Proton exchange membrane fuel cell performance degradation prediction method based on deep learning

Abstract

The invention discloses a proton exchange membrane fuel cell performance degradation prediction method based on deep learning, and belongs to the technical field of automation. According to the method, dynamic working condition degradation data obtained by a fuel cell system in a period of continuous time is taken as an original input signal, a Pearson's correlation coefficient method is adopted for processing, and then an output voltage variable in the preprocessed data is smoothed by adopting wavelet transform; and decomposing a single batch of input by the trained multi-scale decomposition module, and inputting the long-term modal sequence and the short-term modal sequence of each batch into the trained performance degradation prediction model to obtain a final fuel cell degradation data prediction result. By utilizing the method, long-term modal and short-term modal data characteristics of the proton exchange membrane fuel cell under the dynamic working condition can be deeply researched, and the prediction accuracy of the degradation data is powerfully enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of automation technology, and in particular relates to a proton exchange membrane fuel cell performance degradation prediction method based on deep learning. Background Art

[0002] As the global demand for energy continues to increase, the environmental pollution and energy shortage caused by traditional energy sources are becoming more and more serious. Under this situation, people are eager to find renewable and clean energy. Proton exchange membrane fuel cells (PEMFCs) use hydrogen as fuel and have the characteristics of high energy conversion efficiency, low operating temperature, high specific power and good environmental friendliness. They are considered to be one of the effective energy solutions, and therefore have attracted much attention and received support from many policies.

[0003] However, despite the many advantages of PEMFC, its reliability and durability face challenges in practical applications. The performance of PEMFC will gradually decline over time, which is mainly affected by factors such as mechanical stress under dynamic conditions, temperature changes during operation, membrane aging, dynamic load cycles, and start-stop operations. Therefore, accurate and efficient PEMFC performance degradation prediction technology can provide strong information support for system maintenance strategy design, operating condition optimization, and health status assessment.

[0004] Since PEMFC is a complex electrochemical system involving multiple physical quantities, multiple components and multiple factors that are strongly coupled, its performance decay process shows strong nonlinearity, randomness and uncertainty, which greatly increases the difficulty of predicting its performance degradation. At present, the existing prediction methods of proton exchange membrane fuel cells have shortcomings: first, it is difficult to accurately predict its remaining life under dynamic operating conditions; second, the correlation between fuel cell operating variables and the multimodal information of variables on the time scale are not fully considered; third, the existing data-driven and hybrid prediction methods rely on complex nonlinear connections and basic model combinations of many internal parameters, resulting in poor model interpretability, high computational complexity and easy overfitting.

[0005] Therefore, a proton exchange membrane fuel cell dynamic operating data prediction method is urgently needed to fill the research gaps in prediction accuracy, multimodal information mining and model structure optimization. Summary of the invention

[0006] The purpose of the present invention is to solve the defects of low prediction accuracy under dynamic working conditions, inability to effectively mine multimodal information in working data, complex model structure, etc. in the prior art, and provide a proton exchange membrane fuel cell performance degradation prediction method based on deep learning. The method of the present invention screens out characteristic variables highly correlated with output voltage variables from the original data according to the working condition characteristics of the proton exchange membrane fuel cell and the Pearson correlation coefficient method. The screened data is smoothed by wavelet transform to reduce noise such as spikes in the data. The pre-processed data is decomposed into long-term and short-term modes by multi-scale decomposition method. For the short-term mode, a combined short-term prediction module including encoding, internal decomposition, equidistant convolution blocks and other parts is used to extract the nonlinear deep features of the data, and a weight encoder, a position encoder and an equidistant convolution block are used to further explore local nonlinear features. For the long-term mode, a gated recurrent unit is used to extract the long-term change trend of the data to obtain the predicted value of the long-term mode. The predicted values ​​of the two modal outputs are integrated through a fully connected layer to obtain the final prediction result.

