault diagnosis method and system based on cloud platform

Info

Publication number: CN117930815A
Application number: CN202410107281.4A
Authority: CN
Inventors: 翁秀奇
Original assignee: Jiangsu Qingmiao Environmental Protection Co ltd; Nanjing Vocational University of Industry Technology NUIT
Current assignee: Nanjing Vocational University of Industry Technology NUIT
Priority date: 2024-01-25
Filing date: 2024-01-25
Publication date: 2024-04-26
Anticipated expiration: 2044-01-25
Also published as: CN117930815B

Abstract

The invention relates to the technical field of remote fault diagnosis, in particular to a wind turbine generator remote fault diagnosis method and system based on a cloud platform, comprising the following steps: based on design parameters and real-time operation data of the wind turbine generator, a data fusion technology is adopted, and the data from the sensor, the operation log and the historical maintenance record are subjected to formatting processing, noise removal and missing value processing to generate a comprehensive data set. According to the invention, through integrating the data fusion technology of data from different sources, the interaction among components can be understood by the graph neural network algorithm, the long-term and short-term memory network is helpful for identifying and predicting the component performance change trend, so that the fault point prediction capability is enhanced, the isolated forest algorithm is excellent in identifying abnormal modes, the abnormal detection accuracy is improved, the Bayesian network provides a new view on the probability inference of fault causes, and the genetic algorithm shows excellent performance in the aspect of system parameter optimization, so that the diagnosis process is more accurate and efficient.

Description

Wind turbine generator remote fault diagnosis method and system based on cloud platform

Technical Field

The invention relates to the technical field of remote fault diagnosis, in particular to a wind turbine generator remote fault diagnosis method and system based on a cloud platform.

Background

Remote fault diagnosis technology is a technical field focused on identifying, analyzing and solving equipment or system faults by remote methods, and is particularly important in the wind power industry, because wind turbines are usually located in remote areas, such as mountaintops or seas, which makes on-site inspection and maintenance difficult and costly. The remote fault diagnosis technology can monitor the running state of the wind turbine generator in real time, and predict and diagnose potential faults by collecting and analyzing data. The technology generally relates to data analysis, machine learning, sensor technology and a communication network, and aims to improve the reliability and efficiency of the wind turbine and reduce maintenance cost.

The remote fault diagnosis method for the wind turbine generator is a method for monitoring and diagnosing faults of the wind turbine generator by using a remote technology. The method aims to improve the operation efficiency and reliability of the wind turbine generator and reduce the dependence on field maintenance personnel. Through implementing remote fault diagnosis, the running condition of the wind turbine generator can be monitored in real time, faults can be rapidly excavated and solved, and therefore the downtime is reduced, and the energy output efficiency is improved. In addition, the method is also beneficial to optimizing maintenance plans and resource allocation, and reducing the overall operation and maintenance cost.

Traditional wind turbine generator fault diagnosis methods mainly rely on empirical judgment and limited data analysis, which results in defects in fault detection and prediction. Due to the lack of efficient data integration capability, conventional methods have difficulty in fully capturing and utilizing multi-source data, limiting the accuracy of the diagnostic process. In addition, conventional methods often have limitations in dealing with complex system dynamics and component interactions, and are unable to effectively identify and predict potential failure points. These limitations result in delays in maintenance reactions, increased operating and maintenance costs, and even the risk of equipment damage or operational interruption due to inaccurate fault diagnosis.

Disclosure of Invention

The invention aims to solve the defects in the prior art, and provides a remote fault diagnosis method and system for a wind turbine generator based on a cloud platform.

In order to achieve the above purpose, the present invention adopts the following technical scheme: a wind turbine generator remote fault diagnosis method based on a cloud platform comprises the following steps:

S1: based on design parameters and real-time operation data of the wind turbine generator, a data fusion technology is adopted, and a comprehensive data set is generated by carrying out formatting processing, noise removal and missing value processing on data from a sensor, an operation log and a history maintenance record;

S2: based on the comprehensive data set, a graph neural network algorithm is adopted, a plurality of components of the wind turbine generator are represented by constructing nodes, dynamic connection and influence among the components are represented by edges, and then the interrelationship among the components is mapped in a graph structure, so that a component interaction relation graph is generated;

S3: based on the component interaction relation map, a long-period memory network is adopted to analyze sequence data of component characteristics changing along with time, potential change trend and periodic mode are identified, and then performance change and fault occurrence points in a future time period are predicted, so that performance trend prediction analysis is generated;

S4: based on the performance trend prediction analysis, constructing a plurality of decision trees by using an isolated forest algorithm, evaluating the path length from each data point to the leaves, distinguishing normal data from abnormal data, and further identifying a key abnormal mode;

S5: analyzing the key abnormal mode through a Bayesian network, carrying out probability inference on the fault cause by utilizing the conditional probability distribution of the network, and simultaneously carrying out prediction on the fault cause and the fault position by referring to historical fault data and real-time observation values to generate fault cause depth analysis;

S6: by utilizing the fault cause deep analysis, adopting a genetic algorithm, performing iterative optimization on parameters of a fault diagnosis system by simulating natural selection and a crossover and mutation mechanism in a genetic process, capturing an optimal parameter combination, and generating diagnosis strategy optimization;

S7: based on the diagnosis strategy optimization, a decision support system algorithm is adopted, real-time operation data of the wind turbine generator and the optimized diagnosis strategy are integrated on a cloud platform, the real-time operation data and the optimized diagnosis strategy comprise continuous monitoring of the real-time data and real-time updating of fault prediction results, and maintenance plans and emergency response measures are formulated according to the fault prediction and historical maintenance data to generate a maintenance decision scheme.

As a further aspect of the present invention, the integrated data set includes temperature, vibration frequency, current intensity of sensor readings, and time stamping, maintenance history, and equipment status records of operation logs, the component interaction relationship map includes action force, action frequency, and action pattern among components, and dependency and probability estimation of interaction among components, the performance trend prediction analysis includes prediction of operation status, identification of potential fault points, and estimation of expected maintenance time in a future period of multiple components, the critical anomaly pattern includes abnormal vibration pattern, target anomaly index of temperature anomaly, and energy consumption deviation, the fault cause depth analysis includes probability estimation of fault occurrence, classification of fault cause, and severity prediction of fault influence, the diagnosis strategy optimization includes adjustment of diagnosis threshold, adjustment of model parameters, and efficiency optimization of diagnosis flow, and the maintenance decision scheme includes prioritization of maintenance tasks, resource allocation scheme, and emergency response scheme.

As a further scheme of the invention, based on design parameters and real-time operation data of the wind turbine, a data fusion technology is adopted, and the data from the sensor, the operation log and the history maintenance record are subjected to formatting treatment and noise and missing value removal treatment, so that the steps for generating the comprehensive data set are specifically as follows:

S101: based on design parameters and real-time operation data of the wind turbine, a Z-score standardization algorithm is adopted, and the data are converted into a format with uniform scale and measurement by calculating deviation of each data point and an average value and dividing the deviation by a standard deviation, so that a standardized data set is generated;

S102: based on the standardized data set, identifying and filtering abnormal values and noise in the data by using a box graph method, determining abnormal values by comparing deviations between data points and quartiles, eliminating error data points of a distortion analysis result, and generating abnormal value filtering data;

S103: filling the missing values in the data set by using a K-nearest neighbor filling method based on the outlier filtering data, and estimating the missing values by analyzing the average value of K nearest data points associated with the missing points to generate missing value processing data;

S104: and processing data based on the missing values, performing deep analysis and fusion on the multi-source data by adopting a characteristic data fusion method, eliminating the difference between the data by unifying the format and the structure of the differential data source, and synthesizing multiple types of characteristic information, including sensor reading, operation log and history maintenance record, to generate a comprehensive data set.

As a further scheme of the invention, based on the comprehensive data set, a graph neural network algorithm is adopted, a plurality of components of the wind turbine generator are represented by constructing nodes, dynamic connection and influence among the components are represented by edges, and then interrelationships among the components are mapped in a graph structure, and the step of generating a component interaction relation graph specifically comprises the following steps:

S201: based on the comprehensive data set, identifying a plurality of components in the wind turbine generator by utilizing an entity identification algorithm, analyzing the characteristics, the operation parameters and the running states of the components, identifying each physical component, establishing independent entities and generating component entity mapping;

S202: based on the component entity mapping, a relation network analysis method is adopted to determine the dynamic connection and the mutual influence among the components, and the functional relation, the signal interaction and the control dependency relation among the components are evaluated to construct the interactive network connection and generate a component interaction network;

s203: based on the component interaction network, converting the components and the interrelationships thereof into a graph structure by using a graph construction method, and simultaneously taking each component as a node and the interaction between the components as an edge to generate a preliminary interaction map;

S204: based on the preliminary interaction map, the nodes and edges in the map are learned by using a map neural network, key features and modes of interaction among the components are identified and extracted, the map is optimized, the relation among the components is reflected, and a component interaction relation map is generated.

