Fuel cell life prediction method based on recoverable fault avoidance and ESN combination

Abstract

The invention discloses a fuel cell life prediction method based on recoverable fault avoidance and ESN combination. The method comprises the following specific steps: firstly, extracting a characteristic value of HI, clustering by using a K average algorithm and generating a state label, performing data enhancement by using TimeGAN, training a K proximity algorithm model by using real data and the generated data, and generating the state label for the input fuel cell HI by using the trained model; secondly, customizing a recoverable fault avoidance strategy, and training the ESN to predict under the guidance of the custom strategy; and finally, calculating the degradation trend, the confidence interval and the remaining service life thereof according to a prediction result. According to the invention, the accuracy and robustness of identification are improved, and the recoverable fault can be accurately identified. The median of the relative accuracy is improved by more than 19%, and the symmetric average absolute percentage error is reduced by more than 11%.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault prediction and health management, and in particular relates to a fuel cell life prediction method based on recoverable fault avoidance combined with ESN. Background Art

[0002] As a clean and green hydrogen energy application device, proton exchange membrane fuel cells can efficiently convert the chemical energy stored in hydrogen directly into electrical energy. They have been applied in many fields including transportation. However, the lack of durability still restricts the large-scale commercialization of proton exchange membrane fuel cells. In order to solve this problem, the remaining useful life (RUL) prediction has become one of the key technologies.

[0003] RUL prediction can predict the future health status by evaluating and analyzing the current and historical conditions of the system, mainly by predicting the degradation trend through the analysis of the health indicator (HI). HI refers to the dimensionless characteristic value of the output voltage or power of the battery stack after eliminating the dynamic operating condition interference and normalization, which can show the changes in the operating conditions of the battery, abnormal conditions and degradation behavior of the system. The performance degradation shown by HI can generally be divided into recoverable and irreversible, both of which will shorten the life of the fuel cell. Among them, recoverable faults refer to the phenomenon that the performance of the fuel cell recovers to a certain extent after polarization testing, shutdown and other operations, and will affect the accuracy of RUL prediction.

[0004] At present, the main research methods for the life prediction of fuel cells are divided into two categories: model-based methods and data-driven methods. Model-based methods describe the degradation process through physical models, while data-driven methods learn key features from historical data and do not rely on physical models. As a typical data-driven model, ESN is widely used in this field with its efficient learning ability and good generalization performance. The document with patent number "CN202310576482.4" discloses "a method, device and system for thermal management of fuel cells based on neural networks", and establishes an ESN model to predict data. However, since its solution does not fully consider the dynamic changes of the system, recoverable faults will cause the long-term prediction accuracy of the ESN model to be low. The document with patent number "CN20131031104.3" discloses "a method for predicting the remaining life of a monotonic echo state network", and establishes a model with monotonically increasing or decreasing output weights. However, due to the limitations of the model in dealing with complex nonlinear dynamic systems, it is easily disturbed by recoverable faults, resulting in unsatisfactory long-term prediction results. The common problem of the above two methods is that they fail to effectively deal with the impact of recoverable faults on long-term predictions, the RUL prediction error is large, and the prediction accuracy is low. Summary of the invention

[0005] In view of the problems that the prior art cannot effectively handle the impact of recoverable faults on long-term prediction, the RUL prediction has large errors and low prediction accuracy, the present invention proposes a fuel cell life prediction method based on recoverable fault avoidance combined with ESN.

[0006] To achieve the above object, the technical solution of the present invention is as follows: A fuel cell life prediction method based on recoverable fault avoidance combined with ESN, comprising the following steps:

[0007] Step 1: Input HI and extract feature values, use K-means algorithm for clustering, generate state labels, use time series generative adversarial network for data enhancement, train K-nearest neighbor algorithm model with real data and generated data, and use the trained model to generate state labels for input fuel cell HI;

[0008] Step 2: Customize recoverable fault avoidance strategies for different operating states and train ESN to make predictions under the guidance of customized strategies;

[0009] Step 3: Calculate the degradation trend and its confidence interval based on the prediction results, and further estimate the RUL.

[0010] Furthermore, the characteristic value in step 1 includes three characteristics, and the formulas of the three characteristic values ​​are:

[0011]

[0012]

[0013] Among them, t k represents the time corresponding to the kth point, Indicates that at t k HI at the moment, Indicates t k The first characteristic of the moment is Indicated in An extreme point in the neighborhood, Indicates t k The second characteristic of the moment is Indicates t k Characteristics of the moment three.

