PEMFC degradation prediction method and system based on spatial dynamic non-stationary reconstruction attention

Abstract

The invention belongs to the technical field of fuel cell equipment degradation prediction, and provides a PEMFC degradation prediction method and system based on spatial dynamic non-stationary reconstruction attention, and the technical scheme is as follows: obtaining multi-sensing historical operation data of a fuel cell PEMFC; training the constructed PEMFC degradation prediction model based on the multi-sensing historical operation data of the fuel cell PEMFC to obtain a trained PEMFC degradation prediction model; the construction process of the PEMFC degradation prediction model comprises the steps that statistical information of multi-sensor time sequence data in the spatial dimension is acquired, and the statistical information in the spatial dimension is combined and mapped to a non-stationary factor; obtaining battery degradation characteristics in a time dimension; based on a non-stationary factor, dynamically correcting an attention weight in a self-attention mechanism during spatial feature extraction to obtain a corrected battery degradation spatial feature; the method can better cope with the non-stationarity of multiple sensor signals at different moments and under different working conditions, thereby capturing more critical and richer degradation feature expressions.

Description

Technical Field

[0001] The present invention belongs to the technical field of fuel cell device degradation prediction, and particularly relates to a PEMFC degradation prediction method and system based on spatial dynamic non-stationary reconstruction attention. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] As a clean energy source that directly converts chemical energy into electrical energy, fuel cells often use fuels such as hydrogen and methanol. In particular, proton exchange membrane fuel cells (PEMFCs) have received extensive attention due to their advantages such as compact structure and fast startup speed. However, the current fuel cells still have limitations in terms of lifespan and are difficult to be commercialized on a large scale, which hinders the popularization of related technologies.

[0004] To extend the service life of fuel cells and reduce maintenance costs, existing technologies have proposed solutions from multiple aspects such as material improvement, stack design to predictive health management (PHM). Among them, PHM emphasizes using data-driven or a combination of models and data for fault prediction and lifespan management, which can provide early warnings before faults occur and guide optimal maintenance strategies. Existing fuel cell degradation prediction methods mainly include model-driven methods, hybrid methods, and data-driven methods. Model-driven methods have certain advantages in theoretical interpretability, but have high requirements for prior knowledge and modeling experience; hybrid methods are also restricted by the disadvantages of the model part; data-driven methods rely on machine learning (especially deep learning) to automatically extract non-linear features by analyzing and mining the collected monitoring data. In recent years, methods such as recurrent neural networks (RNN) and their variants (LSTM, GRU) and convolutional neural networks (CNN) in deep learning have made much progress. However, when dealing with complex fuel cell degradation processes, long-time series dependencies, and multi-variable couplings, the inventors found that the above technical solutions have the following problems:

[0005] (1) Insufficient modeling of time and space coupling: Since the degradation process of fuel cells is closely coupled with time series dynamics and various sensor information, existing methods mainly focus on sequence modeling in the time dimension and lack consideration of the "space" correlation between different sensors or different variables, often unable to accurately capture complex coupling characteristics, resulting in limited degradation prediction accuracy in dynamic environments.

[0006] (2) Lack of effective processing for non - stationary features: Fuel cells exhibit significant non - stationary features under different operating conditions and time periods. Traditional methods usually adopt stationary processing to simplify the training process, but this will ignore the dynamic change rules in the real environment, reduce the ability to depict complex degradation patterns, and it is difficult to maintain a high level of prediction accuracy and robustness in practice.

[0007] (3) Insufficient analysis of the contribution of key sensors: There are obvious differences in the influence of different sensors on the degradation process of fuel cells. If the specific contributions of key sensors to the degradation results can be identified, it will greatly promote the optimization of equipment maintenance and monitoring strategies. Most existing methods can only give the final degradation prediction value, lacking an interpretable analysis method for the contribution degree of different sensors or feature dimensions, and it is difficult to form targeted maintenance suggestions.

