A dynamic prediction method for PEMFC remaining life based on attention neural network

Abstract

The present invention relates to a PEMFC remaining life dynamic prediction method based on an attention neural network, which belongs to the technical field of PEMFC remaining life dynamic prediction, and solves the problem of poor accuracy of the remaining life dynamic prediction result. The method includes: collecting dynamic operation data of training PEMFC in the whole life cycle and preprocessing and screening key influencing factors; obtaining the correlation matrix between key influencing factors and stack voltage, and constructing a data sample set; constructing a multi-layer neural network CNN-BiLSTM-Attention model, and using it as a remaining life dynamic prediction model; using the data sample set to train and test the remaining life dynamic prediction model; real-time collection of the dynamic operation data of the PEMFC to be predicted in an evaluation cycle, and constructing corresponding input data, and the remaining life dynamic prediction model processes the input data and predicts and outputs the predicted value of the stack voltage after the preset time period ends in the current evaluation cycle.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic prediction of PEMFC remaining life, and in particular to a dynamic prediction method of PEMFC remaining life based on attention neural network (CNN-BiLSTM-Attention). Background Art

[0002] Proton Exchange Membrane Fuel Cell (PEMFC) is widely used in aerospace, automobiles, aircraft and other transportation vehicles due to its high efficiency, no pollution, low noise and other advantages. However, limited durability, high cost and low reliability are still the main obstacles to its large-scale commercial application. Predicting the remaining useful life (RUL) of PEMFC plays an important role in improving the reliability and durability of PEMFC.

[0003] At present, the remaining service life prediction of fuel cells can be divided into three categories according to different prediction methods:

[0004] The first is a model-driven prediction method, which requires the establishment of a mathematical degradation model. The accuracy of the prediction depends on the accuracy of the degradation model. However, since the proton exchange membrane fuel cell system has the characteristics of multi-physics field, multi-time scale and nonlinearity, and its internal degradation mechanism is still unclear, it is difficult to establish an accurate and reliable degradation model.

[0005] The second is a data-driven prediction method, which generally uses statistical techniques and fitting prediction methods through algorithms such as particle filtering and local weighted projection regression, but has a low prediction accuracy.

[0006] The third is a hybrid prediction method based on the joint drive of model and data, which needs to balance the prediction accuracy and complexity. At present, many researchers have used long short-term memory recurrent neural network to predict the remaining life of PEMFC. Some scholars integrated the Convolutional Neural Network (CNN) and the Bidirectional Recurrent Neural Network (BiRNN) into a new network (CNN-BiRNN) model for the voltage decay and remaining life prediction of PEMFC, which improved the accuracy of the prediction. Other scholars have proposed a PEMFC remaining life prediction method based on CNN and long short-term memory network (LSTM), which further improved the prediction accuracy and prediction speed. However, the research on the dynamic remaining life of PEMFC is still in its infancy. Under dynamic conditions, the physical fields such as voltage, temperature, and pressure of the PEMFC system fluctuate rapidly, and its performance degradation shows randomness and strong nonlinearity, which makes life prediction difficult.

[0007] In summary, the factors related to the dynamic prediction of the remaining life of PEMFC and the selection of the prediction model will affect the accuracy of the prediction. In addition, in the prior art, the directly collected data is used for model training, and it is difficult to dig out the deeper impact of the relevant data on the life prediction, which ultimately affects the prediction accuracy. Therefore, how to effectively improve the accuracy of the dynamic prediction results of the remaining life of PEMFC is still a technical problem that needs to be solved urgently. Summary of the invention

[0008] In view of the above analysis, an embodiment of the present invention aims to provide a PEMFC remaining life dynamic prediction method based on an attention neural network (CNN-BiLSTM-Attention), which is a PEMFC dynamic remaining life prediction method that combines a convolutional neural network with an attention mechanism (Attention) and a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, BiLSTM), so as to solve the problem of poor accuracy of the dynamic prediction results of the remaining life of PEMFC in the prior art.