[0014] Furthermore, the specific steps of using the K-means algorithm to perform clustering and generate state labels in step 1 are as follows:

[0015] Use the K-means algorithm to cluster HI: First, randomly select K objects as the initial center points; then, according to the distance from each data point to the center point, assign it to the cluster with the closest center point. Then, based on the current cluster assignment, recalculate the position of each cluster center; the new center is the mean of all data points in the current cluster. The entire process will continue to iterate and update until the center point position of each cluster tends to be stable;

[0016] According to the clustering results, the operating status is mainly divided into the rising stage, the stable stage, the declining stage one and the declining stage two, among which the declining degree of the declining stage two is deeper than that of the declining stage one.

[0017] Furthermore, in step 1, the time series generative adversarial network is composed of an embedding function, a recovery function, a sequence generator and a sequence discriminator;

[0018] The expression of the embedded function e is:

[0019]

[0020] in, and Represents the vector space of static features and dynamic features, then and Representation and feature space and The corresponding latent space. t represents time, and Embedding network representing static and dynamic features, s and v t It is the feature space and A random vector in .

[0021] The expression of the recovery function r is:

[0022]

[0023]

[0024] in, and Representing static and dynamic embedded recovery networks;

[0025] The formula for the generation function g of the sequence generator is:

[0026]

[0027] in, and A generator network that represents static and dynamic features;

[0028] The formula of the discriminant function d of the sequence discriminator is:

[0029]

[0030] in, and denote the forward and backward hidden state sequences, respectively, and is the classification function of the output layer.

[0031] Furthermore, the recoverable fault avoidance strategy customized in step 2 is specifically:

[0032] Strategy 1: Time t corresponding to the kth point k , judge the running status according to the label. If it is in the stable stage, use t1 to t k All historical HIs are used as training sets, and multiple ESN models are trained. t1 is the time corresponding to the first point of HI;

[0033] Strategy 2: The current recoverable fault consists of the jth monotonically decreasing interval, and the time corresponding to its maximum point is t mj If the running state is "Descending Phase 1", then backtrack to t p (t1 <t p <t mj ,p∈Z * ), t p is the last moment of the most recent stable phase, from t1 to t p HI is used as the training set, and t p Set as the starting point of prediction, not t k In this case, the prediction process consists of two stages, starting from t p to kIt is called "simulation prediction". k to n It is called "conventional prediction", t n is the time corresponding to the last point of HI. Another case is that the operation state belongs to "Descending Phase 2". At this time, trace back to the last moment of the most recent "Descending Phase 1" t q (t mj <t q <t k ,q∈Z * ), will be from t1 to t q HI is used as the training set, and t q Set as the prediction starting point. Therefore, from t q to k This is the “simulation prediction” stage;

[0034] Strategy 3: The current recoverable fault consists of the jth and j+1th monotonic intervals. In the rising phase of recoverable faults, if t k HI at time t is less than mj At the moment HI, we go back to the most recent descending stage and repeat the steps of “Strategy 2”. On the contrary, in the j+1th monotonically increasing interval, suppose it contains a subinterval t r ~t k (t m(j+1) <t r <t k , r∈Z * ), t r is the left boundary of the subinterval, which determines whether HI rises or recovers to the level above the level at the beginning of the current recoverable fault. Specifically, if t r HI at time t is greater than or equal to mj If the HI at time t1 is reached, it can be considered that the current recoverable fault has ended. k The HI during this period is used as the training set.

[0035] Furthermore, the ESN used in step 2 consists of three different parts: input layer, reserve pool, and output layer. The formula of ESN is:

[0036]

[0037] y(t)=W out x(t)(13)

[0038] Among them, the input vector u(t) is the historical data at time t, W in is the connection weight from the input layer to the reserve layer, W out is the connection weight from the reservoir to the output layer, x(t-1) is the output of the previous state of the reservoir, is the state feedback weight, x(t) represents the state of the activated neurons in the reserve pool at time t-1, and y(t) is the output result of the output layer.

[0039] Furthermore, the calculation formula of RUL in step 3 is:

[0040] RUL=EOL-t a (14)

[0041] Among them, t a is the time when prediction starts, and EOL is the end of life, that is, the time when HI reaches the failure threshold.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention constructs an operation status recognition model, which effectively divides different categories through the K-means algorithm, and uses TimeGAN to perform data enhancement according to categories, making up for the problem of insufficient and unbalanced data. The K-nearest neighbor algorithm model is trained by real data and generated data. The enhanced training data is more comprehensive and balanced, thereby improving the accuracy and robustness of recognition. The model can accurately identify the occurrence of recoverable faults, providing strong support for customizing recoverable fault avoidance strategies.

