CN118188342A - Fan-oriented fault early warning and life prediction method - Google Patents

Abstract

The invention discloses a fan-oriented fault early warning and life predicting method, which comprises the following steps: acquiring original operation data in the operation process of the fan through a sensor or a data acquisition device; carrying out data preprocessing on the original data; dividing the preprocessed data into a training set and a testing set; predicting by using OAKR model and performing fault early warning based on MSE index and multivariate Bayesian index; and after fault early warning, predicting the residual life by using an LSTM-WPHM model. According to the invention, the OAKR model, the LSTM model and the WPHM model are combined to perform fault early warning and residual life prediction on the fan, the fault early warning and residual life prediction can be effectively realized through actual data verification, and maintenance personnel can be helped to judge the current state of the unit, so that the occurrence of unit faults and excessive maintenance is avoided, and the longest effective service life is achieved due to the least consumption of resources.

Description

A fault warning and life prediction method for wind turbines

Technical Field

The present invention relates to the technical field of fault warning and life prediction of wind turbines, and in particular to a fault warning and life prediction method for wind turbines.

Background technique

Wind turbines have very strict requirements on the natural environment. Considering the distribution of wind resources and the purpose of maximizing their utilization, wind turbines are usually built in remote and wide areas such as offshore, Gobi Desert and grassland. Affected by the extremely harsh operating environment and complex and changeable operating conditions, wind turbines are more prone to failure than traditional power generation equipment. This leads to high initial construction investment and very expensive later operation and maintenance costs. Due to the special operating environment of wind turbines, their maintenance usually requires a long time to prepare. When the unit has an unplanned shutdown due to a sudden failure, the long maintenance time will lead to a reduction in the power generation of the wind farm, further affecting the economic benefits of the wind farm. With the continuous development of science and technology, wind turbines are gradually developing towards large-scale, complex and mass production, which makes wind turbine maintenance more difficult and expensive. Therefore, how to reduce the operation and maintenance costs of wind turbines has become a major challenge faced by improving the economic benefits of wind power generation and promoting the healthy development of the wind power industry.

At present, there are two main maintenance methods for wind turbines: post-maintenance and regular maintenance. Post-maintenance, also known as corrective maintenance, refers to a series of maintenance work carried out to restore the unit to a state where it can perform the specified functions after a failure. Regular maintenance, also known as preventive maintenance, refers to regular systematic inspection, testing and parts replacement of the unit to reduce the probability of failure and mitigate the consequences of failure. In recent years, with the rapid development of the wind power industry, commonly used maintenance methods have gradually exposed their shortcomings and are difficult to meet the requirements of operators. In order to reduce operation and maintenance costs and improve the availability of wind turbines, condition monitoring maintenance has become a research focus for formulating wind turbine maintenance strategies. Condition monitoring maintenance, also known as predictive maintenance, is to monitor performance status data, timely and accurately evaluate the operating status of the unit, judge the occurrence of failures, and predict the trend of performance status, so as to formulate a maintenance plan in advance and determine the time and content of unit maintenance. Based on the identification and early warning of abnormal conditions of the unit before a failure occurs, as well as the prediction of the remaining life of the unit, maintenance personnel can determine the current operating status of the unit, determine the necessity of implementing maintenance or parts replacement, and formulate the optimal maintenance plan to minimize the occurrence of unit failures and excessive maintenance, and ensure that there is sufficient time to carry out repairs or obtain replacement parts, so that the components and even the entire wind turbine unit can consume the least resources and achieve the longest effective service life while ensuring safety and reliability.

The data-driven life prediction method is to establish a model that can describe the law of equipment operation status changes based on the historical operation data obtained, and then analyze the current operation status of the equipment and predict the remaining life of the equipment before failure. This method does not require a systematic understanding of the mechanical structure and failure mechanism of the equipment. Instead, it uses statistical theory, machine learning technology, pattern recognition and other methods and principles to analyze the historical operation data of the equipment, establish a functional mapping relationship between the prediction model and the operation monitoring variables, so as to realize the equipment's operation status monitoring and remaining life estimation, and provide corresponding decision-making information for the formulation of maintenance plans. Data-driven life prediction methods include reliability methods, random processes, time series and neural networks.

