WO2023231995A1 - 一种基于迁移学习的航空发动机寿命预测与健康评估方法 - Google Patents
一种基于迁移学习的航空发动机寿命预测与健康评估方法 Download PDFInfo
- Publication number
- WO2023231995A1 WO2023231995A1 PCT/CN2023/096973 CN2023096973W WO2023231995A1 WO 2023231995 A1 WO2023231995 A1 WO 2023231995A1 CN 2023096973 W CN2023096973 W CN 2023096973W WO 2023231995 A1 WO2023231995 A1 WO 2023231995A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- health
- data
- neural network
- aeroengine
- network model
- Prior art date
Links
- 230000036541 health Effects 0.000 title claims abstract description 95
- 238000000034 method Methods 0.000 title claims abstract description 85
- 238000013526 transfer learning Methods 0.000 title claims abstract description 24
- 230000008569 process Effects 0.000 claims abstract description 45
- 238000013480 data collection Methods 0.000 claims abstract description 13
- 238000010276 construction Methods 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 11
- 230000009467 reduction Effects 0.000 claims abstract description 10
- 238000000605 extraction Methods 0.000 claims abstract description 8
- 238000013500 data storage Methods 0.000 claims abstract description 7
- 238000003062 neural network model Methods 0.000 claims description 51
- 230000007423 decrease Effects 0.000 claims description 27
- 238000006731 degradation reaction Methods 0.000 claims description 20
- 230000003862 health status Effects 0.000 claims description 20
- 230000007613 environmental effect Effects 0.000 claims description 19
- 238000012549 training Methods 0.000 claims description 19
- 230000015556 catabolic process Effects 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 12
- 230000002159 abnormal effect Effects 0.000 claims description 11
- 238000013508 migration Methods 0.000 claims description 11
- 230000005012 migration Effects 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 239000002826 coolant Substances 0.000 claims description 8
- 239000000446 fuel Substances 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 7
- 238000002485 combustion reaction Methods 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 230000003068 static effect Effects 0.000 claims description 4
- 239000000654 additive Substances 0.000 claims description 3
- 230000000996 additive effect Effects 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000011425 standardization method Methods 0.000 claims description 3
- 238000012935 Averaging Methods 0.000 claims description 2
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 claims description 2
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 claims description 2
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 230000000306 recurrent effect Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 6
- 210000002569 neuron Anatomy 0.000 description 5
- 230000000875 corresponding effect Effects 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000010006 flight Effects 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Definitions
- the present invention relates to the field of complex equipment life prediction, and in particular to an aerospace engine life prediction and health assessment method based on transfer learning.
- the remaining useful life (RUL) prediction and health status assessment of key life-limiting parts are urgent problems to be solved; the relevant methods for predicting the remaining useful life (RUL) of key life-limiting parts can be roughly divided into model-based approach and data-driven approach.
- the model-based method mainly relies on the physical decay model of the system to analyze the dynamic characteristics, mechanical structural characteristics and material characteristics of mechanical equipment.
- establishing a complete physical model of an aeroengine requires consideration of complex physical, chemical and aerothermodynamic processes, which requires huge human resources.
- the data-driven method no longer relies on mechanical knowledge, but can evaluate the health status and predict the life based on the working status parameters and environmental parameters of the equipment.
- the data-driven method converts rough sensor data into useful information that can be learned, and establishes the corresponding relationship between data and tags, making it easier to predict the remaining service life and health status assessment of critical life-limited parts.
- the purpose of the present invention is to overcome the deficiencies in the existing technology and provide an aerospace engine life prediction and health assessment method based on transfer learning.
- This aerospace engine life prediction and health assessment method based on transfer learning includes the following steps:
- Step 1 Data collection and data storage: Arrange sensors in the data collection module, and the sensors collect the working status parameters and environmental parameters of key life-limiting parts of the aeroengine; the data storage module uses sensor categories as columns and data collection time as rows. Make the collected working status parameters and environmental parameters into tables and store them in the aeroengine working status database;
- Step 2 Data preprocessing: The data preprocessing module processes the working status parameters and environmental parameters collected by the data collection module for missing values, outliers and standardization; the flight status of each aircraft is divided based on the single flight duration of each aircraft as a standard; Flight status refers to engine damage at different flight lengths;
- Step 3 Construction of health factors: First, select the efficiency and air mass flow of key life-limiting parts of the aeroengine as performance indicators; then normalize the performance indicators, select the minimum value among all performance indicators as the health factor, and obtain the health factor Curve; finally smooth the health factor curve;
- Step 4 The data set construction module uses the XGBoost model to extract key performance indicators and perform dimensionality reduction on the key performance indicators;
- Step 5 Intersect the health factor with the normalized reduction speed v, and divide the health state of the aeroengine into an initial decline process, a normal decline process and an abnormal decline process; reduce The small speed v is the slope value of the health factor curve at the current moment; establish life and health status labels, and finally slice the data to create a time series data set;
- Step 6 Build and train a stacked GRU neural network model to extract features and predict time series data
- Step 7 Use the automatic feature extraction capability of the stacked GRU neural network model to mine multi-dimensional time series data to mine useful features related to life and health status, and predict the remaining life and health assessment of the aeroengine;
- Step 8 Use the transfer learning strategy to generalize the stacked GRU neural network model to different flight conditions: train the stacked GRU neural network model for one of the operating conditions, and then use the underlying neural network layer parameters of the trained stacked GRU neural network model to Freeze, fine-tune the parameters of the high-level network and output layer for data under different working conditions, and compare the recognition accuracy with the non-fine-tuned stacked GRU neural network model to test the success of the migration.
- the key life-limiting parts in step 1 include fans, high-pressure compressors, low-pressure compressors, high-pressure turbines, and low-pressure turbines;
- Working state parameters and environmental parameters include fan inlet pressure P 1 , fan inlet temperature T 1 , fan outlet flow rate W 2 , fan outlet pressure P 2 , branch duct pressure P 3 , LPC outlet pressure P 4 , LPC outlet temperature T 4 , HPC outlet pressure P 5 , HPC outlet static pressure P s5 , HPC outlet temperature T 5 , fuel flow rate W f , HPT coolant flow rate W 6 , LPT coolant flow rate W 7 , combustion chamber pressure P b , combustion chamber temperature T b , LPC inlet flow W 8 , HPC inlet flow W 9 , HPT outlet flow W 10 , HPT outlet temperature T 10 , HPT outlet pressure P 10 , LPT outlet flow W 11 , LPT outlet pressure P 11 , LPT outlet temperature T 11 , fan speed N f , physical core speed Nc, fan assembly margin, LPC assembly margin, HPC assembly margin, HPC inlet fuel flow ratio, aircraft flight altitude, Mach number and throttle resolver angle.
- the data preprocessing module in step 2 uses the mean completion method to process missing values for the collected working status parameters and environmental parameters, and fills in the gaps by averaging the values on both sides of the missing values; for the collected working status parameters
- the outliers are discarded directly, where outliers refer to values that deviate from the range of the state parameters when the equipment is working normally;
- the Z-score standardization method is used for standardization, and the formula is:
- ⁇ and ⁇ are the mean and variance of sample X respectively, X represents the data before standardization, and X ⁇ represents the data after standardization.
- step 3 As a preference, in step 3:
- the performance indicators are normalized using max-min normalization, and the calculation formula is:
- x is the original value of the sample
- x ⁇ is the normalized result
- x max and x min are the minimum and maximum values of all performance indicators respectively;
- ⁇ a (t) is the health factor at time t
- t b represents the b-th power at time t
- t s represents the time when the aircraft engine is put into service
- ⁇ n (t s ) is the initial wear amount of the aircraft engine
- the XGBoost model is an additive model composed of n base models. Assuming that the tree model to be trained at the t-th iteration is f t (x), then the prediction result at the t-th iteration satisfy:
- step 5 the step 5:
- the initial decline process refers to the process of mild initial wear and tear of key parts of the aeroengine;
- the normal decline process refers to the process of aeroengine performance decline caused by mild initial wear;
- the abnormal decline process refers to the performance degradation of the aeroengine due to the failure of key equipment the process of speeding up;
- T s When slicing the data, take T s as a time step, and conduct data on different health states respectively. Perform the slicing operation, and finally obtain an input sample set of N ⁇ T s ⁇ C, where N is the number of samples and C is the sample dimension.
- step 6 specifically includes the following steps:
- Step 6.1 Use Python language to build a stacked GRU neural network model:
- the GRU neural network model includes an input layer, multiple intermediate layers and an output layer, which are connected in turn to the input layer, intermediate layers and output layers;
- the input layer feature map group is a multi-dimensional array, and the input sample format is N ⁇ T s ⁇ C; T s is a time step, N is the number of samples, and C is the sample dimension;
- the hidden layer contains 3 pairs of stacked GRU-Dropout layers and a Flatten layer;
- the structural unit of the GRU neural network model includes update gate Z t and reset gate R t , and update gate Z t is used to control the current state H t from the historical state H The amount of information retained in t-1 and the amount of new information accepted from the candidate state H t ⁇ ; the reset gate R t is used to control whether the calculation of the candidate state H t ⁇ depends on the historical state H t-1 ;
- _ _ The input weight, historical state weight and bias of the state; W r , U r and b r are the input weight, historical state weight and bias of the reset gate respectively; ⁇ and tanh are nonlinear activation functions;
- the stacked GRU-Dropout layer randomly discards the parameters of the upper layer; finally, the output of the stacked GRU-Dropout layer is input into the Flatten layer, and after the dimensionality is reduced, it is converted into a one-dimensional vector; the one-dimensional vector is output through the output layer to predict the sequence result. ;
- Step 6.2 Train the stacked GRU neural network model to perform feature extraction and prediction on time series data:
- the root mean square error e RMSE the mean absolute error e MAPE and the correlation coefficient R 2 are used to measure the difference between the predicted value and the true value of the aero-engine life prediction.
- the differences are calculated as follows:
- y i is the real value, is the predicted value, is the mean, N is the number of predicted values;
- the accuracy rate is used to characterize the proportion of the number of correctly classified samples of the GRU neural network model to the total number of samples.
- the beneficial effects of the present invention are: firstly, it proposes a construction process of health status dividing factors; secondly, it uses the extreme gradient boosting (XGBoost) regression model, which is less affected by extreme bias values and has higher generalization, to extract key performance parameters and perform data analysis. Dimensionality reduction; then use the automatic feature extraction capability of the Gated Recurrent Unit (GRU) network for multi-dimensional time series data to mine useful features related to lifespan and health status to achieve remaining lifespan prediction and health assessment; and finally use transfer learning strategies to generalize the model To achieve different flight conditions, multi-operating mode model migration is realized; the invention efficiently utilizes the historical operation data resources of the entire life cycle of the aeroengine to provide a reliable basis for the life prediction and health assessment of the aeroengine.