[0008] (4) There are gradient and parallelization bottlenecks in long - sequence dependencies: Traditional recurrent neural networks often suffer from the problems of vanishing gradients or exploding gradients when processing long time series, and convolutional networks are also limited by the receptive field size and difficult to capture ultra - long - time dependencies; for the data - processing requirements of non - stationary long sequences, existing models either require a large amount of computing resources or have problems such as information forgetting and low training efficiency, and they cannot well balance prediction accuracy and computational cost. Summary of the Invention

[0009] To solve at least one of the technical problems existing in the above - mentioned background technology, the present invention provides a PEMFC degradation prediction method and system based on spatial - dynamic non - stationary reconstruction attention, which can better cope with the non - stationarity of multi - sensor signals at different times and different operating conditions, so as to capture more critical and rich degradation feature expressions; at the same time, the parallel time - dimension decomposition and space - dimension modeling can effectively prevent interference between different feature information, taking into account both global correlation and local fluctuation features, thus significantly improving the prediction accuracy of the complex degradation process of fuel cells.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] The first aspect of the present invention provides a PEMFC degradation prediction method based on spatial - dynamic non - stationary reconstruction attention, including the following steps:

[0012] Obtain the multi - sensor historical operation data of the proton exchange membrane fuel cell (PEMFC);

[0013] Based on the multi - sensor historical operation data of the proton exchange membrane fuel cell (PEMFC), train the constructed PEMFC degradation prediction model to obtain the trained PEMFC degradation prediction model; wherein, the construction process of the PEMFC degradation prediction model includes:

[0014] Obtain the statistical information of multi-sensor time-series data in the spatial dimension, combine the multi-sensor time-series data and its statistical information in the spatial dimension, and map the current time-series distribution information to non-stationary factors;

[0015] Obtain the battery degradation characteristics of multi-sensor time-series data in the time dimension;

[0016] Based on the non-stationary factors, dynamically correct the attention weights in the self-attention mechanism during spatial feature extraction to obtain the corrected battery degradation spatial features;

[0017] Fuse the battery degradation characteristics in the time dimension and the corrected battery degradation spatial features to obtain degradation prediction features;

[0018] Based on the trained PEMFC degradation prediction model, predict the operation data of the fuel cell PEMFC under actual working conditions to obtain the degradation prediction results.

[0019] Furthermore, use the multi-sensor historical operation data of the fuel cell PEMFC as training data, and use the corresponding degradation index or remaining service life as labels to form a training set.

[0020] Furthermore, the statistical information of multi-sensor time-series data in the spatial dimension includes the mean vector and standard deviation vector in the spatial dimension:

[0021]

[0022] Among them, respectively represent the mean vector and standard deviation vector in the spatial dimension, represents the row vector obtained after transposing the i-th sensor observation value in the input X, and N is the sensor dimension.

[0023] Furthermore, combine the multi-sensor time-series data and its statistical information in the spatial dimension, and map the current time-series distribution information to non-stationary factors. The calculation formula is:

[0024]

[0025] Among them, α is the scaling factor, β is the bias factor, is the transpose of the multi-sensor time-series sequence, and are the mean vector and standard deviation vector in the spatial dimension, is a learnable weight matrix, and the output dimension c is determined by the parameters α and β, and d model is the input spatial dimension. When the input statistical parameter s is , the value of c is set to 1. At this time, α represents a positive scalar; and when the input statistical parameter s is When c is set to N, β represents a bias vector containing N elements at this time.

[0026] Furthermore, battery degradation features in the time dimension of multi-sensor time series data are obtained, including:

[0027] Average pooling is used to extract fluctuating trend features and seasonal features;

[0028] The fluctuating trend features and seasonal features are input into a linear layer to obtain battery degradation features in the time dimension.

[0029] Furthermore, based on non-stationary factors, the attention weights in the self-attention mechanism during spatial feature extraction are dynamically corrected. When obtaining the corrected battery degradation spatial features, it includes adopting a multi-head strategy in the spatial self-attention mechanism, splitting Q, K, and V into multiple sub-spaces for parallel computing and reconstructing attention, concatenating and outputting the output features of the multi-head attention mechanism, and using a feed-forward neural network for propagation based on the output features of the multi-head attention mechanism to obtain the corrected battery degradation spatial features.

