Vehicle-mounted PEMFC health index extraction and RUL prediction method

Abstract

The invention provides a health index extraction and RUL prediction method for a vehicle-mounted PEMFC, and the method employs a relative voltage loss rate as a health index, and comprises the steps: S1, obtaining PEMFC operation data, and the operation data comprises vehicle information and PEMFC related information; s2, carrying out conversion and preprocessing on the acquired operation data; s3, analyzing the working current distribution condition of the PEMFC through the reconstruction data, and analyzing the voltage change condition under different currents; s4, fitting an initial polarization curve of the PEMFC, extracting RVLR under all currents, and obtaining an RVLR change curve through a sliding window strategy and a local weighted regression algorithm; s5, building a prediction model of the RUL based on deep learning, and optimizing model parameters; and S6, the RUL of the vehicle-mounted PEMFC is predicted. A simple, accurate and universal health index-RVLR representing the vehicle-mounted PEMFC is extracted for real vehicle data, a sliding window strategy is designed to represent the working performance of the PEMFC, and the method has high universality under different data sets.

Description

Technical Field

[0001] The present application relates to the technical field of battery health status analysis, and in particular to a method for extracting health indicators and predicting RUL for on-board PEMFC. Background Art

[0002] Proton exchange membrane fuel cell (PEMFC) has the advantages of stable performance, zero emission, high energy conversion efficiency, wide fuel source and flexible use of renewable energy, and is one of the ideal solutions to global energy and environmental problems. However, in actual vehicle applications, the short life and high cost of PEMFC remain the technical bottlenecks in the commercialization of fuel cell vehicles. Affected by harsh environments and dynamic working conditions, the degradation of PEMFC includes complex physical and chemical mechanisms such as corrosion and dissolution of catalysts, structural damage of gas diffusion layers, and decomposition of polymers in proton exchange membranes. It is difficult to accurately estimate the health status and remaining useful life (RUL) of fuel cells in practice. Among them, fuel cell life prediction technology is of great significance to the improvement of fuel cell performance and service life. Establishing effective health indicators and accurately predicting the life of fuel cells have become key issues in improving their reliability and economy.

[0003] Health indicators are the benchmark for evaluating the degradation state of PEMFC. In the RUL prediction research of PEMFC, health indicators can be mainly divided into static indicators and dynamic indicators. Static indicators generally refer to physical quantities that are easy to measure in steady-state conditions, mainly including voltage, power and their combination. However, in the actual operation of vehicle-mounted fuel cells, the measurement data are affected by operating conditions and internal system parameters, voltage and power are prone to violent fluctuations, and the degradation trend is not obvious. Under dynamic operating conditions, scholars have proposed dynamic health indicators such as electrochemical catalytic surface area, electrochemical impedance spectrum, virtual steady-state voltage, and relative power loss rate. Among them, the measurement of electrochemical catalytic surface area or electrochemical impedance spectrum can only be performed offline, the surrounding magnetic field changes require special magnetic field measurement equipment, the extraction of virtual steady-state voltage is relatively complicated, and the relative power loss rate cannot completely eliminate the impact of current changes on its trend.

[0004] The static health indicators currently used fluctuate greatly in actual vehicle PEMFC, and the dynamic health indicators are difficult to extract, making it difficult to accurately and efficiently characterize the health status of vehicle-mounted PEMFC.

[0005] There are three main methods for predicting the life of PEMFC: model-based, data-based, and hybrid-based. The model-based method is to establish a complex electrochemical model through the internal reaction mechanism of PEMFC, but the model is difficult to build and the calculation is complex, and the accuracy and versatility of the model are poor. The hybrid-based method studies the degradation mechanism of PEMFC, extracts the internal degradation characteristics of the stack and makes predictions, but its model is relatively complex and the calculation is large, making it difficult to implement online application.

