A fuel cell life prediction method and system based on multi-source feature fusion
By using a multi-source feature fusion method, a multi-task LSTM network and a heterogeneous feature extraction network are constructed. Combined with the particle swarm optimization algorithm, the problems of insufficient accuracy and stability in fuel cell lifetime prediction are solved, and higher accuracy and stable lifetime prediction are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV OF TECH
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies suffer from insufficient accuracy and stability in fuel cell life prediction, particularly in their inadequate ability to characterize long-term aging evolution patterns, making it difficult to meet the life requirements of applications such as transportation and stationary power generation.
A multi-source feature fusion method is adopted. By collecting multi-dimensional health index data during the operation of fuel cells, data preprocessing and correlation verification are performed. A multi-task LSTM network and a heterogeneous feature extraction network are constructed. Particle swarm optimization algorithm is combined for global optimization to obtain the optimal combination of input dimensions and health indexes, thereby improving the accuracy and stability of fuel cell life prediction.
It effectively reduces noise interference, weakens the impact of outliers, improves feature representation capabilities, enhances the consistency and integrity of time series data, and can more comprehensively characterize the nonlinearity and dynamics of fuel cell life decay process, thereby improving prediction accuracy and stability.
Smart Images
Figure CN122246183A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fuel cell technology, and in particular to a method and system for predicting fuel cell lifetime based on multi-source feature fusion. Background Technology
[0002] The lifespan and durability of fuel cells during long-term operation remain a major bottleneck restricting their commercialization. Currently, proton exchange membrane fuel cells (PEMFCs) typically achieve a lifespan of 5,000–10,000 hours under laboratory conditions, but there is still a significant gap to meet the lifespan requirements of over 20,000 hours or even 40,000 hours for applications such as transportation and stationary power generation. With the expansion of application scale, fuel cell lifespan is gradually becoming one of the core technological challenges restricting the industrialization of hydrogen energy. Fuel cell lifespan is affected by various factors, including membrane electrode aging, catalyst degradation, improper hydrothermal management, and electrochemical stress caused by start-stop cycles. Achieving long-term stable operation of fuel cells is not only a necessary condition for improving hydrogen energy utilization efficiency and reducing operating costs, but also a key link in promoting the industrialization of fuel cells and realizing a green and low-carbon energy system.
[0003] Chinese patent CN119493002A discloses a method, system, device, and medium for predicting fuel cell performance degradation. This scheme includes acquiring a battery durability dataset, scaling the data in the dataset using standard scores to obtain preprocessed data, dividing the preprocessed data into training and test sets, constructing a degradation prediction model based on a graph neural network, optimizing the hyperparameters of the degradation prediction model using a random search algorithm, and training the degradation prediction model using the training set data to obtain a trained prediction model, inputting the test set data into the trained prediction model, and outputting the prediction results. However, the above scheme mainly focuses on short-step prediction of output voltage degradation and uses graph neural networks and random search modeling, which lacks sufficient ability to represent the long-term aging evolution of fuel cells, resulting in insufficient accuracy and stability in fuel cell life prediction. Therefore, it is essential to provide a fuel cell life prediction method and system based on multi-source feature fusion to improve the accuracy and stability of fuel cell life prediction. Summary of the Invention
[0004] In view of this, the present invention proposes a fuel cell lifetime prediction method and system based on multi-source feature fusion, which helps to improve the accuracy and stability of fuel cell lifetime prediction.
[0005] This invention provides a fuel cell lifetime prediction method based on multi-source feature fusion, the method comprising: Collect multidimensional health indicator data during the operation of the fuel cell, and perform data preprocessing on the multidimensional health indicator data to obtain multidimensional health time series data; The correlation of health indicators in the multidimensional health time series data is verified, and the health indicator data that meet the correlation indicator conditions are determined as candidate health indicators. Based on each candidate health indicator and a multi-objective optimization function, the optimal input dimension and the optimal combination of health indicators are obtained. Based on the optimal input dimension and the optimal combination of health indicators, a heterogeneous prediction model for fuel cells is constructed. The heterogeneous prediction model for fuel cells includes a multi-task LSTM network and a heterogeneous feature extraction network. The fuel cell heterogeneous prediction model is globally optimized using the particle swarm optimization algorithm. The multidimensional health index data to be predicted is input into the optimized fuel cell heterogeneous prediction model to obtain the fuel cell life degradation prediction results.
