A PEMFC Degradation Prediction Method and System Based on Spatial Dynamic Non-stationary Reconstruction Attention

Abstract

This invention belongs to the field of fuel cell device degradation prediction technology, and provides a PEMFC degradation prediction method and system based on spatial dynamic non-stationary reconfiguration attention. The technical solution involves acquiring multi-sensor historical operating data of a fuel cell PEMFC; training a constructed PEMFC degradation prediction model based on the multi-sensor historical operating data to obtain a trained PEMFC degradation prediction model; wherein the construction process of the PEMFC degradation prediction model includes: acquiring statistical information of multi-sensor time-series data in the spatial dimension, and mapping the spatial statistical information to a non-stationary factor; acquiring battery degradation features in the time dimension; and dynamically correcting the attention weights in the self-attention mechanism during spatial feature extraction based on the non-stationary factor to obtain the corrected battery degradation spatial features; this method can better address the non-stationarity of multi-sensor signals under different times and operating conditions, thereby capturing more critical and richer degradation feature expressions.

Description

Technical Field

[0001] This invention belongs to the field of fuel cell equipment degradation prediction technology, and particularly relates to a PEMFC degradation prediction method and system based on spatial dynamic non-stationary reconfiguration attention. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Fuel cells, as a clean energy source that directly converts chemical energy into electrical energy, commonly use fuels such as hydrogen and methanol. In particular, proton exchange membrane fuel cells (PEMFCs) have attracted widespread attention due to their compact structure and fast start-up speed. However, current fuel cells still have limitations in terms of lifespan, hindering large-scale commercialization and impeding the promotion of related technologies.

[0004] To extend the lifespan of fuel cells and reduce maintenance costs, existing technologies have proposed solutions from multiple perspectives, including material improvement, stack design, and Predictive Health Management (PHM). PHM emphasizes using data-driven or combined model-and-data approaches for fault prediction and lifespan management, providing early warnings and guiding optimal maintenance strategies before failures occur. Existing fuel cell degradation prediction methods mainly include model-driven, hybrid, and data-driven methods. Model-driven methods have certain advantages in theoretical interpretability but require a high level of prior knowledge and modeling experience; hybrid methods are similarly limited by the shortcomings of the model component; data-driven methods leverage machine learning (especially deep learning) to automatically extract nonlinear features by analyzing and mining collected monitoring data. In recent years, recurrent neural networks (RNNs) and their variants (LSTM, GRU) and convolutional neural networks (CNNs) in deep learning have made significant progress. However, when dealing with complex fuel cell degradation processes, long-term dependencies, and multivariate coupling, the inventors have found that the above-mentioned technical solutions have the following problems:

[0005] (1) Insufficient modeling of temporal and spatial coupling: Since the degradation process of fuel cells is closely coupled with time series dynamics and information from multiple sensors, existing methods focus more on time-series modeling and do not adequately consider the spatial correlation between different sensors or variables. This often fails to accurately capture complex coupling characteristics, resulting in limited accuracy of degradation prediction in dynamic environments.

[0006] (2) Lack of effective handling of non-stationary characteristics: Fuel cells exhibit significant non-stationary characteristics under different operating conditions and time periods. Traditional methods usually use stationarization to simplify the training process, but this ignores the dynamic changes in the real environment and reduces the ability to characterize complex degradation patterns. In practice, it is difficult to maintain a high level of prediction accuracy and robustness.

[0007] (3) Insufficient analysis of the contribution of key sensors: Different sensors have significantly different impacts on the degradation process of fuel cells. Identifying the specific contribution of key sensors to the degradation results would greatly promote the optimization of equipment maintenance and monitoring strategies. Most existing methods can only provide the final degradation prediction value and lack interpretable analysis methods for the contribution of different sensors or feature dimensions, making it difficult to form targeted maintenance recommendations.