[0007] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0008] A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning comprises the following steps:

[0009] S1. taking the dynamic operating degradation data of the fuel cell system acquired over a continuous period of time as the original input signal;

[0010] S2. Using the Pearson correlation coefficient method to screen the characteristic variables that are highly correlated with the output voltage variable in the original input signal and deleting the redundant variables, the preprocessed data is obtained;

[0011] S3. Smoothing the output voltage variable in the preprocessed data using wavelet transform to obtain smoothed data and generate an input data set;

[0012] S4. After batch processing and dividing the input data set, a single batch input is obtained, and the single batch input is decomposed using the trained multi-scale decomposition module to obtain a long-term modal sequence and a short-term modal sequence;

[0013] S5. The long-term modal sequence and short-term modal sequence of each batch are input into the trained performance degradation prediction model. The long-term modal sequence is processed by the gated recurrent unit to obtain the long-term modal sequence prediction result. The short-term modal sequence is processed by the short-term prediction module to obtain the short-term modal sequence prediction result. The long-term modal sequence prediction result and the short-term modal sequence prediction result are integrated through the fully connected layer to obtain the final fuel cell degradation data prediction result.

[0014] Based on the above scheme, each step can be implemented in the following preferred specific manner.

[0015] Preferably, the specific implementation method of step S2 is as follows:

[0016] S21. Taking the output voltage variable as the target variable, the Pearson correlation coefficient between each characteristic variable in the original input signal and the target variable is calculated;

[0017] S22. Sort all feature variables in the original input signal by the absolute value of the Pearson correlation coefficient, and only retain feature variables whose absolute value of the Pearson correlation coefficient is higher than a preset screening threshold;

[0018] S23. According to the actual mechanism of the fuel cell operating condition, if the characteristic variables in the original input signal contain redundant variables, the redundant variables are directly deleted, otherwise no processing is performed.

[0019] Preferably, the specific implementation method of step S3 is as follows:

[0020] S31. Selecting a wavelet basis function according to the signal characteristics of the output voltage variable to perform multi-layer decomposition on the preprocessed data to obtain a wavelet coefficient sequence;

[0021] S32. Performing soft threshold processing on the wavelet coefficient sequence to obtain a processed wavelet coefficient sequence;

[0022] S33. Reconstruct the output voltage signal using the processed wavelet coefficient sequence to obtain smoothed data and construct an input data set from the smoothed data.

[0023] Preferably, the specific implementation method of step S4 is as follows:

[0024] S41. Treat a single batch input as a long-term modal sequence;

[0025] S42. Perform zero-padding operation on the single batch input to obtain first padding data, use the first convolution kernel to perform average pooling on the first padding data to obtain first average pooling data, use the second convolution kernel to perform average pooling on the first padding data to obtain second average pooling data, average the first average pooling data and the second average pooling data to obtain first mean data, and use the difference between the single batch input and the first mean data as the short-term modal sequence.

[0026] Preferably, in step S5, the specific processing process in the short-term prediction module is as follows:

[0027] S51. First, concatenate the input short-term modal sequence with the 0-filled sequence with the same feature dimension to obtain a concatenated short-term modal sequence, and encode the concatenated short-term modal sequence using a weight encoder and a position encoder to obtain an encoded sequence;

[0028] S52. Use a multi-scale decomposition function to internally decompose the encoded sequence to obtain two internal decomposition subsequences;

[0029] S53. In the equidistant convolution block, firstly, each of the internal decomposition subsequences is downsampled, and each internal decomposition subsequence corresponds to a sampled local sequence. The deep features of each sampled local sequence are extracted by an equidistant convolution layer, and each sampled local sequence corresponds to a deep feature extracted by an equidistant convolution. Each deep feature is processed by an upsampling convolution layer, and each deep feature corresponds to an output feature of the upsampling convolution layer. The output features of the upsampling convolution layer are fused by a two-dimensional convolution layer to obtain a complete output feature of the equidistant convolution layer. Finally, the complete output feature of the equidistant convolution layer is projected by a single-layer FNN network to obtain the output result of the equidistant convolution block.

[0030] S64. Project and truncate the output results of the equidistant convolution block to obtain the short-term modal sequence prediction results.

[0031] Preferably, in step S52, the specific processing process of internal decomposition by the multi-scale decomposition function is as follows: perform a 0-filling operation on the encoded sequence to obtain second filled data, use a third convolution kernel to average pool the second filled data to obtain third average pooled data, use a fourth convolution kernel to average pool the second filled data to obtain fourth average pooled data, average the third average pooled data and the fourth average pooled data to obtain second mean data, use the difference between the encoded sequence and the second mean data as the second internal decomposition subsequence, and use the encoded sequence as the first internal decomposition subsequence.