With the continuous development of science and technology, mechanical equipment is becoming increasingly large-scale and complex, and it is becoming more difficult to obtain its complete failure mechanism and physical structure, so the life prediction method based on physical models is greatly limited in application and research. With the development of the Internet of Things, the Internet and computer technology, it has become easier to obtain equipment monitoring data, and data-driven methods have gradually become a more effective way to solve the life prediction of mechanical equipment. Therefore, there is an urgent need for a method to realize fault warning and life prediction of key components taking into account the special performance of wind turbines in harsh operating environments and complex and changeable working conditions.

Summary of the invention

In order to solve the problems existing in the prior art, the present invention provides a fault warning and life prediction method for wind turbines. According to the operating characteristics of each key component of the wind turbine, an optimized autoassociative kernel regression model, a long short-term memory network model and a Weibull proportional risk model are developed to achieve an organic combination of fault warning and life prediction of key components, thereby solving the problems mentioned in the above background technology.

To achieve the above object, the present invention provides the following technical solution: a fault warning and life prediction method for a wind turbine, comprising the following steps:

S1. Obtaining original operation data of the fan during operation through sensors or data acquisition devices;

S2, preprocessing the original data;

S3, dividing the preprocessed data into a training set and a test set, wherein the training set and the test set are obtained by dividing the historical health data of the unit in a ratio of 85% to 15%;

S4. Use the OAKR model to make predictions and perform fault warnings based on the MSE index and multivariate Bayesian index;

S5. After a fault warning occurs, the LSTM-WPHM model is used to predict the remaining life.

Preferably, in step S2, the data preprocessing includes processing of missing values and infinite values, deleting start and stop data, eliminating seasonal factors, data standardization and wavelet packet Bayesian denoising.

Preferably, in step S4, the following is specifically included:

S41, model training: divide 85% of the historical health data for model training, use the equal-interval sampling method to sample the first 70% of the data to generate a memory matrix; use the simplex optimization method to optimize the kernel function bandwidth h in the OAKR algorithm for the last 15% of the data;

S42, Model testing: Use the trained OAKR model for model testing to obtain the performance parameters of the model in a healthy state, providing a basis for model early warning;

S43, model prediction: input the preprocessed data to be tested into the trained OAKR model, output the predicted value, calculate the mean square error (MSE) between the predicted value and the true value and the Bayes factor alarm index, and when the threshold is exceeded, the model will alarm.

Preferably, in step S4, the OAKR method steps include:

The first step is to calculate the distance between the monitoring vector v and each memory vector _Xi to obtain an n×1 distance vector d, using Euclidean distance calculation:

The second step is to calculate the weight w through the obtained distance matrix d and Gaussian kernel function. w is an n×1 vector matrix, and each element is calculated by the following formula:

Where h is the kernel function bandwidth, which determines the smoothness of the kernel function. A small bandwidth h can reflect more details but often leads to undersmoothing of the tail. On the contrary, a large bandwidth h often loses details, resulting in oversmoothing of the parts with drastic changes.

Step 3: Use the simplex algorithm to optimize the bandwidth h;

Step 4: Use the obtained weight w to make the predicted value of the monitoring vector v It is calculated by the weighted average of each memory vector _Xi , and the calculation formula is as follows:

Preferably, in step S5, the following is specifically included:

S51. Related variable screening: Based on expert knowledge and fault mechanism, screen out the related variables that affect the bearing temperature change, and determine the degree of correlation between each variable and the bearing temperature through correlation analysis;

S52, LSTM model training: Use the variable data under the unit health status to train the LSTM bearing temperature prediction model and fit the nonlinear mapping relationship between the relevant variables and the bearing temperature;

S53, LSTM model verification: verify the trained LSTM bearing temperature prediction model, analyze whether the model training process converges, and evaluate whether the mean square error and determination coefficient of the model meet the requirements. If the model does not meet the accuracy requirements, retrain the model;

S54, bearing status index determination: if the LSTM bearing temperature prediction model meets the requirements, the predicted value of the bearing temperature in a healthy state is calculated, and the residual between the actual value and the predicted value is used as the bearing status index;

S55, WPHM modeling: Based on the unit fault information, the bearing status indicators of the faulty units and healthy units in the training set are selected to establish the WPHM model;

S56, model threshold determination: Use the LSTM-WPHM model to perform reliability analysis on the units in the training set, and give the failure threshold _Ff and the alarm threshold _Fa by comparing the differences between the faulty units and the healthy units;

S57, bearing fault alarm: Use the LSTM-WPHM model to analyze the unit, calculate the cumulative failure probability F(t,X) of the unit at each time, and combine the threshold setting to give a model alarm;

S58. Bearing life prediction: When F(t,X) exceeds the alarm threshold _Fa , the remaining life of the unit fault is predicted. According to the change trend of F(t,X) before the model alarm, the unit degradation failure curve is fitted to predict the moment when F(t,X) reaches the failure threshold _Ff . Combined with the confidence interval estimated by the fitting curve parameters, the remaining life interval is given.