As a further scheme of the invention, based on the component interaction relation map, a long-term and short-term memory network is adopted to analyze sequence data of component characteristics changing along with time, and potential change trend and periodic mode are identified, so that performance change and fault occurrence points in a future time period are predicted, and the step of generating performance trend prediction analysis is specifically as follows:

S301: based on the component interaction relation graph, carrying out time sequence analysis by using a statistical method, and generating time sequence characteristic analysis by carrying out trend analysis, periodic detection and relevance evaluation on the operation data of multiple components, wherein the trend analysis comprises calculation of moving average, seasonal decomposition and autocorrelation coefficients of a time sequence, and identification of a time dependent mode of component behaviors;

S302: based on the time sequence characteristic analysis, a long-period memory network model is adopted to deeply learn the identified time dependence mode, a long-period dependence relation is captured through a network layer, a hidden mode and a trend in a time sequence are learned, and an LSTM learning result is generated;

S303: based on the LSTM learning result, predicting performance change and fault points in a future time period, performing prediction analysis by utilizing the output of an LSTM model, and simultaneously evaluating the running state of the component in the future time period, predicting a potential performance degradation area and the occurrence time of faults to generate future performance fault prediction;

S304: based on the future performance fault prediction, performing multidimensional analysis on the running state of the whole wind turbine generator in a future time period, wherein the multidimensional analysis comprises the steps of combining a prediction result with a historical maintenance record, evaluating potential maintenance requirements and providing a risk management strategy, and generating performance trend prediction analysis.

As a further scheme of the invention, based on the performance trend prediction analysis, a plurality of decision trees are constructed by using an isolated forest algorithm, the path length from each data point to the tree leaves is estimated, normal data and abnormal data are distinguished, and the step of identifying a key abnormal mode is specifically as follows:

s401: based on the performance trend prediction analysis, constructing a plurality of independent decision trees by using an isolated forest algorithm, constructing a decision tree for each subset by randomly selecting data subsets and features, and dividing data by the randomly selected features to generate a decision tree forest;

S402: based on the decision tree forest, estimating the path length of each data point reaching a leaf node in the decision tree, and generating path length analysis by calculating the number of splitting steps from a root node to the leaf node and analyzing data isolation risks based on the path distance;

s403: distinguishing normal points and abnormal points in the data set based on the path length analysis, identifying and marking abnormality based on the path length of the data points by comparing the path length with a preset threshold value, mining potential faults or performance problems, and generating an abnormal point detection result;

s404: based on the outlier detection result, analyzing the data points marked as outliers, capturing common characteristics and behavior patterns of the outlier data points, identifying an outlier pattern causing a performance problem or failure, and generating a key outlier pattern analysis.

As a further scheme of the present invention, the key abnormal pattern is analyzed through a bayesian network, probability inference is performed on the fault cause by using the conditional probability distribution of the network, and meanwhile, the prediction of the fault cause and the fault location is performed by referring to the historical fault data and the real-time observation value, and the step of generating the fault cause depth analysis specifically comprises:

S501: based on the key abnormal modes, a Bayesian network construction method is utilized to create a relation between network model representation abnormal modes, nodes in the network are set based on the abnormal modes, edges are defined according to mode relevance, a network is further formed, probability relation between differentiated abnormal modes is reflected, and a Bayesian network structure is generated;

S502: based on the Bayesian network structure, performing conditional probabilistic reasoning, and generating a conditional dependence analysis result by calculating probabilities of occurrence of other modes when a given abnormal mode exists and analyzing and quantifying a dependence relationship between the abnormal modes;

S503: based on the condition-dependent analysis result, by combining the historical fault data and the real-time observation value, deducing a fault cause, and evaluating the similarity between a plurality of abnormal modes and the historical fault case by using a Bayesian network to generate a fault cause reasoning result;

s504: and predicting the fault position and type based on the fault cause reasoning result, analyzing the correlation between the fault cause and a target component in the wind turbine, selecting a part with a problem and potential fault types, and generating fault cause deep analysis.

As a further scheme of the invention, by utilizing the fault cause depth analysis and adopting a genetic algorithm, the parameters of a fault diagnosis system are subjected to iterative optimization by simulating a natural selection and crossing and mutation mechanism in a genetic process, and the optimal parameter combination is captured, so that the diagnosis strategy optimization is generated specifically by the steps of:

S601: based on the fault cause depth analysis, initializing parameters of a fault diagnosis system by adopting a genetic algorithm, randomly generating an initial population from a plurality of parameter combinations by adopting an algorithm simulation natural selection mechanism, wherein each parameter combination represents a solution, and generating the initial parameter population;

S602: based on the initial parameter population, executing crossover operation of a genetic algorithm, and forming a new parameter set by exchanging parameter parts of various solutions to generate a crossly generated parameter combination;

s603: based on the parameter combination generated by the intersection, performing mutation operation of a genetic algorithm, introducing new mutation operation by randomly changing part of parameter values, avoiding the local optimal solution of the algorithm, and generating a mutated parameter combination;

s604: and determining a final optimized parameter combination through the selection operation of a genetic algorithm based on the mutated parameter combination, evaluating the fitness of a plurality of solution sets, and selecting an optimal parameter configuration based on the fitness to generate diagnosis strategy optimization.

As a further scheme of the invention, based on the diagnosis strategy optimization, a decision support system algorithm is adopted, real-time operation data of the wind turbine generator and the optimized diagnosis strategy are integrated on a cloud platform, the method comprises the steps of continuously monitoring the real-time data, updating a fault prediction result in real time, and making a maintenance plan and an emergency response measure according to the fault prediction and historical maintenance data, wherein the step of generating a maintenance decision scheme comprises the following specific steps:

S701: integrating real-time operation data of the wind turbine generator and the diagnosis strategy optimization based on a cloud platform, extracting data from multiple data sources by adopting a data warehouse construction method, unifying data formats, cleaning and removing inconsistent or erroneous data, and loading the processed data into a centralized data warehouse to generate a real-time data set;

s702: based on the real-time data set, adopting time sequence analysis, and generating a fault prediction model output by identifying time correlation modes and trends in the data and combining long-term and short-term memory network algorithms to learn long-term dependency relations among data points so as to predict data trend in a future time period;

S703: based on the fault prediction model output and the historical maintenance data, adopting a multi-criterion decision analysis method, comprehensively referencing various decision factors including cost, risk and benefit, and carrying out weight distribution and comprehensive evaluation on the multi-decision factors by using a hierarchical analysis method so as to generate a maintenance plan scheme;

S704: based on the maintenance planning scheme, a resource optimization and linear programming algorithm is adopted to analyze the resource requirement of the maintenance activity, the required manpower, material and time cost is calculated by establishing a mathematical model of resource allocation and cost, and then the resource allocation is adjusted and optimized to generate a maintenance decision scheme.

The wind turbine remote fault diagnosis system based on the cloud platform is used for executing the wind turbine remote fault diagnosis method based on the cloud platform, and comprises a data integration module, an interactive relation modeling module, a performance trend analysis module, an abnormal mode detection module, a cause analysis and prediction module and a maintenance strategy decision module;

the data integration module extracts multi-source data by adopting an ETL algorithm based on design parameters and real-time operation data of the wind turbine generator, converts the difference of data format matching multi-data sources, loads the multi-source data into a unified data platform, and generates a comprehensive data environment through data cleaning and format standardization;

the interaction relation modeling module is used for constructing a component relation map of the wind turbine generator by using a graph neural network algorithm based on the comprehensive data environment, mapping the components and the relation thereof into nodes and edges in the graph by analyzing the characteristics and interaction of the components, revealing the dynamic interaction among the components and generating a component interaction map;

The performance trend analysis module is used for deep learning of time series data of the components by applying a long-short-term memory network based on the component interaction map, analyzing the variation trend and the periodicity pattern, predicting performance fluctuation and potential fault points in a future time period and generating performance trend prediction;

the abnormal pattern detection module is used for detecting an abnormal pattern by utilizing an isolated forest algorithm based on performance trend prediction, isolating data points by constructing a plurality of decision trees, analyzing the path length of the data points to identify the abnormal pattern, and generating abnormal pattern analysis;

The cause analysis and prediction module is used for deducing the cause of the fault by adopting a Bayesian network algorithm based on abnormal mode analysis, analyzing the association and influence between abnormal modes by using a probability model, and generating a fault deducing result by combining historical fault data;

The maintenance strategy decision module optimizes the maintenance strategy by using a genetic algorithm and multi-criterion decision analysis based on fault inference results, captures an optimal maintenance plan and an emergency response scheme by simulating a natural selection mechanism and cross mutation operation, and establishes a maintenance strategy scheme.

Compared with the prior art, the invention has the advantages and positive effects that:

According to the invention, the data from different sources can be effectively integrated by applying the data fusion technology, so that the integrity and quality of the data are improved. The graph neural network algorithm enables the complex interaction among the components to be deeply understood, and the understanding of fault influence factors is improved. The introduction of the long-term and short-term memory network provides a powerful tool for identifying and predicting the variation trend of the component performance, and enhances the prediction capability of future fault points. The use of the isolated forest algorithm is excellent in identifying abnormal patterns, and improves the accuracy of abnormality detection. In addition, the application of the Bayesian network provides a new view on probability inference of fault reasons, and improves the comprehensiveness of diagnosis. The use of genetic algorithms exhibits efficient performance in system parameter optimization, making the whole diagnostic process more accurate and efficient.