[0006] The data-based method does not rely on empirical formulas and PEMFC aging mechanisms, that is, it models through historical degradation data and quickly predicts future aging trends based on input data. Deep learning shows better performance in processing big data, multi-feature extraction and transfer learning, such as convolutional neural network (CNN), recurrent neural network, long short-term memory (LSTM) network, etc., but common deep models accumulate errors in PEMFC long-term series prediction and require a large amount of data. At the same time, some designs combine traditional models to establish life prediction models, such as CNN-LSTM, CNN-bidirectional LSTM, etc. Although multiple combinations of traditional models can improve prediction accuracy, this will cause the computational complexity of the model to increase rapidly, making it difficult to apply in practical applications.

[0007] Therefore, the current life prediction model is less effective in long time series prediction and requires a large amount of data. The common combination model has high computational complexity. Summary of the invention

[0008] In view of this, the present application provides a method for extracting health indicators and predicting RUL of on-board PEMFC, so as to extract a universal health indicator from real vehicle data to characterize the health status of PEMFC, and realize simple and efficient PEMFC life prediction.

[0009] To achieve the above objectives, the technical solutions adopted in this application are as follows: A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC, wherein the method uses relative voltage loss rate (RVLR) as a health indicator, and includes: S1: Acquire PEMFC operation data, wherein the operation data includes vehicle information and PEMFC related information, wherein the vehicle information includes vehicle mileage, vehicle operating condition and vehicle speed, and the PEMFC related information includes voltage, current and power; S2: converting and preprocessing the acquired operation data, wherein the conversion and preprocessing include reconstructing the acquired operation data at fixed time intervals to obtain reconstructed data; S3: Analyze the operating current distribution of PEMFC by reconstructing data, and analyze the voltage changes under different currents; S4: Fit the initial polarization curve of PEMFC, extract the RVLR under all currents, and obtain the RVLR variation curve through the sliding window strategy and the locally estimated scatterplot smoothing (LOESS) algorithm; S5: Build a prediction model for RUL based on deep learning and optimize the model parameters; S6: Based on the RUL prediction model, the RUL of the vehicle-mounted PEMFC is predicted, and error analysis is performed under different data sets to verify the accuracy and versatility of the model.

[0010] Furthermore, the step S2 is specifically as follows: S2.1: converting the running data types into valid operation values, identifying and deleting abnormal values, and obtaining processed original data A; S2.2: Reconstruct the original data A at 0.5 h intervals to obtain the reconstructed data B = [B 0 , B 1 , …,B t ], that is, B 0 and B 1 The time interval is 0.5 h.

[0011] Furthermore, the step S3 is specifically as follows: S3.1: sort and analyze the frequency of the current data in the reconstructed data B, and extract all data changes under different currents; S3.2: Determine the main working current range of PEMFC, analyze the voltage changes in the high current working and idle current ranges, the main working current range is the current range formed by the up and down fluctuations of the main working current, sort them in order from high to low according to the frequency of current operation, and take the first current value as the main working current.

[0012] Furthermore, the step S4 is specifically as follows: S4.1: The first day's operation data is extracted from the original data A. The initial polarization curve is fitted nonlinearly to characterize the initial working performance of PEMFC at different currents. The nonlinear fitting formula is: in Vis the voltage, I is the current, β 1 - β 6 is a constant, then we get V and I The corresponding change curve of S4.2: Based on the fitting of the initial polarization curve, the initial voltage at different currents can be obtained V 0 , so that according to the real-time voltage value V i The original RVLR is obtained, and its calculation formula is: ; S4.3: Modify the original RVLR by: S4.3.1: Calculate the relative voltage loss rate for all currents within the window interval R s , which is determined by the relative voltage loss rate at each current R in composition; S4.3.2: Calculate the relative voltage loss of the main operating current R i1 The difference between the average change and the average change of relative voltage loss of other currents is recorded as the determination coefficient M n , sort the currents in descending order according to their working frequency, and take the first current value as the main working current; S4.3.3: If M n If the absolute value of is greater than the threshold, it means that the current has a greater impact on the original RVLR, then the judgment coefficient M n Relative voltage loss rate under adjusted current R in , and obtain the relative voltage loss within the corrected window interval R s_m , and then get the modified RVLR; if M n If the absolute value of is less than or equal to the threshold, no adjustment is required and the original RVLR is directly output. The threshold is equal to R i1 The average fluctuation value of S4.4: Use LOESS fitting to correct the actual changes in RVLR data and obtain the actual RVLR change curve.