[0044] 2. The present invention optimizes the ESN strategy for predicting HI degradation trends by identifying and isolating recoverable faults through customized strategies. Specifically, different strategies are selected for different operating states. When recoverable faults are identified, these three strategies start predictions by going back to the time point before the fault occurs, thereby avoiding the interference of recoverable faults on the long-term prediction of fuel cells, reducing the error of RUL prediction, and improving the accuracy of fuel cell life prediction. Compared with the monotonic ESN prediction model, the median relative accuracy of the present invention is improved by more than 19%, and the symmetric mean absolute percentage error is reduced by more than 11%. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0046] Figure 2 This is a rendering of the overall degradation trend of HI predicted by the present invention;

[0047] Figure 3 This is a comparison diagram of the RUL estimation results of the present invention and the monotonic ESN when a recoverable fault occurs. DETAILED DESCRIPTION

[0048] like Figure 1The figure shows the overall process diagram of the present invention. First, extract the characteristic value reflecting the operating state, build the operating state recognition model based on the K-means clustering algorithm and the K-nearest neighbor algorithm, and generate a state label for the input fuel cell HI; second, customize the recoverable fault avoidance strategy, and train the ESN to make predictions under the guidance of the customized strategy; finally, calculate the degradation trend and its confidence interval and RUL based on the prediction results. In order to realize the degradation prediction of the fuel cell, the specific implementation steps are as follows.

[0049] Step 1: Input HI and extract feature values, use K-means algorithm for clustering, generate state labels, use time series generative adversarial network (TimeGAN) for data enhancement, train K-nearest neighbor algorithm model with real data and generated data, and use the trained model to generate state labels for input fuel cell HI. It mainly includes the following steps.

[0050] Step 1.1: Extract three eigenvalues ​​reflecting the operating status. Feature 1 is used to reflect whether HI is in a decreasing state, an increasing state, or a stable state. Feature 2 is used to reflect the monotonicity of HI in a monotonic interval. Feature 3 is used to reflect the rate of change of HI in a monotonic interval.

[0051] The main formulas for the three eigenvalues ​​are:

[0052]

[0053] Among them, t k represents the time corresponding to the kth point, Indicates that at t k HI at the moment, Indicates t k The first characteristic of the moment is Indicated in An extreme point in the neighborhood, Indicates t k The second characteristic of the moment is Indicates t k Characteristics of the moment three.

[0054] Step 1.2: Use the K-means algorithm to cluster the HI of the fuel cell and generate a state label.

[0055] Use the K-means algorithm to cluster HI. First, randomly select K objects as the initial center points. Then, according to the distance from each data point to the center point, assign it to the cluster with the closest center point. Then, based on the current cluster assignment, recalculate the position of each cluster center. The new center is the mean of all data points in the current cluster. The whole process will continue to iterate and update until the center point position of each cluster tends to be stable.

[0056] According to the clustering results, the operating status is mainly divided into the rising stage, the stable stage, the falling stage 1 and the falling stage 2, among which the falling stage 2 is deeper than the falling stage 1. According to the analysis of the clustering results, the falling stage 1 and the falling stage 2 do not necessarily appear consecutively. However, when the falling stage 2 appears, the previous state must be the falling stage 1.

[0057] Step 1.3: Use TimeGAN to perform data enhancement on each type of running status, generate more sample data, solve the problem of data imbalance between different categories, and thus improve the training effect of the classification model.

[0058] TimeGAN mainly consists of an embedding function, a recovery function, a sequence generator, and a sequence discriminator. Usually, data consists of two elements, static features that do not change over time and dynamic features that change over time. The purpose of the embedding function e is to embed static features and dynamic features into their latent codes. With h t The expression of the embedded function e is:

[0059]

[0060]

[0061] in, and Represents the vector space of static features and dynamic features, then and Representation and feature space and The corresponding latent space. t represents time, and Embedding network representing static and dynamic features, s and v t It is the feature space and The recovery function can recover the feature representation from the latent code and The expression of the recovery function r is:

[0062]

[0063] in, and Represents the recovery network for static and dynamic embeddings. The sequence generator is constructed from a known vector space and Extract vector z from s and z t , and use the generating function to synthesize it into a latent code, the generated result is and The formula for the generating function g is:

[0064]

[0065] in, and The generator network represents static features and dynamic features. The sequence discriminator calculates the category information corresponding to the latent vector through the discriminant function and The formula of the discriminant function d is:

[0066]

[0067] in, and denote the forward and backward hidden state sequences, respectively, and is the classification function of the output layer.