- XGBoost extreme gradient boosting
- Figure 1 is a flow chart of aircraft engine life prediction and health assessment
- Figure 2 is a schematic diagram of the sensor layout points of the turbocharged engine
- Figure 3 is a diagram of the health factor construction process
- Figure 4 is a histogram of the importance analysis results of key performance parameters based on XGBoost
- Figure 5 is a diagram of the health status classification process
- Figure 6 is a simplified diagram of the stacked GRU network structure
- Figure 7-1 shows the prediction results of the remaining service life of aeroengines during short-distance flights
- Figure 7-2 shows the prediction results of the remaining service life of the aero-engine during mid-flight
- Figure 7-3 shows the prediction results of the remaining service life of aeroengines during long-distance flights
- Figure 8 is a diagram of health status assessment results obtained by the embodiment of the present invention.
- Embodiment 1 of this application provides an aerospace engine life prediction and health assessment method based on transfer learning as shown in Figure 1:
- Step 1 Data collection and data storage: Arrange sensors in the data collection module, and the sensors collect the working status parameters and environmental parameters of key life-limiting parts of the aeroengine; the data storage module The block uses Sql server database technology, with sensor categories as columns and data collection time as rows. The collected working status parameters and environmental parameters are tabulated and stored in the aero-engine working status database to achieve data interaction and effective storage; aero-engine On the one hand, the working status database interacts with users and the cloud to accept data from users, cache data for users in advance, and upload data to the cloud. On the other hand, it stores some historical data and provides training samples for the deep learning module.
- Blades Key life-limiting parts include fans, high-pressure compressors, low-pressure compressors, high-pressure turbines, and low-pressure turbines; working state parameters and environmental parameters include fan inlet pressure P 1 , fan inlet temperature T 1 , fan outlet flow rate W 2 , and fan outlet pressure P 2.
- LPT coolant flow W 7 combustion chamber pressure P b , combustion chamber temperature T b , LPC inlet flow W 8 , HPC inlet flow W 9 , HPT outlet flow W 10 , HPT outlet temperature T 10 , HPT outlet pressure P 10 , LPT outlet flow W 11 , LPT outlet pressure P 11 , LPT outlet temperature T 11 , fan speed N f , physical core speed Nc, fan assembly margin (Smfan), LPC assembly margin (SmLPC), HPC assembly margin ( SmHPC), HPC inlet fuel flow ratio (phi), aircraft flight altitude (Alt), Mach number (Mach) and throttle resolver angle (TRA).
- Smfan fan assembly margin
- SmLPC LPC assembly margin
- SmHPC HPC inlet fuel flow ratio
- phi aircraft flight altitude
- Mach number Mach number
- TRA throttle resolver angle
- the data preprocessing module processes the working status parameters and environmental parameters collected by the data acquisition module for missing values, abnormal values and standardization processing; the flight status is determined by the total flight duration and flight time of the aircraft in one flight cycle. Determined by altitude, and since there is a positive correlation between flight altitude and flight duration in the data set, the present invention uses the flight duration of a single journey of each aircraft as a standard to divide the flight status, thereby distinguishing the damage conditions between engines with different flight mileage; The status is the damage of the engine with different flight lengths; the data preprocessing module uses the mean completion method to process the missing values of the collected working status parameters and environmental parameters, and fills in the gaps through the average of the values on both sides of the missing value; the collected values are When outlier processing is performed on the working status parameters and environmental parameters, the outliers are discarded directly, where the outliers refer to values that deviate from the range of the status parameters when the equipment is working normally; the Z-score standardization method is used for standardization processing, and the
- ⁇ and ⁇ are the mean and variance of sample X respectively, and X represents standardization processing.
- the data before, X ⁇ represents the data after normalization; the data under different damage modes and flight conditions need to be standardized.
- Step 3 Construction of health factors: First, select the efficiency (e) and air mass flow (f) of the key life-limiting parts of the aeroengine as performance indicators. As the service time of the aeroengine increases, the 10 performance indicators have degradation trends at different speeds. Among them, the current performance index with the largest amount of degradation determines the performance of the aeroengine; then the performance index is normalized, and the minimum value among all performance indexes (10 performance indexes) is selected as the health factor to obtain the health factor curve; finally, the The health factor curve is smoothed.
- the performance indicators are normalized using max-min normalization, and the calculation formula is:
- x is the original value of the sample
- x ⁇ is the normalized result
- x max and x min are the minimum and maximum values of all performance indicators respectively.
- ⁇ a (t) is the health factor at time t
- t b represents the b-th power at time t
- t s represents the time when the aircraft engine is put into service
- the data set construction module uses the XGBoost model (extreme gradient boosting regression model), which is less affected by extreme bias values and has higher generalization, to extract key performance indicators and perform dimensionality reduction on the key performance indicators; the XGBoost model fits each Based on the importance of performance parameters on health factors, the highest 12 key performance parameters are extracted as the input of the subsequent deep learning model; the XGBoost model is an optimized distributed gradient boosting library, and the internal decision tree uses a regression tree, which is efficient and Flexible and portable; it classifies elements in intervals according to the data characteristics of the elements in the sample space. Each classification forms tree branches, and the combination of branches after several iterations forms a regression tree model; each tree model Each contains several internal nodes and leaf nodes. The head node divides the current space into two parts, and the leaf nodes are the corresponding space results after the division.
- XGBoost model extreme gradient boosting regression model
- the XGBoost model is an additive model composed of n base models. Assuming that the tree model to be trained at the t-th iteration is f t (x), then the prediction result at the t-th iteration satisfy:
- the remaining service life refers to the time period that an aero engine goes through after a major overhaul and the aircraft flies normally for several rounds to the next major overhaul.
- the present invention takes 100 rounds as the longest interval point for major overhaul.
- n n ⁇ 100 rounds of aircraft flight and cannot guarantee the normal flight of the aircraft
- the end of life is taken as n.
- the aircraft flight reaches 100 rounds, but The engine still needs to be overhauled when it can still meet the flight requirements. At this time, 100 rounds is taken as the end of life.
- the starting point of life is recorded as 100%, the end of life is recorded as 0, and the lifespan degradation curve is positively correlated with the health factor.
- the health factor is intersected with the normalized reduction speed v, and the health status of the aeroengine is divided into an initial decline process, a normal decline process and an abnormal decline process;
- the reduction speed v is the current
- the slope value of the health factor curve at any time when the slope is larger, it indicates that the health factor decreases faster, which in turn indicates that the performance of the aeroengine declines faster at this time; establish life and health status labels, and finally slice the data to create a time series Data set;
- the initial decay process refers to the inevitable mild initial wear of key parts of the aeroengine due to manufacturing and assembly tolerances;
- the normal decay process refers to the aeroengine performance decline process caused by mild initial wear. This process of engine It has not received serious damage and is in normal working condition;
- the abnormal degradation process refers to the process in which the performance degradation of aeroengines accelerates due to the failure of key equipment.
- T s When slicing the data, take T s as a time step, perform slicing operations on the data of different health states, and finally obtain an input sample set of N ⁇ T s ⁇ C, where N is the number of samples, C is the sample dimension, and the label Corresponds to the RUL and health status at the end of the time series data.
- Step 6 Build and train a stacked GRU neural network model to extract features and predict time series data.
- Step 6.1 Use Python language to build a stacked GRU neural network model:
- the GRU neural network model includes an input layer, multiple intermediate layers (hidden layers) and an output layer (the specific network structure can be adjusted according to the specific data scale), which are connected in turn to the input layer, intermediate layers and output layers.
- the input layer feature mapping group is a multi-dimensional array (the dimension is the number of key performance parameters), and the input sample format is N ⁇ T s ⁇ C; T s is a time step, N is the number of samples, and C is the sample dimension.
- the hidden layer contains 3 pairs of stacked GRU-Dropout layers and a Flatten layer;
- GRU is an automatic mining time series data feature, which is widely used in various equipment PHM systems, such as the identification of vibration, acoustic and temperature signals;
- GRU is relatively Compared with the traditional LSTM network, the structure is more simplified and the training time is less, which avoids the information redundancy between the "doors" inside the LSTM and is more conducive to processing aeronautical data sets with a long time;
- the structural units of the GRU neural network model include Update gate Z t and reset gate R t , update gate Z t is used to control the amount of information retained by the current state H t from the historical state H t-1 and the amount of new information accepted from the candidate state H t ⁇ ; reset Gate R t is used to control whether the calculation of candidate state H t ⁇ depends on the historical state H t-1 .
- _ _ The input weight, historical state weight and bias of the state; W r , U r and br are the input weight, historical state weight and bias of the reset gate respectively; ⁇ and tanh are nonlinear activation functions.
- Stacking GRU-Dropout layers randomly discards the parameters of the upper layer, thereby reducing the neural network The complexity of the network improves the training efficiency; finally, the output of the stacked GRU-Dropout layer is input into the Flatten layer, and after the dimensionality is reduced, it is converted into a one-dimensional vector; the one-dimensional vector is output through the output layer to predict the sequence result; the output layer is A fully connected layer (Dense).
- the Dense layer is set with 1 neuron and no nonlinear activation function is set; for the health assessment module, the Dense layer is set with 3 neurons (corresponding to 3 health states) and nonlinear
- the activation function is softmax; L1 and L2 regularization terms are added to the Dense layer to optimize the model convergence process and prevent overfitting.
- Step 6.2 Train the stacked GRU neural network model to perform feature extraction and prediction on time series data:
- the set ratio usually 4:1, when the amount of data is large, the test set ratio can be appropriately increased
- set a single Send data batches 32 samples, or integer multiples of 32
- an early stopping command early stopping
- the test set accuracy does not increase, stop training
- save the stacked GRU neural network model monitor the changes in the loss function value of the stacked G
- Step 7 Use the automatic feature extraction capability of the stacked GRU neural network model to mine multi-dimensional time series data to mine useful features related to life and health status, and predict the remaining life and health assessment of the aeroengine; aviation engine life prediction is a regression problem, and health assessment It is a classification problem.
- the root mean square error eRMSE , the mean absolute error eMAPE and the correlation coefficient R2 are used to measure the difference between the predicted value and the true value of the aeroengine life prediction (regression problem).