[0009] The present invention discloses a method for dynamically predicting the remaining life of a PEMFC based on an attention neural network, the method comprising:

[0010] Collect and preprocess the dynamic operation data of the training PEMFC at each operating time in the entire life cycle; screen the key influencing factors according to the influence of various parameters in the preprocessed dynamic operation data on the dynamic prediction of the remaining life of PEMFC;

[0011] The normalized dynamic operation data is trimmed to construct a trimmed data set of the corresponding relationship between the key influencing factors and the stack voltage; a correlation matrix between the key influencing factors and the stack voltage is obtained in combination with the influence degree; a data sample set is constructed based on the trimmed data set and the correlation matrix;

[0012] Constructing a multi-layer neural network CNN-BiLSTM-Attention model and using it as a remaining life dynamic prediction model; using the data sample set to train and test the remaining life dynamic prediction model to obtain a remaining life dynamic prediction model that passes the test;

[0013] The dynamic operation data of the PEMFC to be predicted in an evaluation cycle is collected in real time, and the corresponding input data is constructed. The remaining life dynamic prediction model processes the input data and predicts and outputs the predicted value of the stack voltage after the preset length of time at the end of the current evaluation cycle.

[0014] On the basis of the above scheme, the present invention also makes the following improvements:

[0015] Furthermore, the dynamic running data is preprocessed by performing the following operations:

[0016] It is determined whether there are outliers in the various parameter data of the dynamic operation data of each operation time of the PEMFC in the whole life cycle. If there are outliers, the dynamic operation data of the corresponding operation time are eliminated, and the dynamic operation data of the corresponding operation time are supplemented by using an interpolation algorithm.

[0017] Afterwards, the Hampel filter is used to smooth and linearly normalize the parameter data in the dynamic operation data, and the parameter data are mapped to the range of [0,1] to obtain the normalized dynamic operation data of PEMFC at each operating time in its entire life cycle.

[0018] Further, filter the key influencing factors by performing the following operations:

[0019] Based on the normalized dynamic operation data of PEMFC at various operating times during its entire life cycle, the influence of various parameters and operating time other than the stack voltage on the stack voltage is quantified, and the parameter categories that have an effective impact on the dynamic prediction of the remaining life of PEMFC are selected as the key influencing factors from various parameters and operating time other than the stack voltage.

[0020] Furthermore, the Pearson correlation coefficient is introduced to quantify the influence of each influencing factor on the key measurement variables.

[0021] Further, construct a data sample set by performing the following operations:

[0022] The influence degrees of all key influencing factors are summed up and normalized so that the sum is 1; the relative weight of the influence degree of each key influencing factor is calculated as the weight of each key influencing factor;

[0023] The whole life cycle is evenly divided into multiple evaluation cycles; in each evaluation cycle, the stack voltage of the current evaluation cycle is selected from the trimmed data set to construct a state matrix, and the key influencing factors of the current evaluation cycle are selected to construct a coupling matrix, and the coupling degree matrix between the key influencing factors of the current evaluation cycle and the stack voltage is determined based on the coupling relationship between the state matrix and the coupling matrix;

[0024] The coupling degree of each key influencing factor in the coupling degree matrix of the key influencing factors of the current evaluation cycle and the stack voltage is multiplied by the weight to obtain the correlation matrix of the key influencing factors of the current evaluation cycle and the stack voltage;

[0025] The trimmed data set and the correlation matrix of each evaluation cycle are used as input data, and the stack voltage after the evaluation cycle ends for a preset time is used as the label of the input data; thereby constructing a data sample set.

[0026] Further, The state matrix of an evaluation cycle is expressed as , the coupling matrix is ​​expressed as ; Indicates the number of key influencing factors, represents the number of sampling times in an evaluation cycle,

[0027] No. The coupling matrix of the evaluation cycle is expressed as ,according to ,get .