[0030] The second aspect of the present invention provides a PEMFC degradation prediction system based on spatially dynamic non-stationary reconstructed attention, including:

[0031] A data acquisition module, which is used to acquire multi-sensor historical operation data of the fuel cell PEMFC;

[0032] A model training module, which is used to train the constructed PEMFC degradation prediction model based on the multi-sensor historical operation data of the fuel cell PEMFC to obtain a trained PEMFC degradation prediction model; among them, the construction process of the PEMFC degradation prediction model includes:

[0033] Obtain the statistical information of multi-sensor time series data in the spatial dimension, combine the multi-sensor time series data and its statistical information in the spatial dimension, and map the current time series distribution information to non-stationary factors;

[0034] Obtain battery degradation features in the time dimension of multi-sensor time series data;

[0035] Based on non-stationary factors, dynamically correct the attention weights in the self-attention mechanism during spatial feature extraction to obtain corrected battery degradation spatial features;

[0036] Fuse the battery degradation features in the time dimension and the corrected battery degradation spatial features to obtain degradation prediction features;

[0037] A degradation prediction module, which is used to predict the operation data of the fuel cell PEMFC under actual working conditions based on the trained PEMFC degradation prediction model to obtain a degradation prediction result.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] In the present invention, the parallel spatial self-attention network and trend-seasonal decomposition linear focus on degradation information in different dimensions respectively. In the spatial self-attention network, an adaptive non-stationary factor is introduced to dynamically reconstruct the attention, which can better cope with the non-stationarity of multi-sensor signals at different times and under different working conditions, so as to capture more critical and rich degradation feature expressions. At the same time, the parallel time dimension decomposition and spatial dimension modeling can effectively prevent interference between different feature information, taking into account both global correlation and local fluctuation characteristics, thus significantly improving the prediction accuracy of the complex degradation process of fuel cells. The method of the present invention has good application prospects in fuel cell health management and remaining useful life prediction.

[0040] Advantages of additional aspects of the present invention will be partly given in the following description, partly will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The specification drawings forming a part of the present invention are used to provide further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0042] Figure 1 is the overall flowchart of fuel cell degradation prediction provided by an embodiment of the present invention;

[0043] Figure 2 is the network structure diagram of the fuel cell degradation model provided by an embodiment of the present invention;

[0044] Figure 3 is the structure diagram of the adaptive non-stationary reconstruction attention mechanism provided by an embodiment of the present invention;

[0045] Figure 4 is the comparison result with four representative methods under the FC1 dataset provided by an embodiment of the present invention; among them, (a) is the result comparison of 1-step prediction, (b) is the error comparison of 1-step prediction, (c) is the result comparison of 3-step prediction, (d) is the error comparison of 3-step prediction, (e) is the result comparison of 5-step prediction, and (f) is the error comparison of 5-step prediction;

[0046] Figure 5 is the comparison result with four representative methods under the FC2 dataset provided by an embodiment of the present invention; among them, (a) is the result comparison of 1-step prediction, (b) is the error comparison of 1-step prediction, (c) is the result comparison of 3-step prediction, (d) is the error comparison of 3-step prediction, (e) is the result comparison of 5-step prediction, and (f) is the error comparison of 5-step prediction;

[0047] Figure 6 It is the comparison result of the performance of different models with different training lengths under the FC2 dataset provided by the embodiments of the present invention; among them, (a) is the result comparison when the training, validation, and test set splitting ratios are 3:1:6 for 1-step prediction, (b) is the error comparison when the training, validation, and test set splitting ratios are 3:1:6 for 1-step prediction, (c) is the result comparison when the training, validation, and test set splitting ratios are 5:1:4 for 1-step prediction, (d) is the error comparison when the training, validation, and test set splitting ratios are 5:1:4 for 1-step prediction, (e) is the result comparison when the training, validation, and test set splitting ratios are 7:1:2 for 1-step prediction, and (f) is the error comparison when the training, validation, and test set splitting ratios are 7:1:2 for 1-step prediction. Detailed implementation manners

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0050] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0051] As the operation time extends, the output performance of the fuel cell gradually decays, and the deterioration process usually shows a voltage drop, a decrease in output power, etc. To address this challenge, many studies evaluate the life of fuel cells or predict the deterioration trend based on time series data. However, these methods usually only focus on the deterioration process in the time dimension and ignore the significant spatial correlation that often exists between the measurements of multiple sensors in the fuel cell system. Without fully modeling the coupling relationship between these sensors, the understanding of the overall deterioration mechanism is inevitably one-sided.