[0008] (4) There are gradient and parallelization bottlenecks in long sequence dependencies: Traditional recurrent neural networks often encounter gradient vanishing or gradient explosion problems when processing long sequences, and convolutional networks are also limited by the size of the receptive field and have difficulty capturing ultra-long-term dependencies; for the data processing needs of non-stationary long sequences, existing models either require a lot of computing resources or have problems such as information forgetting and low training efficiency, and cannot yet balance prediction accuracy and computing cost well. Summary of the Invention

[0009] To address at least one of the technical problems mentioned above, this invention provides a PEMFC degradation prediction method and system based on spatial dynamic non-stationary reconfiguration attention. This method can better address the non-stationarity of multi-sensor signals under different times and operating conditions, thereby capturing more critical and richer degradation feature expressions. At the same time, the parallel temporal dimension decomposition and spatial dimension modeling can effectively prevent interference between different feature information, taking into account both global correlation and local fluctuation features, thereby significantly improving the prediction accuracy of complex degradation processes of fuel cells.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

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

[0012] Acquire historical operating data from multiple sensors in the PEMFC fuel cell;

[0013] The PEMFC degradation prediction model is trained using multi-sensor historical operating data of the fuel cell PEMFC to obtain the trained PEMFC degradation prediction model; the construction process of the PEMFC degradation prediction model includes:

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

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

[0016] Based on the non-stationary factor, the attention weight in the self-attention mechanism during spatial feature extraction is dynamically corrected to obtain the corrected spatial features of battery degradation.

[0017] By fusing battery degradation features in the time dimension and the corrected battery degradation spatial features, degradation prediction features are obtained;

[0018] The degradation prediction results are obtained by predicting the PEMFC operating data under actual operating conditions based on the trained PEMFC degradation prediction model.

[0019] Furthermore, the historical operating data of the multi-sensor PEMFC fuel cell 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.

[0020] Furthermore, 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:

[0021]

[0022] in, Let these represent the mean vector and standard deviation vector of the spatial dimension, respectively. Let N represent the row vector obtained by transposing the i-th sensor observation in input X, where N is the sensor dimension.

[0023] Furthermore, by combining multi-sensor time-series data and its spatial statistical information, the current time-series distribution information is mapped to a non-stationary factor, calculated using the following formula:

[0024]

[0025] Where α is the scaling factor and β is the bias factor. For the transpose of a multi-sensor time series, and Let be the mean vector and standard deviation vector of the spatial dimension. It is a learnable weight matrix, and the output dimension c is determined by parameters α and β, d model Given the input spatial dimension, when the input statistical parameter s is... When c is set to 1, α represents a positive scalar; while when the input statistical parameter s is... When c is set to N, β represents a bias vector containing N elements.

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

[0027] Average pooling is used to extract the trend and seasonal characteristics of fluctuations;

[0028] By inputting the trend and seasonal characteristics of fluctuations into a linear layer, the battery degradation characteristics in the time dimension are obtained.

[0029] Furthermore, based on the non-stationary factor, the attention weights in the self-attention mechanism during spatial feature extraction are dynamically modified to obtain the modified battery degradation spatial features. This includes using a multi-head strategy in the spatial self-attention mechanism to divide Q, K, and V into multiple subspaces for parallel computation and reconstruction of attention. The output features of the multi-head attention mechanism are then concatenated and output. Based on the output features of the multi-head attention mechanism, a feedforward neural network is used for propagation to obtain the modified battery degradation spatial features.

[0030] A second aspect of the present invention provides a PEMFC degradation prediction system based on spatial dynamic nonstationary reconstruction attention, comprising:

[0031] The data acquisition module is used to acquire historical operating data from multiple sensors in the PEMFC fuel cell.

[0032] The model training module is used to train the constructed PEMFC degradation prediction model based on the multi-sensor historical operating data of the fuel cell PEMFC to obtain the trained PEMFC degradation prediction model; the construction process of the PEMFC degradation prediction model includes:

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

[0034] Acquire battery degradation characteristics in the time dimension from multi-sensor time-series data;

[0035] Based on the non-stationary factor, the attention weight in the self-attention mechanism during spatial feature extraction is dynamically corrected to obtain the corrected spatial features of battery degradation.