[0032] Preferably, in step S53, the specific processing process of downsampling each internal decomposition subsequence is as follows: perform a 0-filling operation on the first internal decomposition subsequence to obtain third filling data, perform a 0-filling operation on the second internal decomposition subsequence to obtain fourth filling data, use a convolution kernel with a length of i to perform average pooling on the third filling data to obtain fifth average pooling data, use a convolution kernel with a length of i to perform average pooling on the fourth filling data to obtain sixth average pooling data, and then use a convolution kernel with a length of i to perform one-dimensional convolution on the fifth average pooling data to obtain a first local sequence after sampling corresponding to the first internal decomposition subsequence, and then use a convolution kernel with a length of i to perform one-dimensional convolution on the sixth average pooling data to obtain a second local sequence after sampling corresponding to the second internal decomposition subsequence.

[0033] Preferably, in step S53, the specific processing procedure of extracting deep features by the equidistant convolution layer is as follows: after performing equidistant convolution, Tanh activation function and Dropout operations on the first local sequence in turn, a first initial feature is obtained, the first initial feature and the first local sequence are added and layer regularization is performed to obtain a first deep feature corresponding to the first local sequence, and after performing equidistant convolution, Tanh activation function and Dropout operations on the second local sequence in turn, a second initial feature is obtained, and the second initial feature and the second local sequence are added and layer regularization is performed to obtain a second deep feature corresponding to the second local sequence.

[0034] Preferably, in step S53, the process of processing the deep features by the upsampling convolution layer is as follows: after performing upsampling convolution, Tanh activation function and Dropout operations on the first deep feature in sequence, a third initial feature is obtained, the third initial feature and the first deep feature are added and layer regularization is performed to obtain the output feature corresponding to the first deep feature, and the second deep feature is successively performed upsampling convolution, Tanh activation function and Dropout operations to obtain the fourth initial feature, and the fourth initial feature and the second deep feature are added and layer regularization is performed to obtain the output feature corresponding to the second deep feature.

[0035] Preferably, in step S53, the specific process of using a single-layer FNN network for projection transformation is as follows: the complete output features of the equidistant convolution layer are input into the single-layer FNN network to obtain projection output features, the projection output features are added to the complete output features of the equidistant convolution layer, and then layer regularization is performed to obtain the output result of the equidistant convolution block.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) The present invention provides a proton exchange membrane fuel cell performance degradation prediction method based on deep learning, which can better play the advantages of each module and achieve complementary advantages. Under dynamic operating conditions, the prediction accuracy is significantly improved.

[0038] (2) The present invention adopts a feature selection method, selects feature variables according to the intrinsic degradation mechanism and the Pearson correlation coefficient method, and uses wavelet transform to reduce the influence of noise such as internal spikes, effectively reducing the influence of unimportant data and internal noise.

[0039] (3) The present invention proposes a multi-scale decomposition module to extract long-term and short-term patterns, which are processed by different methods. The multi-scale decomposition module can assist subsequent modules in extracting temporal information from degradation data of different scales.

[0040] (4) The present invention proposes a short-term prediction module, which includes four parts: encoding, internal decomposition, equidistant convolution block and projection interception, which can further explore the local nonlinear characteristics of the data.

[0041] (5) The long-term modality uses gated recurrent units to collect information on longer time scales, which also helps to balance the local overfitting caused by short-term prediction results, thereby improving the overall accuracy of the performance degradation prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall architecture of the method of the present invention;

[0043] Figure 2 It is a schematic diagram of the Pearson correlation coefficient matrix of each characteristic variable of the present invention;

[0044] Figure 3 is a schematic diagram of the architecture of the gated recurrent unit used in the present invention;

[0045] Figure 4 It is a schematic diagram of the overall architecture of the short-term prediction module of the present invention;

[0046] Figure 5 It is a schematic diagram for comparing the values ​​of the real data and the predicted results of the present invention;

[0047] Figure 6 It is a schematic diagram for comparing the errors between the real data and the predicted results of the present invention. DETAILED DESCRIPTION

[0048] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0049] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0050] The embodiment of the present invention is based on a section of continuous dynamic operating condition data collected by a fuel cell system. The data set is composed of continuous dynamic operating condition data, and there are 29 input dimensions and 10,000 discrete time recording points in the data set. The 29 input dimensions are: working time, total battery voltage, relative humidity, DC input current feedback, DC output current feedback, single chip sub-low voltage, air inlet temperature, maximum allowable current, hydrogen inlet temperature, air inlet pressure, air outlet pressure, hydrogen inlet pressure, hydrogen outlet pressure, cooling water inlet temperature, cooling water outlet temperature, ambient temperature, DC input voltage feedback, process load current, water pump power, DC temperature feedback, cooling temperature difference, average single chip voltage, water pump voltage, single chip maximum voltage, cooling water inlet pressure, stack response current, actual air flow, single chip minimum voltage.