[0068] Step 1.4: Use the real data and the generated data as training sets to train the K-nearest neighbor algorithm model.

[0069] Step 1.5: Generate a state label for the input fuel cell HI using the K-nearest neighbor algorithm model.

[0070] The K-nearest neighbor algorithm calculates the distance between different data points and predicts the category of unknown operating status by analyzing the similarity of known category features.

[0071] Use the trained model to classify the samples to be classified. According to the distance between the sample to be classified and all training samples, find the K training samples closest to the sample to be classified. The category that appears most frequently will be assigned to the sample to be classified to generate a state label.

[0072] Step 2: Customize a recoverable fault avoidance strategy and train the ESN to make predictions under the guidance of the customized strategy. The specific steps are as follows:

[0073] Step 2.1: Customize the following three recoverable fault avoidance strategies for different operating states:

[0074] Strategy 1: Time t corresponding to the kth point k , judge the running status according to the label. If it is in the stable stage, use t1 to t k All historical HIs are used as training sets, and multiple ESN models are trained. t1 is the time corresponding to the first point of HI.

[0075] Strategy 2: The current recoverable fault consists of the jth monotonically decreasing interval, and the time corresponding to its maximum point is t mj If the running state is "Descending Phase 1", then backtrack to t p (t1<tp <t mj , p∈Z * ), t p is the last moment of the most recent stable phase, from t1 to t p HI is used as the training set, and t p Set as the starting point of prediction, not t k In this case, the prediction process consists of two stages, starting from t p to k It is called "simulation prediction". k to n It is called "conventional prediction", t n is the time corresponding to the last point of HI. Another case is that the operation state belongs to "Descending Phase 2". At this time, trace back to the last moment of the most recent "Descending Phase 1" t q (t mj <t q <t k , q∈Z * ), will be from t1 to t q HI is used as the training set, and t q Set as the prediction starting point. Therefore, from t q to k This is the “simulation prediction” stage.

[0076] Strategy 3: The current recoverable fault consists of the jth and j+1th monotonic intervals. In the rising phase of recoverable faults, if t k HI at time t is less than mj At the moment HI, we go back to the most recent descending stage and repeat the steps of “Strategy 2”. On the contrary, in the j+1th monotonically increasing interval, suppose it contains a subinterval t r ~t k (t m(j+1 ) <t r <t k , r∈Z * ), t r is the left boundary of the subinterval, which determines whether HI rises or recovers to the level above the level at the beginning of the current recoverable fault. Specifically, if t r HI at time t is greater than or equal to mj If the HI at time t1 is reached, it can be considered that the current recoverable fault has ended. k The HI during this period is used as the training set.

[0077] Step 2.2: According to the selected strategy, use the specified data to train the ESN network to start prediction from the backtracking time point under the guidance of the customized strategy to obtain the prediction result of HI.

[0078] The ESN specified for predicting HI mainly includes three different parts: input layer, reserve pool and output layer. The main formula of ESN is:

[0079]

[0080] y(t)=W out x(t) (13)

[0081] Among them, the input vector u(t) is the historical data at time t, W in is the connection weight from the input layer to the reserve layer, W out is the connection weight from the reservoir to the output layer, x(t-1) is the output of the previous state of the reservoir, is the state feedback weight, x(t) represents the state of the activated neurons in the reserve pool at time t-1, and y(t) is the output result of the output layer.

[0082] Step 3: Calculate the degradation trend and its confidence interval based on the prediction results, and further estimate the RUL.

[0083] The requirements for confidence intervals are different for different strategies: for strategy one, this embodiment sets a corresponding 98% confidence interval; and for strategies two and three, this embodiment sets a corresponding 95% confidence interval.

[0084] The formula for calculating RUL is:

[0085] RUL=EOL-t a (14)

[0086] Among them, t a is the time when prediction starts, and EOL is the end of life, that is, the time when HI reaches the failure threshold.

[0087] In order to verify the classification effect of the state recognition model of the present invention, precision and accuracy are used as indicators to evaluate the performance of the classification model. Precision represents the proportion of samples that are actually positive among all samples predicted by the model to be positive. The calculation formula of precision is as follows:

[0088]

[0089] Among them, Precision represents the accuracy rate. TP represents the number of samples correctly predicted by the model as positive, and FP represents the number of samples that the model incorrectly predicts as positive from negative classes.