[0006] Based on the above technical solutions, preferably, the step of preprocessing the multidimensional health indicator data to obtain multidimensional health time-series data specifically includes: Gaussian filtering is applied to each health indicator in the multidimensional health indicator data to obtain standard health indicator data after removing outliers. The standard health indicator data are compensated using cubic interpolation to obtain a multidimensional time series dataset. The health indicator data corresponding to each dimension in the multidimensional time series dataset are normalized to obtain the multidimensional health time series data.
[0007] Based on the above technical solutions, preferably, the step of verifying the correlation between health indicators in the multidimensional health time-series data and determining the health indicator data that meets the correlation indicator conditions as candidate health indicators specifically includes: Using the voltage change of the fuel cell as a benchmark, the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient between each health indicator data in the multidimensional health time series data and the voltage change are calculated respectively. If the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is greater than a preset correlation index value, the health index data corresponding to the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is determined as a candidate health index.
[0008] More preferably, the step of obtaining the optimal input dimension and optimal combination of health indicators based on each candidate health indicator and the multi-objective optimization function specifically includes: A multi-objective optimization function is constructed by taking the prediction target accuracy function, the optimal permutation and combination function of health indicators, and the model computational complexity function as optimization content. The optimal input dimension and the optimal health indicator combination corresponding to the optimal input dimension are obtained by optimizing each candidate health indicator combination under different input dimensions.
[0009] More preferably, the multi-task LSTM network includes a multi-index allocation layer, a multi-branch feature extraction layer, a shared fully connected layer, and a transition output layer connected in sequence, wherein, The multi-indicator allocation layer is used to receive multi-dimensional health time-series data determined by the optimal combination of health indicators, and allocate each health indicator data in the multi-dimensional health time-series data to the corresponding task branch according to the health indicator category. The multi-branch feature extraction layer includes multiple parallel LSTM task branches. Each LSTM task branch is used to perform time-series dependency modeling and aging feature extraction on the corresponding health indicator data, and outputs the corresponding aging feature vector. The shared fully connected layer is used to score the importance of the aging feature vectors output by each LSTM task branch, and converts the importance score results into normalized weights corresponding to each LSTM task branch through the Softmax function. Based on the normalized weights corresponding to each LSTM task branch, the aging feature vectors are weighted and aggregated to generate a multidimensional aging feature vector. The transition output layer is used to output the multidimensional aging feature vector to the heterogeneous feature extraction network.
[0010] More preferably, the heterogeneous feature extraction network includes, in sequence, an input mapping layer, a position encoding layer, a multi-head self-attention feature extraction layer, a residual normalization layer, and a prediction output layer, wherein, The input mapping layer is used to receive the multidimensional aging features output by the multi-task LSTM network and perform dimensional mapping on the multidimensional aging features to generate a unified dimension input feature sequence. The position encoding layer is used to introduce temporal position information into the input feature sequence to characterize the sequential relationship between features at each time point; The multi-head self-attention feature extraction layer is used to perform multi-head parallel attention calculation on the input feature sequence after introducing temporal location information, so as to extract global correlation features and coupled aging features between different health indicators and between different time periods; The residual normalization layer is used to perform residual connection and normalization processing on the global correlation features and the coupled aging features output by the multi-head self-attention feature extraction layer; The prediction output layer is used to output the fuel cell lifespan degradation prediction results.
[0011] More preferably, the particle swarm optimization algorithm performs secondary optimization on the heterogeneous feature extraction network in the fuel cell heterogeneous prediction model based on the aging characteristics of the multi-task LSTM network in the fuel cell heterogeneous prediction model, so as to balance the prediction accuracy of different aging indicators in the fuel cell heterogeneous prediction model.