[0036] By fusing battery degradation features in the time dimension and the corrected battery degradation spatial features, degradation prediction features are obtained;

[0037] The degradation prediction module is used to predict the degradation results of fuel cell PEMFC operating data under actual operating conditions based on the trained PEMFC degradation prediction model.

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

[0039] This invention focuses on degradation information from different dimensions through parallel spatial self-attention networks and trend-seasonal decomposition linearization. In the spatial self-attention network, an adaptive non-stationary factor is introduced to dynamically reconstruct attention, better addressing the non-stationarity of multi-sensor signals under different times and operating conditions, thereby capturing more critical and richer degradation feature expressions. Simultaneously, parallel temporal decomposition and spatial modeling effectively prevent interference between different feature information, taking into account both global correlations and local fluctuations, thus significantly improving the prediction accuracy of complex fuel cell degradation processes. This invention's method shows promising application prospects in fuel cell health management and remaining life prediction.

[0040] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

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

[0043] Figure 2 This is a network structure diagram of the fuel cell degradation model provided in this embodiment of the invention;

[0044] Figure 3 This is a structural diagram of the adaptive non-stationary reconstruction attention mechanism provided in an embodiment of the present invention;

[0045] Figure 4 The results are compared with four representative methods on the FC1 dataset provided in this embodiment of the invention; where (a) is the comparison of the results of 1-step prediction, (b) is the comparison of the error of 1-step prediction, (c) is the comparison of the results of 3-step prediction, (d) is the comparison of the error of 3-step prediction, (e) is the comparison of the results of 5-step prediction, and (f) is the comparison of the error of 5-step prediction.

[0046] Figure 5 The results are compared with four representative methods on the FC2 dataset provided in this embodiment of the invention; where (a) is the comparison of the results of 1-step prediction, (b) is the comparison of the error of 1-step prediction, (c) is the comparison of the results of 3-step prediction, (d) is the comparison of the error of 3-step prediction, (e) is the comparison of the results of 5-step prediction, and (f) is the comparison of the error of 5-step prediction.

[0047] Figure 6 These are performance comparisons of different models with different training lengths on the FC2 dataset provided in this embodiment of the invention; wherein, (a) is a comparison of results when the ratio of training, validation, and test set splitting is 3:1:6 under 1-step prediction, (b) is a comparison of errors when the ratio of training, validation, and test set splitting is 3:1:6 under 1-step prediction, (c) is a comparison of results when the ratio of training, validation, and test set splitting is 5:1:4 under 1-step prediction, (d) is a comparison of errors when the ratio of training, validation, and test set splitting is 5:1:4 under 1-step prediction, (e) is a comparison of results when the ratio of training, validation, and test set splitting is 7:1:2 under 1-step prediction, and (f) is a comparison of errors when the ratio of training, validation, and test set splitting is 7:1:2 under 1-step prediction. Detailed Implementation

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

[0049] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, 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 operating time increases, the output performance of fuel cells gradually declines, with degradation typically manifesting as voltage drop and reduced output power. To address this challenge, many studies have assessed fuel cell lifetime or predicted degradation trends based on time-series data. However, these methods usually only focus on the degradation process over time, neglecting the significant spatial correlations that often exist between measurements from multiple sensors in a fuel cell system. Without adequately modeling the coupling relationships between these sensors, the understanding of the overall degradation mechanism is inevitably one-sided.