- the calculation formulas are as follows:
- y i is the real value
- N is the number of predicted values.
- the accuracy rate is used to characterize the proportion of the number of correctly classified samples of the GRU neural network model to the total number of samples.
- Step 8 Transfer learning refers to transferring the knowledge learned in a certain field (source domain) to the machine learning strategy applied in another similar field (target domain); the present invention uses data in a single flight state to train the model, using The transfer learning strategy generalizes the stacked GRU neural network model to different flight conditions to achieve GRU network model migration under multiple operating conditions: treat different aero-engine operating conditions as different fields, and train the stack for one of the operating conditions.
- the GRU neural network model and then use the data of other working conditions to fine-tune the parameters of the high-level neural network layer of the model based on the model fine-tuning method, thereby achieving the purpose of saving computer resources and improving efficiency;
- the underlying neural network layer of the stacked GRU neural network model that has been trained is The parameters of the network layer (input layer, part of the GRU layer) are frozen. According to the data of different working conditions, the parameters of the high-level network and output layer are fine-tuned, and the recognition accuracy of the stacked GRU neural network model without fine-tuning is compared to test the success of the migration. , improve calculation efficiency.
- Embodiment 2 of this application provides specific applications of the method in Embodiment 1:
- Step1 Data collection (this article is the whole life cycle operation data of the turbocharged engine, 32 sensors are deployed), and the collected raw data is stored in the Sql server database.
- Step2 Data preprocessing, perform outlier processing, missing value completion, and standardization processing on the data, convert the data into data types that can be used for supervised learning, and classify flight states according to different flight parameters.
- the specific standards are as shown in Table 1:
- Step3 Construct a health factor curve based on the health factor establishment standards to characterize the health state degradation process of the aeroengine, and use the XGBoost model to analyze the importance of performance parameters, extract key performance parameters, divide the health state and establish life labels.
- Step4 Divide the source domain data set (single working condition data) into a training set at a ratio of 4:1.
- the test set and the target domain data set (other working condition data) are set as the validation set.
- Step5 Build a stacked GRU model and substitute the training set data into the model for training. Use the test set to check the changes in error and accuracy, and save the trained model.
- Step 6 Lock the low-level network layer of the saved model, and use the small batch data of the verification set to fine-tune the parameters of the high-level network layer to achieve the effect of model migration and verify the success of the migration.
- fan assembly margin Smfan
- LPC assembly margin SmLPC
- HPC assembly margin SmHPC
- phi HPC inlet fuel flow ratio
- Alt aircraft flight altitude
- Mach number Mach number
- throttle throttle
- TRA resolver angle
- FIG. 3 it is a process diagram for constructing health factors.
- the present invention uses the efficiency (e) and air mass flow (f) of five key life-limiting parts as performance indicators, normalizes them, and then takes the minimum value at each moment. The value is the current health factor value. Finally, the obtained health factor discrete value is fitted into a smooth curve according to the aeroengine degradation relationship.
- FIG. 4 it is a diagram showing the importance analysis results of XGBoost performance parameters of the present invention.
- the main hyperparameters that need to be set in the XGBoost model are maximum depth, learning rate and parameter amount.
- the hyperparameter optimization method of the present invention when establishing the regression model adopts grid search.
- the search space is (maximum depth: [5,11], learning rate: [0.01,0.1], parameter amount: [100,400]), and adopts the average
- the root square error is used as the objective function, and three-fold cross-validation is implemented. It was finally determined that the optimal hyperparameters were the maximum depth of 8, the learning rate of 0.1, and the number of parameters of 300.
- the health state of the aeroengine can be divided into initial decline, normal decline and abnormal decline processes.
- Initial deterioration refers to the inevitable degradation of critical engine components due to manufacturing and assembly tolerances.
- the initial wear amount; normal decline refers to the process of engine performance decline caused by the initial wear amount.
- abnormal decline refers to the engine performance decline caused by the failure of key equipment.
- the present invention proposes a classification standard that combines health factors with degradation rate. The intersection of the two is used to establish the standard as shown in Table 2, in which the degradation rate is also standardized:
- Figure 6 shows the stacked GRU network model structure.
- the number of neurons is 64, 128, and 64.
- Each GRU layer is followed by a Dropout layer, which is used to discard 20% of the hyperparameters.
- the last layer is Dropout.