[0051] like Figure 1 As shown, in a preferred implementation of the present invention, the above-mentioned proton exchange membrane fuel cell performance degradation prediction method based on deep learning includes the following steps S1 to S5. The specific implementation process is described in detail below.

[0052] 1. Data Processing

[0053] S1. The dynamic operating condition degradation data of the fuel cell system acquired over a continuous period of time is used as an original input signal.

[0054] It should be noted that in step S1, the dynamic working condition degradation data has been cleaned, and the time interval for collecting dynamic working condition degradation data is the same. In the present invention, the original input signal can be recorded as X in ∈R m×c , m is the total time length of the original input signal, and c is the input dimension of the original input signal. In this embodiment, m=10000, c=29.

[0055] S2. Use Pearson correlation coefficient method to filter the original input signal X in After selecting the characteristic variables that are highly correlated with the output voltage variable and deleting the redundant variables, the preprocessed data is obtained.

[0056] It should be noted that, in the present invention, the preprocessed data can be recorded as X Vrel ∈R m×d , d is the input dimension of the preprocessed data. In this embodiment, a total of 22 redundant variables are deleted, so the input dimension of the preprocessed data is d=7.

[0057] In step S2, the specific implementation method of using the Pearson correlation coefficient method to screen the characteristic variables is as follows:

[0058] S21. Take the output voltage variable as the target variable T and calculate the original input signal X using the following formula in The Pearson correlation coefficient r between each feature variable F and the target variable T ft :

[0059]

[0060] Among them, F i represents the observed value of the characteristic variable; T i represents the observed value of the target variable; F represents the observed value of the characteristic variable i The average value of represents the observed value T of the target variable i The average value of .

[0061] S22. The absolute value of the Pearson correlation coefficient is used to compare the original input signal X in All characteristic variables in the dataset are sorted and only characteristic variables with absolute values ​​of Pearson correlation coefficients higher than the preset screening threshold are retained.

[0062] S23. According to the actual mechanism of fuel cell operation, if the original input signal X in If the feature variables in contain redundant variables, they are directly deleted, otherwise no processing is performed.

[0063] Figure 2 The Pearson correlation coefficient between the variables is displayed, and the variables are screened according to the coefficient, and the characteristic variables are selected according to the pre-set threshold.

[0064] S3. Preprocessed data X Vrel The output voltage variable is smoothed by wavelet transform to obtain the smoothed data X∈R m×d And generate the input dataset.

[0065] In step S3, the preprocessed data X Vrel The output voltage variable is smoothed by wavelet transform, and the specific implementation method is as follows:

[0066] S31. Select the wavelet basis function according to the signal characteristics of the output voltage variable to process the preprocessed data X Vrel Perform multi-layer decomposition to obtain a sequence of wavelet coefficients.

[0067] In this embodiment, the Dobesi wavelet with three vanishing moments is selected to process the preprocessed data X Vrel Perform 4-layer decomposition.

[0068] S32. Perform soft threshold processing on the wavelet coefficient sequence to obtain a processed wavelet coefficient sequence.

[0069] In this embodiment, the Donoho-Johnstone threshold denoising method is specifically used to set a dynamically adjustable threshold to obtain a processed wavelet coefficient sequence.

[0070] S33. Use the processed wavelet coefficient sequence to reconstruct the output voltage signal and obtain the smoothed data X∈R m×d And construct the input dataset from the smoothed data.

[0071] 2. Multi-scale Decomposition

[0072] S4. After batch processing of the input data set, a single batch input X is obtained. i ∈R I×d , use the trained multi-scale decomposition module to analyze the single batch input X i Decompose and obtain the long-term modal sequence X t ∈R I×d and the short-term modal series X s ∈R I×d , where I is the batch length of a single batch input.