[0052] Based on the above considerations, this invention first introduces a spatial Transformer to capture the correlations between multiple sensors. Unlike traditional Transformers, only the encoder part is retained, and the decoder structure is removed to reduce model complexity. When processing temporal information, this invention performs trend-seasonal 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 result from using only spatial information. However, the voltage and operating conditions of fuel cells under real-world conditions often fluctuate drastically and asynchronously, making it difficult for traditional attention mechanisms to output discriminative and interpretable attention maps, as they lose information after normalization. To address this problem, this invention proposes an Adaptive Non-stationary Reconstruct Attention (ANRA), which incorporates non-stationary features into the attention reconstruction process, thereby obtaining more interpretable inter-sensor coupling relationships.

[0053] Example 1

[0054] like Figure 1 As 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 multi-sensor historical operating 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] The test set is obtained by collecting multi-sensor operating data of the fuel cell system under real-time operating conditions.

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

[0058] In this embodiment, modeling is performed by integrating temporal and spatial perspectives, and training is conducted using a training set.

[0059] Figure 2 This is a schematic diagram of the structure of the Spatial Adaptive Non-Stationary Reconstruction Attention Transformer (SANformer) model proposed in this invention, as shown below. Figure 2 As shown, the construction process of the degradation prediction model for spatial adaptive nonstationary reconstruction attention mainly consists of the following key steps:

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

[0061] Spatiotemporal feature analysis: Extracting temporal and spatial features from PEMFC data collected from multiple sensors.

[0062] On the one hand, the original information X is transposed and encoded, and then input into a 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 raw data X undergoes seasonal-trend decomposition, and the separated seasonal components are then modeled in a linear layer to capture 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 multilayer perceptron (MLP); this synthesis method captures the spatial coupling between sensor measurements and preserves highly fluctuating temporal patterns, thereby enabling more accurate and interpretable prediction of fuel cell degradation trends.

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

[0066] like Figure 2 and Figure 3 As shown, in order to more clearly illustrate the technical solution of the present invention, the processing procedure is described in detail below:

[0067] Assume the observed multi-sensor time-series data are as follows:

[0068]

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

[0070] After X undergoes preprocessing such as filtering and normalization, it is transposed as follows:

[0071]

[0072] It is mapped to a high-dimensional representation E by a linear mapping. s The dimension can be set to d. model :

[0073] E s =X T ·W E ,

[0074] in, E is a learnable parameter. s It will serve as the raw input for the attention mechanism.

[0075] Self-attention mechanisms can capture global dependencies between all positional features in an input sequence and support efficient parallel computation because they do not depend on the order of sequence processing.

[0076] In traditional self-attention mechanisms, the following is calculated first:

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

[0078] Then, the attention weights (AW) are obtained by scaling the dot product:

[0079]

[0080] However, directly inputting continuously changing fuel cell data can reduce 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 recovered by denormalization, which makes it difficult for the model to generate a focused attention map.

[0081] To address the non-stationary distribution of sensor data in fuel cells under dynamic operating conditions, this invention utilizes an ANR module in spatial attention calculation, with improvements to the attention mechanism as follows: Figure 3 As shown, the calculation details are as follows:

[0082] (1) Statistical feature extraction: for the original input Calculate its mean value over the sensor dimension N. and standard deviation It can be defined as:

[0083]

[0084] in, Let these represent the mean vector and standard deviation vector of the spatial dimension, respectively. Since... The distribution may become unstable over time and under different operating conditions, and it is necessary to capture this dynamic change. This represents the row vector obtained by transposing the i-th sensor observation value in the input X.

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

[0086] To enable α and β to be dynamically adjusted over time, this invention designs two parallel bidirectional gated recurrent network (Bi-GRU) submodules in the ANR, which are respectively input to... and The learning process maps the current temporal distribution information to a scaling factor α and a bias factor β. This can be denoted as:

[0087]

[0088] in, It is a learnable weight matrix, and the output dimension c is determined by parameters α and β. Specifically, when the input statistical parameters s are... When c is set to 1, α represents a positive scalar; while when the input parameter s is... When c is set to N, β represents a bias vector containing N elements.