- the layer is connected to a Flatten layer, which straightens the feature map into a one-dimensional sequence, and the input and output layer (Dense) performs prediction and recognition.
- the Dense layer is set with 1 neuron and no nonlinear activation function is set; for the health assessment module, the Dense layer is set with 3 neurons (corresponding to 3 health states), and the nonlinear activation function is softmax.
- the Dense layer adds L1 and L2 regularization terms to optimize the model convergence process and prevent overfitting.
- the model optimization method uses the Adam algorithm and sets an early stopping command. When the test set accuracy does not increase, the training is stopped and the model is saved.
- the main hyperparameters that need to be set for the stacked GRU model are the learning rate and the mini-batch training batch size. Grid search is also used to optimize the parameters.
- the search space is (learning rate: [0.001, 0.01], batch size: [32,256]), and the root mean square error is used as the objective function.
- the optimal hyperparameters were finally determined: the learning rate was 0.001, and the mini-batch training batch size was 128.
- the state identification results are shown in Figure 8.
- the vertical direction is the true label of the sample, and the horizontal direction is the model predicted label.
- the health assessment accuracy rates of the model under the three working conditions are 94.894%, 96.418%, and 99.216% respectively.
- the recognition accuracy of the three working conditions is average. Both reached 99.036%, almost 100%, indicating that the performance parameters of aeroengines in the abnormal degradation stage are highly distinguishable.
- the system is still in good working condition, so the difference in data characteristics is small, but the model can still distinguish them with a high accuracy.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Computational Mathematics (AREA)
- Mathematical Physics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Software Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Computer Hardware Design (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Algebra (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Databases & Information Systems (AREA)
- Operations Research (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
本发明涉及一种基于迁移学习的航空发动机寿命预测与健康评估方法,包括步骤:数据采集和数据储存;数据预处理;健康因子构建。本发明的有益效果是:本发明首先提出健康状态划分因子的构建流程;其次采用受极端偏值影响较小、泛化性更高的极端梯度提升回归模型提取关键性能参数,对数据进行降维;然后利用门控循环单元网络对多维时序数据的自动特征提取能力挖掘关联寿命与健康状态的有用特征,实现剩余使用寿命预测与健康评估;最终利用迁移学习策略将模型泛化至不同的飞行状态下,实现多工况模型迁移;本发明高效利用航空发动机全寿命周期历史运行数据资源,为航空发动机的寿命预测与健康评估提供可靠的依据。
Description
本申请要求于2022年05月30日提交中国专利局、申请号为202210599974.0、发明名称为“一种基于迁移学习的航空发动机寿命预测与健康评估方法”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
本发明涉及复杂设备寿命预测领域,特别是涉及一种基于迁移学习的航空发动机寿命预测与健康评估方法。
作为近年来重要的民用出行与军事保障手段,航空飞机以及直升机的使用愈加广泛,由于航空发动机关键限寿件故障或衰退引起的通航一般事故也屡见不鲜。而直升机由于其复杂的动力传输结构和极端的服役环境,发生事故的概率远高于固定翼飞机。因此,开发和应用航空发动机智能运维与健康管理系统(Prognostics and health management,PHM)对于我国的航空安全事业十分重要。
在当前航空发动机PHM系统中,关键限寿件的剩余使用寿命(RUL)预测与健康状态评估是急需解决的问题;关键限寿件的剩余使用寿命(RUL)预测的相关方法可粗略划分为基于模型的方法和基于数据驱动的方法。
基于模型的方法主要是依托于系统的物理衰退模型,分析机械设备的动力学特性、机械结构特性以及材料特性。然而,建立一个完备的航空发动机物理模型需要考虑复杂的物理、化学和空气热力学过程,需要耗费巨大的人力资源。相反,基于数据驱动的方法则不再依托于机械知识,根据设备的工作状态参数和环境参数即可评估健康状态和预测寿命。基于数据驱动的方法将粗糙的传感器数据转化为可以学习的有用信息,建立数据与标签的对应关系,使关键限寿件的剩余使用寿命预测与健康状态评估更加容易实现。
随着人工智能技术的发展和应用,机器学习方法凭借其灵活高效的优点,逐渐被PHM系统研发人员采用。传统机器学习方法极依赖复杂的特征工程技术,而深度学习则免去了这一过程。深度学习依托的神经网络能够自动提取原始数据的深层特征。当前国内航空发动机PHM系统的研究主要着眼于航空发动机单个零部件的诊断,对发动机整体性能衰退评估的研究较少。对于航空发动机这类复杂的机械系统,几乎无研究提出一套完整的健康状态划分标准。因此建立健全的健康状态影响因子对于模型评估有着极大的促进作用。
发明内容
本发明的目的是克服现有技术中的不足,提供一种基于迁移学习的航空发动机寿命预测与健康评估方法。
这种基于迁移学习的航空发动机寿命预测与健康评估方法,包括以下步骤:
步骤1、数据采集和数据储存:在数据采集模块内布置传感器,由传感器采集航空发动机关键限寿件的工作状态参数与环境参数;数据存储模块以传感器类别为列,以采集数据时间为行,将采集的工作状态参数与环境参数制成表格存入航空发动机工作状态数据库;
步骤2、数据预处理:数据预处理模块将数据采集模块采集得到的工作状态参数与环境参数进行缺失值、异常值及标准化处理;以各个飞机的单次旅程飞行时长作为标准来划分飞行状态;飞行状态为不同飞行长度发动机的损伤情况;
步骤3、健康因子构建:首先选择航空发动机关键限寿件的效率和空气质量流量作为性能指标;然后对性能指标进行归一化处理,选取所有性能指标中的最小值作为健康因子,得到健康因子曲线;最终对健康因子曲线进行平滑处理;
步骤4、数据集构建模块采用XGBoost模型提取关键性能指标,对关键性能指标进行降维处理;
步骤5、将健康因子与标准化处理后的减小速度v取交集,将航空发动机的健康状态划分为初始衰退过程、正常衰退过程和异常衰退过程;减
小速度v为当前时刻健康因子曲线的斜率值;建立寿命与健康状态标签,最终对数据切片,制作时序数据集;
步骤6、搭建并训练堆叠GRU神经网络模型,对时间序列数据进行特征提取和预测;
步骤7、利用堆叠GRU神经网络模型对多维时序数据的自动特征提取能力挖掘关联寿命与健康状态的有用特征,对航空发动机进行剩余寿命预测与健康评估;
步骤8、利用迁移学习策略将堆叠GRU神经网络模型泛化至不同的飞行状态下:针对其中一个工况训练堆叠GRU神经网络模型,然后将训练完成的堆叠GRU神经网络模型的底层神经网络层参数冻结,针对不同工况的数据,对高层网络和输出层进行参数微调,并与未微调堆叠GRU神经网络模型的识别准确率对比,检验迁移的成功性。
作为优选,步骤1中关键限寿件包括风扇、高压压缩机、低压压缩机、高压涡轮和低压涡轮;
工作状态参数与环境参数包括风扇入口压力P1、风扇入口温度T1、风扇出口流量W2、风扇出口压力P2、支路导管压力P3、LPC出口压力P4、LPC出口温度T4、HPC出口压力P5、HPC出口静压力Ps5、HPC出口温度T5、燃料流量Wf、HPT冷却液流量W6、LPT冷却液流量W7、燃烧室压力Pb、燃烧室温度Tb、LPC入口流量W8、HPC入口流量W9、HPT出口流量W10、HPT出口温度T10、HPT出口压力P10、LPT出口流量W11、LPT出口压力P11、LPT出口温度T11、风扇速度Nf、物理核心速度Nc、风扇装配余量、LPC装配余量、HPC装配余量、HPC入口燃料流量比例、飞机飞行高度、马赫数和油门旋转变压器角度。
作为优选,步骤2中数据预处理模块采用均值补全法对采集得到的工作状态参数与环境参数进行缺失值处理,通过缺失值两侧值的平均数补全空缺;对采集得到的工作状态参数与环境参数进行异常值处理时,将异常值直接舍弃,其中异常值指偏离设备正常工作时状态参数范围的值;进行标准化处理时采用Z-score标准化方式,公式为:
上式中,μ和σ分别是样本X的均值和方差,X表示进行标准化处理前的数据,X`表示进行标准化处理后的数据。
作为优选,步骤3中:
采用max-min归一化对性能指标进行归一化处理,计算公式为:
上式中,x为样本原值,x`为归一化结果,xmax和xmin分别是所有性能指标中的最小值和最大值;
采用最小二乘法插值对离散的健康因子进行拟合,根据航空发动机性能衰退关系式对健康因子曲线进行平滑处理,得到最终健康因子平滑曲线;其中航空发动机性能衰退关系式为:
δa(t)=1-exp(atb)+δn(ts)+ξ
δa(t)=1-exp(atb)+δn(ts)+ξ
上式中,δa(t)为t时刻的健康因子,tb表示t时刻的b次方,ts表示航空发动机性投入使用的时刻;δn(ts)为航空发动机初始磨损量;其中a=U(0.001,0.003),b=U(1.4,1.6),ξ=N(0,0.001)。
作为优选,步骤4中,XGBoost模型是由n个基模型组成的加法模型,假设第t次迭代要训练的树模型是ft(x),则第t次迭代时的预测结果满足:
上式中:为前t-1棵树的预测结果,ft(xi)为第t棵树模型。
作为优选,步骤5中:
初始衰退过程指航空发动机关键零部件出现的轻度初始磨损的过程;正常衰退过程指由于轻度初始磨损引发的航空发动机性能衰退的过程;异常衰退过程指航空发动机由于关键设备故障引起的性能退化速度加快的过程;
对数据切片时,取Ts为一个时间步长,分别对不同健康状态的数据进
行切片操作,最终得到N×Ts×C的输入样本集合,其中N为样本数量,C为样本维度。
作为优选,步骤6具体包括以下步骤:
步骤6.1、采用Python语言搭建堆叠GRU神经网络模型:
GRU神经网络模型包括一个输入层,多个中间层和一个输出层,依次连接输入层、中间层和输出层;
输入层特征映射组为多维数组,输入样本格式为N×Ts×C;Ts为一个时间步长,N为样本数量,C为样本维度;