[0028] Furthermore, the structure of the multi-layer neural network CNN-BiLSTM-Attention model is as follows:

[0029] The first layer is the input layer, which is used to receive input data;

[0030] The second layer is the convolutional neural network layer; the convolutional neural network layer first sets a one-dimensional convolutional layer, which is used to extract features from the input data; the input end of the one-dimensional convolutional layer is connected to the output end of the input layer, and the output end of the one-dimensional convolutional layer is connected to the input end of the pooling layer; the pooling layer is used to perform pooling processing on the features extracted by the one-dimensional convolutional layer, and the output end of the pooling layer is connected to the input end of the Dropout layer;

[0031] The third layer is a bidirectional long short-term memory neural network layer. In the bidirectional long short-term memory neural network layer, the output of the Dropout layer is connected to the input of the BiLSTM hidden layer. 64 hidden layer nodes are set in the BiLSTM hidden layer, and the activation function uses the tanh function.

[0032] The fourth layer is the attention layer; the output of the BiLSTM hidden layer is connected to the input of the attention layer. The attention layer is trained through a fully connected layer, and the output of the fully connected layer is scaled and normalized using the sigmoid function.

[0033] The last layer is the output layer, which outputs the remaining life prediction results of PEMFC.

[0034] Furthermore, the data dimension of the input layer is the same as the data dimension of the input data.

[0035] Furthermore, the BiLSTM hidden layer is a double-layer structure, the upper layer is a forward LSTM, and the lower layer is a reverse LSTM.

[0036] Furthermore, the dynamic operation data of the PEMFC to be predicted in an evaluation cycle is collected in real time, and the corresponding input data is constructed and executed:

[0037] The dynamic operation data of the PEMFC to be predicted in an evaluation cycle is preprocessed, and the corresponding trimmed data set is constructed, and then the correlation matrix is ​​calculated to obtain the corresponding input data.

[0038] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0039] The PEMFC remaining life dynamic prediction method based on the attention neural network provided by the present invention has the following advantages:

[0040] (1) In terms of information preprocessing, different from traditional data processing methods, the Hampel filter is first used for data smoothing and linear normalization, and then the Pearson correlation coefficient is introduced to select key influencing factors. The correlation matrix between the key influencing factors and the stack voltage is obtained based on the degree of influence. Based on the associated influencing factors and the correlation matrix, the impact on the stack voltage can be better quantified and the prediction accuracy can be improved.

[0041] (2) The CNN-BiLSTM-Attention model is used to predict the life of the fuel cell. After data preprocessing, the combination of CNN and BiLSTM is introduced in the dynamic prediction part of the model. The attention mechanism is added to the output of the model, and finally the dynamic remaining life prediction result of PEMFC is output. The overall design selects the appropriate number of network layers according to the timing and data set size of the dynamic operation data, and solves the impact of multi-physical field nonlinearity and volatility on the prediction. Specifically, in the input layer, the operating data set of PEMFC under dynamic conditions is input, and the data dimension can be selected according to the type of PEMFC training set data, so as to achieve rational screening of data. In addition, the CNN-BiLSTM-Attention model established in this embodiment takes into account the characteristics of PEMFC dynamic operation data, and uses CNN to extract data from the divided training set. Convolution is performed first and then pooling operation is performed to extract dynamic operation features, so as to achieve the extraction of feature matrix. The dropout layer is added to prevent overfitting, and the filtered and extracted information is connected to the BiLSTM network, taking into account the past and future time series information, increasing the network's information extraction ability and deep mining of time series information. The Attention layer is added to the output of the BiLSTM network to improve the efficiency of information processing through differentiated weight distribution, thereby solving the problem of information loss caused by long sequences in BiLSTM.