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

[0090]

[0091] The above modifications enable the model to adaptively strengthen or suppress the influence of certain sensors on the degradation process during attention calculation, thereby still being able to output a relatively focused and interpretable attention map in non-stationary, multi-condition data, and intuitively give the attention weights of different features on degradation.

[0092] Furthermore, to enhance the model's ability to capture multidimensional degradation features, this invention employs a multi-head attention strategy in the spatial self-attention mechanism. This involves dividing Q, K, and V into multiple subspaces, computing and reconstructing the attention in parallel, and then concatenating them through a learnable mapping W. O Restore to the original dimension. This process can be represented as:

[0093] MultiHead(Q ′ ,K ′ V ′ ,α,β)=Concat(head1,head2,…,head h W O ,

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

[0095] Among them, head i It is every attention head, In the embodiment, using h represents the number of attention heads. The output mapping represents the attention mechanism.

[0096] After the multi-head attention mechanism outputs high-dimensional features, a feedforward neural network is used for further propagation to enhance the model's non-linear expressive power. The FFN is designed as a two-layer feedforward neural network, as shown in the following formula:

[0097] FFN(x) = ReLU(xW1+b1)W2+b2,

[0098] in, x represents the output of the aforementioned spatial multi-head attention mechanism. The structure of the FFN output serves as the output of this layer's encoder, while the information extracted repeatedly through N layers of encoders serves as the output of the spatial features.

[0099] To avoid neglecting temporal information in spatial self-attention modeling, this invention obtains the periodic or irregular fluctuation characteristics of fuel cell degradation from the temporal dimension through seasonal-trend decomposition combined with linear mapping in the parallel branch. The specific operation is as follows:

[0100] First, average pooling is used to extract seasonal information about fluctuations. This process is defined as follows:

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

[0102] Where X represents the original input sequence, Trend, and represent the trend component and the seasonal component, respectively. 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, and the fluctuating seasonal feature is input into the linear layer, as shown in the following formula:

[0104]

[0105] in, It is worth noting that the S entered here TIt does not model any spatial relationships. The time-modeling output O... T The information fused with spatial features is ultimately used to output a degradation prediction via a multilayer perceptron.

[0106]

[0107] in, This represents the model's predicted output.

[0108] To verify the effectiveness of this invention, the dataset from the IEEE 2014 PHM Data Challenge was used for training, validation, and testing. This dataset consists of two parts, FC1 and FC2. The PEMFC used in the experiment consists of five battery cells. FC1 and FC2 were operated 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. FC1 ran for 1154 hours, and FC2 ran for 1020 hours. Sensors recorded key operating parameters of the battery during operation, such as voltage, current, temperature, gas flow rate, and pressure of each stack.

[0109] Since the data comes from different sensors, the scales vary, the sampling time is uneven, and there is a lot of noise and spikes, which may bring incorrect input signals to the model and mislead the model training. Therefore, the raw data is averaged at 1-hour intervals and smoothed using the LOESS method. The filter window is set to 20 and 10 respectively (smaller window for quasi-dynamic testing to capture more details).

[0110] To more intuitively evaluate the degradation prediction performance of the SANRA-Transformer model, this invention uses the FC1 and FC2 datasets for testing and compares it with four representative methods, including CNN, LSTM, traditional Transformer, and non-stationary Transformer. First, experiments were conducted with 1, 3, and 5 time steps respectively, targeting different temporal fine-grained predictions. The experimental comparison results on the FC1 dataset are shown in Table 1 and... Figure 4 As shown, Figure 4 In the table, (a) compares the results of 1-step prediction, (b) compares the errors of 1-step prediction, (c) compares the results of 3-step prediction, (d) compares the errors of 3-step prediction, (e) compares the results of 5-step prediction, and (f) compares the errors of 5-step prediction.

[0111] The experimental comparison results on the FC2 dataset are shown in Table 2 and Figure 5 As shown, Figure 5In the table, (a) compares the results of 1-step prediction, (b) compares the errors of 1-step prediction, (c) compares the results of 3-step prediction, (d) compares the errors of 3-step prediction, (e) compares the results of 5-step prediction, and (f) compares the errors of 5-step prediction.