[0073] Step S4, the specific process of decomposition by the multi-scale decomposition module is as follows:

[0074] S41. Treat a single batch input as a long-term modal sequence: X t =X i ;

[0075] S42. Perform a zero padding operation on the single batch input to obtain first padding data, use the first convolution kernel to perform average pooling on the first padding data to obtain first average pooling data, use the second convolution kernel to perform average pooling on the first padding data to obtain second average pooling data, average the first average pooling data and the second average pooling data to obtain first mean data, and use the difference between the single batch input and the first mean data as the short-term modal sequence:

[0076] X s =X i -mean(AvgPool(Padding(X i )) kernel1 ,AvgPool(Padding(X i )) kemel2 )

[0077] Among them, Padding() represents the zero padding operation; mean() represents the mean operation; AvgPool() kernel1 Indicates average pooling by the first convolution kernel, AvgPool() kemel2 Indicates average pooling by the second convolution kernel.

[0078] In the present invention, after the input data set is divided into batches, each batch of input data is obtained, that is, a single batch input X i The data batch processing can select the data length I of each batch of input. In this embodiment, each batch of input data takes 11 time points, corresponding to I=11. After the input data set is decomposed by the multi-scale decomposition module, the long-term modal sequence X is obtained. t ∈R 11×7 and the short-term modal series X s ∈R 11×7 .

[0079] S5. Input the long-term modal sequence and short-term modal sequence of each batch into the trained performance degradation prediction model, and process the long-term modal sequence by the gated recurrent unit to obtain the long-term modal sequence prediction result Y t ∈R O×d , the short-term prediction module processes the short-term modal sequence to obtain the short-term modal sequence prediction result Y s ∈R O×d , the long-term modal sequence prediction results and the short-term modal sequence prediction results are integrated through the fully connected layer to obtain the final fuel cell degradation data prediction result Y∈R O×d , where O represents the time length of the prediction result.

[0080] It should be noted that, in this embodiment, the size of O is set to 1, representing a single-step prediction, and the fuel cell degradation data prediction result is a vector of 1×7. In addition, since the entire input data set is divided into several batches, the several inputs in each batch are decomposed by the exponential decomposition module, and the performance degradation prediction model completes the prediction, and this process is repeated continuously to obtain the fuel cell degradation data prediction results of all batches of the input data set.

[0081] It should be noted that, in step S5 of the present invention, for the long-term modal sequence X t , we can use the gated recurrent unit to get the corresponding prediction result Y t =GRU(X t ). In this embodiment, the structural diagram of the gated recurrent unit is as follows: Figure 3 As shown, the implementation method thereof belongs to the prior art and will not be described in detail.

[0082] It should be noted that in step S5 of the present invention, for the short-term modal sequence X s , using the short-term prediction module consisting of four parts: encoding, internal decomposition, equidistant convolution, and projection truncation to obtain the short-term modal sequence prediction result Y s ∈R O×d .

[0083] Further, in the present invention, if Figure 4 As shown in Figure 2, the specific processing process of the short-term prediction module is as follows:

[0084] 1) Encoding process

[0085] S51. First, concatenate the input short-term modal sequence with the 0-filled sequence with the same feature dimension to obtain a concatenated short-term modal sequence, and use a weight encoder and a position encoder to encode the concatenated short-term modal sequence to obtain an encoded sequence Where D represents the high-dimensional features of the encoded sequence and can be any integer.

[0086] The encoding process is implemented as follows:

[0087]

[0088] Among them, VE represents the weight encoder, PE represents the position encoder, and X zero It represents a 0-filled sequence of length O, and Concat() represents a sequence concatenation operation.

[0089] In this embodiment, the size of D is 64, so the dimension of the encoded sequence is

[0090] 2) Internal decomposition process

[0091] S52. Use the multi-scale decomposition function M() to implement the encoded sequence Perform internal decomposition and obtain two internal decomposition subsequences.

[0092] The specific process of internal decomposition is as follows:

[0093] Perform 0 padding operation on the encoded sequence to obtain second padding data, use the third convolution kernel to perform average pooling on the second padding data to obtain third average pooling data, use the fourth convolution kernel to perform average pooling on the second padding data to obtain fourth average pooling data, average the third average pooling data and the fourth average pooling data to obtain second mean data, and use the difference between the encoded sequence and the second mean data as the second internal decomposition subsequence The encoded sequence As the first inner decomposition subsequence

[0094]

[0095]

[0096] in, Indicates that the third convolution kernel performs average pooling. Indicates that average pooling is performed by the fourth convolution kernel.

[0097] In this embodiment, the first internal decomposition subsequence and the second inner decomposition subsequence The R (11 +1)×64 Vector.