[0042] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0044] Figure 1 A flowchart of a method for dynamically predicting the remaining life of a PEMFC based on an attention neural network provided in an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of the structure of a multi-layer neural network CNN-BiLSTM-Attention model provided in an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of the dynamic test results of the CNN-LSTM model provided in an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of the dynamic test results of the CNN-BiLSTM model provided in an embodiment of the present invention;

[0048] Figure 5 A schematic diagram of the dynamic test results of the CNN-LSTM-Attention model provided in an embodiment of the present invention;

[0049] Figure 6 A schematic diagram of dynamic test results of the multi-layer neural network CNN-BiLSTM-Attention model provided in an embodiment of the present invention;

[0050] Figure 7 A test evaluation comparison chart of the CNN-BiLSTM-Attention model, the CNN-LSTM model, and the multi-layer neural network CNN-BiLSTM-Attention model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0052] A specific embodiment of the present invention discloses a method for dynamically predicting the remaining life of a PEMFC based on an attention neural network, the flow chart of which is as follows: Figure 1 As shown, the following steps are included.

[0053] Step S1: collecting and preprocessing the dynamic operation data of the training PEMFC at each operation time in the whole life cycle; then, screening the key influencing factors according to the influence of various parameters in the preprocessed dynamic operation data on the dynamic prediction of the remaining life of the PEMFC.

[0054] In this embodiment, the dynamic operation data of PEMFC at each operation time in the whole life cycle includes at least: stack voltage, stack current, battery temperature, fuel gas pressure, fuel gas temperature, and fuel gas humidity. For example, in the specific implementation process, the sampling frequency can be set to 1Hz. Among the various parameter data of the above dynamic operation data, the stack voltage directly reflects the remaining life of PEMFC and can be used as the prediction output of the later prediction model. Therefore, in the later model training and prediction process, the stack voltage is used as the output data to evaluate the remaining life of PEMFC. In addition to the stack voltage, the remaining parameter data have a certain influence on the dynamic prediction of the remaining life of PEMFC. However, due to the complex structure of PEMFC, the influence of these parameter data on the remaining life of PEMFC is difficult to express with an exact expression, and the degree of influence is also different. Therefore, this step also screens out parameter data with a greater degree of influence by calculating the degree of influence, which is used for later model training and prediction. The specific process is described as follows.

[0055] In this embodiment, in view of the high dimension and complex structure of the dynamic operation data of the fuel cell, and considering the abnormal data and noise interference that occur during the parameter acquisition process of the sensor, the following preprocessing method is formulated.

[0056] Step S11: pre-processing the dynamic operation data: determine whether there are outlier data in the various parameter data of the dynamic operation data of each operation time of the PEMFC in the whole life cycle. If so, the dynamic operation data of the corresponding operation time are eliminated, and the dynamic operation data of the corresponding operation time are completed by using an interpolation algorithm; then, the Hampel filter is used to smooth and linearly normalize the various parameter data in the dynamic operation data, and the various parameter data are mapped to the range of [0,1] to obtain the normalized dynamic operation data of the PEMFC at each operation time in the whole life cycle.

[0057] Step S12: Based on the normalized dynamic operation data of PEMFC at each operating time during its entire life cycle, quantify the influence of various parameters and operating time other than the stack voltage on the stack voltage, and select parameter categories that have an effective influence on the dynamic prediction of the remaining life of PEMFC as key influencing factors from various parameters and operating time other than the stack voltage.

[0058] Preferably, in this embodiment, the Pearson correlation coefficient is introduced to quantify the influence of each influencing factor on the key measurement variable, so as to screen out parameter categories whose absolute value of the influence degree is greater than or equal to a preset influence degree threshold.

[0059] Step S2: trimming the normalized dynamic operation data to construct a trimmed data set of the corresponding relationship between key influencing factors and the stack voltage; combining the influence degree, obtaining a correlation matrix between the key influencing factors and the stack voltage; constructing a data sample set based on the trimmed data set and the correlation matrix.

[0060] Step S21: sum up the influence degrees of all key influencing factors and normalize them so that the sum is 1; calculate the relative weight of the influence degree of each key influencing factor as the weight of each key influencing factor.