[0052] Based on the above considerations, the present invention first introduces a spatial Transformer to capture the correlation between multiple sensors. Different from traditional Transformers, only the encoder part is retained, and the decoder structure is removed to reduce the model complexity. When processing temporal dimension information, the present invention performs trend-season decomposition on the observed data and directly uses linear mapping in the model to capture local features, thereby avoiding the loss of details that may occur when only using spatial information. However, the voltage and operating conditions of fuel cells under actual conditions often fluctuate violently and asynchronously, which makes it difficult for traditional attention mechanisms to output distinguishable and interpretable attention maps because they lose information after normalization. To solve the above problems, the present invention proposes an Adaptive Non-stationary Reconstruct Attention (ANRA), which incorporates non-stationary features into the attention reconstruction process to obtain a more interpretable coupling relationship between sensors.

[0053] Embodiment 1

[0054] As Figure 1 shown, this embodiment provides a PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention, including the following steps:

[0055] Step 1: Obtain the multi-sensor historical operation data of the fuel cell PEMFC as training data, and use the degradation index or remaining useful life (RUL) corresponding to the detection data as labels to form a training set;

[0056] Obtain the multi-sensor operation data collected by the fuel cell system under real-time working conditions as a test set;

[0057] Step 2: Train the constructed PEMFC degradation prediction model based on the training set to obtain a trained PEMFC degradation prediction model;

[0058] In this embodiment, modeling is carried out by integrating the temporal and spatial aspects and training is performed using the training set.

[0059] Figure 2 is a schematic structural diagram of the Spatial Adaptive Non-stationary Reconstruct Attention Transformer model (SANformer) proposed by the present invention. As Figure 2 shown, the construction process of the degradation prediction model of spatial adaptive non-stationary reconstruction attention is mainly divided into the following key steps:

[0060] Non-stationary feature extraction: First, the original data is input into the Adaptive Non-stationary Reconstruct (ANR) layer to extract high-dimensional features from the statistical information. This step preserves the key fluctuation details and enhances the predictability of the fuel cell degradation process.

[0061] Spatio-temporal feature analysis: Temporal and spatial features are extracted from the PEMFC data collected by multiple sensors.

[0062] On the one hand, the original information X is transposed, encoded, and then input into the spatial Transformer with only an encoder to obtain the spatial correlation between different sensors, and the attention map is reconstructed by using the non-stationary information extracted by ANR;

[0063] On the other hand, the same original data X undergoes seasonal-trend decomposition, and the separated seasonal components are then modeled in a linear layer to capture the time dependence;

[0064] Degradation prediction output: Finally, the output layer integrates the fused temporal and spatial representations, and then generates the predicted degradation voltage through a multi-layer perceptron (MLP); this comprehensive method captures the spatial coupling relationship between sensor measurements and retains the highly fluctuating temporal patterns, thus achieving a more accurate and interpretable prediction of the fuel cell degradation trend.

[0065] Step 3: Obtain the multi-sensor operation data collected by the fuel cell system under real-time working conditions as the test set, use the test set to predict the performance of the model, and predict the operation data of the fuel cell PEMFC to be detected based on the PEMFC degradation prediction model after passing the test to obtain the degradation prediction result.

[0066] As Figure 2 and Figure 3 shown, in order to more clearly express the technical solution of the present invention, the following details the processing process:

[0067] Assume that the observed multi-sensor time series data is:

[0068]

[0069] where T represents the time series length and N represents the number of sensors.