Preferably, in step S52, the LSTM model consists of 2 layers of LSTM neural networks, 2 layers of Dropout layers and 1 layer of Dense layer;

The number of nodes in the first layer of LSTM neural network in the LSTM model is 64, and the activation function is the Relu function. The number of nodes in the second layer of LSTM neural network is 128, and the activation function is the Relu function. The number of nodes in the Dense layer is 1, and the Dropout layer is added after each layer of LSTM network.

Preferably, in step S54, the residual δ between the predicted value and the actual value is used as an index to characterize the bearing state, and the bearing state index δ is used as an input of the WPHM model, and the calculation formula is as follows:

δ＝y _true -y _pre

Where y _true is the actual value of the bearing temperature, and y _pre is the predicted value of the bearing temperature.

Preferably, in step S58, an exponential function is used to fit its changing trend, and the expression is:

y(t)＝ae ^t +b

Where t is the running time of the unit bearing. When the failure threshold F _f is reached, the remaining life t _RUL of the bearing is:

t _RUL = t _pf - _ta ;

Where t _a is the bearing failure warning time, and t _pf is the predicted unit downtime.

The beneficial effect of the present invention is that the present invention combines the OAKR model, the LSTM model, and the WPHM model to perform fault warning and remaining life prediction for the fan. The OAKR model is used to predict and develop fault warning based on the MSE indicator and the multivariate Bayesian indicator. After the fault warning occurs, the LSTM-WPHM model is used to predict the remaining life. It has been verified by actual data that fault warning and remaining life prediction can be effectively realized, which can help maintenance personnel judge the current state of the unit, avoid unit failure and excessive maintenance, and consume the least resources to achieve the longest effective service life.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of the steps of the method of the present invention;

FIG2 is a schematic diagram of the structure of the LSTM bearing temperature prediction model of the present invention;

FIG3 is a schematic diagram of fault warning and life prediction of the LSTM-WPHM model of the present invention;

FIG4 is a schematic diagram of two verification analyses of missed reporting units in a test set according to an embodiment of the present invention, wherein (a) is missed reporting unit 1 and (b) is missed reporting unit 2.

Detailed ways

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

The present invention provides a technical solution: a fault warning and life prediction method for a wind turbine, as shown in FIG1 , which is a step flow of the present invention, specifically comprising the following steps:

1) Obtaining the operation data of the fan during operation through sensors and other data acquisition devices;

2) Preprocess the original data: handle missing values and infinite values; delete start and stop data; eliminate seasonal factors; standardize data; and use wavelet packet Bayesian denoising;

3) Divide the training set and test set: The training set and test set are obtained by dividing the historical health data of the crew in a ratio of 85% to 15%;

4) Model training: 85% of the historical health data for model training is further divided, of which the first 70% of the data is used to generate the memory matrix, and the last 15% of the data is used to optimize the kernel function bandwidth h in the OAKR algorithm. The first 70% of the data is sampled using the equal-interval sampling method to generate the memory matrix; the last 15% of the data is optimized using the simplex optimization method to optimize the kernel function bandwidth h in the OAKR algorithm;

5) Model testing: Use the trained OAKR model for model testing to obtain the performance parameters of the model in a healthy state, providing a basis for model early warning;

6) Model prediction: Input the preprocessed data to be tested into the trained OAKR model, output the predicted value, calculate the mean square error (MSE) and Bayes factor alarm index between the predicted value and the true value, and the model will alarm when the threshold is exceeded;

7) After the model alarm occurs, life prediction is performed;

8) Related variable screening: Based on expert knowledge and fault mechanism, screen out the related variables that affect the bearing temperature change, and determine the degree of correlation between each variable and the bearing temperature through correlation analysis;

9) LSTM model training: Use the variable data under the healthy state of the unit to train the LSTM bearing temperature prediction model and fit the nonlinear mapping relationship between the relevant variables and the bearing temperature;

10) LSTM model verification: Verify the trained LSTM bearing temperature prediction model, analyze whether the model training process has converged, and evaluate whether the mean square error and determination coefficient of the model meet the requirements. If the model does not meet the accuracy requirements, retrain the model.