Drawings

FIG. 1 is a schematic workflow diagram of the present invention;

FIG. 2 is a S1 refinement flowchart of the present invention;

FIG. 3 is a S2 refinement flowchart of the present invention;

FIG. 4 is a S3 refinement flowchart of the present invention;

FIG. 5 is a S4 refinement flowchart of the present invention;

FIG. 6 is a S5 refinement flowchart of the present invention;

FIG. 7 is a S6 refinement flowchart of the present invention;

FIG. 8 is a S7 refinement flowchart of the present invention;

Fig. 9 is a system flow diagram of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

In the description of the present invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like indicate orientations or positional relationships based on the orientation or positional relationships shown in the drawings, merely to facilitate describing the present invention and simplify the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and therefore should not be construed as limiting the present invention. Furthermore, in the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

Example 1

Referring to fig. 1, the present invention provides a technical solution: a wind turbine generator remote fault diagnosis method based on a cloud platform comprises the following steps:

s3: based on the component interaction relation graph, a long-period memory network is adopted to analyze sequence data of component characteristics changing along with time, potential change trend and periodic mode are identified, and then performance change and fault occurrence points in a future time period are predicted, so that performance trend prediction analysis is generated;

S4: based on performance trend prediction analysis, constructing a plurality of decision trees by using an isolated forest algorithm, evaluating the path length from each data point to leaves, distinguishing normal data from abnormal data, and further identifying a key abnormal mode;

S5: analyzing the key abnormal mode through a Bayesian network, carrying out probability inference on the fault cause by utilizing the conditional probability distribution of the network, and simultaneously referring to historical fault data and real-time observation values, carrying out prediction on the fault cause and the fault position, and generating fault cause depth analysis;

s6: by utilizing fault cause deep analysis, adopting a genetic algorithm, performing iterative optimization on parameters of a fault diagnosis system by simulating natural selection and a crossover and mutation mechanism in a genetic process, capturing an optimal parameter combination, and generating diagnosis strategy optimization;

S7: based on diagnosis strategy optimization, a decision support system algorithm is adopted, real-time operation data of the wind turbine generator and the optimized diagnosis strategy are integrated on the cloud platform, the real-time operation data comprise continuous monitoring of the real-time data and real-time updating of fault prediction results, and a maintenance plan and emergency response measures are formulated according to the fault prediction and historical maintenance data to generate a maintenance decision scheme.

The comprehensive data set comprises temperature, vibration frequency, current intensity of sensor readings, time marks of operation logs, maintenance history and equipment state records, the component interaction relation map comprises action intensity, action frequency and action mode among components, and dependency and interaction probability estimation among components, the performance trend prediction analysis comprises prediction of running states, identification of potential fault points and estimation of expected maintenance time in a future time period of a plurality of components, the key abnormal mode comprises target abnormal indexes of abnormal vibration modes, temperature abnormality and energy consumption deviation, the fault cause deep analysis comprises probability assessment of fault occurrence, classification of fault cause and severity prediction of fault influence, the diagnosis strategy optimization comprises adjustment of diagnosis threshold values, adjustment of model parameters and efficiency optimization of diagnosis flow, and the maintenance decision scheme comprises priority ordering of maintenance tasks, resource allocation scheme and emergency response scheme.

The accuracy and the efficiency of fault diagnosis are greatly improved, misjudgment and missed judgment are reduced, and problem positioning is accelerated, so that the maintenance and repair efficiency is improved. Sudden faults are reduced through predictive maintenance, maintenance plans are optimized, and equipment life is prolonged. The method also enhances the decision support capability, provides scientific basis for maintenance decision and reduces the operation and maintenance cost. The system can also optimize resource allocation and emergency response based on the forecast and historical data, improving resource utilization efficiency and reducing potential losses. Through continuous parameter optimization and model iteration, the method can be matched with new data and running conditions, and the advancement of diagnosis performance is maintained. In the long term, the method is beneficial to remarkably reducing the overall operation cost of the wind turbine generator and brings remarkable economic benefit. Predictive maintenance reduces energy waste caused by equipment failure, prolongs equipment service life, reduces waste and pollutant generation, and promotes environmental sustainability.

Referring to fig. 2, based on design parameters and real-time operation data of a wind turbine, a data fusion technology is adopted, and the steps of generating a comprehensive data set by formatting, removing noise and missing values from data of a sensor, an operation log and a history maintenance record are specifically as follows:

s102: based on the standardized data set, identifying and filtering abnormal values and noise in the data by using a box graph method, determining the abnormal values by comparing the deviation between the data points and quartiles, eliminating error data points of the distortion analysis result, and generating abnormal value filtering data;

S104: based on the missing value processing data, the characteristic data fusion method is adopted to carry out deep analysis and fusion on the multi-source data, the format and structure of the differential data sources are unified, the difference among the data is eliminated, and the multi-type characteristic information including sensor reading, operation log and history maintenance record is synthesized to generate a comprehensive data set.

In the S101 substep, design parameters and real-time operation data of the wind turbine are processed, and a Z-score normalization algorithm is used. This process first involves calculating the difference between each data point and the average of the overall dataset, and dividing by the standard deviation. In this way, the raw data is converted into standardized data so that data of different magnitudes and units can be compared and analyzed under the same standard.

The specific operation is as follows: first, the average value and standard deviation of each item of data are calculated. For each data point, the average value is subtracted from its value and then divided by the standard deviation. Thus, each value in the dataset is converted to a Z-score representing the relative distance of the data point from the average of the dataset, expressed as a multiple of the standard deviation. The purpose of this step is to eliminate differences between different data sets due to differences in dimension or magnitude, making the data more suitable for subsequent statistical analysis.

In the S102 substep, the normalized data is processed using the bin graph method to identify and filter outliers and noise. The box plot is a graph showing the distribution of data, showing the maximum, minimum, median, and upper and lower quartiles of the data. By comparing the deviation between the data points and the quartile, it can be determined which data points are outliers.

In operation, a bin graph is first drawn and then outliers are determined based on the boundaries of the bin graph. Typically, data points that are beyond the box whisker (i.e., the portion outside the upper and lower quartiles) are considered outliers. These outliers are caused by erroneous data entry, measurement errors, or other anomalies, and need to be removed from the dataset to avoid misleading in subsequent analysis.

In the S103 substep, the missing values in the outlier filtered data are processed using the K-nearest neighbor padding method. The core of this approach is to find K data points that are similar to the missing data points and use the average of these data points to estimate the missing value.

In particular, the value of K (K is a positive integer) is first determined, and then for each missing value, the nearest K data points in the dataset are found. The selection of these data points is determined by a distance metric (e.g., euclidean distance) based on similarity to the location of the missing values. After determining these points, an average is calculated and the missing values are filled with this average. The method effectively estimates the missing values, so that the data set is more complete, and a more reliable data basis is provided for subsequent analysis.

In the sub-step S104, feature level data fusion is performed on the processed data, which involves integrating multi-source data (such as sensor readings, operation logs, and history maintenance records), unifying the format and structure thereof, and eliminating the variability between data. By means of the fusion, information of different data sources can be synthesized, and more comprehensive characteristics can be extracted.

Operationally, the format and structure of each data source is first analyzed, and then appropriate conversion and adjustment methods are adopted to unify the data sources. This involves conversion of data formats, unification of units, alignment of time stamps, etc. After the steps are completed, the data are subjected to deep analysis, and the characteristics of different data sources are fused. For example, sensor data is associated with the oplog to more fully understand the operational state of the aggregate. The end result of this fusion is a comprehensive data set that provides a richer, more accurate basis for further analysis and decision making.

Assume that there is a set of wind turbine data, including design parameters and real-time operational data. These data include wind speed, rotor speed, power generation, etc., from different sensors and recording systems, respectively. These data will be processed through the four sub-steps described above.

In the wind speed data, the readings from the wind speed sensor are in meters per second.

In the rotating speed data of the rotating wheel, the unit of the rotating wheel is the revolution per minute.

In the power generation amount data, the unit of power generation amount per hour is kilowatt-hour.

In S101, data normalization

The original dataset is assumed to include the following values:

wind speed: [3.5,4.8,3.9,...]

Rotating speed of the rotating wheel: [250,300,280,...]

Generating capacity: [500,650,600,...]

The mean and standard deviation of each data were calculated. For example, assume that the average value of wind speed is 4.0 m/s and the standard deviation is 0.5 m/s. For each data point, a Z-score normalization is applied, for example, with a Z-score of (3.5-4.0)/0.5 = -1 (3.5-4.0)/0.5 = -1 for a wind speed of 3.5 meters/second. In this way, the raw data is converted into a normalized dataset, wherein each data point is represented by its deviation from the mean.