[0098] 3) Equidistant convolution process

[0099] S53. In the equidistant convolution block, the specific processing process is as follows:

[0100] a) First, each internal decomposition subsequence is downsampled, and each internal decomposition subsequence corresponds to a sampled local sequence.

[0101] In step S53-a), the specific processing process of downsampling each internal decomposition subsequence is as follows: performing a 0-filling operation on the first internal decomposition subsequence to obtain third filling data, performing a 0-filling operation on the second internal decomposition subsequence to obtain fourth filling data, using a convolution kernel with a length of i to perform average pooling on the third filling data to obtain fifth average pooling data, using a convolution kernel with a length of i to perform average pooling on the fourth filling data to obtain sixth average pooling data, and then using a convolution kernel with a length of i to perform one-dimensional convolution on the fifth average pooling data to obtain a sampled first local sequence corresponding to the first internal decomposition subsequence. Then use the convolution kernel with a length of i to perform a one-dimensional convolution on the sixth average pooling data to obtain the second local sequence after sampling corresponding to the second internal decomposition subsequence The specific implementation is as follows:

[0102]

[0103]

[0104] Among them, Conv1d() kernel＝i Represents a one-dimensional convolution operation using a convolution kernel of length i.

[0105] In this embodiment, the length of the convolution kernel is i=3, and the first local sequence dimension is The second local sequence dimension is

[0106] b) An equidistant convolution layer is used to extract the deep features of each sampled local sequence, and each sampled local sequence corresponds to a deep feature extracted by an equidistant convolution.

[0107] In S53-b) of the present invention, the specific processing process of extracting deep features by the equidistant convolution layer is as follows:

[0108] After performing equidistant convolution, Tanh activation function and Dropout operations on the first local sequence in turn, the first initial feature is obtained. The first initial feature and the first local sequence are added and layer regularization is performed to obtain the first deep feature corresponding to the first local sequence. After performing equidistant convolution, Tanh activation function and Dropout operations on the second local sequence in turn, the second initial feature is obtained. The second initial feature and the second local sequence are added and layer regularization is performed to obtain the second deep feature corresponding to the second local sequence. The specific implementation is as follows:

[0109]

[0110]

[0111] Among them, IC1d() represents equidistant convolution; Tanh() represents activation function; Dropout() represents random inactivation of the layer; Norm() represents regularization of the layer.

[0112] In this embodiment, the first deep feature and the second deep feature Both Vector. and The superscript ' represents and is a local variable in the equidistant convolution block.

[0113] c) Each deep feature is processed by an upsampling convolution layer to restore the standard length of the target sequence, and each deep feature corresponds to an output feature of the upsampling convolution layer.

[0114] In S53-c) of the present invention, the process of processing the deep features by the upsampling convolution layer is as follows: after performing upsampling convolution, Tanh activation function and Dropout operations on the first deep feature in turn, a third initial feature is obtained, and the third initial feature and the first deep feature are added and layer regularization is performed to obtain an output feature corresponding to the first deep feature After upsampling convolution, Tanh activation function and Dropout operations are performed on the second deep feature in turn, the fourth initial feature is obtained. The fourth initial feature and the second deep feature are added and layer regularization is performed to obtain the output feature corresponding to the second deep feature. The specific implementation is as follows:

[0115]

[0116]

[0117] Among them, Conv1dTrans() represents upsampling convolution.

[0118] In this embodiment, and The superscript U represents that this vector is an upsampled output vector.

[0119] d) The output features of the upsampling convolution layer are fused by the two-dimensional convolution layer to obtain the complete output features of the equidistant convolution layer. This is accomplished using the following formula:

[0120]

[0121] Among them, Conv2d represents a two-dimensional convolutional layer.

[0122] In this embodiment, The superscript m represents that this vector is the output variable after fusion.

[0123] e) Finally, a single-layer FNN network is used to project the complete output features of the equidistant convolution layer to obtain the output results of the equidistant convolution block.

[0124] In step S53-e), the specific process of using a single-layer FNN network to perform projection transformation is as follows: the complete output features of the equidistant convolutional layer are input into the single-layer FNN network to obtain the projection output features The output features of the projection are added to the complete output features of the equidistant convolution layer and then layer regularization is performed to obtain the output result Y of the equidistant convolution block. p ∈R (I+O)×D , implemented as follows:

[0125]

[0126] Among them, FNN() represents a single-layer FNN network.