[0070] After preprocessing X through filtering, normalization, etc., it is transposed to:

[0071]

[0072] It is mapped to the high-dimensional space representation E through a linear mapping s , and the dimension can be set to d model :

[0073] E s = X T ·W E ,

[0074] Among them, is a learnable parameter, and E s will be used as the original input of the attention mechanism.

[0075] The self-attention mechanism can capture the global dependencies between the features at all positions in the input sequence, and since it does not depend on the sequential processing of the sequence, it supports efficient parallel computing.

[0076] In the traditional self-attention mechanism, first calculate:

[0077] Q = E s ·W Q , K = E s ·W K , V = E s ·W V ,

[0078] Then obtain the attention weight AW through the scaled dot product form:

[0079]

[0080] However, directly inputting continuously changing fuel cell data may weaken the predictability of the model. Commonly used normalization and denormalization are used to alleviate this problem, but this operation occurs in the time dimension, and the spatial non-stationarity of the original sequence cannot be fully restored through denormalization, which makes it difficult for the model to generate a focused attention map.

[0081] To address the non-stationary distribution of different sensor data of fuel cells under dynamic working conditions, the present invention uses the ANR module in spatial attention calculation, and its improvement to the attention mechanism is as Figure 3 shown, and its calculation details are as follows:

[0082] (1) Statistical feature extraction: Calculate the mean and standard deviation of the original input on the sensor dimension N, which can be defined as:

[0083]

[0084] Among them, respectively represent the mean vector and standard deviation vector of the spatial dimension. Since may be distributed unstably with time and working conditions, it is necessary to capture this dynamic change. It represents the row vector obtained after transposing the i-th sensor observation value in the input X.

[0085] (2) The ANR layer learns the non-stationary factors:

[0086] To enable α and β to be dynamically adjusted over time, the present invention designs two parallel bidirectional gated recurrent network (Bi-GRU) sub-modules in ANR, and inputs and respectively to learn to map the current temporal distribution information to the scaling factor α and the bias factor β. It can be denoted as:

[0087]

[0088] where, is a learnable weight matrix, and the output dimension c is determined by the parameters α and β. Specifically, when the input statistical parameter s is , the value of c is set to 1, and at this time α represents a positive scalar; while when the input parameter s is , the value of c is set to N, and at this time β represents a bias vector containing N elements.

[0089] After obtaining the two non-stationary factors α and β, the present invention reconstructs the original scaled dot-product self-attention to adapt to the dynamic distribution characteristics. The attention output obtained from the correction formula of the attention weight is as follows:

[0090]

[0091] The above modifications can enable the model to adaptively strengthen or suppress the influence of certain sensors on the degradation process in the attention calculation, so that in non-stationary and multi-condition data, it can still output a more focused and interpretable attention map, intuitively giving the attention weights of different features to the degradation.

[0092] In addition, to improve the model's ability to capture multi-dimensional degradation features, the present invention adopts a multi-head strategy (Multi-Head Attention) in the spatial self-attention mechanism, that is, splitting Q, K, and V into multiple sub-spaces for parallel calculation and reconstruction of attention, and then splicing them and passing them through a learnable mapping W O to restore them to the original dimension. This process can be expressed as:

[0093] MultiHead(Q ′ , K ′ , V ′ , α, β) = Concat(head 1 , head 2 , …, head h )W O ,

[0094] head i = Attention(QW i Q , KW i K , VW i V , α, β),

[0095] where head i is each attention head, In the embodiment, h is used to represent the number of attention heads, representing the output mapping of the attention mechanism.