11) Bearing status index: If the LSTM bearing temperature prediction model meets the requirements, the predicted value of the bearing temperature in a healthy state is calculated, and the residual between the actual value and the predicted value is used as the bearing status index.

12) WPHM modeling: Based on the unit fault information, the bearing status indicators of the faulty units and healthy units in the training set are selected to establish the WPHM model.

13) Model threshold determination: The LSTM-WPHM model is used to perform reliability analysis on the units in the training set. By comparing the differences between faulty units and healthy units, the failure threshold _Ff and alarm threshold _Fa are given.

14) Bearing fault alarm: Use the LSTM-WPHM model to analyze the unit, calculate the cumulative failure probability F(t,X) of the unit at each time, and combine the threshold setting to give the corresponding model alarm.

15) Bearing life prediction: When F(t,X) exceeds the alarm threshold _Fa , the remaining life of the unit fault is predicted. According to the change trend of F(t,X) before the model alarm, the unit degradation failure curve is fitted to predict the moment when F(t,X) reaches the failure threshold _Ff . Combined with the confidence interval estimated by the fitting curve parameters, the remaining life interval is given.

Optimized Auto-Associative Kernel Regression (OAKR)

Auto-associative Kernel Regression (AAKR) is a non-parametric modeling technique based on the principle of similarity between two-column arrays. It uses the similarity between the monitoring vector and the memory vector to infer the response of the model. The model training is simple and does not depend on the equipment and fault type information. It is suitable for multi-variable equipment operation monitoring and fault alarm. The OAKR method is based on this and obtains the best model parameters through optimization, so that it can be better used for equipment status monitoring and fault warning.

Extract the multi-dimensional monitoring data of the equipment and store the sample memory vectors used to create the OAKR model in a matrix X, where Xi _,j represents the i-th vector value of the j-th key variable. For n memory vectors, the memory matrix X of p variables can be expressed as:

The monitoring vector is represented by a 1×p matrix V:

V＝[v ₁ ,v ₂ ,…,v _p ] (2)

The predicted value of the model can be obtained by weighted averaging each memory vector of the memory matrix X, where the weighted average parameter is estimated using the unit health data. The OAKR method consists of four steps: First, the distance between the monitoring vector v and each memory vector _Xi is calculated to obtain an n×1 distance vector d. The present invention uses the most commonly used Euclidean distance calculation:

In the second step, the weight w is calculated by the obtained distance matrix d and Gaussian kernel function. w is also an n×1 vector matrix, and each element is calculated by the following formula:

Where h is the kernel function bandwidth, which determines the smoothness of the kernel function. A small bandwidth h can reflect more details but often leads to less smooth tails; on the contrary, a large bandwidth h often loses details and leads to overly smooth parts with drastic changes.

The third step is to optimize the bandwidth h using the simplex algorithm.

The fourth step is to use the obtained weight w to make the predicted value of the monitoring vector v It is calculated by the weighted average of each memory vector _Xi , and the calculation formula is as follows:

LSTM-WPHM model

LSTM temperature prediction model

The relevant variables affecting the bearing temperature are selected as the model input, and the bearing temperature is selected as the model output to train the LSTM bearing temperature prediction model. Figure 2 shows the structure of the LSTM bearing temperature prediction model, which consists of 2 layers of LSTM neural networks, 2 layers of Dropout layers, and 1 layer of Dense layer. The number of nodes in the first layer of the LSTM neural network in the model is 64, and the activation function is the Relu function. The number of nodes in the second layer of the LSTM neural network is 128, and the activation function is the Relu function. The number of nodes in the Dense layer is 1. After the Dropout layer is added to each layer of the LSTM network, 20% of the neurons are randomly removed in each iterative training to reduce the probability of overfitting of the model.