[0090] The accuracy rate indicates the ratio of the number of samples correctly classified by the model to the total number of samples. The accuracy rate is calculated as follows:

[0091]

[0092] Among them, ACC represents the accuracy, TN represents the number of samples correctly predicted by the model as negative classes, and FN represents the number of samples that the model incorrectly predicts as negative classes from positive classes.

[0093] The data shows that after data enhancement by TimeGAN, the accuracy of each stage has increased. The overall classification accuracy has increased from 95.40% to 97.24%. The specific results of the comparison of the accuracy and precision of each stage before and after data enhancement of the present invention are shown in Table 1 and Table 2:

[0094] Table 1

[0095] Descending stage 1 Descending stage 2 Rising stage Stable phase Before data augmentation 93.9％ 87.9％ 85.5％ 97.7％ After data enhancement 97.8％ 89.1％ 94.5％ 98.5％

[0096] Table 2

[0097] Before data augmentation After data enhancement Accuracy 95.40％ 97.24％

[0098] like Figure 2 The figure shows the effect of the present invention on predicting the overall degradation trend of HI. The solid line is the training set, the dotted line is the test set, the solid line with triangles is the predicted value, the dotted lines above and below the predicted value represent the range of the confidence interval, and the part enclosed by the ellipse is the area of ​​recoverable faults. As can be seen from the figure, the present invention can identify the "decline phase 1" of the current recoverable fault, and then trace back to the time point of the "stable phase", use the data before this time point to train the model, and reduce the impact of recoverable faults.

[0099] In order to verify the prediction effect of the present invention, it is compared with the prediction effect of the monotone ESN, which can more clearly demonstrate the advantages of the present invention in improving the prediction accuracy.

[0100] like Figure 3 The figure shows a comparison of the results of RUL estimation when a recoverable fault occurs between the present invention and the monotonic ESN. The time from 750 hours to 880 hours when the recoverable fault occurs is selected, and the RUL of each point in the interval is estimated. The solid line with triangles is the RUL predicted by the present invention, the solid line with circles is the RUL predicted by the monotonic ESN, the straight line is the true value of the RUL, the dotted line is the range of the prognostic field of view, the horizontal axis is time, and the vertical axis is RUL. It can be seen that in this recoverable fault area, compared with the monotonic ESN, the present invention can accurately identify the different stages of recoverable faults and avoid them, making the prediction results more stable, and significantly improving the prediction accuracy of RUL.

[0101] In the same area, the symmetric mean absolute percentage error and the median relative accuracy are used as indicators to reflect the difference in the performance of the two methods. The symmetric mean absolute percentage error is the average of the absolute error percentages of multiple predictions within the same prediction range, and the calculation formula is as follows:

[0102]

[0103] Where sMAPE represents the symmetric mean absolute percentage error, i represents time, and RUL i represents the actual remaining service life at time i, It represents the remaining useful life predicted by the model at time i, and n represents the number of samples evaluated.

[0104] Relative accuracy can be used to quantitatively evaluate the prediction performance. The relative accuracy formula at time i is as follows:

[0105]

[0106] Among them, RA i is the relative accuracy at time i.

[0107] The data show that the median relative accuracy of the present invention is improved by more than 19%, and the symmetric mean absolute percentage error is reduced by more than 11%. The specific results of the comparison of the symmetric mean absolute percentage error and relative accuracy median of the present invention with the monotonic ESN are shown in Table 3:

[0108] Table 3

[0109]

[0110]

[0111] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited thereto. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, including the combination of various technical features in any other suitable manner, and these simple modifications and combinations should also be regarded as the contents disclosed by the present invention and belong to the protection scope of the present invention.

Claims

1. A fuel cell life prediction method based on recoverable fault avoidance and ESN, characterized in that: The following steps are included Step 1: Input HI and extract feature values, use K-means algorithm for clustering, generate state labels, use time series generative adversarial network for data enhancement, train K-nearest neighbor algorithm model with real data and generated data, and use the trained model to generate state labels for input fuel cell HI; Step 2: Customize recoverable fault avoidance strategies for different operating states and train ESN to make predictions under the guidance of customized strategies; Step 3: Calculate the degradation trend and its confidence interval based on the prediction results, and further estimate the RUL.