隐含层包含3对堆叠GRU-Dropout层和一个Flatten层;GRU神经网络模型的结构单元包括更新门Zt和重置门Rt,更新门Zt用来控制当前状态Ht从历史状态Ht-1中保留的信息量和从候选状态Ht`中接受新信息的量;重置门Rt用来控制候选状态Ht`的计算是否依赖历史状态Ht-1;
GRU神经网络模型的状态更新方式为:
Ht=Zt⊙Ht-1+(1-Zt)⊙Ht`
Ht=Zt⊙Ht-1+(1-Zt)⊙Ht`
其中更新门输出为:
Zt=σ(WzXt+UzHt-1+bz)
Zt=σ(WzXt+UzHt-1+bz)
候选状态Ht`为
Ht`=tanh(WhXt+Uh(Rt⊙Ht-1)+bh)
Ht`=tanh(WhXt+Uh(Rt⊙Ht-1)+bh)
其中重置门输出为:
Rt=σ(WrXt+UrHt-1+br)
Rt=σ(WrXt+UrHt-1+br)
上式中,Xt为t时刻GRU神经网络模型的输入,Wz、Uz和bz分别为更新门的输入权重、历史状态权重和偏置;Wh、Uh和bh分别为候选状态的输入权重、历史状态权重和偏置;Wr、Ur和br分别为重置门的输入权重、历史状态权重和偏置;σ和tanh为非线性激活函数;
堆叠GRU-Dropout层对上层的参数进行随机舍弃;最后将堆叠GRU-Dropout层的输出输入Flatten层,进行维数的削减后,转化为一维向量;将一维向量经过输出层输出预测序列结果;
步骤6.2、训练堆叠GRU神经网络模型,对时间序列数据进行特征提取和预测:
将建立的时序数据集输入待训练的GRU神经网络模型中,按设定比例划分训练集和测试集;设置单次训练送入数据批量;采用Adam算法对堆叠GRU神经网络模型进行优化,并设置提前中止命令;当测试集准确率不上升时,停止训练并保存堆叠GRU神经网络模型;实时监测堆叠GRU神经网络模型的损失函数值变化,最后以折线图的形式输出堆叠GRU神经网络模型的预测误差与准确率。
作为优选,步骤7中对航空发动机进行剩余寿命预测与健康评估时,采用均方根误差eRMSE、平均绝对误差eMAPE以及相关系数R2来衡量航空发动机寿命预测的预测值与真实值之间的差异,计算式分别如下:
上式中,yi为真实值,为预测值,为均值,N是预测值的个数;
进行航空发动机健康评估时,使用准确率来表征GRU神经网络模型的分类正确的样本个数占整体样本个数的比例。
本发明的有益效果是:首先提出一个健康状态划分因子的构建流程;其次采用受极端偏值影响较小、泛化性更高的极端梯度提升(XGBoost)回归模型提取关键性能参数,对数据进行降维;然后利用门控循环单元(GRU)网络对多维时序数据的自动特征提取能力挖掘关联寿命与健康状态的有用特征,实现剩余使用寿命预测与健康评估;最终利用迁移学习策略将模型泛化至不同的飞行状态下,实现多工况模型迁移;本发明高效利用航空发动机全寿命周期历史运行数据资源,为航空发动机的寿命预测与健康评估提供可靠的依据。
说明书附图
为了更清楚地说明本发明实施例或现有技术中的技术方案,下面将对实施例中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1为航空发动机寿命预测与健康评估的流程图;
图2为涡轮增压发动机传感器布置点示意图;
图3为健康因子构建过程图;
图4为基于XGBoost关键性能参数重要性分析结果直方图;
图5为健康状态划分过程图;
图6为堆叠GRU网络结构简图;
图7-1为短途飞行中航空发动机剩余使用寿命预测结果图;
图7-2为中途飞行中航空发动机剩余使用寿命预测结果图;
图7-3为长途飞行中航空发动机剩余使用寿命预测结果图;
图8为本发明实施例得到的健康状态评估结果图。
符号说明:
1-低压压缩机、2-高压压缩机、3-高压涡轮、4-低压涡轮
1-低压压缩机、2-高压压缩机、3-高压涡轮、4-低压涡轮
下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。
为使本发明的上述目的、特征和优点能够更加明显易懂,下面结合附图和具体实施方式对本发明作进一步详细的说明。
实施例一
本申请实施例一提供了一种如图1所示基于迁移学习的航空发动机寿命预测与健康评估方法:
步骤1、数据采集和数据储存:在数据采集模块内布置传感器,由传感器采集航空发动机关键限寿件的工作状态参数与环境参数;数据存储模
块利用Sql server数据库技术,以传感器类别为列,以采集数据时间为行,将采集的工作状态参数与环境参数制成表格存入航空发动机工作状态数据库,实现数据的交互和有效存储;航空发动机工作状态数据库一方面与用户和云端进行数据交互,实现接受来自用户的数据、提前为用户缓存数据、向云端上传数据,另一方面存储部分历史数据,为深度学习模块提供训练样本。
关键限寿件包括风扇、高压压缩机、低压压缩机、高压涡轮和低压涡轮;工作状态参数与环境参数包括风扇入口压力P1、风扇入口温度T1、风扇出口流量W2、风扇出口压力P2、支路导管压力P3、LPC出口压力P4、LPC出口温度T4、HPC出口压力P5、HPC出口静压力Ps5、HPC出口温度T5、燃料流量Wf、HPT冷却液流量W6、LPT冷却液流量W7、燃烧室压力Pb、燃烧室温度Tb、LPC入口流量W8、HPC入口流量W9、HPT出口流量W10、HPT出口温度T10、HPT出口压力P10、LPT出口流量W11、LPT出口压力P11、LPT出口温度T11、风扇速度Nf、物理核心速度Nc、风扇装配余量(Smfan)、LPC装配余量(SmLPC)、HPC装配余量(SmHPC)、HPC入口燃料流量比例(phi)、飞机飞行高度(Alt)、马赫数(Mach)和油门旋转变压器角度(TRA)。
步骤2、数据预处理:数据预处理模块将数据采集模块采集得到的工作状态参数与环境参数进行缺失值、异常值及标准化处理;飞行状态是由该航空飞机一次飞行周期的总飞行时长和飞行高度来决定的,又由于数据集中飞行高度与飞行时长存在正相关,本发明以各个飞机的单次旅程飞行时长作为标准来划分飞行状态,从而区别不同飞行里程的发动机之间的损伤情况;飞行状态为不同飞行长度发动机的损伤情况;数据预处理模块采用均值补全法对采集得到的工作状态参数与环境参数进行缺失值处理,通过缺失值两侧值的平均数补全空缺;对采集得到的工作状态参数与环境参数进行异常值处理时,将异常值直接舍弃,其中异常值指偏离设备正常工作时状态参数范围的值;进行标准化处理时采用Z-score标准化方式,公式为:
上式中,μ和σ分别是样本X的均值和方差,X表示进行标准化处理
前的数据,X`表示进行标准化处理后的数据;不同损伤模式和飞行状态下的数据均需要进行标准化操作。
步骤3、健康因子构建:首先选择航空发动机关键限寿件的效率(e)和空气质量流量(f)作为性能指标,伴随航空发动机的服役时长增加,10个性能指标存在不同速度的退化趋势,其中退化量最大的当前性能指标决定了航空发动机的性能;然后对性能指标进行归一化处理,选取所有性能指标(10个性能指标)中的最小值作为健康因子,得到健康因子曲线;最终对健康因子曲线进行平滑处理。
采用max-min归一化对性能指标进行归一化处理,计算公式为:
上式中,x为样本原值,x`为归一化结果,xmax和xmin分别是所有性能指标中的最小值和最大值。
采用最小二乘法插值对离散的健康因子进行拟合,根据航空发动机性能衰退关系式对健康因子曲线进行平滑处理,得到最终健康因子平滑曲线;其中航空发动机性能衰退关系式为:
δa(t)=1-exp(atb)+δn(ts)+ξ
δa(t)=1-exp(atb)+δn(ts)+ξ
上式中,δa(t)为t时刻的健康因子,tb表示t时刻的b次方,ts表示航空发动机性投入使用的时刻;δn(ts)为航空发动机初始磨损量,是发动机出厂即存在的不可避免的轻微损伤;其中a=U(0.001,0.003),b=U(1.4,1.6),ξ=N(0,0.001)。
步骤4、数据集构建模块采用受极端偏值影响较小、泛化性更高的XGBoost模型(极端梯度提升回归模型)提取关键性能指标,对关键性能指标进行降维处理;XGBoost模型拟合各个性能参数对健康因子的重要性影响,提取最高的12个关键性能参数作为后续深度学习模型的输入;XGBoost模型是一种经优化的分布式梯度提升库,内部决策树采用回归树,具有高效、灵活可移植的特点;其根据样本空间内元素的数据特征对元素进行区间分类,每一次分类均形成树状分枝,若干次迭代后的分枝组合便形成了回归树模型;每个树模型均包含若干个内部节点和叶节点,内
部节点将当前空间进行二分,叶节点则是划分后对应的空间结果。
XGBoost模型是由n个基模型组成的加法模型,假设第t次迭代要训练的树模型是ft(x),则第t次迭代时的预测结果满足:
上式中:为前t-1棵树的预测结果,ft(xi)为第t棵树模型。
步骤5、剩余使用寿命是指航空发动机在经过一次重大检修后经飞机正常飞行若干轮次后至下一次重大检修时所经历的时间周期。当关键限寿件的损伤或出现设备故障时,航空发动机的性能衰退加快,其RUL也相对变短。本发明以100轮为重大检修的最长间隔点,当发动机在飞机飞行n(n<100)轮出现故障无法保证飞机正常飞行时,取寿命终点为n,当飞机飞行轮次达到100轮但发动机依旧能够满足飞行要求时也需要检修,此时取100轮为寿命终点。以起始点寿命记作100%,寿命终止记作0,寿命退化曲线与健康因子成正相关。进一步地,健康状态标签构建方面,将健康因子与标准化处理后的减小速度v取交集,将航空发动机的健康状态划分为初始衰退过程、正常衰退过程和异常衰退过程;减小速度v为当前时刻健康因子曲线的斜率值;当斜率越大,表明健康因子减小速度越快,进而说明航空发动机此时性能衰退的速度也更快;建立寿命与健康状态标签,最终对数据切片,制作时序数据集;初始衰退过程指航空发动机关键零部件由于制造和装配公差出现的不可避免的轻度初始磨损的过程;正常衰退过程指由于轻度初始磨损引发的航空发动机性能衰退的过程,该过程发动机还未收到严重的损伤,处于正常工作状态;异常衰退过程指航空发动机由于关键设备故障引起的性能退化速度加快的过程。
对数据切片时,取Ts为一个时间步长,分别对不同健康状态的数据进行切片操作,最终得到N×Ts×C的输入样本集合,其中N为样本数量,C为样本维度,标签对应时序数据最末端时刻RUL和健康状态。
步骤6、搭建并训练堆叠GRU神经网络模型,对时间序列数据进行特征提取和预测。
步骤6.1、采用Python语言搭建堆叠GRU神经网络模型:
GRU神经网络模型包括一个输入层,多个中间层(隐含层)和一个输出层(具体网络结构可以根据特定的数据规模进行调整),依次连接输入层、中间层和输出层。
输入层特征映射组为多维数组(维度即为关键性能参数个数),输入样本格式为N×Ts×C;Ts为一个时间步长,N为样本数量,C为样本维度。
隐含层包含3对堆叠GRU-Dropout层和一个Flatten层;GRU是一种自动挖掘时序数据特征,被广泛地用于各类设备PHM系统中,例如振动、声学与温度信号的识别;GRU相对于传统的LSTM网络来说结构更加简化,训练时间较少,避免了LSTM内部“门”之间的信息冗余,更有利于处理时间长度大的航空数据集;GRU神经网络模型的结构单元包括更新门Zt和重置门Rt,更新门Zt用来控制当前状态Ht从历史状态Ht-1中保留的信息量和从候选状态Ht`中接受新信息的量;重置门Rt用来控制候选状态Ht`的计算是否依赖历史状态Ht-1。
GRU神经网络模型的状态更新方式为:
Ht=Zt⊙Ht-1+(1-Zt)⊙Ht`
Ht=Zt⊙Ht-1+(1-Zt)⊙Ht`
其中更新门输出为:
Zt=σ(WzXt+UzHt-1+bz)
Zt=σ(WzXt+UzHt-1+bz)
候选状态Ht`为
Ht`=tanh(WhXt+Uh(Rt⊙Ht-1)+bh)
Ht`=tanh(WhXt+Uh(Rt⊙Ht-1)+bh)
其中重置门输出为:
Rt=σ(WrXt+UrHt-1+br)
Rt=σ(WrXt+UrHt-1+br)
上式中,Xt为t时刻GRU神经网络模型的输入,Wz、Uz和bz分别为更新门的输入权重、历史状态权重和偏置;Wh、Uh和bh分别为候选状态的输入权重、历史状态权重和偏置;Wr、Ur和br分别为重置门的输入权重、历史状态权重和偏置;σ和tanh为非线性激活函数。
堆叠GRU-Dropout层对上层的参数进行随机舍弃,从而降低神经网
络的复杂度,提高训练效率;最后将堆叠GRU-Dropout层的输出输入Flatten层,进行维数的削减后,转化为一维向量;将一维向量经过输出层输出预测序列结果;输出层为一个全连接层(Dense),对于寿命预测模块,Dense层设置1个神经元,不设置非线性激活函数;对于健康评估模块,Dense层设置3个神经元(对应3中健康状态),非线性激活函数为softmax;Dense层增添L1、L2正则化项,优化模型收敛过程,防止过拟合。
步骤6.2、训练堆叠GRU神经网络模型,对时间序列数据进行特征提取和预测:
将建立的时序数据集输入待训练的GRU神经网络模型中,按设定比例划分训练集和测试集(一般为4:1,数据量较大时可以适当提高测试集占比);设置单次训练送入数据批量(32个样本,或32的整数倍);采用Adam算法对堆叠GRU神经网络模型进行优化,并设置提前中止命令(early stopping);当测试集准确率不上升时,停止训练并保存堆叠GRU神经网络模型;实时监测堆叠GRU神经网络模型的损失函数值变化,最后以折线图的形式输出堆叠GRU神经网络模型的预测误差与准确率。