[0013] Furthermore, the step S5 is specifically as follows: S5.1: Introduce the self-attention (SA) mechanism into the gated recurrent unit (GRU) network model to form the self-attention gated recurrent unit (SAGRU) lifespan prediction model; S5.2: Input sequence RVLR=[X t+1 , X t+2 , X t+3 , X t+4 , …, X t+n ] Input to SAGRU model for training; S5.3: Design an Adam optimizer to optimize the hyperparameters of the SAGRU model, including the model learning rate, weights, and bias terms.

[0014] Furthermore, the step S6 is specifically as follows: S6.1: predicting the RVLR value at the next moment by using the trained SAGRU model and the RVLR value at the previous moment, and predicting the RUL of the on-board PEMFC according to the RVLR value at the next moment; S6.2: Calculate the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R) of the SAGRU model. 2 ), according to the RMSE, MAE, MAPE and R 2 Evaluate the model prediction performance.

[0015] Compared with the prior art, the beneficial effects of this application are: 1. Extract a simple, accurate and universal health indicator for vehicle-mounted PEMFC - RVLR based on real vehicle data, and design a sliding window strategy to characterize the working performance of PEMFC. It can effectively characterize the health status changes of PEMFC and identify the end of life and RUL of the fuel cell stack, and has high versatility under different data sets; 2. A long-time series prediction model based on deep learning has been established. Compared with traditional models, it has higher accuracy and stability and is expected to be applied to cloud platforms or real vehicle controllers; 3. The SAGRU RUL prediction model proposed in this paper, in which the SA mechanism can effectively improve the prediction accuracy and long time series prediction ability of the GRU model. Compared with the five models of backpropagation (BP) neural network, CNN, LSTM, GRU and SALSTM, the model error is the smallest under 60% of the training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 This is a flow chart of a method for extracting health indicators and predicting RUL of vehicle-mounted PEMFC in this application; Figure 2 This is the flow chart of RVLR extraction in this application; Figure 3 This is the structure diagram of the GRU model in this application; Figure 4 This is the RUL prediction framework diagram in this application; Figure 5 Figure 2 is the RUL prediction error diagram for different data sets in this application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0019] like Figure 1 As shown, a method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC is provided. The method uses relative voltage loss rate (RVLR) as a health indicator, including: S1: Acquire PEMFC operation data, wherein the operation data includes vehicle information and PEMFC related information, wherein the vehicle information includes vehicle mileage, vehicle operating condition and vehicle speed, and the PEMFC related information includes voltage, current and power; Specifically, in step S1, the real vehicle operation data can be obtained by downloading from a big data cloud platform or collecting data from a real vehicle CAN card, wherein the cloud platform data collection time interval is 10 s, and the real vehicle CAN card data collection time interval is 1 s.

[0020] S2: converting and preprocessing the acquired operation data, wherein the conversion and preprocessing include reconstructing the acquired operation data at fixed time intervals to obtain reconstructed data; As a further implementation, the step S2 is specifically: S2.1: converting the running data types into valid operation values, identifying and deleting abnormal values, and obtaining processed original data A; The data collected by the real vehicle CAN card is CAN protocol data, which needs to be converted according to the data protocol; the data downloaded from the cloud platform is a discontinuous string, which needs to be summarized and converted into numerical values. The abnormal value is a value where the current or voltage is less than 0, or a value where the PEMFC is not working properly. The processing of abnormal values ​​includes deletion, replacement (average), etc.