[0127] 4) Projection truncation process

[0128] Step S6: The output result Y of the equidistant convolution block p Project and truncate to obtain the short-term modal sequence prediction result Y s ∈R O×d , implemented as follows:

[0129] Y s =Truncate(Project(Y p ))

[0130] Among them, Project() represents the projection operation, and Truncate() represents the truncation operation.

[0131] It should be noted that in step S5 of the present invention, after receiving the long-term modal sequence prediction results and the short-term modal sequence prediction results, the long-term and short-term modal sequences are added in the fully connected layer to output the final fuel cell degradation data prediction results.

[0132] In order to better demonstrate the specific implementation and technical effects of the present invention, the deep learning-based proton exchange membrane fuel cell performance degradation prediction method shown in steps S1 to S5 in the above preferred implementation method is applied to a specific example.

[0133] Example

[0134] The specific implementation process of the deep learning-based proton exchange membrane fuel cell performance degradation prediction method used in this embodiment is as described above and will not be repeated here.

[0135] like Figure 5 The figure shows the comparison between the real data curve and the predicted data curve, where Real Data represents the real data curve and Predictions represents the predicted data curve. Figure 5 The horizontal axis represents time, and the vertical axis represents the corresponding values ​​of the real data and the predicted data. It can be seen that the predicted data curve fits the real data curve well. Figure 6 , which is the distribution of the prediction error of the performance degradation model proposed by the method of the present invention, wherein the black line represents the fitting value and the red dot represents the observed value. Figure 6 The horizontal axis represents the real data, and the vertical axis represents the predicted value. It can be seen that the performance degradation model has a good prediction effect and can effectively explain the performance degradation of proton exchange membrane fuel cells.

[0136] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A proton exchange membrane fuel cell performance degradation prediction method based on deep learning, characterized in that: The following steps are involved: S1. Taking the dynamic operating degradation data of the fuel cell system acquired over a continuous period of time as the original input signal; S2. Using the Pearson correlation coefficient method to screen the characteristic variables that are highly correlated with the output voltage variable in the original input signal and delete the redundant variables, the preprocessed data is obtained; S3. Smoothing the output voltage variable in the preprocessed data using wavelet transform to obtain smoothed data and generate an input data set; S4. After batch processing and dividing the input data set, a single batch input is obtained, and the single batch input is decomposed using the trained multi-scale decomposition module to obtain a long-term modal sequence and a short-term modal sequence; S5. The long-term modal sequence and short-term modal sequence of each batch are input into the trained performance degradation prediction model. The long-term modal sequence is processed by the gated recurrent unit to obtain the long-term modal sequence prediction result. The short-term modal sequence is processed by the short-term prediction module to obtain the short-term modal sequence prediction result. The long-term modal sequence prediction result and the short-term modal sequence prediction result are integrated through the fully connected layer to obtain the final fuel cell degradation data prediction result.

2. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 1, characterized in that: The specific implementation method of step S2 is as follows: S21. Taking the output voltage variable as the target variable, the Pearson correlation coefficient between each characteristic variable in the original input signal and the target variable is calculated; S22. Sort all feature variables in the original input signal by the absolute value of the Pearson correlation coefficient, and only retain feature variables whose absolute value of the Pearson correlation coefficient is higher than a preset screening threshold; S23. According to the actual mechanism of the fuel cell operating condition, if the characteristic variables in the original input signal contain redundant variables, the redundant variables are directly deleted, otherwise no processing is performed.

3. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 1, characterized in that: The specific implementation method of step S3 is as follows: S31. Selecting a wavelet basis function according to the signal characteristics of the output voltage variable to perform multi-layer decomposition on the preprocessed data to obtain a wavelet coefficient sequence; S32. Performing soft threshold processing on the wavelet coefficient sequence to obtain a processed wavelet coefficient sequence; S33. Reconstruct the output voltage signal using the processed wavelet coefficient sequence to obtain smoothed data and construct an input data set from the smoothed data.

4. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 1, characterized in that: The specific implementation method of step S4 is as follows: S41. Treat a single batch input as a long-term modal sequence; S42. Perform zero-padding operation on the single batch input to obtain first padding data, use the first convolution kernel to perform average pooling on the first padding data to obtain first average pooling data, use the second convolution kernel to perform average pooling on the first padding data to obtain second average pooling data, average the first average pooling data and the second average pooling data to obtain first mean data, and use the difference between the single batch input and the first mean data as the short-term modal sequence.

5. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 1, characterized in that: In step S5, the specific processing process in the short-term prediction module is as follows: S51. First, concatenate the input short-term modal sequence with the 0-filled sequence with the same feature dimension to obtain a concatenated short-term modal sequence, and encode the concatenated short-term modal sequence using a weight encoder and a position encoder to obtain an encoded sequence; S52. Use a multi-scale decomposition function to internally decompose the encoded sequence to obtain two internal decomposition subsequences; S53. In the equidistant convolution block, firstly, each of the internal decomposition subsequences is downsampled, and each internal decomposition subsequence corresponds to a sampled local sequence. The deep features of each sampled local sequence are extracted by an equidistant convolution layer, and each sampled local sequence corresponds to a deep feature extracted by an equidistant convolution. Each deep feature is processed by an upsampling convolution layer, and each deep feature corresponds to an output feature of the upsampling convolution layer. The output features of the upsampling convolution layer are fused by a two-dimensional convolution layer to obtain a complete output feature of the equidistant convolution layer. Finally, the complete output feature of the equidistant convolution layer is projected by a single-layer FNN network to obtain the output result of the equidistant convolution block. S64. Project and truncate the output results of the equidistant convolution block to obtain the short-term modal sequence prediction results.

6. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 5, characterized in that: In step S52, the specific processing process of internal decomposition by the multi-scale decomposition function is as follows: perform a 0-filling operation on the encoded sequence to obtain second filling data, use a third convolution kernel to average pool the second filling data to obtain third average pooling data, use a fourth convolution kernel to average pool the second filling data to obtain fourth average pooling data, average the third average pooling data and the fourth average pooling data to obtain second mean data, use the difference between the encoded sequence and the second mean data as the second internal decomposition subsequence, and use the encoded sequence as the first internal decomposition subsequence.

7. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 6, characterized in that: In step S53, the specific processing process of downsampling each internal decomposition subsequence is as follows: perform a 0-filling operation on the first internal decomposition subsequence to obtain third filling data, perform a 0-filling operation on the second internal decomposition subsequence to obtain fourth filling data, use a convolution kernel with a length of i to perform average pooling on the third filling data to obtain fifth average pooling data, use a convolution kernel with a length of i to perform average pooling on the fourth filling data to obtain sixth average pooling data, and then use a convolution kernel with a length of i to perform one-dimensional convolution on the fifth average pooling data to obtain a first local sequence after sampling corresponding to the first internal decomposition subsequence, and then use a convolution kernel with a length of i to perform one-dimensional convolution on the sixth average pooling data to obtain a second local sequence after sampling corresponding to the second internal decomposition subsequence.

8. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 7, characterized in that: In step S53, the specific processing process of extracting deep features by the equidistant convolution layer is as follows: after performing equidistant convolution, Tanh activation function and Dropout operations on the first local sequence in turn, a first initial feature is obtained, the first initial feature and the first local sequence are added and layer regularization is performed to obtain a first deep feature corresponding to the first local sequence, and the second local sequence is successively subjected to equidistant convolution, Tanh activation function and Dropout operations to obtain a second initial feature, and the second initial feature and the second local sequence are added and layer regularization is performed to obtain a second deep feature corresponding to the second local sequence.

9. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 8, characterized in that: In step S53, the process of processing the deep features by the upsampling convolution layer is as follows: after performing upsampling convolution, Tanh activation function and Dropout operations on the first deep feature in turn, a third initial feature is obtained, the third initial feature and the first deep feature are added and layer regularization is performed to obtain the output feature corresponding to the first deep feature, and the second deep feature is successively performed upsampling convolution, Tanh activation function and Dropout operations to obtain the fourth initial feature, and the fourth initial feature and the second deep feature are added and layer regularization is performed to obtain the output feature corresponding to the second deep feature.

10. A method for predicting performance degradation of a proton exchange membrane fuel cell based on deep learning as claimed in claim 9, characterized in that: In step S53, the specific process of using a single-layer FNN network for projection transformation is as follows: the complete output features of the equidistant convolution layer are input into the single-layer FNN network to obtain the projection output features, the projection output features are added to the complete output features of the equidistant convolution layer, and layer regularization is performed to obtain the output result of the equidistant convolution block.

Images