In S102, outlier and noise filtering is performed, and based on the normalized data, a box graph is drawn and outliers are identified. It is assumed that 4.8 m/s is recognized as an abnormal value in the wind speed data because the upper limit range of the box graph is exceeded. Such outliers will be removed from the dataset.

In S103, the missing value processing, assuming that there is a missing value in the wheel rotation speed data, selects K value as 2, captures two nearest valid data points (for example, 280 and 300 rpm), calculates the average value of these two points (290 rpm) and fills the missing value.

And S104, feature level data fusion is carried out, and the processed data are fused. For example, there is a certain correlation between the excavation wind speed and the power generation amount. By fusing the data, the operation efficiency of the wind turbine generator is more comprehensively understood. Finally, a comprehensive data set is obtained, wherein the comprehensive data set comprises data after normalization, denoising, missing value processing and feature fusion. The data set provides a more accurate and comprehensive view angle for further analysis, and can be used for optimizing the performance and maintenance strategy of the wind turbine.

Referring to fig. 3, based on a comprehensive data set, a graph neural network algorithm is adopted, a plurality of components of a wind turbine generator are represented by constructing nodes, edges represent dynamic connection and influence among the components, and then interrelationships among the components are mapped in a graph structure, and a component interaction relation graph is generated specifically by the steps of:

S202: based on the component entity mapping, a relation network analysis method is adopted to determine the dynamic connection and the mutual influence among the components, and the functional relation, the signal interaction and the control dependency relation among the components are evaluated to construct the interactive network connection, so as to generate a component interaction network;

S204: based on the preliminary interaction map, the nodes and edges in the map are learned by using the map neural network, key features and modes of interaction among the components are identified and extracted, the map is optimized, the relation among the components is reflected, and a component interaction relation map is generated.

In the S201 substep, a plurality of components in the wind turbine generator are identified through an entity identification algorithm. The process begins with the analysis of a comprehensive data set that includes characteristics, operating parameters, and operating conditions of the components. For example, algorithms use machine learning models, such as random forests or support vector machines, to classify components. These models learn features that identify different components through a training set.

Specifically, the algorithm first receives data from the sensors, such as temperature, pressure, or vibration levels, and then identifies the data based on known component characteristics. For example, if the sensor data shows high vibration levels and temperature rises, the algorithm identifies that this mode is associated with a particular gearbox fault. In this way, a separate entity representation is built for each physical component and a component entity mapping file is generated that includes the identity, state, and associated parameters of each component.

In the sub-step S202, dynamic connections and interactions between components are determined using a relational network analysis method. This process involves the evaluation of functional relationships, signal interactions, and control dependencies between components. For example, algorithms analyze time series in sensor data to identify inter-dependencies among components.

To achieve this, time series analysis and causal relationship modeling methods are used. For example, if a failure of one component results in a degradation of another component, such a relationship may be determined by correlation analysis and causal inference algorithms (e.g., a gland causal relationship analysis). Ultimately, these relationships are mapped into a network where nodes represent components and edges represent interactions with each other. This component interaction network not only shows direct connections between components, but also reveals potential indirect interactions.

In a sub-step S203, the components and their interrelationships are converted into graph structures using a graph construction method. This process involves converting information in a component interaction network into a graph data model, where each component is treated as a node and interactions between components are treated as edges.

In this step, it is critical to ensure that the graph structure accurately reflects the relationships between the components. Thus, the weight and direction of the edges will be set according to the degree and direction of influence between the components. For example, if a failure of one component often results in a performance degradation of another component, the impact relationship may be marked as a directional and higher weight edge. The result of the process is a preliminary interaction map, which lays a foundation for the next graph neural network analysis.

In the step S204, the nodes and edges in the preliminary interaction map are learned by using a Graph Neural Network (GNN). The graphic neural network is particularly suitable for processing graphic structure data, and can effectively identify and extract key characteristics and modes of interaction among components.

In step, the graph neural network learns complex interaction relations among components through the characteristics of nodes and edges. For example, GNNs analyze path lengths between nodes, the number of neighbors of a node, or the weights of edges to understand dependencies and impact forces between components. From these analyses, GNNs can identify key modes of interaction, such as which combinations of components are most prone to failure, or which interactions between components are critical to overall system stability. The finally generated component interaction relation map not only reflects the physical and functional relation among components, but also reveals a deeper dynamic relation among the components.

The integrated dataset is assumed to include the following components and their associated data:

temperature, vibration level in the gearbox (Gearbox).

In the Blade (Blade), the curvature and the rotational speed.

In generators (generators), current, voltage, temperature.

In the Control System (Control System), a response time and a status code are instructed.

In S201, component entity mapping, data characteristics of each component are analyzed using an entity recognition algorithm. For example, a Support Vector Machine (SVM) algorithm is applied to the temperature and vibration level data of the gearbox, and a model is trained from the historical data and fault records. If the vibration level of the gearbox abnormally increases, the model identifies this as a potential fault. After each component is identified, a component entity mapping file is generated listing all components and their current state and operating parameters.

In S202, the component interaction network construction uses time series analysis to determine the dynamic connection between components. For example, if the vibration level of the gearbox increases, the blade rotational speed decreases, which indicates that there is a functional link between the two. Through this analysis, an interaction network between components is constructed, including direct and indirect interactions between the different components.

In S203, the preliminary interaction map is constructed, the interaction network of the components is converted into a graph structure, each component becomes a node, and the interactions between the components become edges. For example, the interaction between the gearbox and the blade appears as one edge, the weight of which reflects the extent of the effect. This diagram structure clearly shows the dependency and interaction pattern between components.

In S204, the component interaction relationship graph is generated using the graph neural network to analyze the preliminary interaction graph. GNNs identify key interaction patterns by learning the characteristics of nodes (components) and edges (interactions). For example, GNN finds that when the gearbox temperature exceeds a certain threshold, the efficiency of the generator drops significantly. From these analyses, GNN optimization maps were generated and final component interaction relationship maps were generated.

Referring to fig. 4, based on a component interaction relationship graph, a long-short term memory network is adopted to analyze sequence data of component characteristics changing along with time, identify potential variation trend and periodicity pattern, and further predict performance variation and failure occurrence points in a future time period, and the step of generating performance trend prediction analysis is specifically as follows:

S301: based on the component interaction relation graph, carrying out time sequence analysis by using a statistical method, and generating time sequence characteristic analysis by carrying out trend analysis, periodic detection and relevance evaluation on operation data of multiple components, wherein the trend analysis comprises calculation of moving average, seasonal decomposition and autocorrelation coefficients of a time sequence, and identification of a time dependent mode of component behaviors;

S302: based on time sequence characteristic analysis, a long-term memory network model is adopted to deeply learn the identified time dependence mode, a long-term dependence relationship is captured through a network layer, a hidden mode and a trend in a time sequence are learned, and an LSTM learning result is generated;

S303: based on LSTM learning results, predicting performance change and fault points in a future time period, performing prediction analysis by utilizing the output of an LSTM model, and simultaneously evaluating the running state of the component in the future time period, predicting potential performance degradation areas and fault occurrence time to generate future performance fault prediction;

s304: based on future performance fault prediction, multidimensional analysis is carried out on the running state of the whole wind turbine generator in a future time period, and the method comprises the steps of combining a prediction result with a historical maintenance record, evaluating potential maintenance requirements and providing a risk management strategy to generate performance trend prediction analysis.

In the step S301, through time series analysis based on the component interaction relationship map, specific operations include deep analysis of component operation data by using a statistical method. First, operational data of a plurality of components, which are time-series data of various physical parameters such as temperature, speed, voltage, etc., are collected. Then, smoothing is performed on the data by a moving average method to reduce the influence of short-term fluctuations on trend analysis. For example, a moving average of 7 or 30 days is calculated to reveal long-term trends in component performance. Next, the periodic patterns of the data are analyzed using seasonal decomposition techniques, such as decomposing out trend, seasonal, and residual components by the STL (Seasonal and Trend decomposition using Loess) method. In addition, the calculation of the autocorrelation coefficients helps to identify self-similarity in the time series, the time-dependent patterns are revealed by calculating the degree of correlation of the data points with their past values, the final yield of this stage is an exhaustive analysis of the time series characteristics of each component, and these analysis results reveal the time-dependent patterns of component behavior, providing a basis for the next deep learning.

In the sub-step S302, a long short term memory network (LSTM) is employed to learn deeply the time dependent patterns in the time series. Based on the analysis result of S301, an LSTM network model is designed, which has a plurality of hidden layers, each layer containing a number of LSTM cells. The structure of the network is optimized by adjusting the number of LSTM units and the number of hidden layers. In the training process, the time sequence data is divided into a series of time windows, and each window contains sequence data with a certain time length as network input. The LSTM network learns long-term dependencies in the time series data through its unique gating mechanism and captures hidden patterns and trends on the basis of this. After training, the LSTM model can predict future trend according to past data, and the main result of the steps is an LSTM learning result, which represents deep understanding of the model on the time dependent mode of the component behavior.