[0061] Step S22: Divide the entire life cycle into multiple evaluation cycles on average; in each evaluation cycle, select the stack voltage of the current evaluation cycle from the trimmed data set to construct a state matrix, select the key influencing factors of the current evaluation cycle to construct a coupling matrix, and determine the coupling degree matrix between the key influencing factors of the current evaluation cycle and the stack voltage based on the coupling relationship between the state matrix and the coupling matrix.

[0062] In an exemplary embodiment, The state matrix of an evaluation cycle is expressed as , the coupling matrix is ​​expressed as ; Indicates the number of key influencing factors, represents the number of sampling times in an evaluation cycle,

[0063] Assume The coupling matrix of the evaluation cycle is expressed as ,according to ,get .

[0064] Step S23: multiply the coupling degree of each key influencing factor in the coupling degree matrix between the key influencing factors of the current evaluation cycle and the stack voltage by the weight to obtain the correlation matrix between the key influencing factors of the current evaluation cycle and the stack voltage.

[0065] Step S24: taking the trimmed data set and the correlation matrix of each evaluation cycle as input data, and taking the stack voltage after the evaluation cycle ends for a preset time as the label of the input data; thereby constructing a data sample set.

[0066] Step S3: construct a multi-layer neural network CNN-BiLSTM-Attention model and use it as a remaining life dynamic prediction model; use the data sample set to train and test the remaining life dynamic prediction model to obtain a remaining life dynamic prediction model that passes the test.

[0067] In this embodiment, the structure of the multi-layer neural network CNN-BiLSTM-Attention model is as follows Figure 2 As shown, they are:

[0068] (1) The first layer is the input layer, which is used to receive input data. The data dimension of the input layer is the same as the data dimension of the input data.

[0069] (2) The second layer is the convolutional neural network (CNN) layer. The convolutional neural network layer first sets up a one-dimensional convolutional layer, which is used to extract features from the input data. The input of the one-dimensional convolutional layer is connected to the output of the input layer, and the output of the one-dimensional convolutional layer is connected to the input of the pooling layer. The pooling layer is used to pool the features extracted by the one-dimensional convolutional layer, and the output of the pooling layer is connected to the input of the Dropout layer. During the training process of the convolutional neural network, the Dropout layer temporarily discards some neural network units from the network according to a certain probability.

[0070] In this embodiment, the one-dimensional convolution layer (Conv1D) is provided with 64 convolution kernels, each of which has a size of 2 and a moving step of 1, and the ReLU function is used as the activation function of the convolution layer. After feature extraction of the input data, the feature extraction result is transmitted to the pooling layer. The pooling layer adopts the maximum pooling strategy, and the size of the pooling window is 3. After retaining the key information of the data and reducing the dimension of the data through the pooling layer, in order to prevent overfitting, the Dropout layer is added to the convolutional neural network layer, and the parameter is set to 0.2.

[0071] (3) The third layer is a bidirectional long short-term memory neural network (BiLSTM) layer. In the bidirectional long short-term memory neural network layer, the output of the Dropout layer is connected to the input of the BiLSTM hidden layer. 64 hidden layer nodes are set in the BiLSTM hidden layer, and the activation function uses the tanh function. Specifically, the BiLSTM hidden layer is a two-layer structure, the upper layer is a forward LSTM, the lower layer is a reverse LSTM, and the number of neurons is 32.

[0072] (4) The fourth layer is the attention layer. The output of the BiLSTM hidden layer is connected to the input of the attention layer. The attention layer is trained through a fully connected layer, and the output of the fully connected layer is scaled and normalized using the sigmoid function to obtain the distribution weight of the vector corresponding to each BiLSTM hidden layer.

[0073] (5) The last layer is the output layer, which outputs the remaining life prediction result of PEMFC. The remaining life prediction result is the stack voltage of PEMFC.