[0012] A second aspect of this application provides a fuel cell lifetime prediction system based on multi-source feature fusion. The fuel cell lifetime prediction system includes a data acquisition module, a data processing module, and a lifetime prediction module. The data acquisition module is used to collect multidimensional health indicator data during the operation of the fuel cell, and to preprocess the multidimensional health indicator data to obtain multidimensional health time series data. The data processing module is used to verify the correlation of health indicators in the multidimensional health time series data, and to determine the health indicator data that meet the correlation indicator conditions as candidate health indicators. Based on each candidate health indicator and the multi-objective optimization function, the optimal input dimension and the optimal combination of health indicators are obtained. Based on the optimal input dimension and the optimal combination of health indicators, a fuel cell heterogeneous prediction model is constructed. The fuel cell heterogeneous prediction model includes a multi-task LSTM network and a heterogeneous feature extraction network. The lifespan prediction module is used to perform global optimization of the fuel cell heterogeneous prediction model according to the particle swarm optimization algorithm, and input the multidimensional health index data to be predicted into the optimized fuel cell heterogeneous prediction model to obtain the fuel cell lifespan degradation prediction results.
[0013] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory.
[0014] A fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a fuel cell lifetime prediction method based on multi-source feature fusion.
[0015] The fuel cell lifetime prediction method and system based on multi-source feature fusion provided by this invention have the following advantages over existing technologies: (1) By preprocessing the multidimensional health index data during the operation of fuel cells, noise interference can be effectively reduced, the impact of outliers can be weakened, and the consistency and integrity of time series data can be enhanced, thus providing a reliable data foundation for subsequent life prediction. The correlation of each health index can be verified, and candidate health indices that meet the conditions can be screened out. Indicators with low correlation to life decay and high redundancy can be effectively eliminated, invalid feature input can be reduced, and feature representation ability can be improved. Furthermore, the optimal input dimension and optimal health index combination can be obtained based on candidate health indices and multi-objective optimization functions, which can achieve a balance between prediction accuracy, model complexity and computational efficiency, and avoid overfitting or prediction performance degradation caused by excessive dimensionality or improper feature selection. At the same time, by constructing a fuel cell heterogeneous prediction model that includes a multi-task LSTM network and a heterogeneous feature extraction network, the temporal evolution relationship between multidimensional health indices and the coupling relationship between heterogeneous features can be explored simultaneously, which can more fully characterize the nonlinearity, dynamics and multi-source correlation in the fuel cell life decay process, thereby improving the accuracy and stability of fuel cell life prediction.
[0016] (2) Using the voltage change of the fuel cell as a benchmark, the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient between each health index data and the voltage change are calculated respectively. This allows for a comprehensive evaluation of the relationship between health indexes and fuel cell performance degradation from both linear and nonlinear correlation perspectives, thereby improving the accuracy of candidate health index identification. Furthermore, by simultaneously introducing the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient, it is possible not only to identify health indexes with a significant linear relationship with the voltage change, but also to identify health indexes with complex nonlinear coupling relationships, thus avoiding the problem of missing effective features due to using only a single linear analysis method. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating the fuel cell lifetime prediction method based on multi-source feature fusion provided by this invention; Figure 2 This is a schematic diagram of the structure of the multi-task LSTM network provided by the present invention; Figure 3 This is a schematic diagram of the heterogeneous feature extraction network provided by the present invention; Figure 4A schematic diagram of the structure of the fuel cell life prediction system provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention.
[0019] Explanation of reference numerals in the attached figures: 1. Fuel cell life prediction system; 11. Data acquisition module; 12. Data processing module; 13. Life prediction module; 2. Electronic equipment; 21. Processor; 22. Communication bus; 23. User interface; 24. Network interface; 25. Memory. Detailed Implementation
[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0021] This invention discloses a fuel cell lifetime prediction method based on multi-source feature fusion, with reference to... Figure 1 and Figure 2 The steps of this method include S1 to S4.
[0022] Step S1: Collect multidimensional health indicator data during the operation of the fuel cell, and perform data preprocessing on the multidimensional health indicator data to obtain multidimensional health time series data.
[0023] This step also includes steps S11 to S13.
[0024] Step S11: Perform Gaussian filtering on each health indicator data in the multidimensional health indicator data to obtain standard health indicator data after removing abnormal data.
[0025] In this step, Gaussian filtering is applied to each dimension of the multidimensional health indicator data to remove outliers. The filtering width of the Gaussian filter is set to ±2σ, where σ is the standard deviation of the current dimension data. Outliers outside the Gaussian distribution curve under this dimension are removed.