步骤7、利用堆叠GRU神经网络模型对多维时序数据的自动特征提取能力挖掘关联寿命与健康状态的有用特征,对航空发动机进行剩余寿命预测与健康评估;航空发动机寿命预测属于回归问题,而健康评估属于分类问题,采用均方根误差eRMSE、平均绝对误差eMAPE以及相关系数R2来衡量航空发动机寿命预测(回归问题)的预测值与真实值之间的差异,计算式分别如下:
上式中,yi为真实值,为预测值,为均值,N是预测值的个数。
进行航空发动机健康评估(分类问题)时,使用准确率(Acc)来表征GRU神经网络模型的分类正确的样本个数占整体样本个数的比例。
步骤8、迁移学习是指将某一领域(源域)学习到的知识迁移至另一相似领域(目标域)中应用的机器学习策略;本发明采用单一飞行状态下的数据来训练模型,利用迁移学习策略将堆叠GRU神经网络模型泛化至不同的飞行状态下,实现多工况下的GRU网络模型迁移:将不同的航空发动机运行工况视作不同的领域,针对其中一个工况训练堆叠GRU神经网络模型,然后基于模型微调的方法利用其他工况的数据对模型高层神经网络层进行参数微调,进而达到节省计算机资源,提高效率的目的;将训练完成的堆叠GRU神经网络模型的底层神经网络层(输入层、部分GRU层)参数冻结,针对不同工况的数据,对高层网络和输出层进行参数微调,并与未微调堆叠GRU神经网络模型的识别准确率对比,检验迁移的成功性,提高计算效率。
实施例二
在实施例一的基础上,本申请实施例二提供了实施例一中方法的具体应用:
Step1:数据采集(本文为涡轮增压发动机全寿命周期运行数据,布设32个传感器),并将采集的原始数据存储入Sql server数据库。
Step2:数据预处理,对数据进行异常值处理、缺失值补全,标准化处理,并将数据转换为可用于监督学习的数据类型,并针对不同的飞行参数划分飞行状态,具体标准如表1:
表1飞行状态划分依据表
Step3:依据健康因子建立标准构建健康因子曲线,表征航空发动机的健康状态退化过程,并利用XGBoost模型对性能参数进行重要性分析,提取关键性能参数,划分健康状态并建立寿命标签。
Step4:将源域数据集(单一工况数据)以4:1的比例划分为训练集,
测试集,目标域数据集(其他工况数据)设置为验证集。
Step5:搭建堆叠GRU模型并将训练集数据代入模型进行训练,利用测试集检验误差与准确率变化,保存训练完成的模型。
Step6:将保存的模型低级网络层锁定,利用验证集小批量数据对高级网络层进行参数微调,达到模型迁移的效果,并验证迁移成功性。
如图2所示,为本发明传感器布设点,其中P为总压力,Ps为静压力,T为总温度,W为流量,下标1~11分别代表风扇入口、风扇出口、支路导管、LPC出口、HPC出口、HPT冷却液、LPT冷却液、LPC入口、HPC入口、HPT出口、LPT出口,b代表燃烧室,Wf为燃料流量,Nf为风扇速度,Nc为物理核心速度。此外还包括风扇装配余量(Smfan)、LPC装配余量(SmLPC)、HPC装配余量(SmHPC)、HPC入口燃料流量比例(phi)以及飞机飞行高度(Alt)、马赫数(Mach)和油门旋转变压器角度(TRA)一共32个参数。
如图3所示,为健康因子构建过程图,本发明以5种关键限寿件的效率(e)和空气质量流量(f)作为性能指标,并归一化,然后取每个时刻的最小值为当前健康因子值,最后根据航空发动机退化关系式对所得健康因子离散值进行拟合为一条光滑曲线。
如图4所示,为本发明XGBoost性能参数重要性分析结果图,XGBoost模型需要设定的主要超参数为最大深度、学习率和参数量。本发明在回归模型建立时的超参数优化方法采取网格化搜索,搜索空间为(最大深度:[5,11],学习率:[0.01,0.1],参数量:[100,400]),采用均方根误差作为目标函数,并实施三折交叉验证。最终确定最优超参数为最大深度为8,学习率为0.1,参数量为300。
通过XGBoost回归树拟合后输出各个元素对健康因子的重要性程度如图4,其中虚线(2.5%)以上的部分表示对健康因子影响较大的性能参数,加上健康因子共13个特征,作为后续寿命预测与健康评估的输入变量。
如图5所示,为健康状态划分过程图,根据健康因子的衰减速度可以将航空发动机的健康状态划分为初始衰退、正常衰退与异常衰退过程。初始衰退是指发动机关键零部件由于制造和装配公差引起的不可避免的轻
度初始磨损量;正常衰退是指由于初始磨损量引发的发动机性能衰退的过程,该过程发动机还未收到严重的损伤,处于正常工作状态;而异常衰退则是指发动机由于关键设备故障引起的性能退化速度加快的阶段。本发明提出将健康因子与退化速率相结合的划分标准,两者取交集,建立标准如表2,其中退化速度同样进行了标准化处理:
表2健康状态划分依据表
图6为堆叠GRU网络模型结构,共3层GRU层,神经元个数依次为64,128,64,每层GRU层后接一个Dropout层,用于舍弃20%的超参数,最后一层Dropout层接一个Flatten层,将特征映射拉直为一维序列,输入输出层(Dense)进行预测与识别。对于寿命预测模块,Dense层设置1个神经元,不设置非线性激活函数;对于健康评估模块,Dense层设置3个神经元(对应3种健康状态),非线性激活函数为softmax。Dense层增添L1、L2正则化项,优化模型收敛过程,防止过拟合,模型优化方法采用Adam算法,并设置提前中止(early stopping)命令,当测试集准确率不上升时停止训练保存模型。堆叠GRU模型需要设定的主要超参数为学习率和小批量训练批次大小。同样采取网格化搜索对参数进行优化,搜索空间为(学习率:[0.001,0.01],批次大小:[32,256]),采用均方根误差作为目标函数。最终确定最优超参数:学习率为0.001,小批量训练批次大小为128。
如图7-1至图7-3所示,为分别使用3种单一工况训练模型,以同工况数据置入模型测试后,输出的预测结果。由图可知,堆叠GRU模型在拟合发动机关键性能参数和RUL之间的联系时,效果出众,预测结果均分布在真实值邻近,均方根误差小于0.1,平均绝对误差均保持在4%左右,预测值与真实值的相关性较高,接近0.95。
对于健康评估模块,状态识别结果如图8,纵向为样本真实标签,横向为模型预测标签。三个工况下模型的健康评估准确率分别为94.894%、96.418%、99.216%。其中对于异常退化状态,三个工况的识别准确率平
均达到了99.036%,近乎100%,表明异常退化阶段航空发动机表现出来的性能参数具备较强的可区分性。而其它两个衰退过程由于系统依旧处于良好的工作状态,因此数据特征差异性较小,但模型依旧能够以较高的准确率区分。
将上述评价指标取均值,与传统模型CNN和LSTM对比如下表3。GRU在寿命预测和健康评估方面效果均优于其他两者,LSTM由于对时序数据的敏感性在寿命预测方面效果较CNN好,但在健康评估方面效果较劣,可能为数据维度大引起的模型学习困难。
表3模型效果对比表
因为不同工况下数据之间的浅层特征具有一定的相似性,而其深层特征则是差异性的表现。通过对单一工况的模型高级层进行参数微调(Fine tune),即锁定前两个GRU层,以其他两个工况的小批量数据训练第三个GRU层和输出层。设定迁移组合为Fs1-Fs2,Fs1-Fs3,Fs2-Fs1,Fs2-Fs3,Fs3-Fs1,Fs3-Fs2,前者为源域,后者为目标域。采用迁移策略和未采用迁移策略的结果对比如下表4。结果显示,在跨工况预测和识别方面,未采用迁移学习策略前,单一工况模型对其他工况数据的识别效果很差,甚至存在低于60%,采用了迁移学习后误差明显降低,准确率明显上升,且相邻两个工况由于数据相似度较高,识别效果也更好。这表明,在实际情况下数据稀缺或分布不均衡的情况下采用迁移学习策略对模型进行更新,能够明显地提升预测和识别效果。
表4迁移效果比较表
本说明书中各个实施例采用递进的方式描述,每个实施例重点说明的都是与其他实施例的不同之处,各个实施例之间相同相似部分互相参见即可。对于实施例公开的系统而言,由于其与实施例公开的方法相对应,所以描述的比较简单,相关之处参见方法部分说明即可。
本文中应用了具体个例对本发明的原理及实施方式进行了阐述,以上实施例的说明只是用于帮助理解本发明的方法及其核心思想;同时,对于本领域的一般技术人员,依据本发明的思想,在具体实施方式及应用范围上均会有改变之处。综上所述,本说明书内容不应理解为对本发明的限制。
Claims (8)
- 一种基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于,包括以下步骤:步骤1、数据采集和数据储存:在数据采集模块内布置传感器,由传感器采集航空发动机关键限寿件的工作状态参数与环境参数;数据存储模块以传感器类别为列,以采集数据时间为行,将采集的工作状态参数与环境参数制成表格存入航空发动机工作状态数据库;步骤2、数据预处理:数据预处理模块将数据采集模块采集得到的工作状态参数与环境参数进行缺失值、异常值及标准化处理;以各个飞机的单次旅程飞行时长作为标准来划分飞行状态;飞行状态为不同飞行长度发动机的损伤情况;步骤3、健康因子构建:首先选择航空发动机关键限寿件的效率和空气质量流量作为性能指标;然后对性能指标进行归一化处理,选取所有性能指标中的最小值作为健康因子,得到健康因子曲线;最终对健康因子曲线进行平滑处理;步骤4、数据集构建模块采用XGBoost模型提取关键性能指标,对关键性能指标进行降维处理;步骤5、将健康因子与标准化处理后的减小速度v取交集,将航空发动机的健康状态划分为初始衰退过程、正常衰退过程和异常衰退过程;减小速度v为当前时刻健康因子曲线的斜率值;建立寿命与健康状态标签,最终对数据切片,制作时序数据集;步骤6、搭建并训练堆叠GRU神经网络模型,对时间序列数据进行特征提取和预测;步骤7、利用堆叠GRU神经网络模型对多维时序数据的自动特征提取能力挖掘关联寿命与健康状态的有用特征,对航空发动机进行剩余寿命预测与健康评估;步骤8、利用迁移学习策略将堆叠GRU神经网络模型泛化至不同的飞行状态下:针对其中一个工况训练堆叠GRU神经网络模型,然后将训练完成的堆叠GRU神经网络模型的底层神经网络层参数冻结,针对不同工况的数据,对高层网络和输出层进行参数微调,并与未微调堆叠GRU神经网络模型的识别准确率对比,检验迁移的成功性。
- 根据权利要求1所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于:步骤1中关键限寿件包括风扇、高压压缩机、低压压缩机、高压涡轮和低压涡轮;工作状态参数与环境参数包括风扇入口压力P1、风扇入口温度T1、风扇出口流量W2、风扇出口压力P2、支路导管压力P3、LPC出口压力P4、LPC出口温度T4、HPC出口压力P5、HPC出口静压力Ps5、HPC出口温度T5、燃料流量Wf、HPT冷却液流量W6、LPT冷却液流量W7、燃烧室压力Pb、燃烧室温度Tb、LPC入口流量W8、HPC入口流量W9、HPT出口流量W10、HPT出口温度T10、HPT出口压力P10、LPT出口流量W11、LPT出口压力P11、LPT出口温度T11、风扇速度Nf、物理核心速度Nc、风扇装配余量、LPC装配余量、HPC装配余量、HPC入口燃料流量比例、飞机飞行高度、马赫数和油门旋转变压器角度。
- 根据权利要求1所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于:步骤2中数据预处理模块采用均值补全法对采集得到的工作状态参数与环境参数进行缺失值处理,通过缺失值两侧值的平均数补全空缺;对采集得到的工作状态参数与环境参数进行异常值处理时,将异常值直接舍弃,其中异常值指偏离设备正常工作时状态参数范围的值;进行标准化处理时采用Z-score标准化方式,公式为:
上式中,μ和σ分别是样本X的均值和方差,X表示进行标准化处理前的数据,X`表示进行标准化处理后的数据。 - 根据权利要求1所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于,步骤3中:采用max-min归一化对性能指标进行归一化处理,计算公式为:
上式中,x为样本原值,x`为归一化结果,xmax和xmin分别是所有性能指标中的最小值和最大值;采用最小二乘法插值对离散的健康因子进行拟合,根据航空发动机性 能衰退关系式对健康因子曲线进行平滑处理,得到最终健康因子平滑曲线;其中航空发动机性能衰退关系式为:
δa(t)=1-exp(atb)+δn(ts)+ξ上式中,δa(t)为t时刻的健康因子,tb表示t时刻的b次方,ts表示航空发动机性投入使用的时刻;δn(ts)为航空发动机初始磨损量;其中a=U(0.001,0.003),b=U(1.4,1.6),ξ=N(0,0.001)。 - 根据权利要求1所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于,步骤4中,XGBoost模型是由n个基模型组成的加法模型,假设第t次迭代要训练的树模型是ft(x),则第t次迭代时的预测结果满足:
上式中:为前t-1棵树的预测结果,ft(xi)为第t棵树模型。 - 根据权利要求1所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于,步骤5中:初始衰退过程指航空发动机关键零部件出现的轻度初始磨损的过程;正常衰退过程指由于轻度初始磨损引发的航空发动机性能衰退的过程;异常衰退过程指航空发动机由于关键设备故障引起的性能退化速度加快的过程;对数据切片时,取Ts为一个时间步长,分别对不同健康状态的数据进行切片操作,最终得到N×Ts×C的输入样本集合,其中N为样本数量,C为样本维度。