[0074] Based on the above description, the multi-layer neural network CNN-BiLSTM-Attention model constructed in this embodiment can select the appropriate number of network layers according to the characteristics of PEMFC dynamic operation data, and for the timing of dynamic operation data and data set size. The weight coefficient of the neural network CNN-BiLSTM-Attention model is learned and optimized using the training set. The input data in the training set is input into the input layer of the neural network CNN-BiLSTM-Attention model, and CNN is used for data extraction. The dynamic operation features are extracted by convolution and then pooling. The Dropout layer is added to prevent overfitting caused by repeated data training, and the filtered and extracted information is then connected to the BiLSTM network, taking into account past and future time series information. The output of the BiLSTM network is input into the Attention layer, and the efficiency of information processing is improved by differentiated weight distribution. Finally, the attention layer is added, and the output of the fully connected layer is normalized using the sigmoid function. For example, the number of iterations is set to 50 times.

[0075] In the specific implementation process, the data sample set is divided into a training sample set and a test sample set. The training sample set is used to train the dynamic prediction model of remaining life, and the weight coefficient in the dynamic prediction model of remaining life is optimized; then the trained dynamic prediction model of remaining life is tested with the test sample set, and the mean square error, mean absolute error and root mean square error are calculated according to the test results. If the mean square error, mean absolute error and root mean square error all meet the expected requirements, the test passes, and the dynamic prediction model of remaining life that passes the test is obtained; otherwise, jump to the step of training the dynamic prediction model of remaining life with the training sample set. The error function includes the calculation formula of mean square error (MSE), mean absolute error (MAE) and root mean square error (RMSE) as follows:

[0076] (1) ;

[0077] (2) ;

[0078] (3);

[0079] Where: is the total number of test samples in the test sample set. A set of input data and output data with an input-output correspondence relationship in the test sample set is regarded as a test sample; For the test sample set The stack voltage output by the test sample; For the test sample set The stack voltage in each test sample.

[0080] Step S4: collect the dynamic operation data of the PEMFC to be predicted in an evaluation cycle in real time, and construct the corresponding input data; input the constructed input data into the remaining life dynamic prediction model that has passed the test, and the remaining life dynamic prediction model predicts and outputs the predicted value of the stack voltage after the preset time period at the end of the current evaluation cycle, so as to evaluate the remaining life of the PEMFC to be predicted.

[0081] The process of constructing and obtaining the corresponding input data in step S4 is implemented by referring to steps S1 and S2. First, the dynamic operation data of the predicted PEMFC in an evaluation cycle is preprocessed, and the corresponding trimmed data set is constructed, and then the correlation matrix is ​​calculated to obtain the corresponding input data, so that the remaining life of the predicted PEMFC can be evaluated.

[0082] The PEMFC remaining life dynamic prediction method based on the attention neural network provided in this embodiment has the following advantages:

[0083] (1) In terms of information preprocessing, different from traditional data processing methods, the Hampel filter is first used for data smoothing and linear normalization, and then the Pearson correlation coefficient is introduced to select key influencing factors. The correlation matrix between the key influencing factors and the stack voltage is obtained based on the degree of influence. Based on the associated influencing factors and the correlation matrix, the impact on the stack voltage can be better quantified and the prediction accuracy can be improved.

[0084] (2) The CNN-BiLSTM-Attention model is used to predict the life of the fuel cell. After data preprocessing, the combination of CNN and BiLSTM is introduced in the dynamic prediction part of the model. The attention mechanism is added to the output of the model, and finally the dynamic remaining life prediction result of PEMFC is output. The overall design selects the appropriate number of network layers according to the timing and data set size of the dynamic operation data, and solves the impact of multi-physical field nonlinearity and volatility on the prediction. Specifically, in the input layer, the operating data set of PEMFC under dynamic conditions is input, and the data dimension can be selected according to the type of PEMFC training set data, so as to achieve rational screening of data. In addition, the CNN-BiLSTM-Attention model established in this embodiment takes into account the characteristics of PEMFC dynamic operation data, and uses CNN to extract data from the divided training set. Convolution is performed first and then pooling operation is performed to extract dynamic operation features, so as to achieve the extraction of feature matrix. The dropout layer is added to prevent overfitting, and the filtered and extracted information is connected to the BiLSTM network, taking into account the past and future time series information, increasing the network's information extraction ability and deep mining of time series information. The Attention layer is added to the output of the BiLSTM network to improve the efficiency of information processing through differentiated weight distribution, thereby solving the problem of information loss caused by long sequences in BiLSTM.