[0112] Table 1. Comparison of experimental results of various prediction methods on the FC1 dataset.

[0113]

[0114] Table 2 shows the experimental comparison results of various prediction methods on the FC2 dataset.

[0115]

[0116] Experimental results show that the proposed model performs highly consistent with actual voltage in degradation prediction, and significantly outperforms LSTM, CNN, and Transformer-based models. It is worth noting that the unmodified Transformer model loses most of the fluctuation information due to over-smoothing, resulting in poor performance. Furthermore, CNN still faces significant challenges in handling long-term time series dependencies, especially in dynamic models.

[0117] Subsequently, the performance of the FC 2 one-step prediction experiment was analyzed under different training lengths, and the results were compared with those of the Bi-LSTM and Non-stationary Transformer models in the above experiments. Figure 6 Shown in Figure 6 In the figures, (a) compares the results when the training, validation, and test set split ratio is 3:1:6 under one-step prediction; (b) compares the errors when the training, validation, and test set split ratio is 3:1:6 under one-step prediction; (c) compares the results when the training, validation, and test set split ratio is 5:1:4 under one-step prediction; (d) compares the errors when the training, validation, and test set split ratio is 5:1:4 under one-step prediction; (e) compares the results when the training, validation, and test set split ratio is 7:1:2 under one-step prediction; and (f) compares the errors when the training, validation, and test set split ratio is 7:1:2 under one-step prediction. A shorter training length means the model can enter the working state earlier. Experimental results show that the model of this invention is not very sensitive to the length of the training set, while the accuracy of the LSTM-based method is limited to some extent by the amount of training information.

[0118] Finally, ablation experiments were used to evaluate the effectiveness of two internal components in the module of this invention: the effectiveness of ANAR in reconstructing spatial attention and the effectiveness of the trend decomposition module in temporal modeling. To this end, three ablation experiments were designed: spatial attention alone, spatial attention reconstructed by ANAR, and ANAR reconstructed spatial attention with added trend decomposition (SANRA Transformer). Taking the single-step degradation prediction of FC1 as an example, the ablation experiment results are shown in Table 3.

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

[0120]

[0121] Table 3 visually demonstrates that the enhanced spatial attention focuses more intently on different types of information. For different degradation trends, the contributions of different sensors to degradation are clearly shown on the attention map. Taking the 5-step degradation prediction of FC1 as an example, it illustrates the different weights of trend decomposition in temporal modeling. Ablation experiments show that the SANRA Transformer's prediction performance is superior to the original spatial attention; removing any module significantly reduces model performance, indicating that both temporal and spatial perspectives are meaningful for modeling degradation.

[0122] Example 2

[0123] This embodiment provides a PEMFC degradation prediction system based on spatial dynamic non-stationary reconstruction attention, including:

[0124] The data acquisition module is used to acquire historical operating data from multiple sensors in the PEMFC fuel cell.

[0125] The model training module is used to train the constructed PEMFC degradation prediction model based on the multi-sensor historical operating data of the fuel cell PEMFC to obtain the trained PEMFC degradation prediction model; the construction process of the PEMFC degradation prediction model includes:

[0126] Obtain statistical information of multi-sensor time-series data in the spatial dimension, and combine multi-sensor time-series data with its statistical information in the spatial dimension to map the current time-series distribution information to a non-stationary factor;

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

[0128] Based on the non-stationary factor, the attention weight in the self-attention mechanism during spatial feature extraction is dynamically corrected to obtain the corrected spatial features of battery degradation.

[0129] By fusing battery degradation features in the time dimension and the corrected battery degradation spatial features, degradation prediction features are obtained;

[0130] The degradation prediction module is used to predict the degradation results of fuel cell PEMFC operating data under actual operating conditions based on the trained PEMFC degradation prediction model.