[0096] After the multi-head attention mechanism outputs high-dimensional features, in order to enhance the non-linear expression ability of the model, a feed-forward neural network is used for further propagation. The FFN is designed as a two-layer feed-forward neural network, and the formula is as follows:

[0097] FFN(x) = ReLU(xW 1 + b 1 )W 2 + b 2 ,

[0098] where x represents the output of the above spatial multi-head attention mechanism. The output structure of FFN is used as the output of this layer of encoder, and the information extracted by repeating through N layers of encoders is used as the output of spatial features

[0099] To avoid ignoring the information in the time dimension in the modeling of spatial self-attention, the present invention obtains the periodic or irregular fluctuation characteristics of fuel cell degradation from the time dimension through seasonal-trend decomposition and combined with linear mapping in the parallel branch. The specific operations are as follows:

[0100] First, average pooling is used to extract the seasonal information of the fluctuation, and the process is defined as follows:

[0101] Trend = AvgPool(Padding(X)), S = X - Trend,

[0102] where X represents the input original sequence, Trend, and respectively represent the trend part and the seasonal part. In this embodiment, Padding is used to ensure that the sequence length remains unchanged, and the moving average kernel size of the average pooling is the input time window length T.

[0103] The decomposed seasonal feature S is input into the linear layer for the seasonal feature of the fluctuation, and the formula is as follows:

[0104]

[0105] Among them, It should be noted that the input S here T does not model any spatial relationships. The output O of the time modeling is fused with the information concerned by the spatial features, and finally the degradation prediction is output through a multi-layer perceptron: T

[0106]

[0107] Among them, represents the predicted output of the model.

[0108] To verify the effectiveness of the present invention, the dataset of the IEEE 2014 PHM data challenge is used for training, validation and testing. This dataset consists of two parts, FC1 and FC2. The PEMFC used in the experiment is composed of 5 battery cells. FC1 and FC2 operate under steady-state and quasi-dynamic conditions with an additional frequency of 5 kHz and an amplitude of 7 A respectively to simulate aging under different conditions. The running time of FC1 is 1154 hours, and the running time of FC2 is 1020 hours. The sensors record the key operating parameters of the battery during operation, such as the voltage, current, temperature, gas flow rate and pressure of each stack, etc.

[0109] Since the data comes from different sensors, there are differences in scale, the sampling time is uneven, and there is a large amount of noise and spikes, which may bring incorrect input signals to mislead the model training. Therefore, the original data is averaged and sampled at 1-hour intervals, and the LOESS method is used for smoothing. The window settings of the filter are set to 20 and 10 respectively (a smaller window for the quasi-dynamic test to capture more details).

[0110] To more intuitively evaluate the degradation prediction performance of the SANRA-Transformer model, the present invention uses two datasets, FC1 and FC2, for testing and compares it with four representative methods, including CNN, LSTM, traditional Transformer, and non-stationary Transformer. First, 1, 3, and 5 time steps are selected respectively for experimental analysis of predictions with different time granularities. The experimental comparison results under the FC1 dataset are shown in Table 1 and Figure 4 as shown, Figure 4 Among them, (a) is the result comparison of 1-step prediction, (b) is the error comparison of 1-step prediction, (c) is the result comparison of 3-step prediction, (d) is the error comparison of 3-step prediction, (e) is the result comparison of 5-step prediction, and (f) is the error comparison of 5-step prediction;

[0111] The experimental comparison results under the FC2 dataset are shown in Table 2 and Figure 5 as shown,​Figure 5 Among them, (a) is the result comparison of 1-step prediction, (b) is the error comparison of 1-step prediction, (c) is the result comparison of 3-step prediction, (d) is the error comparison of 3-step prediction, (e) is the result comparison of 5-step prediction, and (f) is the error comparison of 5-step prediction;

[0112] Table 1 Experimental comparison results of various prediction methods under the FC1 dataset

[0113]

[0114] Table 2 Experimental comparison results of various prediction methods under the FC2 dataset

[0115]

[0116] From the experimental results, it can be observed that the proposed model shows a high degree of consistency with the actual voltage in degradation prediction and is significantly better than the results of LSTM, CNN, and those based on Transformer. It should be noted that the unmodified Transformer loses most of the fluctuation information due to over-smoothing, resulting in poor results. In addition, CNN still faces great challenges in long-time series dependencies, especially in the performance under dynamic models.