[0085] Embodiment 2 of the present invention also provides a specific example of a method for dynamically predicting the remaining life of a PEMFC based on an attention neural network. In this example, the dynamic operation data of the first 600 hours of training the PEMFC is collected, and the corresponding data sample set is obtained according to the processing method of step S1-step S2 to train and test the dynamic prediction model for the remaining life. In this process, the loss function of the model can be set to a mean square error function, and the parameters of the dynamic prediction model for the remaining life are saved when the loss function is the smallest. Table 1 shows the specific parameters of the multi-layer neural network CNN-BiLSTM-Attention model.

[0086] Table 1 Specific parameters of the multi-layer neural network CNN-BiLSTM-Attention model

[0087]

[0088] In order to verify the effectiveness of the improved model in this embodiment, after the model training is completed, the test sample sets are brought into different models for testing and comparative experiments are carried out. Figure 3 , Figure 4 , Figure 5 , Figure 6The comparison between the test values ​​of the four models, CNN-LSTM, CNN-BiLSTM neural network, CNN-LSTM-Attention and CNN-BiLSTM-Attention, on the test sample set and the true values ​​of the test sample set. The solid line is the actual stack voltage and the dotted line is the stack voltage predicted by the corresponding model. Figure 7 This is a test evaluation comparison chart of the CNN-BiLSTM-Attention model, the CNN-LSTM model, and the multi-layer neural network CNN-BiLSTM-Attention model provided in the embodiment of the present invention. Figure 3-Figure 7 It can be seen that the test results of the multi-layer neural network CNN-BiLSTM-Attention model are closest to the true value. After the test passes, the prediction of the remaining life of the PEMFC can be used.

[0089] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0090] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A PEMFC remaining life dynamic prediction method based on attention neural network, characterized in that: The method comprises: Collect and preprocess the dynamic operation data of the training PEMFC at each operating time in the entire life cycle; screen the key influencing factors according to the influence of various parameters in the preprocessed dynamic operation data on the dynamic prediction of the remaining life of PEMFC; The normalized dynamic operation data is trimmed to construct a trimmed data set of the corresponding relationship between the key influencing factors and the stack voltage; a correlation matrix between the key influencing factors and the stack voltage is obtained in combination with the influence degree; a data sample set is constructed based on the trimmed data set and the correlation matrix; Constructing a multi-layer neural network CNN-BiLSTM-Attention model and using it as a remaining life dynamic prediction model; using the data sample set to train and test the remaining life dynamic prediction model to obtain a remaining life dynamic prediction model that passes the test; The dynamic operation data of the PEMFC to be predicted in an evaluation cycle is collected in real time, and the corresponding input data is constructed. The remaining life dynamic prediction model processes the input data and predicts and outputs the predicted value of the stack voltage after the preset length of time at the end of the current evaluation cycle; The structure of the multi-layer neural network CNN-BiLSTM-Attention model is as follows: The first layer is the input layer, which is used to receive input data; The second layer is the convolutional neural network layer; the convolutional neural network layer first sets a one-dimensional convolutional layer, which is used to extract features from the input data; the input end of the one-dimensional convolutional layer is connected to the output end of the input layer, and the output end of the one-dimensional convolutional layer is connected to the input end of the pooling layer; the pooling layer is used to perform pooling processing on the features extracted by the one-dimensional convolutional layer, and the output end of the pooling layer is connected to the input end of the Dropout layer; The third layer is a bidirectional long short-term memory neural network layer. In the bidirectional long short-term memory neural network layer, the output of the Dropout layer is connected to the input of the BiLSTM hidden layer. 64 hidden layer nodes are set in the BiLSTM hidden layer, and the activation function uses the tanh function. The fourth layer is the attention layer; the output of the BiLSTM hidden layer is connected to the input of the attention layer. The attention layer is trained through a fully connected layer, and the output of the fully connected layer is scaled and normalized using the sigmoid function. The last layer is the output layer, which outputs the remaining life prediction results of PEMFC.

2. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 1 is characterized in that: Preprocess the dynamic run data by doing the following: It is determined whether there are outliers in the various parameter data of the dynamic operation data of each operation time of the PEMFC in the whole life cycle. If there are outliers, the dynamic operation data of the corresponding operation time are eliminated, and the dynamic operation data of the corresponding operation time are supplemented by using an interpolation algorithm. Afterwards, the Hampel filter is used to smooth and linearly normalize the parameter data in the dynamic operation data, and the parameter data are mapped to the range of [0,1] to obtain the normalized dynamic operation data of PEMFC at each operating time in its entire life cycle.

3. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 2 is characterized in that: Filter key influencers by doing the following: Based on the normalized dynamic operation data of PEMFC at various operating times during its entire life cycle, the influence of various parameters and operating time other than the stack voltage on the stack voltage is quantified, and the parameter categories that have an effective impact on the dynamic prediction of the remaining life of PEMFC are selected as the key influencing factors from various parameters and operating time other than the stack voltage.

4. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 3 is characterized in that: The Pearson correlation coefficient is introduced to quantify the influence of each influencing factor on the key measurement variables.

5. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to any one of claims 1 to 4, characterized in that: Construct a data sample set by doing the following: The influence degrees of all key influencing factors are summed up and normalized so that the sum is 1; the relative weight of the influence degree of each key influencing factor is calculated as the weight of each key influencing factor; The whole life cycle is evenly divided into multiple evaluation cycles; in each evaluation cycle, the stack voltage of the current evaluation cycle is selected from the trimmed data set to construct a state matrix, and the key influencing factors of the current evaluation cycle are selected to construct a coupling matrix, and the coupling degree matrix between the key influencing factors of the current evaluation cycle and the stack voltage is determined based on the coupling relationship between the state matrix and the coupling matrix; The coupling degree of each key influencing factor in the coupling degree matrix of the key influencing factors of the current evaluation cycle and the stack voltage is multiplied by the weight to obtain the correlation matrix of the key influencing factors of the current evaluation cycle and the stack voltage; The trimmed data set and the correlation matrix of each evaluation cycle are used as input data, and the stack voltage after the evaluation cycle ends for a preset time is used as the label of the input data; Thus, a data sample set is constructed.

6. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 5 is characterized in that: No. The state matrix of an evaluation cycle is expressed as , the coupling matrix is ​​expressed as ; Indicates the number of key influencing factors, represents the number of sampling times in an evaluation cycle, No. The coupling matrix of the evaluation cycle is expressed as ,according to ,get .

7. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 1 is characterized in that: The data dimension of the input layer is the same as the data dimension of the input data.

8. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 7 is characterized in that: The BiLSTM hidden layer is a double-layer structure, the upper layer is a forward LSTM, and the lower layer is a reverse LSTM.

9. The method for dynamic prediction of PEMFC remaining life based on attention neural network according to claim 8, characterized in that: Real-time collection of the dynamic operation data of the PEMFC to be predicted in an evaluation cycle, and the corresponding input data are constructed and executed: The dynamic operation data of the PEMFC to be predicted in an evaluation cycle is preprocessed, and the corresponding trimmed data set is constructed, and then the correlation matrix is ​​calculated to obtain the corresponding input data.

Images