In the S303 substep, predictions of future performance changes and failure points are made based on LSTM learning results. With the trained LSTM model, the most recent time series data is input and the model will output predictions of component performance over a period of time in the future. These predictions include potential performance degradation areas and failure occurrence moments. The prediction results not only reflect the operational state of the component, but can also reveal potential risk points, the main outcome of the steps being the predictive analysis of the operational state of the component over a future period of time, which is critical to the implementation of effective maintenance strategies and risk management.

In the step S304, the future operation state of the whole wind turbine generator is subjected to multidimensional analysis, and in the step, the prediction result obtained in the step S303 is combined with the historical maintenance record so as to comprehensively evaluate the future maintenance requirement of the component. For example, by comparing the predicted performance degradation area to the historical fault record, it can be identified which components require priority maintenance. Meanwhile, by analyzing risk points and historical fault modes in the prediction result, a risk management strategy aiming at a specific component can be developed, the final output of the steps is an exhaustive performance trend prediction analysis report, precious insights are provided for the operation of the wind turbine generator, and future maintenance work is guided.

It is assumed that there is a set of historical operating data of the wind power generator, including time series data of wind speed, temperature, voltage, etc. These data will be processed according to the four sub-steps presented.

In S301, time series characteristic analysis is performed, wind speed (m/S), temperature (°c) and voltage (V) data recorded every hour over the past year are collected, 7-day moving average values are calculated for each time series, short-term fluctuations are smoothed, the time series are decomposed using the STL method, and trends, seasonal and random fluctuations are identified and separated. The autocorrelation coefficients of each time series are calculated, correlations between the data points and their past values are determined, and time series characteristic results for each component are generated, exhibiting trend, periodicity, and time dependent patterns.

In S302, LSTM deep learning designs an LSTM network comprising 3 hidden layers, each layer has 50 units, the time sequence is segmented into windows comprising a week of data, the LSTM model is trained by using historical data as network input, long-term dependency and hidden modes in the time sequence are learned, and a trained LSTM model is obtained, so that the behavior mode of each component can be understood and revealed.

In S303, the performance change and the fault point are predicted, the latest week data is input to the LSTM model, the performance trend of one month in the future is predicted, the predicted result is analyzed, the potential performance degradation area and the fault moment are identified, the predicted result is generated, and the predicted performance and the potential fault point of each component in one month in the future are displayed in detail.

In S304, the multidimensional performance trend prediction analysis combines the prediction result with the history maintenance record, analyzes future maintenance requirements, proposes a targeted risk management strategy according to the prediction and the history fault mode, compiles a comprehensive result and provides detailed performance trend prediction and maintenance suggestions.

In the simulation data and results, assuming that wind speed data shows significant seasonal changes and temperature and voltage data shows steady upward trends, the model identifies the increase in voltage fluctuations at high temperatures, predicts that some units will experience voltage instability during the next high temperature season, and finally reports a maintenance recommendation including a predicted trend graph, a time stamp of potential failure points, and maintenance records based on the predictions and histories.

Referring to fig. 5, based on performance trend prediction analysis, using an isolated forest algorithm, a plurality of decision trees are constructed and path lengths from each data point to leaves are evaluated, normal data and abnormal data are distinguished, and the step of identifying a key abnormal mode is specifically as follows:

s401: based on performance trend prediction analysis, constructing a plurality of independent decision trees by using an isolated forest algorithm, constructing a decision tree for each subset by randomly selecting data subsets and features, and dividing the data by the randomly selected features to generate a decision tree forest;

s402: based on the decision tree forest, estimating the path length of each data point reaching the leaf node in the decision tree, and generating path length analysis by calculating the number of splitting steps from the root node to the leaf node and analyzing the data isolation risk based on the path distance;

S403: distinguishing normal points and abnormal points in the data set based on path length analysis, identifying and marking abnormality based on the path length of the data points by comparing the path length with a preset threshold value, mining potential faults or performance problems, and generating an abnormal point detection result;

S404: based on the outlier detection results, data points marked as outliers are analyzed, common characteristics and behavior patterns of the outlier data points are captured, outlier patterns that lead to performance problems or faults are identified, and key outlier pattern analysis is generated.

In the S401 substep, the performance trend prediction data of the wind turbine generator is processed by using an isolated forest algorithm. Firstly, collecting performance data of the wind turbine, wherein the data comprise various physical parameters such as wind speed, temperature, voltage and the like, and the data are in a time sequence. Next, a plurality of decision trees are constructed using an isolated forest algorithm. The isolated forest algorithm is an unsupervised learning algorithm based on random forests and is specially used for detecting abnormal values. The core is to randomly select subsets and features of data and construct a decision tree for each subset. Each tree segments the data by randomly selected features until each data point is isolated as a separate node, forming a decision tree forest. In the construction process, it is critical to adjust the number and depth of trees. In general, the greater the number of trees, the more stable the result; the depth of the tree is then adjusted according to the complexity of the data. After the step is completed, the generated decision tree forest can represent the distribution condition in the data set, and a basis is provided for the next step of abnormality detection.

In the S402 substep, the path length of each data point to the leaf is estimated based on the decision tree forest in the isolated forest. The path length of each data point refers to the number of splitting steps undergone from the root node to the leaf node. In an isolated forest, outliers are often more easily isolated and therefore their path length is often shorter. Whereas normal data points have a relatively long path length due to their distribution closer to the whole data set. In carrying out the steps, for each tree in the forest, the path length of each data point to the leaf node is calculated, and then the average path length of the corresponding data points on all the trees is taken as the final path length of the point. This process involves a large number of computations, often requiring efficient data structures and algorithms to implement. Through the steps, an analysis result comprising all data points and path lengths thereof can be obtained, and the result provides important quantification basis for distinguishing normal data from abnormal data.

In the step S403, normal points and abnormal points in the data set are distinguished by comparing the path length of the data point with a preset threshold. The core of the isolated forest algorithm is to use the distribution of path lengths to identify anomalies. Outliers can be marked as outliers due to shorter path lengths. A threshold is set and all data points for which the path length is less than this threshold are considered outliers. In practice, the setting of the threshold is typically based on an understanding and a priori knowledge of the dataset. Once the threshold is set, the outlier can be quickly identified by comparing the path lengths of each data point, the result of the step being an outlier detection result that includes the data points marked as outliers. These outliers are indicative of potential faults or performance problems, and have important significance for maintenance and fault prevention of the wind turbine generator.

In the S404 substep, the data points marked as outliers are further analyzed to capture common characteristics and behavior patterns of the outlier data points. The goal of this step is to identify key anomaly patterns that lead to performance problems or failures. Analysis work typically involves clustering and association rule mining of outlier data to find common features and potential links between outlier data. For example, certain combinations of abnormally high temperatures and low wind speeds are found to occur frequently in abnormal data by cluster analysis, revealing a potential risk of failure under certain environmental conditions. Finally, the generated key abnormal mode analysis result not only describes various abnormal modes in detail, but also provides valuable references for operation and maintenance decisions of the wind turbine generator.

Given historical performance data for a group of wind turbines, including wind speed per hour (m/s), temperature (C.) and voltage (V) data, these data will be processed in four sub-steps.

In S401, constructing an isolated forest decision tree, formatting wind speed, temperature and voltage data into a time sequence format, and selecting 100 decision trees to construct an isolated forest. For each tree, a subset of data (e.g., 1000 data points) and features (e.g., wind speed and temperature) are randomly selected, and for each selected subset, a segmentation value is randomly selected to segment the data until each data point is isolated or reaches a preset tree depth limit (e.g., 20), and the above steps are repeated until all 100 trees are built, forming a complete isolated forest that can represent the distribution and structural features of the entire data set.

In S402, the path length of the data point is evaluated, for each tree in the isolated forest, the path length of each data point from the root node to the leaf node is calculated, the path lengths of each data point in all the trees are averaged to obtain the average path length of the point, and a result including all the data points and the average path length thereof is obtained, which reflects the difficulty of the data points being isolated.

In S403, normal points and abnormal points are distinguished, a threshold value of the path length is set based on the characteristics and the priori knowledge of the data set, for example, 0.6 times the average length, the data points with the path length smaller than the threshold value are marked as abnormal points, an abnormal point detection result is generated, and all the data points identified as abnormal are explicitly marked.

In S404, the key anomaly pattern is analyzed to perform clustering and association rule analysis on the data points marked as anomalies, identify common features and behavior patterns in the anomaly points, such as voltage anomalies under high temperature and low wind speed conditions, and generate a key anomaly pattern analysis result, which details the identified anomaly pattern and its potential causes and effects.

Simulation data and results, wind speeds of 5-15m/s, temperatures of-10 ℃ to 35 ℃, voltages of 200V to 300V, found that at high temperatures (> 30 ℃) and low wind speeds (< 7 m/s) the voltages often rise abnormally, marking all data points at these conditions as abnormal, revealing that at certain high temperatures and low wind speeds the cooling efficiency of the wind turbine is reduced resulting in a voltage rise, which requires additional maintenance or system upgrades.