- 根据权利要求5所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于,步骤6具体包括以下步骤:步骤6.1、采用Python语言搭建堆叠GRU神经网络模型:GRU神经网络模型包括一个输入层,多个中间层和一个输出层,依次连接输入层、中间层和输出层;输入层特征映射组为多维数组,输入样本格式为N×Ts×C;Ts为一个时间步长,N为样本数量,C为样本维度;隐含层包含3对堆叠GRU-Dropout层和一个Flatten层GRU神经网络模型的结构单元包括更新门Zt和重置门Rt,更新门Zt用来控制当前状态Ht从历史状态Ht-1中保留的信息量和从候选状态Ht`中接受新信息的量;重置门Rt用来控制候选状态Ht`的计算是否依赖历史状态Ht-1;GRU神经网络模型的状态更新方式为:
Ht=Zt⊙Ht-1+(1-Zt)⊙Ht`其中更新门输出为:
Zt=σ(WzXt+UzHt-1+bz)候选状态Ht`为
Ht`=tanh(WhXt+Uh(Rt⊙Ht-1)+bh)其中重置门输出为:
Rt=σ(WrXt+UrHt-1+br)上式中,Xt为t时刻GRU神经网络模型的输入,Wz、Uz和bz分别为更新门的输入权重、历史状态权重和偏置;Wh、Uh和bh分别为候选状态的输入权重、历史状态权重和偏置;Wr、Ur和br分别为重置门的输入权重、历史状态权重和偏置;σ和tanh为非线性激活函数;堆叠GRU-Dropout层对上层的参数进行随机舍弃;最后将堆叠GRU-Dropout层的输出输入Flatten层,进行维数的削减后,转化为一维向量;将一维向量经过输出层输出预测序列结果;步骤6.2、训练堆叠GRU神经网络模型,对时间序列数据进行特征提取和预测:将建立的时序数据集输入待训练的GRU神经网络模型中,按设定比例划分训练集和测试集;设置单次训练送入数据批量;采用Adam算法对堆叠GRU神经网络模型进行优化,并设置提前中止命令;当测试集准确率不上升时,停止训练并保存堆叠GRU神经网络模型;实时监测堆叠 GRU神经网络模型的损失函数值变化,最后以折线图的形式输出堆叠GRU神经网络模型的预测误差与准确率。 - 根据权利要求7所述基于迁移学习的航空发动机寿命预测与健康评估方法,其特征在于,步骤7中对航空发动机进行剩余寿命预测与健康评估时,采用均方根误差eRMSE、平均绝对误差eMAPE以及相关系数R2来衡量航空发动机寿命预测的预测值与真实值之间的差异,计算式分别如下:
上式中,yi为真实值,为预测值,为均值,N是预测值的个数;进行航空发动机健康评估时,使用准确率来表征GRU神经网络模型的分类正确的样本个数占整体样本个数的比例。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
GB2311367.3A GB2629461A (en) | 2022-05-30 | 2023-05-30 | Transfer-learning-based life prediction and health assessment method for aero-engine |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210599974.0 | 2022-05-30 | ||
CN202210599974.0A CN114997051B (zh) | 2022-05-30 | 2022-05-30 | 一种基于迁移学习的航空发动机寿命预测与健康评估方法 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023231995A1 true WO2023231995A1 (zh) | 2023-12-07 |
Family
ID=83028577
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2023/096973 WO2023231995A1 (zh) | 2022-05-30 | 2023-05-30 | 一种基于迁移学习的航空发动机寿命预测与健康评估方法 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN114997051B (zh) |
WO (1) | WO2023231995A1 (zh) |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117370847A (zh) * | 2023-12-08 | 2024-01-09 | 深圳宇翊技术股份有限公司 | 基于深度学习的隔离开关检测方法及装置 |
CN117407698A (zh) * | 2023-12-14 | 2024-01-16 | 青岛明思为科技有限公司 | 一种混合距离引导的领域自适应故障诊断方法 |
CN117574689A (zh) * | 2024-01-16 | 2024-02-20 | 南京航空航天大学 | 一种基于BiGRU的电源状态评估方法及系统 |
CN117592865A (zh) * | 2023-12-21 | 2024-02-23 | 中国人民解放军军事科学院系统工程研究院 | 一种装备零备件质量状态预测方法及装置 |
CN117648643A (zh) * | 2024-01-30 | 2024-03-05 | 山东神力索具有限公司 | 基于人工智能的索具预测性诊断方法及装置 |
CN117807718A (zh) * | 2024-03-01 | 2024-04-02 | 南京工业大学 | 一种基于失效概率预测的航空发动机维护决策设计方法 |
CN117827275A (zh) * | 2023-12-20 | 2024-04-05 | 北京比格大数据有限公司 | 迁移评估方法及装置、设备及存储介质 |
CN117851921A (zh) * | 2024-03-07 | 2024-04-09 | 山东能源数智云科技有限公司 | 基于迁移学习的设备寿命预测方法及装置 |
CN117875138A (zh) * | 2024-03-12 | 2024-04-12 | 大连理工大学 | 一种机理与数据融合驱动的飞行器结构寿命预测方法 |
CN117972600A (zh) * | 2024-01-03 | 2024-05-03 | 国家能源集团山西电力有限公司 | 一种基于多维故障特征学习的风电机组关键部件异常探测方法 |
CN117972989A (zh) * | 2023-12-28 | 2024-05-03 | 天地(常州)自动化股份有限公司北京分公司 | 一种煤机设备的寿命预测方法及系统 |
CN117976148A (zh) * | 2024-04-01 | 2024-05-03 | 成都市双流区妇幼保健院 | 一种基于机器学习的儿童心理健康问题预测方法及系统 |
CN118020669A (zh) * | 2024-04-12 | 2024-05-14 | 济宁安鑫养殖有限公司 | 一种基于大数据的黑猪养殖疾病智能监测管理方法及系统 |
CN118091469A (zh) * | 2024-04-19 | 2024-05-28 | 深圳市普裕时代新能源科技有限公司 | 基于人工智能的储能电池健康状态预测方法及系统 |
CN118333094A (zh) * | 2024-06-11 | 2024-07-12 | 浙江大学 | 湿式离合器接合性能预测方法及系统 |
CN118332934A (zh) * | 2024-06-12 | 2024-07-12 | 西安热工研究院有限公司 | 一种基于储能容量的超级电容寿命预测方法及系统 |
CN118378494A (zh) * | 2024-06-25 | 2024-07-23 | 中国航空工业集团公司金城南京机电液压工程研究中心 | 一种基于自动更新大模型的管路阀门监测方法和装置 |
CN118445736A (zh) * | 2024-07-08 | 2024-08-06 | 厦门市百岗电气有限公司 | 开关柜健康状态监测方法及设备 |
CN118484680A (zh) * | 2024-06-19 | 2024-08-13 | 浙江双峰电气股份有限公司 | 基于机器学习的电容器剩余寿命预测方法及系统 |
CN118520599A (zh) * | 2024-07-24 | 2024-08-20 | 中国人民解放军军事科学院战争研究院 | 基于域适应迁移学习的旋翼无人机机动仿真模型生成方法 |
CN118551313A (zh) * | 2024-07-29 | 2024-08-27 | 无锡法拉第电机有限公司 | 一种同步发电机绕组状态评估方法及装置 |
CN118656595A (zh) * | 2024-08-09 | 2024-09-17 | 中国人民解放军国防科技大学 | 基于数实结合数据的发动机剩余寿命预测方法及装置 |
CN118688575A (zh) * | 2024-08-23 | 2024-09-24 | 广东电网有限责任公司茂名供电局 | 一种配电网故障处理的自动化缺陷识别方法 |
CN118686779A (zh) * | 2024-08-22 | 2024-09-24 | 南通广兴气动设备有限公司 | 手动泵性能实时监测系统 |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114997051B (zh) * | 2022-05-30 | 2023-05-09 | 浙大城市学院 | 一种基于迁移学习的航空发动机寿命预测与健康评估方法 |
GB2629461A (en) * | 2022-05-30 | 2024-10-30 | Hangzhou City Univ | Transfer-learning-based life prediction and health assessment method for aero-engine |
CN115828699B (zh) * | 2022-12-19 | 2023-07-21 | 华中科技大学 | 功率半导体模块全生命周期结温预测方法、系统及终端 |
CN115713044B (zh) * | 2023-01-09 | 2023-04-25 | 佰聆数据股份有限公司 | 一种多工况切换下的机电设备剩余寿命分析方法和装置 |
CN116305531B (zh) * | 2023-01-13 | 2023-09-15 | 北京控制工程研究所 | 一种航天器健康演化模型建模方法、装置、设备及介质 |
CN116147928B (zh) * | 2023-04-20 | 2023-07-14 | 清华大学 | 航空发动机热射流装置的健康状态确定方法、装置和设备 |
CN116416884B (zh) * | 2023-06-12 | 2023-08-18 | 深圳市彤兴电子有限公司 | 一种显示器模组的测试装置及其测试方法 |
CN116502544B (zh) * | 2023-06-26 | 2023-09-12 | 武汉新威奇科技有限公司 | 一种基于数据融合的电动螺旋压力机寿命预测方法及系统 |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112100767A (zh) * | 2020-09-02 | 2020-12-18 | 西北工业大学 | 一种基于奇异值分解和gru的航空发动机寿命预测方法 |
CN113869563A (zh) * | 2021-09-14 | 2021-12-31 | 北京化工大学 | 一种基于故障特征迁移的航空涡扇发动机剩余寿命预测方法 |
CN114297910A (zh) * | 2021-11-26 | 2022-04-08 | 中国民航大学 | 一种基于改进lstm的航空发动机寿命预测方法 |
CN114997051A (zh) * | 2022-05-30 | 2022-09-02 | 浙大城市学院 | 一种基于迁移学习的航空发动机寿命预测与健康评估方法 |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102789545B (zh) * | 2012-07-12 | 2015-08-19 | 哈尔滨工业大学 | 基于退化模型匹配的涡轮发动机剩余寿命的预测方法 |
US10229369B2 (en) * | 2016-04-19 | 2019-03-12 | General Electric Company | Creating predictive damage models by transductive transfer learning |
CN108959778B (zh) * | 2018-07-06 | 2020-09-15 | 南京航空航天大学 | 一种基于退化模式一致性的航空发动机剩余寿命预测方法 |