The operating data of the fan bearing in the healthy stage is selected to train the LSTM bearing temperature prediction model, and a nonlinear mapping relationship between the bearing temperature and various influencing factors is established. The trained LSTM model is used to predict the bearing temperature value under normal operating conditions. When the bearing is abnormal, a large deviation will occur between the predicted value and the actual value. Therefore, the present invention uses the residual δ between the predicted value and the actual value as an indicator to characterize the bearing state, and the calculation formula is as follows:

δ＝y _true -y _pre (6)

In the formula, y _true is the actual value of the bearing temperature, and y _pre is the predicted value of the bearing temperature. This index integrates the influence of various factors on the bearing temperature and highlights the abnormal temperature changes caused by bearing failure. The bearing status index δ is used as the input of the WPHM model to perform fault warning and life prediction for the fan bearing.

Weibull proportional hazards model

The traditional proportional hazard model does not assume the distribution form of λ ₀ (t), and only gives a discrete baseline failure rate based on existing data, and cannot give a continuous baseline failure rate that changes with historical life. Considering that different mechanical equipment has its corresponding failure distribution function under specific working conditions, and the analysis accuracy of the full parameter model is usually higher than that of the semi-parametric model, some scholars have proposed to introduce the failure distribution commonly used in reliability into the model and establish a full parameter statistical regression model to analyze mechanical equipment more accurately. In actual engineering applications, the Weibull distribution is a distribution function commonly used to describe the change of the failure degree of mechanical equipment over time. Therefore, it is often used as a baseline failure function to substitute into the proportional hazard model to obtain the Weibull Proportional Hazard Model (WPHM), and its failure rate function expression is:

The corresponding reliability index of the WPHM model is:

The unknown parameters β, η and α in the WPHM model are usually estimated using the maximum likelihood estimation method. The general expression of the likelihood function is:

Where n is the total number of data samples, δ _i is the data sample truncation indicator, a value of 1 indicates sample failure, a value of 0 indicates data truncation, and q is the number of sample failures. Substituting equations (7) and (9) into equation (11), the likelihood function of the WPHM model is:

The corresponding log-likelihood function is:

Taking partial derivatives of the unknown parameters β, η and α of the log-likelihood function lnL(β,η,α) in the above formula, we can get

Setting each partial derivative equal to zero, we get a nonlinear system of equations. Solving the system of equations gives us the estimated values of the parameters. and/> Since the equation belongs to a transcendental equation, it is relatively difficult to obtain an analytical solution, so an iterative algorithm is usually used to approximate the numerical solution. The present invention adopts the BFGS algorithm in the quasi-Newton method to solve it. The BFGS algorithm uses the BFGS matrix as an approximation of the second-order derivative in the Newton method to replace the Hessian matrix, which solves the problems of the non-existence of the descending direction and the difficulty in calculating the Hessian matrix in the Newton method, and has better numerical stability. The formula is as follows:

Wherein, _Bk is the BFGS matrix, _Δxk = _xk+1 - _xk , _Δfk = f( _xk+1 )-f( _xk ), _xk is the function variable, and f(·) is the objective function.

LSTM-WPHM model

According to equation (7), the bearing condition index δ given by the LSTM bearing temperature prediction model is integrated to establish the LSTM-WPHM model. The failure rate function expression of the model is:

According to the Weibull proportional hazard model, the unknown parameters in the LSTM-WPHM bearing reliability model are estimated, and its log-likelihood function is as follows:

The unknown parameters β, η and α of the log-likelihood function lnL(β,η,α) in equation (19) are obtained by partial derivatives, and each partial derivative is set equal to zero to obtain a nonlinear system of equations. The parameter estimates can be obtained by solving the system of equations using the BFGS algorithm: and/>

The corresponding reliability index of the LSTM-WPHM model is:

Among them, the cumulative failure probability F(t,X) can characterize the degree of bearing failure. The larger its value, the greater the degree of bearing failure. At the same time, F(t,X) also reflects the performance degradation of the bearing during operation. As the bearing performance deteriorates, F(t,X) will increase gradually. Therefore, this chapter uses the cumulative failure probability F(t,X) as the failure indicator of the bearing to achieve fault warning and life prediction of the fan bearing. Based on the analysis results of the faulty units and healthy units, the mean and maximum statistical characteristics of each unit are obtained, and the alarm threshold F _a and failure threshold F _f of the unit are determined.