[0021] S2.2: Reconstruct the original data A at 0.5 h intervals to obtain the reconstructed data B = [B 0 , B 1 , …,B t ], that is, B 0 and B 1 The time interval is 0.5 h.

[0022] The short-term degradation of PEMFC is small. In order to improve the processing efficiency or reduce the computational complexity, the original data is reconstructed at 0.5 h time intervals to obtain the reconstructed data B = [B 0 , B 1 , …, B t ].

[0023] Specifically, reconstructing the original data at 0.5 hour time intervals may be performed by taking 0.5 hour time intervals as time intervals, and then calculating the mean of the original data within the 0.5 hour time interval, and taking the mean as the value at the current moment. This processing can reduce data complexity.

[0024] S3: Analyze the operating current distribution of PEMFC by reconstructing data, and analyze the voltage changes under different currents; As a further implementation, the step S3 is specifically: S3.1: sort and analyze the frequency of the current data in the reconstructed data B, and extract all data changes under different current values; S3.2: Determine the main operating current range of PEMFC and analyze the voltage changes in the high current operating and idle current ranges.

[0025] According to the frequency of current operation, the current values ​​are sorted in order from high to low, and the first current value is used as the main working current. The main working current interval is the current interval formed by the up and down fluctuations of the main working current. For example, if the main working current is I, and its up and down fluctuation range is 1A, then the main working current interval is [I-1A, I+1A]. Because the data collected from the actual vehicle fluctuates greatly, if a certain current value (such as 32.3A) is used for analysis, the result will be inaccurate, so the current interval (such as I±1A) is used for analysis and statistics.

[0026] By analyzing the changes in voltage in the high current operation and idle current range, it is found that under high current operation, the voltage loss will increase rapidly in the later stage of PEMFC life, reducing the working performance of PEMFC.

[0027] S4: Fit the initial polarization curve of PEMFC, extract the RVLR under all currents, and obtain the RVLR variation curve through the sliding window strategy and the locally estimated scatterplot smoothing (LOESS) algorithm; As a further embodiment, Figure 2 As shown, the step S4 is specifically as follows: S4.1: The first day's operation data is extracted from the original data A. The initial polarization curve is fitted nonlinearly to characterize the initial working performance of PEMFC at different currents. The nonlinear fitting formula is: in V is the voltage, I is the current, β 1 - β 6 is a constant, then we get V and I The corresponding change curve of In the early stage of operation, the performance of PEMFC degrades slowly, and the polarization curve at this time can accurately characterize the initial output performance of the stack at different currents. In the specific implementation, considering the large discreteness of the actual vehicle operation data points, 8 hours is used as the data for one day. Other time lengths can also be used as the data for one day, and this application does not limit this.

[0028] S4.2: Based on the fitting of the initial polarization curve, the initial voltage at different currents can be obtained V 0 , so that according to the real-time voltage value V i The original RVLR is obtained, and its calculation formula is: S4.3: Modify the original RVLR by: S4.3.1: Calculate the relative voltage loss rate for all currents within the window interval R s , which is determined by the relative voltage loss rate at each current R in composition; , is the RVLR within a window interval, It is the current value sorted from high to low according to the current frequency.

[0029] Specifically, the window interval can be set to 2 days or 16 hours.

[0030] S4.3.2: Calculate the relative voltage loss R of the main operating current i1 The difference between the average change and the average change of relative voltage loss of other currents is recorded as the determination coefficient M n ; The currents are sorted in order from high to low according to their working frequency, and the first current value is used as the main working current.

[0031] S4.3.3: If M n If the absolute value of is greater than the threshold, it means that the current has a greater impact on the original RVLR, then the judgment coefficient M n Relative voltage loss rate under adjusted current R in , and obtain the relative voltage loss within the corrected window interval R s_m , and then get the modified RVLR; if M n If the absolute value of is less than or equal to the threshold, no adjustment is required and the original RVLR is directly output. The threshold is equal to R i1 The average fluctuation value of Specifically, the threshold is approximately 0.01; in, ,and ; S4.4: Use LOESS fitting to correct the actual changes in RVLR data and obtain the actual RVLR change curve.