[0117] Subsequently, the analysis of the influence of different training lengths on the model performance was carried out for the 1-step prediction experiment of FC2, and the results were compared with the Bi-LSTM and Non-stationary Transformer models in the above experiments. The results are shown in Figure 6 as follows, Figure 6 Among them, (a) is the result comparison when the training, validation, and test set splitting ratio is 3:1:6 for 1-step prediction, (b) is the error comparison when the training, validation, and test set splitting ratio is 3:1:6 for 1-step prediction, (c) is the result comparison when the training, validation, and test set splitting ratio is 5:1:4 for 1-step prediction, (d) is the error comparison when the training, validation, and test set splitting ratio is 5:1:4 for 1-step prediction, (e) is the result comparison when the training, validation, and test set splitting ratio is 7:1:2 for 1-step prediction, and (f) is the error comparison when the training, validation, and test set splitting ratio is 7:1:2 for 1-step prediction. A shorter training length means that the model can enter the working state earlier. The experimental results show that the model of the present invention is not very sensitive to the training set length, while the accuracy of the LSTM-based method is limited to a certain extent by the amount of training information.

[0118] Finally, ablation experiments are used to evaluate the effectiveness of the two internal components in the module of the present invention, that is, the effectiveness of ANAR in reconstructing spatial attention and the effectiveness of the trend decomposition module in temporal modeling. For this purpose, the present invention designs three ablation experiments: separate spatial attention, spatially attention reconstructed by ANAR, and ANAR reconstructed spatial attention (SANRA Transformer) with trend decomposition added. Taking the single-step degradation prediction of FC1 as an example, the results of the ablation experiments are shown in Table 3.

[0119] Table 3 Comparison of experimental results of single-step degradation prediction of FC1

[0120]

[0121] Table 3 intuitively shows that the enhanced spatial attention focuses more on different information, and for different degradation trends, the contributions of different sensors to degradation can be clearly shown on the attention map. Taking the 5-step degradation prediction of FC1 as an example, the different weights of trend decomposition for temporal modeling are shown. The ablation experiments show that the prediction performance of SANRA Transformer is better than the original spatial attention, and removing any one module will significantly reduce the model performance, indicating that modeling degradation from either the temporal or spatial perspective is meaningful.

[0122] Embodiment 2

[0123] The present embodiment provides a PEMFC degradation prediction system based on spatially dynamic non-stationary reconstructed attention, including:

[0124] A data acquisition module for acquiring multi-sensor historical operation data of a fuel cell PEMFC;

[0125] A model training module for training the constructed PEMFC degradation prediction model based on the multi-sensor historical operation data of the fuel cell PEMFC to obtain a trained PEMFC degradation prediction model; wherein, the construction process of the PEMFC degradation prediction model includes:

[0126] Obtaining the statistical information of the multi-sensor time series data in the spatial dimension, combining the multi-sensor time series data and its statistical information in the spatial dimension, and mapping the current time series distribution information to non-stationary factors;

[0127] Obtaining the battery degradation characteristics of the multi-sensor time series data in the time dimension;

[0128] Based on the non-stationary factors, dynamically correcting the attention weights in the self-attention mechanism during spatial feature extraction to obtain corrected spatial battery degradation characteristics;

[0129] Fuse the battery degradation characteristics in the time dimension and the corrected battery degradation spatial characteristics to obtain degradation prediction characteristics;

[0130] A degradation prediction module, which is used to predict the operating data of the fuel cell PEMFC under actual working conditions based on the trained PEMFC degradation prediction model to obtain a degradation prediction result.

[0131] Embodiment III

[0132] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the above-mentioned PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention.

[0133] Embodiment IV

[0134] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the above-mentioned PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention.

[0135] Embodiment V

[0136] This embodiment provides a program product, which is a computer program product, including a computer program. The feature is that when the computer program is executed by a processor, it implements the steps in the above-mentioned PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention.