Referring to fig. 6, the key abnormal mode is analyzed through a bayesian network, probability inference is performed on the fault cause by using the conditional probability distribution of the network, meanwhile, the historical fault data and the real-time observation value are referred to, prediction of the fault cause and the fault position is performed, and the step of generating the fault cause deep analysis specifically comprises:

S501: based on the key abnormal modes, a Bayesian network construction method is utilized to create a network model to represent the relation between the abnormal modes, and a Bayesian network structure is generated by setting nodes in the network based on the abnormal modes and defining edges according to mode relevance so as to form a network and reflect the probability relation between the differential abnormal modes;

s502: based on the Bayesian network structure, performing conditional probability reasoning, and generating a conditional dependence analysis result by calculating the probability of occurrence of other modes when a given abnormal mode exists and analyzing and quantifying the dependency relationship between the abnormal modes;

S503: based on the condition-dependent analysis result, by combining the historical fault data and the real-time observation value, deducing the fault cause, and evaluating the similarity between various abnormal modes and the historical fault case by using a Bayesian network to generate a fault cause reasoning result;

S504: based on the fault cause reasoning result, predicting the fault position and type, analyzing the correlation between the fault cause and a target component in the wind turbine, selecting a part with a problem and potential fault types, and generating fault cause depth analysis.

In step S501, a key abnormal pattern is identified.

Example code (Clikit-learn library of Python for cluster analysis):

from sklearn.cluster import KMeansimport numpy as np

# hypothesis anomaly data

data＝np.array([[1,2],[1,4],[1,0],

[10,2],[10,4],[10,0]])

kmeans＝KMeans(n_clusters＝2,random_state＝0).fit(data)print(kmeans.labels_)

Bayesian network construction:

Example code (using pgmpy libraries):

from pgmpy.models import BayesianModelfrom pgmpy.factors.discrete import TabularCPD

# definition node

model＝BayesianModel([('A','B'),('B','C')])

# Definition conditional probability distribution

cpd_A＝TabularCPD(variable＝'A',variable_card＝2,values＝[[0.8],[0.2]])

cpd_B＝TabularCPD(variable＝'B',variable_card＝2,

values＝[[0.7,0.3],[0.3,0.7]],

evidence＝['A'],evidence_card＝[2])

cpd_C＝TabularCPD(variable＝'C',variable_card＝2,

values＝[[0.9,0.1],[0.2,0.8]],

evidence＝['B'],evidence_card＝[2])

model.add_cpds(cpd_A,cpd_B,cpd_C)

model.check_model()

In step S502, a conditional probability is calculated.

Example code (reasoning function using pgmpy):

from pgmpy.inference import VariableElimination

infer＝VariableElimination(model)

result＝infer.query(variables＝['C'],evidence＝{'A':1})print(result)

in step S503, the historical failure data is integrated.

Example code (data pre-processing):

# is just one example here, and actual data processing is more complex

Historical _fault_data=np.array ([ [..]) >) historical fault data # data preprocessing step.

Fault cause reasoning:

example code (using bayesian network model):

The reasoning here for # hypothesis failure data related to model 'A' node # is based on the previously defined model and reasoning engine

fault_reason＝infer.query(variables＝['B','C'],evidence＝{'A':fault_data})print(fault_reason)

In step S504, predictive analysis is performed.

Example code (prediction based on reasoning results):

Suppose that the fault location and type are related to model 'C' nodes

fault_location_type＝infer.query(variables＝['C'])print(fault_location_type)

Failure cause and component correlation analysis:

example code (analyze correlation):

This section # typically involves complex statistical analysis and data processing # example code is just a simple correlation calculation importscipy.stats as stats

Component_data=np

correlation＝stats.pearsonr(fault_reason.values,

component_data)print(correlation)

Referring to fig. 7, by using fault cause deep analysis and adopting a genetic algorithm, iterative optimization is performed on parameters of a fault diagnosis system by simulating a natural selection and a crossover and mutation mechanism in a genetic process, and an optimal parameter combination is captured, so that the steps for generating diagnosis strategy optimization are specifically as follows:

S601: based on fault cause depth analysis, initializing parameters of a fault diagnosis system by adopting a genetic algorithm, randomly generating an initial population from a plurality of parameter combinations by adopting an algorithm simulation natural selection mechanism, wherein each parameter combination represents a solution, and generating the initial parameter population;

S602: based on the initial parameter population, performing crossover operation of a genetic algorithm, and forming a new parameter set by exchanging parameter parts of various solutions to generate a crossly generated parameter combination;

S604: based on the mutated parameter combination, determining a final optimized parameter combination through the selection operation of a genetic algorithm, evaluating the fitness of a plurality of solution sets, selecting an optimal parameter configuration based on the fitness, and generating diagnosis strategy optimization.

In a sub-step S601, data required by the fault diagnosis system is first collected, including machine operating parameters, fault records, sensor readings, etc., formatted as a structured data table. These parameters are then initialized using a genetic algorithm. In this process, the encoding mode of the parameters is defined first, for example, each parameter is encoded as a binary string or real number. Then, a group of initial parameter combinations are randomly generated according to the requirements of the fault diagnosis system to form an initial population. Each parameter combination represents a potential solution, namely a specific set of fault diagnosis system configurations, the key to the procedure being to ensure diversity of the population to cover a wide range of potential solution spaces. After completion, the resulting initial population provides the basis for subsequent genetic manipulation.

In S602 substep, a crossover operation is performed on the initial population. In this process, combinations of parameters (i.e., individuals) in a population are randomly selected as parents, and then part of the parameters are exchanged by some sort of crossover strategy (e.g., single-point crossover, multi-point crossover, or uniform crossover) to create new individuals. The goal of the crossover operation is to explore new solution space by combining the features of different individuals to improve the probability of finding the optimal solution. When performing crossover operations, it is critical to determine crossover points and the probability of crossover, and after the end of the steps, a new set of parameter combinations, i.e., the crossover population, is generated.

In the S603 substep, mutation operation is performed on the crossed population. The mutation operation is to introduce new genetic mutation by randomly changing the values of part of the parameters in the individual, and the purpose of the step is to maintain the diversity of the population and prevent the algorithm from converging to the locally optimal solution prematurely. In performing the mutation operation, it is important to determine the mutation points and mutation probabilities. Typically, the probability of variation is set low to ensure population stability while introducing the necessary randomness. After the mutation operation is completed, the parameter combination comprising the new characteristics is obtained, so that possibility is provided for finding a better solution.

In S604 substep, a final set of optimization parameters is determined by a selection operation of the genetic algorithm. In this process, each individual (i.e., combination of parameters) in the mutated population is evaluated for fitness, which is assessed based on its performance (e.g., diagnostic accuracy, speed, etc.) in fault diagnosis. Then, the individuals with the best performance are selected to form a new generation population based on the fitness, and in the step, the design of the fitness function is crucial, so that the optimization effect of the genetic algorithm is directly influenced. Finally, the selected optimized parameter set is the optimal configuration of the fault diagnosis system, and the accuracy and efficiency of diagnosis can be improved.

The fault diagnosis system is used for detecting faults of the wind turbine generator, and related data comprise indexes such as wind speed, temperature and voltage. The above steps are applied to optimize the diagnostic parameters.

In the initialization population, a plurality of fault diagnosis parameter combinations are generated according to the historical fault data and the operation parameters to form the initial population.

In the crossover operation, combinations of parameters in the population are selected, and new combinations are generated by crossover, for example, combining the temperature threshold of one parameter combination with the voltage threshold of another.

In the mutation operation, some parameter values are randomly changed in the crossed population, such as randomly adjusting the temperature threshold.

In selecting the optimized parameters, an optimal parameter combination is selected based on the diagnostic effect (such as diagnostic accuracy) of each parameter combination.

Referring to fig. 8, based on optimization of diagnostic strategies, a decision support system algorithm is adopted to integrate real-time operation data of a wind turbine and the optimized diagnostic strategies on a cloud platform, wherein the real-time operation data comprise continuous monitoring of the real-time data and real-time updating of fault prediction results, and the steps of generating a maintenance decision scheme are specifically as follows:

S701: integrating real-time operation data and diagnosis strategy optimization of the wind turbine generator based on a cloud platform, adopting a data warehouse construction method to extract data from multiple data sources, unifying data formats, cleaning and removing inconsistent or erroneous data, and loading the processed data into a centralized data warehouse to generate a real-time data set;

S702: based on a real-time data set, adopting time sequence analysis, and generating a fault prediction model output by recognizing a time correlation mode and a trend in data and combining a long-term and short-term memory network algorithm to learn a long-term dependency relationship between data points so as to predict the trend of the data in a future time period;

s703: based on the output of the fault prediction model and the historical maintenance data, a multi-criterion decision analysis method is adopted, various decision factors including cost, risk and benefit are comprehensively referred, and the multi-decision factors are subjected to weight distribution and comprehensive evaluation by using a hierarchical analysis method, so that a maintenance plan scheme is generated;

S704: based on a maintenance planning scheme, a resource optimization and linear programming algorithm is adopted to analyze the resource requirement of maintenance activities, the required manpower, material and time cost is calculated by establishing a mathematical model of resource allocation and cost, and then the resource allocation is adjusted and optimized to generate a maintenance decision scheme.