US12136035B2 (en) * | 2020-06-26 | 2024-11-05 | Tata Consultancy Services Limited | Neural networks for handling variable-dimensional time series data |
CN112257333A (zh) * | 2020-09-24 | 2021-01-22 | 浙江工业大学 | 一种基于深度学习的机械设备内部组件寿命预测方法 |
CN113837464A (zh) * | 2021-09-22 | 2021-12-24 | 浙大城市学院 | 一种基于CNN-LSTM-Attention的热电联产锅炉负荷预测方法 |
CN114021461A (zh) * | 2021-11-04 | 2022-02-08 | 浙大城市学院 | 基于XGBoost的电子膨胀阀质量流量特性预测方法 |
-
2022
- 2022-05-30 CN CN202210599974.0A patent/CN114997051B/zh active Active
-
2023
- 2023-05-30 WO PCT/CN2023/096973 patent/WO2023231995A1/zh unknown
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112100767A (zh) * | 2020-09-02 | 2020-12-18 | 西北工业大学 | 一种基于奇异值分解和gru的航空发动机寿命预测方法 |
CN113869563A (zh) * | 2021-09-14 | 2021-12-31 | 北京化工大学 | 一种基于故障特征迁移的航空涡扇发动机剩余寿命预测方法 |
CN114297910A (zh) * | 2021-11-26 | 2022-04-08 | 中国民航大学 | 一种基于改进lstm的航空发动机寿命预测方法 |
CN114997051A (zh) * | 2022-05-30 | 2022-09-02 | 浙大城市学院 | 一种基于迁移学习的航空发动机寿命预测与健康评估方法 |
Cited By (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117370847B (zh) * | 2023-12-08 | 2024-02-13 | 深圳宇翊技术股份有限公司 | 基于深度学习的隔离开关检测方法及装置 |
CN117370847A (zh) * | 2023-12-08 | 2024-01-09 | 深圳宇翊技术股份有限公司 | 基于深度学习的隔离开关检测方法及装置 |
CN117407698A (zh) * | 2023-12-14 | 2024-01-16 | 青岛明思为科技有限公司 | 一种混合距离引导的领域自适应故障诊断方法 |
CN117407698B (zh) * | 2023-12-14 | 2024-03-08 | 青岛明思为科技有限公司 | 一种混合距离引导的领域自适应故障诊断方法 |
CN117827275A (zh) * | 2023-12-20 | 2024-04-05 | 北京比格大数据有限公司 | 迁移评估方法及装置、设备及存储介质 |
CN117592865A (zh) * | 2023-12-21 | 2024-02-23 | 中国人民解放军军事科学院系统工程研究院 | 一种装备零备件质量状态预测方法及装置 |
CN117592865B (zh) * | 2023-12-21 | 2024-04-05 | 中国人民解放军军事科学院系统工程研究院 | 一种装备零备件质量状态预测方法及装置 |
CN117972989A (zh) * | 2023-12-28 | 2024-05-03 | 天地(常州)自动化股份有限公司北京分公司 | 一种煤机设备的寿命预测方法及系统 |
CN117972600A (zh) * | 2024-01-03 | 2024-05-03 | 国家能源集团山西电力有限公司 | 一种基于多维故障特征学习的风电机组关键部件异常探测方法 |
CN117574689A (zh) * | 2024-01-16 | 2024-02-20 | 南京航空航天大学 | 一种基于BiGRU的电源状态评估方法及系统 |
CN117648643B (zh) * | 2024-01-30 | 2024-04-16 | 山东神力索具有限公司 | 基于人工智能的索具预测性诊断方法及装置 |
CN117648643A (zh) * | 2024-01-30 | 2024-03-05 | 山东神力索具有限公司 | 基于人工智能的索具预测性诊断方法及装置 |
CN117807718A (zh) * | 2024-03-01 | 2024-04-02 | 南京工业大学 | 一种基于失效概率预测的航空发动机维护决策设计方法 |
CN117807718B (zh) * | 2024-03-01 | 2024-05-10 | 南京工业大学 | 一种基于失效概率预测的航空发动机维护决策设计方法 |
CN117851921A (zh) * | 2024-03-07 | 2024-04-09 | 山东能源数智云科技有限公司 | 基于迁移学习的设备寿命预测方法及装置 |
CN117851921B (zh) * | 2024-03-07 | 2024-06-18 | 山东能源数智云科技有限公司 | 基于迁移学习的设备寿命预测方法及装置 |
CN117875138A (zh) * | 2024-03-12 | 2024-04-12 | 大连理工大学 | 一种机理与数据融合驱动的飞行器结构寿命预测方法 |
CN117976148A (zh) * | 2024-04-01 | 2024-05-03 | 成都市双流区妇幼保健院 | 一种基于机器学习的儿童心理健康问题预测方法及系统 |
CN118020669A (zh) * | 2024-04-12 | 2024-05-14 | 济宁安鑫养殖有限公司 | 一种基于大数据的黑猪养殖疾病智能监测管理方法及系统 |
CN118091469A (zh) * | 2024-04-19 | 2024-05-28 | 深圳市普裕时代新能源科技有限公司 | 基于人工智能的储能电池健康状态预测方法及系统 |
CN118333094A (zh) * | 2024-06-11 | 2024-07-12 | 浙江大学 | 湿式离合器接合性能预测方法及系统 |
CN118332934A (zh) * | 2024-06-12 | 2024-07-12 | 西安热工研究院有限公司 | 一种基于储能容量的超级电容寿命预测方法及系统 |
CN118484680A (zh) * | 2024-06-19 | 2024-08-13 | 浙江双峰电气股份有限公司 | 基于机器学习的电容器剩余寿命预测方法及系统 |
CN118378494A (zh) * | 2024-06-25 | 2024-07-23 | 中国航空工业集团公司金城南京机电液压工程研究中心 | 一种基于自动更新大模型的管路阀门监测方法和装置 |
CN118445736A (zh) * | 2024-07-08 | 2024-08-06 | 厦门市百岗电气有限公司 | 开关柜健康状态监测方法及设备 |
CN118520599A (zh) * | 2024-07-24 | 2024-08-20 | 中国人民解放军军事科学院战争研究院 | 基于域适应迁移学习的旋翼无人机机动仿真模型生成方法 |
CN118551313A (zh) * | 2024-07-29 | 2024-08-27 | 无锡法拉第电机有限公司 | 一种同步发电机绕组状态评估方法及装置 |
CN118656595A (zh) * | 2024-08-09 | 2024-09-17 | 中国人民解放军国防科技大学 | 基于数实结合数据的发动机剩余寿命预测方法及装置 |
CN118686779A (zh) * | 2024-08-22 | 2024-09-24 | 南通广兴气动设备有限公司 | 手动泵性能实时监测系统 |
CN118688575A (zh) * | 2024-08-23 | 2024-09-24 | 广东电网有限责任公司茂名供电局 | 一种配电网故障处理的自动化缺陷识别方法 |
Also Published As
Publication number | Publication date |
---|---|
CN114997051B (zh) | 2023-05-09 |
CN114997051A (zh) | 2022-09-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2023231995A1 (zh) | 一种基于迁移学习的航空发动机寿命预测与健康评估方法 | |
CN112766303B (zh) | 一种基于cnn的航空发动机故障诊断方法 | |
WO2024087128A1 (zh) | 针对航空发动机剩余使用寿命预测的多尺度混杂注意力机制建模方法 | |
CN109766583A (zh) | 基于无标签、不均衡、初值不确定数据的航空发动机寿命预测方法 | |
CN114218872B (zh) | 基于dbn-lstm半监督联合模型的剩余使用寿命预测方法 | |
Zheng et al. | Fault diagnosis method based on supervised particle swarm optimization classification algorithm | |
CN110750655A (zh) | 一种智能ietm故障维修辅助系统的知识库优化方法 | |
CN114298183B (zh) | 飞行动作智能识别方法 | |
CN113268370B (zh) | 一种根因告警分析方法、系统、设备及存储介质 | |
CN114880925A (zh) | 基于时间卷积网络和多层自注意力的装备寿命预测方法 | |
CN113341919B (zh) | 一种基于时序数据长度优化的计算系统故障预测方法 | |
CN112329337B (zh) | 基于深度强化学习的航空发动机剩余使用寿命估计方法 | |
Li et al. | Aero-engine exhaust gas temperature prediction based on LightGBM optimized by improved bat algorithm | |
CN115791174B (zh) | 一种滚动轴承异常诊断方法、系统、电子设备及存储介质 | |
Zhang et al. | Online anomaly detection for aeroengine gas path based on piecewise linear representation and support vector data description | |
Song et al. | A novel wind power prediction approach for extreme wind conditions based on TCN-LSTM and transfer learning | |
CN107545112B (zh) | 多源无标签数据机器学习的复杂装备性能评估与预测方法 | |
CN113673774A (zh) | 基于自编码器和时序卷积网络的航空发动机余寿预测方法 | |
CN116521406A (zh) | 一种基于残差门gru-vae模型的航空发动机未超限飞参数据异常检测方法 | |
CN115048873A (zh) | 一种用于飞机发动机的剩余使用寿命预测系统 | |
GB2629461A (en) | Transfer-learning-based life prediction and health assessment method for aero-engine | |
CN116561528B (zh) | 一种旋转机械的rul预测方法 | |
Bai et al. | Remaining Useful Life Prediction of Aircraft Engines based on FA-SDAE-BiGRU | |
CN117421665A (zh) | 一种考虑运行工况的两阶段航空发动机气路故障诊断方法 | |
CN111104986B (zh) | 一种基于部件特征的航空发动机多故障并发诊断方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
ENP | Entry into the national phase |
Ref document number: 202311367 Country of ref document: GB Kind code of ref document: A Free format text: PCT FILING DATE = 20230530 |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23815174 Country of ref document: EP Kind code of ref document: A1 |