[0026] Step S12: Compensate the standard health indicator data based on cubic interpolation to obtain a multidimensional time series dataset.
[0027] In this step, after Gaussian filtering of the multidimensional health indicator data in step S11, standard health indicator data after removing outliers can be obtained. Since there may be local missing data, discontinuous sampling points, inconsistent sampling intervals, or distortion of individual sampling values in the actual collection process of various health indicators, interpolation compensation processing is required on the standard health indicator data to ensure the integrity, continuity, and consistency of the input data for the subsequent life prediction model.
[0028] In this embodiment, the health indicator data for each dimension are arranged according to the sampling time order or the cycle number order to form corresponding time-series data. For positions with missing values or requiring the addition of intermediate sampling points, cubic interpolation is used to fit the changing trend between adjacent known sampling points, and the health indicator values at the missing or resampled positions are calculated accordingly. Cubic interpolation can combine the changing relationship between adjacent sampling points to construct a continuous and smooth fitting curve within each interval. This ensures that the compensated data not only passes through the known sampling points but also maintains good smoothness and curvature continuity at the connection between adjacent intervals, thereby improving the reliability of the interpolation results.
[0029] After interpolating and compensating the health indicator data for each dimension, the data is further aligned and combined according to a unified time axis or cyclic axis to form a complete multidimensional time series dataset. This step effectively compensates for missing information and discontinuities in the original sampled data, improving the completeness, smoothness, and temporal consistency of the multidimensional health indicator data. This provides a stable and reliable data foundation for subsequent data normalization processing and the training of fuel cell life prediction models.
[0030] Step S13: Normalize the health indicator data corresponding to each dimension in the multidimensional time series dataset to obtain multidimensional health time series data.
[0031] In this step, each dimension of the dataset is normalized using the Min-Max method, transforming it to the range [0,1], and used as multidimensional health time series data as input to the model.
[0032] Specifically, after interpolating and compensating the standard health indicator data in step S12, a multidimensional time series dataset aligned with various health indicators on a unified time axis or cyclic axis can be obtained. Since different health indicators typically differ in dimensions, value ranges, and magnitudes of change, directly using them as input to subsequent lifespan prediction models can easily lead to larger health indicators having a strong impact on the model training process, thereby reducing the comparability between features of different dimensions and affecting feature fusion performance. Therefore, it is necessary to normalize the health indicator data corresponding to each dimension in the multidimensional time series dataset.
[0033] In this embodiment, the maximum and minimum values of each health indicator are calculated over the entire time series. Based on these maximum and minimum values, the values of each health indicator at each sampling time are normalized, so that the data of each health indicator are mapped to a unified numerical range, preferably to the range of zero to one. This processing can effectively eliminate the influence of differences in units between different health indicators, improve the comparability and consistency between the input features of each dimension, and thus enhance the stability of subsequent model training.
[0034] Furthermore, when the maximum and minimum values of a certain health indicator are the same within the current sample interval, it indicates that the indicator has not changed significantly within the corresponding interval and belongs to a constant or approximately stationary sequence. In this case, to avoid anomalies during normalization, the normalization result of the health indicator can be set to a preset constant value, preferably zero, to ensure the stability of the normalization process and the consistency of the data processing results. After completing the normalization processing of the health indicator data for each dimension, multidimensional health time-series data can be obtained.
[0035] Step S2: Verify the correlation between health indicators in the multidimensional health time series data, and determine the health indicator data that meet the correlation indicator conditions as candidate health indicators.
[0036] This step also includes steps S21 to S22.
[0037] Step S21: Using the voltage change of the fuel cell as a benchmark, calculate the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient between each health indicator data and the voltage change in the multidimensional health time series data.
[0038] In this step, the voltage change is used as a benchmark to verify its correlation with other acquired data, and the coefficients include the linear Pearson correlation coefficient. r And the nonlinear maximum mutual information coefficient MIC.