[0032] The RVLR extracted from real vehicle data retains a lot of noise. In order to study the actual health status changes of PEMFC and accurately define the RUL of the stack, LOESS is used to fit the actual changes of the data to obtain the actual RVLR change curve.

[0033] S5: Build a prediction model for RUL based on deep learning and optimize the model parameters; As a further embodiment, Figure 4 As shown, the step S5 is specifically as follows: S5.1: Introduce the self-attention (SA) mechanism into the gated recurrent unit (GRU) network model to form the self-attention gated recurrent unit (SAGRU) lifespan prediction model; S5.1.1: The hidden state of the previous moment is controlled by the two gating mechanisms of the reset gate and update gate in the GRU network structure. H t-1 And the input at this time X t Learning is performed to retain the important degradation characteristics of the fuel cell. The reset gate R t and update gate Z t The formulas are: In the formula R t and Z t The range is 0 to 1, σ is the Sigmoid activation function, H t-1 is the candidate hidden layer state at the previous moment, W z and W r is the weight coefficient, X t Enter data for health status; GRU is an improved form of deep learning LSTM neural network. Its model is simpler and has improved training efficiency. This model can solve problems such as effective long-term memory and gradient in back propagation, and is suitable for building a prediction model for PEMFC time series data. The GRU network structure has two key gating mechanisms: reset gate and update gate. The GRU model structure is as follows: Figure 3 As shown. Reset gate R t Determine the amount of information lost at the previous moment and update gate Zt Controls the amount of state information transmitted from the previous moment to the current moment.

[0034] S5.1.2: Effectively capture the long-term and short-term dependencies of the sequence and save them in the hidden state of the GRU model H t Then the dependency is passed to the future time step; the hidden state H t The formulas are: H t is the candidate hidden layer state at the current moment; tanh is the hyperbolic tangent function, which is used to give a nonlinear transformation so that the output range is between [−1,1]; W h is the weight coefficient; b z , b r and b h is the bias term; S5.1.3: The SA mechanism is introduced into the GRU model to form the SAGRU model. The SA mechanism autonomously learns the contribution of input features and targets during training, thereby improving the accuracy of the GRU model in long time series prediction tasks. The SA mechanism formula is as follows: Where Q is the query vector; K is the key vector; V is the value vector; d K It is a scaling factor to prevent the dot product value from being too large and causing the gradient to disappear.

[0035] The actual extracted RVLR contains time series and trend features, which causes the GRU model to lose some information during long-term training, resulting in reduced model accuracy. The SA mechanism can autonomously learn the contribution of input features and targets during training, thereby improving the accuracy of the GRU model in long-term series prediction tasks. The principle of the SA mechanism is to linearly transform the input sequence to obtain Q, K, and V, obtain the attention weight of the sequence through Q and K, normalize it, and multiply it with V to obtain a feature sequence with weights, which is then combined with the original sequence to obtain an input sequence with output vector weights, which is then used as the input of the GRU prediction model to improve the accuracy of the life prediction model.

[0036] S5.2: Input sequence RVLR=[X t+1 , X t+2 , X t+3 , Xt+4 , …, X t+n ] Input to SAGRU model for training; Input sequence RVLR=[X t+1 , X t+2 , X t+3 , X t+4 , …, X t+n ], the SAGRU model predicts the state of the next step by inputting the previous data, that is, [X t+1 , X t+2 , X t+3 ] as input to predict the health status at the next time t+4. In the specific implementation, a single-layer GRU model can be used, with 3 input layers, 1 output layer, and 20 neurons. The constructed SAGGRU model is simpler.

[0037] S5.3: Design an Adam optimizer to optimize the hyperparameters of the SAGRU model, including the model learning rate, weights, and bias terms.