In a sub-step S701, the objective is to create a data warehouse comprising real-time operational data of the wind turbines and optimized diagnostic strategies. First, data is collected from a plurality of data sources (e.g., sensor data, travel logs, maintenance records), the format including time series data, log files, and tabular data. Then, through the data cleansing process, the data format is unified and inconsistent or erroneous data is removed. The processed data is then loaded into the centralized data warehouse using data warehouse construction methods, such as ETL (extraction, transformation, loading) processes, with the goal of generating an accurate, consistent and readily accessible real-time data set that provides a reliable data basis for subsequent fault prediction and maintenance decisions.

In S702 substep, a failure prediction model is generated using time series analysis and long-short term memory (LSTM) algorithms based on the integrated real-time data set. Time series analysis is used to identify patterns and trends in the data, such as periodic, seasonal, or long-term trends. In conjunction with the LSTM algorithm, long-term dependencies between data points can be learned deep, which involves selecting an appropriate model architecture, adjusting network parameters (e.g., number of layers, number of units), and training the model. By this deep learning approach, the trend of the data over a future time period, such as the time and type of potential failure occurrence, can be predicted. After the completion, the generated fault prediction model can update the prediction result in real time, and timely fault early warning is provided.

In S703 substep, a maintenance planning scheme is generated using a multi-criterion decision analysis method in combination with the failure prediction model output and the historical maintenance data. In this process, various decision factors are considered, such as cost, risk, benefit, etc. And (3) carrying out weight distribution and comprehensive evaluation on the decision factors by using a Analytic Hierarchy Process (AHP) and other technologies. This process requires detailed analysis of the potential impact and cost effectiveness of various maintenance activities to formulate an efficient and economical maintenance plan. Finally, the generated maintenance planning scheme aims at optimizing the operation efficiency of the wind turbine generator and reducing the long-term operation and maintenance cost.

In a sub-step S704, the resource requirements of the maintenance activities are analyzed using a resource optimization and linear programming algorithm. By building a mathematical model of the resource allocation and cost, the required manpower, material and time costs are calculated. The resource allocation is then adjusted and optimized to ensure the efficiency and economy of the maintenance activities, the steps involving complex mathematical planning and optimization techniques with the aim of achieving optimal maintenance results under limited resource constraints. The finally generated maintenance decision scheme guides the actual maintenance activities and ensures the efficient operation and reliability of the wind turbine generator.

Real-time operation data of the wind turbine generator set are assumed to comprise wind speed (5-15 m/s), temperature (-10 ℃ to 35 ℃) and voltage (200V to 300V). Through the steps, the data are integrated and analyzed, potential faults are predicted, a targeted maintenance plan is formulated, and resource allocation is optimized. For example, if the predictive model indicates that the voltage would rise abnormally under high temperature conditions, the maintenance schedule would include checking and maintaining the cooling system prior to the hot season, while optimizing the scheduling and resource allocation of the maintenance team to reduce operation and maintenance costs and improve response efficiency.

Referring to fig. 9, a remote fault diagnosis system for a wind turbine based on a cloud platform is used for executing the remote fault diagnosis method for a wind turbine based on the cloud platform, and the system comprises a data integration module, an interactive relation modeling module, a performance trend analysis module, an abnormal mode detection module, a cause analysis and prediction module and a maintenance strategy decision module;

the performance trend analysis module is used for carrying out deep learning on time sequence data of the components by applying a long-term and short-term memory network based on the component interaction map, analyzing the variation trend and the periodicity pattern, predicting performance fluctuation and potential fault points in a future time period, and generating performance trend prediction;

The cause analysis and prediction module is used for deducing the fault cause based on abnormal mode analysis by adopting a Bayesian network algorithm, analyzing the association and influence between abnormal modes through a probability model, and generating a fault deducing result by combining historical fault data;

The maintenance strategy decision module optimizes the maintenance strategy by using a genetic algorithm and multi-criterion decision analysis based on the fault inference result, captures an optimal maintenance plan and an emergency response scheme by simulating a natural selection mechanism and a cross mutation operation, and establishes a maintenance strategy scheme.

The data integration module realizes effective integration and standardization of data through application of an ETL algorithm, and provides accurate and comprehensive data support for the system. The interactive relation modeling module utilizes the component relation map constructed by the graph neural network to visualize the dynamic interaction among components and further understand the influence of each component on the system performance. The performance trend analysis module applies the deep learning of the long-term memory network to the time series data, can accurately predict performance fluctuation and potential fault points, and provides key information for timely fault prevention. The abnormal mode detection module improves the diagnosis precision and response speed of the system through efficient abnormal detection of an isolated forest algorithm. The cause analysis and prediction module uses a Bayesian network algorithm to provide probability-based fault cause inference, so that the accuracy and reliability of fault diagnosis are enhanced. Finally, the maintenance strategy decision module combines a genetic algorithm and multi-criterion decision analysis, optimizes the maintenance strategy and ensures the efficient and sustainable operation of the wind turbine.

The present invention is not limited to the above embodiments, and any equivalent embodiments which can be changed or modified by the technical disclosure described above can be applied to other fields, but any simple modification, equivalent changes and modification made to the above embodiments according to the technical matter of the present invention will still fall within the scope of the technical disclosure.

Claims

1. A remote fault diagnosis method of a wind turbine generator based on a cloud platform is characterized by comprising the following steps:

Based on design parameters and real-time operation data of the wind turbine generator, a data fusion technology is adopted, and a comprehensive data set is generated by carrying out formatting processing, noise removal and missing value processing on data from a sensor, an operation log and a history maintenance record;

Based on the comprehensive data set, a graph neural network algorithm is adopted, a plurality of components of the wind turbine generator are represented by constructing nodes, dynamic connection and influence among the components are represented by edges, and then the interrelationship among the components is mapped in a graph structure, so that a component interaction relation graph is generated;

based on the component interaction relation map, a long-period memory network is adopted to analyze sequence data of component characteristics changing along with time, potential change trend and periodic mode are identified, and then performance change and fault occurrence points in a future time period are predicted, so that performance trend prediction analysis is generated;

Based on the performance trend prediction analysis, constructing a plurality of decision trees by using an isolated forest algorithm, evaluating the path length from each data point to the leaves, distinguishing normal data from abnormal data, and further identifying a key abnormal mode;

Analyzing the key abnormal mode through a Bayesian network, carrying out probability inference on the fault cause by utilizing the conditional probability distribution of the network, and simultaneously carrying out prediction on the fault cause and the fault position by referring to historical fault data and real-time observation values to generate fault cause depth analysis;

By utilizing the fault cause deep analysis, adopting a genetic algorithm, performing iterative optimization on parameters of a fault diagnosis system by simulating natural selection and a crossover and mutation mechanism in a genetic process, capturing an optimal parameter combination, and generating diagnosis strategy optimization;

based on the diagnosis strategy optimization, a decision support system algorithm is adopted, real-time operation data of the wind turbine generator and the optimized diagnosis strategy are integrated on a cloud platform, the real-time operation data and the optimized diagnosis strategy comprise continuous monitoring of the real-time data and real-time updating of fault prediction results, and maintenance plans and emergency response measures are formulated according to the fault prediction and historical maintenance data to generate a maintenance decision scheme.

2. The cloud platform based wind turbine remote fault diagnosis method according to claim 1, wherein the comprehensive data set comprises temperature, vibration frequency, current intensity of sensor readings, and time stamp, maintenance history and equipment state record of operation log, the component interaction relation map comprises action strength, action frequency and action mode among components, and dependency and interaction probability estimation among components, the performance trend prediction analysis comprises prediction of running state, identification of potential fault points and estimation of expected maintenance time in a future time period of a plurality of components, the key abnormality mode comprises target abnormality indexes of abnormal vibration mode, temperature abnormality and energy consumption bias, the fault cause depth analysis comprises probability assessment of fault occurrence, category division of fault cause and severity prediction of fault influence, the diagnosis strategy optimization comprises adjustment of diagnosis threshold values, adjustment of model parameters and efficiency optimization of diagnosis flow, and the maintenance decision scheme comprises priority ordering of maintenance tasks, resource allocation scheme and emergency response scheme.

3. The cloud platform-based wind turbine generator system remote fault diagnosis method according to claim 1, wherein the step of generating the comprehensive data set by formatting, removing noise and missing value processing data from the sensor, the operation log and the history maintenance record by adopting a data fusion technology based on design parameters and real-time operation data of the wind turbine generator system is specifically as follows:

Based on design parameters and real-time operation data of the wind turbine, a Z-score standardization algorithm is adopted, and the data are converted into a format with uniform scale and measurement by calculating deviation of each data point and an average value and dividing the deviation by a standard deviation, so that a standardized data set is generated;

Based on the standardized data set, identifying and filtering abnormal values and noise in the data by using a box graph method, determining abnormal values by comparing deviations between data points and quartiles, eliminating error data points of a distortion analysis result, and generating abnormal value filtering data;

Filling the missing values in the data set by using a K-nearest neighbor filling method based on the outlier filtering data, and estimating the missing values by analyzing the average value of K nearest data points associated with the missing points to generate missing value processing data;

And processing data based on the missing values, performing deep analysis and fusion on the multi-source data by adopting a characteristic data fusion method, eliminating the difference between the data by unifying the format and the structure of the differential data source, and synthesizing multiple types of characteristic information, including sensor reading, operation log and history maintenance record, to generate a comprehensive data set.