[0039] Linear Pearson correlation coefficient r The expression is: The expression for the nonlinear maximum mutual information coefficient MIC is: in, This represents the Pearson correlation coefficient. Indicates the first Health indicator data values corresponding to each sample Indicates the first The baseline variable data values corresponding to each sample Represents a health indicator sequence The average value, Represents the sequence of benchmark variables The average value, Represents the total number of samples. This represents the maximum mutual information coefficient of multidimensional health time series data. This indicates that the data is divided into x × y The maximum mutual information value obtained during gridding. This indicates the number of grid cells along the horizontal axis when the data is partitioned. This indicates the number of grid cells along the vertical axis when the data is partitioned. Represents the total number of samples The relevant upper bound function for mesh partitioning, This represents the normalization term, which normalizes the mutual information value so that the MIC value ranges from [0,1].
[0040] Step S22: If the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is greater than the preset correlation index value, the health index data corresponding to the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is determined as a candidate health index.
[0041] In this step, in step S21, the linear Pearson correlation coefficient and nonlinear maximum mutual information coefficient between each health indicator data in the multidimensional health time series data and the voltage change of the fuel cell have been calculated. Based on the calculation results, the strength of the correlation between each health indicator and the voltage change is determined. If the linear Pearson correlation coefficient corresponding to a certain health indicator is greater than the preset correlation index value, it indicates that there is a strong linear correlation between the health indicator and the voltage change, which can reflect the change characteristics in the fuel cell performance degradation process; if the nonlinear maximum mutual information coefficient corresponding to a certain health indicator is greater than the preset correlation index value, it indicates that there is a strong nonlinear correlation between the health indicator and the voltage change, which can also effectively characterize the evolution law of the fuel cell aging state.
[0042] In this embodiment, a health indicator is considered a candidate health indicator if either the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is greater than a preset correlation index value. In other words, health indicators exhibiting both linear and nonlinear correlations are retained to avoid missing effective features due to reliance on a single correlation evaluation method. By combining linear and nonlinear correlation analysis, health indicators closely related to fuel cell lifespan degradation can be screened more comprehensively, improving the representativeness and reliability of the candidate health indicator set.
[0043] Furthermore, the preset correlation index value can be set according to actual application needs, sample size, number of health indicators, and prediction accuracy requirements. Preferably, the preset correlation index value can be determined through experimental statistics, empirical setting, or cross-validation to balance the effectiveness of candidate health indicator selection and the rationality of subsequent model input dimensions. For example, indicators with a correlation index value > 0.3 can be selected as candidate parameters for the model.
[0044] In this embodiment, the voltage change of the fuel cell is used as a benchmark. The linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient between each health indicator data and the voltage change are calculated. This allows for a comprehensive evaluation of the relationship between health indicators and fuel cell performance degradation from both linear and nonlinear perspectives, thereby improving the accuracy of candidate health indicator identification. Furthermore, by simultaneously introducing the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient, not only can health indicators with a significant linear relationship with the voltage change be identified, but also those with complex nonlinear coupling relationships. This avoids the problem of missing effective features due to using only a single linear analysis method.
[0045] Step S3: Based on each candidate health indicator and the multi-objective optimization function, obtain the optimal input dimension and the optimal combination of health indicators, and construct a heterogeneous prediction model for fuel cells based on the optimal input dimension and the optimal combination of health indicators. The heterogeneous prediction model for fuel cells includes a multi-task LSTM network and a heterogeneous feature extraction network.
[0046] This step also includes steps S31 to S32.
[0047] Step S31: The prediction target accuracy function, the optimal permutation and combination function of health indicators, and the model computational complexity function are all used as optimization content to construct a multi-objective optimization function.
[0048] In this step, the decision variables of the multi-objective optimization function are: Multi-objective optimization function The expression is: Prediction target accuracy function The expression is: Optimal permutation and combination function of health indicators The expression is: Model computational complexity function The expression is: in, The input dimension parameter of the prediction model indicates the number of health indicators used by the model. This represents the weighting parameter for health indicator fusion, used to indicate the weight allocation of different health indicators during the feature fusion process; This represents the weight parameters corresponding to the prediction target accuracy function; The weight parameters represent the function corresponding to the optimal permutation and combination of health indicators; The weight parameters represent the computational complexity function of the model. Indicates the first The model prediction value for each sample; Indicates the first The true value of each sample; Indicates the length of the health indicator sequence; Indicates the first Health indicator values at any given time Represents a symbolic function; Indicates the total number of candidate health indicators; Indicates the first Whether a health indicator is selected as a weight value for the model input.