Figure 3 shows a schematic diagram of the LSTM-WPHM model for unit fault alarm and life prediction. First, the multi-dimensional variable data is integrated through the LSTM model to calculate the bearing status index, and then the cumulative failure probability F(t,X) of the bearing is calculated according to the WPHM model. When F(t,X) reaches the alarm threshold _Fa , a unit fault warning is given, and the life prediction analysis of the unit is started. According to the change trend of F(t,X), curve fitting is performed to predict the future change trend of the unit F(t,X). When the predicted F(t,X) reaches the failure threshold _Ff , the unit is judged to have a fault shutdown, and the remaining life of the unit is given. When the model gives a fault warning, it is generally believed that the bearing has early signs of failure. In the absence of maintenance, the cumulative failure probability of the bearing will show a nonlinear monotonic upward trend with the running time. Therefore, the present invention uses an exponential function to fit its change trend, and the expression is:

y(t)＝ae ^t +b (0.1)

Where t is the running time of the unit bearing. When the failure threshold F _f is reached, the remaining life of the bearing t _RUL is

t _RUL = t _pf - t _a (0.2)

In the formula, _ta is the bearing fault warning time, and _tpf is the predicted unit downtime. Considering the uncertainty of the actual operating life of the wind turbine bearing, the upper and lower limits of the failure prediction trend change are given by fitting the confidence interval of the curve parameter estimation, so as to obtain the remaining life interval of the unit bearing, and then quantify the uncertainty of the remaining life. Therefore, the LSTM-WPHM hybrid model is used to achieve an organic combination of fault warning and remaining life prediction before the wind turbine is shut down.

Implementation verification

The actual operation data of a certain offshore wind farm is used as an example to verify the method proposed in the present invention. The process is as follows:

1. Use the OAKR model for fault warning. After a fault alarm occurs, perform the following steps;

2. Affected by variable speed, variable working conditions and environmental changes, the bearing temperature will fluctuate periodically. Based on expert knowledge and fault mechanism, six variable data affecting the drive-end bearing temperature, including winding temperature, speed, wind speed, active power, ambient temperature and cabin temperature, are selected for analysis;

3. The SCADA data of 24 faulty units and 75 healthy units in 4MW wind turbines from 2016 to 2019 were selected as the training set to train the proposed LSTM-WPHM hybrid model;

4. Use the SCADA data of 24 faulty units and 20 healthy units in 2020 as a test set to verify the model and method;

5. Use the test set data to verify the present invention;

6. Life expectancy prediction results.

As shown in Table 1, the overall accuracy of the unit alarm is 78.7%, the false alarm rate is 4.3%, and the average advance time for fault prediction is 13 days and 7 hours.

Table 1 Test set unit alarm results

As shown in Figure 4, Figure 4 is an analysis case of two units that were missed in the test set. It can be seen that the model did not give an alarm at the time of the unit failure. From the original bearing temperature, it can be seen that the temperature definitely exceeded the system alarm warning. After model analysis, the bearing status index did not change significantly at the time of the failure. The reason may be that the bearing temperature exceeding the limit is mainly affected by other factors, not the influencing factors considered by the model or the unit's own failure, so the model did not give an alarm.

In summary, the above method can realize fault warning and life prediction of the unit and reduce the false alarm of the unit. At the same time, it can identify the system underreporting caused by environmental factors, thereby proving the accuracy and effectiveness of the model and method proposed in the present invention for predicting the remaining life of the unit under abnormal conditions.

Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims (8)

1. A fan-oriented fault early warning and life prediction method is characterized by comprising the following steps:

s1, acquiring original operation data in the operation process of a fan through a sensor or a data acquisition device;

S2, carrying out data preprocessing on the original data;

s3, dividing the preprocessed data into a training set and a testing set, wherein the training set and the testing set are obtained by dividing historical health data of the unit according to the proportion of 85% to 15%;

S4, predicting by using OAKR models and performing fault early warning based on MSE indexes and multivariate Bayesian indexes;

S5, after fault early warning, predicting the residual life by using an LSTM-WPHM model.

2. The fan-oriented fault early warning and life prediction method according to claim 1, wherein: in step S2, the data preprocessing includes null value processing, infinite value processing, start-stop data deletion, seasonal factor elimination, data normalization and wavelet Bao Beishe S denoising.

3. The fan-oriented fault early warning and life prediction method according to claim 1, wherein: in step S4, the method specifically includes the following steps:

S41, model training: dividing 85% of historical health data for model training, sampling the first 70% of data by adopting an equidistant sampling method, and generating a memory matrix; optimizing the kernel function bandwidth h in OAKR algorithm on the last 15% of data by adopting a simplex optimization method;

S42, model test: the trained OAKR model is used for model test to obtain the performance parameters of the model in a healthy state, and a foundation is provided for model early warning;

S43, model prediction: and inputting the preprocessed data to be detected into a trained OAKR model, outputting a predicted value, calculating a mean square error MSE and a Bayesian factor alarm index between the predicted value and a true value, and giving an alarm when the threshold value is exceeded.