[0038] During the model training process, SAGRU converges slowly, so the Adam optimizer is designed to optimize the hyperparameters of the SAGRU model. The Adam optimizer combines the stochastic gradient descent and adaptive learning rate algorithms to converge quickly and reduce the training time, so as to gradually reduce the loss function and achieve model optimization.

[0039] S6: Based on the RUL prediction model, the RUL of the vehicle-mounted PEMFC is predicted, and error analysis is performed under different data sets to verify the accuracy and versatility of the model.

[0040] As a further implementation, the step S6 is specifically: S6.1: predicting the RVLR value at the next moment by using the trained SAGRU model and the RVLR value at the previous moment, and predicting the RUL of the on-board PEMFC according to the RVLR value at the next moment; S6.2: Calculate the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R) of the SAGRU model. 2 ), according to the RMSE, MAE, MAPE and R 2 Evaluate the model prediction performance.

[0041] RMSE and MAE are used to evaluate the absolute size of the model prediction error, MAPE reflects the relative size of the error, which is better for RVLR with a smaller value, and R² evaluates the goodness of fit of the model.

[0042] Experimental verification: The RVLR of two vehicle-mounted PEMFC data sets are extracted, which are marked as FC1 and FC2 respectively. The life of these two stacks is about 2000 h, and the RVLR of FC1 increases faster near the cutoff point, indicating that the performance of the fuel cell decreases rapidly near the end of life, and this is the time for the fuel cell to be effectively maintained. Some differences in the cutoff time are due to factors such as the operating conditions of the PEMFC, driver operation, and the stack itself. After more than 2000 h, the voltage loss increases rapidly and continues until the end of the life cycle. The reason is that under high power, insufficient supply of reactants or accumulation of products will increase concentration polarization, and the increase in current will produce more ohmic polarization losses, resulting in a faster drop in the actual voltage of the fuel cell. However, the overall RVLR change characteristics are highly similar, and the extracted RVLR has a high degree of versatility. Although in actual applications, the life of PEMFC is defined according to the power requirements of the application scenario to define the cutoff point of RVLR, effective maintenance of the cutoff point of RVLR can greatly improve the durability and economy of the stack.

[0043] The SAGRU RUL prediction model proposed in this paper, in which the SA mechanism can effectively improve the prediction accuracy and long time series prediction ability of the GRU model, is compared with the five models of backpropagation (BP) neural network, CNN, LSTM, GRU and SALSTM. The model has the smallest error under 60% of the training data, with RMSE of 0.0002069, MAE of 0.0001715, MAPE of 0.1957%, and R 2 is 0.9998, and the RUL prediction time is within 1 hour. For different data sets, the prediction error is as follows Figure 5 shown.

[0044] Depend on Figure 5 It can be seen that for PEMFCs of different powers, their relative voltage losses have similar characteristics, and the proposed RUL prediction model has better transfer learning ability under different RVLRs.

[0045] This application extracts a simple, accurate and universal health indicator - RVLR - to characterize the on-board PEMFC based on real vehicle data, and designs a sliding window strategy to characterize the working performance of PEMFC, which can effectively characterize the changes in the health status of PEMFC and identify the life end time and RUL of the fuel cell stack, and has high versatility under different data sets; a long time series prediction model is established based on deep learning, which has higher accuracy and stability than traditional models, and is expected to be applied to cloud platforms or real vehicle controllers.

[0046] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for extracting health indicators and predicting RUL of vehicle-mounted PEMFC, characterized in that: The method uses the relative voltage loss rate as a health indicator, including: S1: Acquire PEMFC operation data, wherein the operation data includes vehicle information and PEMFC related information, wherein the vehicle information includes vehicle mileage, vehicle operating condition and vehicle speed, and the PEMFC related information includes voltage, current and power; S2: converting and preprocessing the acquired operation data, wherein the conversion and preprocessing include reconstructing the acquired operation data at fixed time intervals to obtain reconstructed data; S3: Analyze the operating current distribution of PEMFC by reconstructing data, and analyze the voltage changes under different currents; S4: Fit the initial polarization curve of PEMFC, extract the relative voltage loss rate at all currents, and obtain the RVLR change curve through the sliding window strategy and local weighted regression algorithm; S5: Build a prediction model for RUL based on deep learning and optimize the model parameters; S6: Based on the RUL prediction model, the RUL of the vehicle-mounted PEMFC is predicted, and error analysis is performed under different data sets to verify the accuracy and versatility of the model.