[0049] Step S32: Optimize each candidate health indicator combination under different input dimensions to obtain the optimal input dimension and the optimal health indicator combination corresponding to the optimal input dimension.
[0050] Furthermore, the fuel cell heterogeneous prediction model comprises a multi-task LSTM network at the upper layer and a heterogeneous feature extraction network at the lower layer. To fully explore the heterogeneous aging patterns in the multi-dimensional operating data of fuel cells, a multi-task LSTM network architecture was constructed. Unlike the traditional direct feature concatenation, a fully connected fusion layer based on interpretable weights was added to the multi-task LSTM network architecture. This layer introduces the idea of an attention mechanism, dynamically learning the importance weights of the output features of each LSTM subtask. .
[0051] like Figure 2 As shown, the multi-task LSTM network includes a multi-index allocation layer, a multi-branch feature extraction layer, a shared fully connected layer, and a transition output layer connected in sequence. The multi-indicator allocation layer is used to receive multi-dimensional health time-series data determined by the optimal combination of health indicators, and allocate each health indicator data in the multi-dimensional health time-series data to the corresponding task branch according to the health indicator category. The multi-branch feature extraction layer includes multiple parallel LSTM task branches. Each LSTM task branch is used to model the temporal dependency relationship and extract aging features from the corresponding health indicator data, and outputs the corresponding aging feature vector. The shared fully connected layer is used to score the importance of the aging feature vectors output by each LSTM task branch, and the Softmax function is used to convert the importance score results into normalized weights corresponding to each LSTM task branch. Based on the normalized weights corresponding to each LSTM task branch, the aging feature vectors are weighted and aggregated to generate a multi-dimensional aging feature vector. The transition output layer is used to output the multidimensional aging feature vector to the heterogeneous feature extraction network.
[0052] In this embodiment, importance weight The larger the value, the more likely it is that the first... k The higher the contribution of the features extracted from each task branch to the current lifetime prediction.
[0053] Importance weight The expression is: The expression for the fully connected layer decision function is: in, Indicates the first k Attention weights for each task branch Indicates the first k Attention score for each task branch Indicates the total number of task branches. Indicates the first k The hidden feature vectors output by each task branch represent the aging features of the corresponding health indicators. Represents the attention weight matrix. This represents the attention bias vector. This represents the transpose of the attention projection vector. The Softmax function is used to convert the scores into a probability distribution of weights, ensuring that the sum of all weights is 1.
[0054] Using the calculated interpretable weights, K The aging features of each branch are aggregated to obtain the final abstracted model input vector Z, which can be represented as: in, This represents the weight matrix composed of attention weights. This represents the feature matrix composed of the output features of each task branch. The dimension of each feature vector is represented by Z, and the resulting model input vector Z is a multidimensional feature matrix. It fully preserves the high-dimensional manifold structure of all subtasks, while using interpretable weights to suppress the feature representation of noisy branches, enabling the model to adaptively 'filter' the feature subspace most sensitive to aging prediction from the multidimensional input.
[0055] In this embodiment, the multi-index allocation layer receives multi-dimensional health time-series data determined by the optimal combination of health indicators and allocates each health indicator data to the corresponding task branch according to the health indicator category. By selectively processing health indicator data of different categories, interference between different types of features in the same modeling path can be avoided, improving the specificity and effectiveness of extracting time-series information of various health indicators, thereby enhancing the model's ability to finely represent the fuel cell degradation process.
[0056] The multi-branch feature extraction layer includes multiple parallel LSTM task branches. Each LSTM task branch performs temporal dependency modeling and aging feature extraction on the corresponding health indicator data, and outputs the corresponding aging feature vector. By setting multiple LSTM task branches in parallel, the dynamic change patterns and long-term temporal dependencies of different health indicators during the degradation process can be explored separately. This is beneficial for fully extracting aging features from multi-source health data and improving the ability to model complex nonlinear decay behavior.