4. The fan-oriented fault early warning and life prediction method according to claim 1, wherein: in step S4, OAKR method steps include:

The first step, calculate the distance between monitor vector v and each memory vector X _i, get a distance vector d of n×1, calculate by euclidean distance:

The second step, calculating the weight w through the obtained distance matrix d and Gaussian kernel function, wherein w is an n multiplied by 1 vector matrix, and each element is calculated by the following formula:

Where h is the kernel bandwidth, determining the smoothness of the kernel, and small bandwidth h may represent more details but often results in tail undersmooth; conversely, a large bandwidth h often loses detail, resulting in a too smooth portion of the severe variation;

thirdly, optimizing the bandwidth h by using a simplex algorithm;

Fourth, the obtained weight w is used as a predicted value of the monitoring vector v The weighted average value of each memory vector X _i is calculated as follows:

5. The fan-oriented fault early warning and life prediction method according to claim 1, wherein: in step S5, the method specifically includes the following steps:

s51, related variable screening: according to expert knowledge and failure mechanism, screening out related variables influencing the temperature change of the bearing, and determining the degree of correlation between each variable and the temperature of the bearing through correlation analysis;

s52, training an LSTM model: training an LSTM bearing temperature prediction model by utilizing variable data in a unit health state, and fitting a nonlinear mapping relation between related variables and bearing temperature;

S53, LSTM model verification: verifying the trained LSTM bearing temperature prediction model, analyzing whether the training process of the model is converged, evaluating whether the mean square error and the decision coefficient of the model meet the requirements, and if the model does not meet the precision requirements, retraining the model;

s54, determining a bearing state index: if the LSTM bearing temperature prediction model meets the requirement, calculating a predicted value of the bearing temperature in a healthy state, and taking the residual error of the actual value and the predicted value as a bearing state index;

S55, WPHM modeling: according to the unit fault information, selecting bearing state indexes of a training set fault unit and a healthy unit, and establishing WPHM models;

s56, determining a model threshold value: reliability analysis is carried out on units in the training set by utilizing an LSTM-WPHM model, and a failure threshold F _f and an alarm threshold F _a are given by comparing differences of a fault unit and a healthy unit;

s57, bearing fault alarm: analyzing the unit by using an LSTM-WPHM model, calculating the cumulative failure probability F (t, X) of the unit at each moment, and giving out model alarm by combining with threshold setting;

S58, predicting the service life of the bearing: when F (t, X) exceeds an alarm threshold F _a, predicting the residual life of the unit fault, fitting a unit degradation failure curve according to the change trend of F (t, X) before model alarm, predicting the moment when F (t, X) reaches the failure threshold F _f, and combining the confidence interval of the parameter estimation of the fitting curve to give a residual life interval.

6. The fan-oriented fault early warning and life prediction method according to claim 5, wherein: in step S52, the LSTM model is composed of a 2-layer LSTM neural network, a 2-layer Dropout layer, and a 1-layer Dense layer;

The number of nodes of the first layer LSTM neural network in the LSTM model is 64, the activation function is Relu functions, the number of nodes of the second layer LSTM neural network in the LSTM model is 128, the activation function is Relu functions, the number of nodes of the Dense layer is 1, and the Dropout layer is added after each layer of LSTM network.

7. The fan-oriented fault early warning and life prediction method according to claim 5, wherein: in step S54, the residual δ between the predicted value and the actual value is used as an index for representing the bearing state, and the bearing state index δ is used as the input of WPHM model, and the calculation formula is as follows:

δ＝y_true-y_pre

Where y _true is an actual bearing temperature value and y _pre is a predicted bearing temperature value.

8. The fan-oriented fault early warning and life prediction method according to claim 5, wherein: in step S58, an exponential function is used to fit the trend, and the expression is:

y(t)＝ae^t+b

wherein, t is the running time of the unit bearing, and when the failure threshold F _f is reached, the residual life t _RUL of the bearing is as follows:

t_RUL＝t_pf-t_a；

Where t _a is the bearing failure warning time and t _pf is the predicted unit downtime.