2. The method for predicting fuel cell life based on recoverable fault avoidance combined with ESN according to claim 1, characterized in that: The characteristic values ​​in step 1 include three characteristics, and the formulas of the three characteristic values ​​are: Among them, t k represents the time corresponding to the kth point, Indicates that at t k HI at the moment, Indicates t k The first characteristic of the moment is Indicated in An extreme point in the neighborhood, Indicates t k The second characteristic of the moment is Indicates t k Characteristics of the moment three.

3. The fuel cell life prediction method based on recoverable fault avoidance combined with ESN according to claim 1 is characterized in that: The specific steps of using the K-means algorithm to perform clustering and generate state labels in step 1 are: Use the K-means algorithm to cluster HI: first randomly select K objects as the initial center points; Then, according to the distance from each data point to the center point, it is assigned to the cluster where the nearest center point is located; then, according to the current cluster assignment, the position of each cluster center is recalculated, and the new center is the mean of all data points in the current cluster; the whole process is continuously iterated and updated until the position of the center point of each cluster tends to be stable; According to the clustering results, the operating status is mainly divided into the rising stage, the stable stage, the declining stage one and the declining stage two, among which the declining degree of the declining stage two is deeper than that of the declining stage one.

4. A fuel cell life prediction method based on recoverable fault avoidance combined with ESN according to claim 2 or 3, characterized in that: In the step 1, the time series generative adversarial network is composed of an embedding function, a recovery function, a sequence generator and a sequence discriminator; The expression of the embedded function e is: in, and Represents the vector space of static features and dynamic features, then and Representation and feature space and The corresponding latent space. t represents time, and Embedding network representing static and dynamic features, s and v t It is the feature space and A random vector in ; The expression of the recovery function r is: in, and Representing static and dynamic embedded recovery networks; The formula for the generation function g of the sequence generator is: in, and A generator network that represents static and dynamic features; The formula of the discriminant function d of the sequence discriminator is: in, and denote the forward and backward hidden state sequences, respectively, and is the classification function of the output layer.

5. The method for predicting fuel cell life based on recoverable fault avoidance combined with ESN according to claim 4 is characterized in that: The recoverable fault avoidance strategy customized in step 2 is specifically: Strategy 1: Time t corresponding to the kth point k , judge the running status according to the label. If it is in the stable stage, use t1 to t k All historical HIs are used as training sets, and multiple ESN models are trained. t1 is the time corresponding to the first point of HI. Strategy 2: The current recoverable fault consists of the jth monotonically decreasing interval, and the time corresponding to its maximum point is t mj If the running state is "decline phase 1", then backtrack to t p (t1<t p <t mj , p∈Z * ), t p is the last moment of the most recent stable phase, from t1 to t p HI is used as the training set, and t p Set as the prediction starting point; in this case, the prediction process consists of two stages, starting from t p to k It is called "simulation prediction", from t k to n It is called "conventional forecast", t n is the time corresponding to the last point of HI; another case is that the running state belongs to "descending stage 2", in which case, it goes back to the last moment t of the most recent "descending stage 1" q (t mj <t q <t k , q∈Z * ), will be from t1 to t q HI is used as the training set, and t q Set as the prediction starting point, from t q to k This is the "simulation prediction" stage. Strategy 3: The current recoverable fault consists of the jth and j+1th monotonic intervals. In the rising phase of the recoverable fault, if t k HI at time t is less than mj HI at the moment, then go back to the most recent decline stage and repeat the steps of "Strategy 2"; on the contrary, in the j+1th monotonically increasing interval, if t r HI at time t is greater than or equal to mj The HI at time t1 to t k The HI during this period is used as the training set.

6. The method for predicting fuel cell life based on recoverable fault avoidance combined with ESN according to claim 5 is characterized in that: The ESN used in step 2 includes three different parts: input layer, reserve pool and output layer. The formula of ESN is: y(t)=W out x(t) (13) Among them, the input vector u(t) is the historical data at time t, W in is the connection weight from the input layer to the reserve layer, W out is the connection weight from the reservoir to the output layer, x(t-1) is the output of the previous state of the reservoir, is the state feedback weight, x(t) represents the state of the activated neurons in the reserve pool at time t-1, and y(t) is the output result of the output layer.

7. The method for predicting fuel cell life based on recoverable fault avoidance combined with ESN according to claim 6, characterized in that: The calculation formula of RUL in step 3 is: RUL=EOL-t a (14) Among them, t a is the time when prediction starts, and EOL is the end of life, that is, the time when HI reaches the failure threshold.

Images