[0137] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention, characterized in that: The steps include: Obtain multi-sensor historical operating data of fuel cell PEMFC; The constructed PEMFC degradation prediction model is trained based on the multi-sensor historical operation data of the fuel cell PEMFC to obtain a trained PEMFC degradation prediction model; wherein the construction process of the PEMFC degradation prediction model includes: Obtain the statistical information of multi-sensor time series data in the spatial dimension, combine the multi-sensor time series data and its statistical information in the spatial dimension, and map the current time series distribution information to the non-stationary factor; Obtain battery degradation characteristics in the time dimension of multi-sensor time series data; Based on the non-stationary factor, the attention weight in the self-attention mechanism during spatial feature extraction is dynamically modified to obtain the modified battery degradation spatial features; The degradation prediction features are obtained by integrating the battery degradation features in the time dimension and the corrected battery degradation spatial features; Based on the trained PEMFC degradation prediction model, the fuel cell PEMFC operation data under actual working conditions is predicted to obtain the degradation prediction results.

2. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as claimed in claim 1 is characterized in that: The multi-sensor historical operation data of the fuel cell PEMFC is used as training data, and the degradation index or remaining service life corresponding to the training data is used as a label to form a training set.

3. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as claimed in claim 1 is characterized in that: The statistical information of multi-sensor time series data in the spatial dimension includes the mean vector and standard deviation vector of the spatial dimension: in, Represent the mean vector and standard deviation vector of the spatial dimension respectively, represents the row vector of the i-th sensor observation in the input X after transposition, where N is the sensor dimension.

4. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as claimed in claim 1 combines multi-sensor time series data and its statistical information in the spatial dimension, maps the current time series distribution information to the non-stationary factor, and the calculation formula is: ANR=Concat(BiGRU(X T ),BiGRU(s))W A , in, α is the scaling factor, β is the bias factor, and X T is the transpose of the multi-sensor timing sequence, and are the mean vector and standard deviation vector of the spatial dimension, is a learnable weight matrix, the output dimension c is determined by the parameters α and β, d model is the input space dimension, when the input statistical parameter s is When the value of c is set to 1, α represents a positive scalar; when the input statistical parameter s is When , the value of c is set to N, and β represents a bias vector containing N elements.

5. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as claimed in claim 1, obtaining battery degradation characteristics of multi-sensor time series data in the time dimension, including: Average pooling is used to extract the trend and seasonal characteristics of fluctuations; The trend characteristics and seasonal characteristics of fluctuations are input into the linear layer to obtain the battery degradation characteristics in the time dimension.

6. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in claim 1, based on the non-stationary factor, dynamically corrects the attention weight in the self-attention mechanism during spatial feature extraction to obtain the corrected battery degradation spatial features, including adopting a multi-head strategy in the spatial self-attention mechanism, dividing Q, K, and V into multiple subspaces to parallelly calculate the reconstruction attention, and outputting the output features of the multi-head attention mechanism after splicing. Based on the output features of the multi-head attention mechanism, a feedforward neural network is used for propagation to obtain the corrected battery degradation spatial features.

7. A PEMFC degradation prediction system based on spatial dynamic non-stationary reconstruction attention, characterized in that: include: A data acquisition module, which is used to obtain multi-sensor historical operation data of the fuel cell PEMFC; A model training module is used to train the constructed PEMFC degradation prediction model based on the multi-sensor historical operation data of the fuel cell PEMFC to obtain a trained PEMFC degradation prediction model; wherein the construction process of the PEMFC degradation prediction model includes: Obtain the statistical information of multi-sensor time series data in the spatial dimension, combine the multi-sensor time series data and its statistical information in the spatial dimension, and map the current time series distribution information to the non-stationary factor; Obtain battery degradation characteristics in the time dimension of multi-sensor time series data; Based on the non-stationary factor, the attention weight in the self-attention mechanism during spatial feature extraction is dynamically modified to obtain the modified battery degradation spatial features; The degradation prediction features are obtained by integrating the battery degradation features in the time dimension and the corrected battery degradation spatial features; The degradation prediction module is used to predict the fuel cell PEMFC operation data under actual working conditions based on the trained PEMFC degradation prediction model to obtain a degradation prediction result.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1 to 6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1 to 6 are implemented.

10. A program product, the program product being a computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps in the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1 to 6 are implemented.

Images