4. The remote fault diagnosis method for the wind turbine generator based on the cloud platform according to claim 1, wherein based on the comprehensive data set, a graph neural network algorithm is adopted, a plurality of components of the wind turbine generator are represented by constructing nodes, edges represent dynamic connection and influence among the components, and further interrelationships among the components are mapped in a graph structure, and the step of generating a component interaction relation graph is specifically as follows:

Based on the comprehensive data set, identifying a plurality of components in the wind turbine generator by utilizing an entity identification algorithm, analyzing the characteristics, the operation parameters and the running states of the components, identifying each physical component, establishing independent entities and generating component entity mapping;

Based on the component entity mapping, a relation network analysis method is adopted to determine the dynamic connection and the mutual influence among the components, and the functional relation, the signal interaction and the control dependency relation among the components are evaluated to construct the interactive network connection and generate a component interaction network;

Based on the component interaction network, converting the components and the interrelationships thereof into a graph structure by using a graph construction method, and simultaneously taking each component as a node and the interaction between the components as an edge to generate a preliminary interaction map;

Based on the preliminary interaction map, the nodes and edges in the map are learned by using a map neural network, key features and modes of interaction among the components are identified and extracted, the map is optimized, the relation among the components is reflected, and a component interaction relation map is generated.

5. The cloud platform-based wind turbine generator system remote fault diagnosis method according to claim 1, wherein based on the component interaction relation map, a long-term and short-term memory network is adopted to analyze sequence data of component characteristics changing along with time, and a potential change trend and a periodic mode are identified, so that performance change and fault occurrence points in a future time period are predicted, and the step of generating performance trend prediction analysis is specifically as follows:

Based on the component interaction relation graph, carrying out time sequence analysis by using a statistical method, and generating time sequence characteristic analysis by carrying out trend analysis, periodic detection and relevance evaluation on the operation data of multiple components, wherein the trend analysis comprises calculation of moving average, seasonal decomposition and autocorrelation coefficients of a time sequence, and identification of a time dependent mode of component behaviors;

Based on the time sequence characteristic analysis, a long-period memory network model is adopted to deeply learn the identified time dependence mode, a long-period dependence relation is captured through a network layer, a hidden mode and a trend in a time sequence are learned, and an LSTM learning result is generated;

Based on the LSTM learning result, predicting performance change and fault points in a future time period, performing prediction analysis by utilizing the output of an LSTM model, and simultaneously evaluating the running state of the component in the future time period, predicting a potential performance degradation area and the occurrence time of faults to generate future performance fault prediction;

Based on the future performance fault prediction, performing multidimensional analysis on the running state of the whole wind turbine generator in a future time period, wherein the multidimensional analysis comprises the steps of combining a prediction result with a historical maintenance record, evaluating potential maintenance requirements and providing a risk management strategy, and generating performance trend prediction analysis.

6. The remote fault diagnosis method for the wind turbine generator set based on the cloud platform according to claim 1, wherein the steps of constructing a plurality of decision trees and evaluating path lengths from each data point to leaves by using an isolated forest algorithm based on the performance trend prediction analysis, distinguishing normal data from abnormal data, and further identifying key abnormal modes are specifically as follows:

Based on the performance trend prediction analysis, constructing a plurality of independent decision trees by using an isolated forest algorithm, constructing a decision tree for each subset by randomly selecting data subsets and features, and dividing data by the randomly selected features to generate a decision tree forest;

Based on the decision tree forest, estimating the path length of each data point reaching a leaf node in the decision tree, and generating path length analysis by calculating the number of splitting steps from a root node to the leaf node and analyzing data isolation risks based on the path distance;

distinguishing normal points and abnormal points in the data set based on the path length analysis, identifying and marking abnormality based on the path length of the data points by comparing the path length with a preset threshold value, mining potential faults or performance problems, and generating an abnormal point detection result;

based on the outlier detection result, analyzing the data points marked as outliers, capturing common characteristics and behavior patterns of the outlier data points, identifying an outlier pattern causing a performance problem or failure, and generating a key outlier pattern analysis.

7. The cloud platform-based wind turbine generator system remote fault diagnosis method according to claim 1, wherein the key abnormal pattern is analyzed through a bayesian network, probability inference is performed on fault reasons by using conditional probability distribution of the network, and meanwhile, prediction of fault reasons and fault positions is performed by referring to historical fault data and real-time observation values, and the step of generating fault reason depth analysis specifically comprises the following steps:

Based on the key abnormal modes, a Bayesian network construction method is utilized to create a relation between network model representation abnormal modes, nodes in the network are set based on the abnormal modes, edges are defined according to mode relevance, a network is further formed, probability relation between differentiated abnormal modes is reflected, and a Bayesian network structure is generated;

Based on the Bayesian network structure, performing conditional probabilistic reasoning, and generating a conditional dependence analysis result by calculating probabilities of occurrence of other modes when a given abnormal mode exists and analyzing and quantifying a dependence relationship between the abnormal modes;

Based on the condition-dependent analysis result, by combining the historical fault data and the real-time observation value, deducing a fault cause, and evaluating the similarity between a plurality of abnormal modes and the historical fault case by using a Bayesian network to generate a fault cause reasoning result;

And predicting the fault position and type based on the fault cause reasoning result, analyzing the correlation between the fault cause and a target component in the wind turbine, selecting a part with a problem and potential fault types, and generating fault cause deep analysis.

8. The cloud platform-based wind turbine generator system remote fault diagnosis method according to claim 1, wherein the fault cause depth analysis is utilized, a genetic algorithm is adopted, the parameters of a fault diagnosis system are subjected to iterative optimization by simulating a natural selection and crossing and mutation mechanism in a genetic process, an optimal parameter combination is captured, and the step of generating diagnosis strategy optimization is specifically as follows:

Based on the fault cause depth analysis, initializing parameters of a fault diagnosis system by adopting a genetic algorithm, randomly generating an initial population from a plurality of parameter combinations by adopting an algorithm simulation natural selection mechanism, wherein each parameter combination represents a solution, and generating the initial parameter population;

Based on the initial parameter population, executing crossover operation of a genetic algorithm, and forming a new parameter set by exchanging parameter parts of various solutions to generate a crossly generated parameter combination;

Based on the parameter combination generated by the intersection, performing mutation operation of a genetic algorithm, introducing new mutation operation by randomly changing part of parameter values, avoiding the local optimal solution of the algorithm, and generating a mutated parameter combination;

and determining a final optimized parameter combination through the selection operation of a genetic algorithm based on the mutated parameter combination, evaluating the fitness of a plurality of solution sets, and selecting an optimal parameter configuration based on the fitness to generate diagnosis strategy optimization.

9. The remote fault diagnosis method of the wind turbine generator set based on the cloud platform according to claim 1, wherein the optimization of the diagnosis strategy is based on a decision support system algorithm, the integration of real-time operation data of the wind turbine generator set and the optimized diagnosis strategy on the cloud platform comprises the continuous monitoring of the real-time data and the real-time updating of a fault prediction result, and the steps of making a maintenance plan and emergency response measures according to the fault prediction and the historical maintenance data to generate a maintenance decision scheme are specifically as follows:

Integrating real-time operation data of the wind turbine generator and the diagnosis strategy optimization based on a cloud platform, extracting data from multiple data sources by adopting a data warehouse construction method, unifying data formats, cleaning and removing inconsistent or erroneous data, and loading the processed data into a centralized data warehouse to generate a real-time data set;

based on the real-time data set, adopting time sequence analysis, and generating a fault prediction model output by identifying time correlation modes and trends in the data and combining long-term and short-term memory network algorithms to learn long-term dependency relations among data points so as to predict data trend in a future time period;

Based on the fault prediction model output and the historical maintenance data, adopting a multi-criterion decision analysis method, comprehensively referencing various decision factors including cost, risk and benefit, and carrying out weight distribution and comprehensive evaluation on the multi-decision factors by using a hierarchical analysis method so as to generate a maintenance plan scheme;

Based on the maintenance planning scheme, a resource optimization and linear programming algorithm is adopted to analyze the resource requirement of the maintenance activity, the required manpower, material and time cost is calculated by establishing a mathematical model of resource allocation and cost, and then the resource allocation is adjusted and optimized to generate a maintenance decision scheme.

10. The remote fault diagnosis system of the wind turbine generator based on the cloud platform is characterized by comprising a data integration module, an interactive relation modeling module, a performance trend analysis module, an abnormal mode detection module, a cause analysis and prediction module and a maintenance strategy decision module;