[0057] A shared fully connected layer is used to score the importance of the aging feature vectors output by each LSTM task branch. The Softmax function then converts the importance scores into normalized weights for each LSTM task branch. Finally, the aging feature vectors are weighted and aggregated according to these normalized weights to generate a multi-dimensional aging feature vector. By introducing importance scoring and normalized weighting mechanisms, the system adaptively highlights health indicator features that contribute significantly to lifespan decay while suppressing the adverse effects of low-contribution or redundant features on the prediction results. This improves the rationality and effectiveness of feature fusion, enhancing the stability and prediction accuracy of the model output.
[0058] The transition output layer is used to output the multi-dimensional aging feature vector to the heterogeneous feature extraction network. By setting the transition output layer, the multi-task LSTM network and the subsequent heterogeneous feature extraction network can be effectively connected. This provides a unified, compact, and highly discriminative input feature representation for further deep feature fusion and lifespan decline prediction, thereby helping to improve the collaborative modeling effect and engineering application performance of the entire fuel cell lifespan prediction model.
[0059] like Figure 3As shown, the heterogeneous feature extraction network is used to receive and process the multidimensional aging features output by the multi-task LSTM. The heterogeneous feature extraction network utilizes multiple attention heads in parallel to capture the long-term and short-term dependencies between different time steps in the input sequence. By calculating the scaled dot product attention between the query vector, key vector, and value vector, residual connections and LayerNorm are introduced in each sub-layer to prevent network degradation and accelerate convergence, ultimately outputting a hidden layer feature matrix containing rich aging information.
[0060] The heterogeneous feature extraction network comprises, in sequence, an input mapping layer, a position encoding layer, a multi-head self-attention feature extraction layer, a residual normalization layer, and a prediction output layer. The input mapping layer is used to receive the multidimensional aging features output by the multi-task LSTM network and perform dimensional mapping on the multidimensional aging features to generate a sequence of input features with uniform dimensions. The positional coding layer is used to introduce temporal positional information into the input feature sequence to represent the sequential relationship between features at different times. The multi-head self-attention feature extraction layer is used to perform multi-head parallel attention computation on the input feature sequence after introducing temporal location information, so as to extract global correlation features and coupled aging features between different health indicators and between different time points; The residual normalization layer is used to perform residual connection and normalization processing on the global correlation features and coupled aging features output by the multi-head self-attention feature extraction layer; The prediction output layer is used to output the fuel cell lifetime degradation prediction results.
[0061] In this embodiment, the input mapping layer receives the multidimensional aging features output by the multi-task LSTM network and performs dimensional mapping on these features to generate a unified dimensional input feature sequence. By setting the input mapping layer, the multidimensional aging features from the previous network can be uniformly represented and spatially aligned, thereby reducing the impact of differences in feature dimensions on subsequent modeling, improving the consistency and compatibility of heterogeneous feature fusion, and providing standardized input for subsequent global feature extraction.
[0062] The positional encoding layer is used to introduce temporal positional information into the input feature sequence to represent the sequential relationship between features at different times. By introducing positional encoding, the network can retain the temporal sequence information of fuel cell aging evolution during feature modeling, avoiding information loss caused by relying solely on feature values and ignoring temporal positional relationships, thereby enhancing the model's ability to represent the dynamic evolution of the lifespan decay process.
[0063] The multi-head self-attention feature extraction layer performs multi-head parallel attention computation on the input feature sequence after incorporating temporal location information to extract global correlation features and coupled aging features between different health indicators and at different time points. By setting up a multi-head self-attention mechanism, it is possible to mine the interaction relationships and cross-timescale dependencies between multidimensional health indicators in parallel from multiple subspaces, more comprehensively capturing the global, coupled, and nonlinear features in the fuel cell lifespan degradation process, thereby improving the modeling capability of complex degradation mechanisms.
[0064] The residual normalization layer is used to perform residual connections and normalization on the globally correlated features and coupled aging features output by the multi-head self-attention feature extraction layer. Residual connections effectively alleviate the gradient vanishing or feature degradation problems during deep network training, preserving the transmission of original and effective information. Normalization improves the stability and convergence speed of network training, enhances the consistency of feature distribution, and thus improves the robustness and generalization ability of the overall prediction model.