[0131] Example 3

[0132] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described above.

[0133] Example 4

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

[0135] Example 5

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

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A PEMFC degradation prediction method based on spatial dynamic nonstationary reconstruction attention, characterized in that, Includes the following steps: Acquire historical operating data from multiple sensors in the PEMFC fuel cell; The PEMFC degradation prediction model is trained using multi-sensor historical operating data of the fuel cell PEMFC to obtain the trained PEMFC degradation prediction model; the construction process of the PEMFC degradation prediction model includes: Obtain statistical information of multi-sensor time-series data in the spatial dimension, and combine multi-sensor time-series data with its statistical information in the spatial dimension to map the current time-series distribution information to a non-stationary factor; Acquire battery degradation characteristics in the time dimension from 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 corrected to obtain the corrected spatial features of battery degradation. By fusing battery degradation features in the time dimension and the corrected battery degradation spatial features, degradation prediction features are obtained; The degradation prediction results are obtained by predicting the PEMFC operation data under actual working conditions based on the trained PEMFC degradation prediction model. Among them, 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: ， ， in, Let these represent the mean vector and standard deviation vector of the spatial dimension, respectively. Indicates input X The Middle i The row vector obtained by transposing the sensor observations, where N is the sensor dimension; Specifically, by combining multi-sensor time-series data and its spatial statistical information, the current time-series distribution information is mapped to a non-stationary factor, calculated using the following formula: ， ， ， Where α is the scaling factor and β is the bias factor. For the transpose of a multi-sensor time series, and Let be the mean vector and standard deviation vector of the spatial dimension. It is a learnable weight matrix, and the output dimension c is determined by the parameters α and β. Given the input spatial dimension, when the input statistical parameter s is... When c is set to 1, α represents a positive scalar; while when the input statistical parameter s is... When c is set to N, β represents a bias vector containing N elements.

2. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in claim 1, characterized in that, The historical operating data of the multi-sensor PEMFC fuel cell is used as training data, and the corresponding degradation index or remaining service life 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 described in claim 1, acquiring battery degradation characteristics in the time dimension from multi-sensor time-series data, including: Average pooling is used to extract the trend and seasonal characteristics of fluctuations; By inputting the trend and seasonal characteristics of fluctuations into a linear layer, the battery degradation characteristics in the time dimension are obtained.

4. The PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in claim 1, based on a non-stationary factor, dynamically corrects the attention weights in the self-attention mechanism during spatial feature extraction to obtain the corrected battery degradation spatial features, including employing a multi-head strategy in the spatial self-attention mechanism. The attention is divided into multiple subspaces and reconstructed in parallel. After being concatenated, the output features of the multi-head attention mechanism are output. 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.

5. A PEMFC degradation prediction system based on spatial dynamic non-stationary reconstruction attention, employing the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1-4, characterized in that, include: The data acquisition module is used to acquire historical operating data from multiple sensors in the PEMFC fuel cell. The model training module is used to train the constructed PEMFC degradation prediction model based on multi-sensor historical operating data of the fuel cell PEMFC to obtain the trained PEMFC degradation prediction model; the construction process of the PEMFC degradation prediction model includes: Obtain statistical information of multi-sensor time-series data in the spatial dimension, and combine multi-sensor time-series data with its statistical information in the spatial dimension to map the current time-series distribution information to a non-stationary factor; Acquire battery degradation characteristics in the time dimension from 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 corrected to obtain the corrected spatial features of battery degradation. By fusing battery degradation features in the time dimension and the corrected battery degradation spatial features, degradation prediction features are obtained; The degradation prediction module is used to predict the degradation results of fuel cell PEMFC operating data under actual operating conditions based on the trained PEMFC degradation prediction model.

6. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1-4.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1-4.

8. A program product, said program product being a computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps in the PEMFC degradation prediction method based on spatial dynamic non-stationary reconstruction attention as described in any one of claims 1-4.

Images