2. A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC according to claim 1, characterized in that: The step S2 is specifically as follows: S2.1: converting the running data types into valid operation values, identifying and deleting abnormal values, and obtaining processed original data A; S2.2: Reconstruct the original data A at 0.5 h intervals to obtain the reconstructed data B = [B0, B1, …, B t ], that is, the time interval between B0 and B1 is 0.5 h.

3. A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC according to claim 2, characterized in that: The step S3 is specifically as follows: S3.1: sort and analyze the frequency of the current data in the reconstructed data B, and extract all data changes under different currents; S3.2: Determine the main working current range of PEMFC, analyze the voltage changes in the high current working and idle current ranges, the main working current range is the current range formed by the up and down fluctuations of the main working current, sort them in order from high to low according to the frequency of current operation, and take the first current value as the main working current.

4. A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC according to claim 3, characterized in that: The step S4 is specifically as follows: S4.1: The first day's operation data is extracted from the original data A. The initial polarization curve is fitted nonlinearly to characterize the initial working performance of PEMFC at different currents. The nonlinear fitting formula is: ; in V is the voltage, I is the current, β 1- β 6 is a constant, and the result is V and I The corresponding change curve of S4.2: Based on the fitting of the initial polarization curve, the initial voltage at different currents can be obtained V 0, so according to the real-time voltage value V i The original RVLR is obtained, and its calculation formula is: ； S4.3: Modify the original RVLR; S4.4: Use the local weighted regression algorithm to fit the actual changes of the corrected RVLR data to obtain the actual RVLR change curve.

5. A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC according to claim 4, characterized in that: The modification of the original RVLR in step S4.3 is specifically as follows: S4.3.1: Calculate the relative voltage loss rate for all currents within the window interval R s , which is determined by the relative voltage loss rate at each current R in composition; S4.3.2: Calculate the relative voltage loss of the main operating current R i1 The difference between the average change and the average change of relative voltage loss of other currents is recorded as the determination coefficient M n , sort the currents in descending order according to their working frequency, and take the first current value as the main working current; S4.3.3: If M n If the absolute value of is greater than the threshold, it means that the current has a greater impact on the original RVLR, then the judgment coefficient M n Relative voltage loss rate under adjusted current R in , and obtain the relative voltage loss within the corrected window interval R s_m , and then get the modified RVLR; if M n If the absolute value of is less than or equal to the threshold, no adjustment is required and the original RVLR is directly output. The threshold is equal to R i1 The average fluctuation value.

6. A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC according to claim 5, characterized in that: The step S5 is specifically as follows: S5.1: Introduce the self-attention mechanism into the gated neural unit network model to form a self-attention gated neural unit lifespan prediction model; S5.2: Input sequence RVLR=[X t+1 , X t+2 , X t+3 , X t+4 , …, X t+n ] Input into the life prediction model for training; S5.3: Design an Adam optimizer to optimize the hyperparameters of the life prediction model, wherein the hyperparameters include model learning rate, weight and bias term.

7. A method for extracting health indicators and predicting RUL for vehicle-mounted PEMFC according to claim 6, characterized in that: The step S6 is specifically as follows: S6.1: predicting the RVLR value at the next moment by using the trained life prediction model and the RVLR value at the previous moment, and predicting the RUL of the vehicle-mounted PEMFC according to the RVLR value at the next moment; S6.2: Calculate RMSE, MAE, MAPE, and R of the life prediction model 2 , according to the RMSE, MAE, MAPE and R 2 Evaluate the model prediction performance.

Images