[0065] The prediction output layer is used to output the fuel cell life degradation prediction results. By predicting and outputting features after heterogeneous feature extraction and enhancement, the fuel cell life degradation trend can be more accurately characterized, improving the accuracy and stability of life prediction results and providing a reliable basis for fuel cell life assessment, health management, and maintenance decisions.
[0066] By setting up an input mapping layer, a position encoding layer, a multi-head self-attention feature extraction layer, a residual normalization layer, and a prediction output layer, it is possible to achieve unified mapping, temporal position enhancement, global correlation mining, and stabilization processing of multi-dimensional aging features, thereby effectively improving the heterogeneous feature fusion effect and the accuracy, stability, and engineering application value of fuel cell life decay prediction.
[0067] Step S4: Globally optimize the fuel cell heterogeneous prediction model using the particle swarm optimization algorithm, and input the multidimensional health index data to be predicted into the optimized fuel cell heterogeneous prediction model to obtain the fuel cell life degradation prediction results.
[0068] In this step, the particle swarm optimization algorithm performs secondary optimization on the heterogeneous feature extraction network in the fuel cell heterogeneous prediction model based on the aging characteristics of the multi-task LSTM network in the fixed fuel cell heterogeneous prediction model, so as to balance the prediction accuracy of different aging indicators in the fuel cell heterogeneous prediction model.
[0069] In one example, within the IEEE PHM 2014 dataset, the static and dynamic operating condition datasets are designated FC1 and FC2, respectively. This study selects data from the FC1 dataset (0h, 48h, 185h, 348h, 515h, 658h, 823h, 991h) and the FC2 dataset (0h, 35h, 182h, 343h, 515h, 666h, 830h, 1016h) for processing. Gaussian filtering is employed to reduce interference from outlier data.
[0070] The linear correlation of each input parameter is verified by using Pearson correlation coefficient, and the nonlinear correlation between health indicators and target variables is evaluated by using MIC. All qualified indicator values are used as candidate health indicator inputs. A multi-task collaborative prediction model is introduced, and multi-objective optimization is used to calculate the optimal input dimension and the best combination of health indicators.
[0071] Multiple health indicators were input as parameters, and network hyperparameters were initialized, including learning rate, number of input and output nodes, batch size, number of hidden units, and maximum number of training epochs. MSE was set as the loss function. After secondary optimization in the heterogeneous feature extraction network, the hyperparameters of the upper LSTM network, including learning rate and number of hidden units, were iteratively adjusted using particle swarm optimization to find the optimal hyperparameter configuration that minimized the loss function. Simultaneously, the decay feature parameters of the upper prediction network were abstracted using fully connected layers and then input into the lower prediction network. Mean squared error (MSE), mean absolute error (MAE), and R² coefficient were used to measure the accuracy of decay prediction. Comparison of accuracy with different prediction methods demonstrated that using the characteristics of multidimensional input health indicators of fuel cells as decay prediction parameters is feasible and has high accuracy.
[0072] This embodiment is implemented using Python on the PyCharm simulation platform. The parameters related to the fuel cell during the simulation are shown in Table 1.
[0073] Table 1 The method was validated using 300 hours of short-term training data from the FC1 and FC2 operating condition datasets. The experimental results are shown in Table 1. Quantitative indicators show that the proposed method achieves high-precision RUL prediction with a small sample size: the MAE for FC1 is only 0.0044, and the RMSE is 0.0055; the MAE for FC2 is 0.0052, and the RMSE is 0.0076. This indicates that the RUL error during short-term training is around 0.5% (measured with normalized data), demonstrating extremely high prediction accuracy. Furthermore, the R² coefficient, representing the goodness of fit, shows that the R² values for FC1 and FC2 reached 0.9471 and 0.9141, respectively, indicating a significant improvement over the comparative algorithms and demonstrating the ability to effectively capture the aging trend of fuel cells.
[0074] In this embodiment, by preprocessing the multidimensional health index data during fuel cell operation, noise interference can be effectively reduced, the impact of outliers can be weakened, and the consistency and integrity of time series data can be enhanced, thus providing a reliable data foundation for subsequent lifespan prediction. Correlation verification of each health index and selection of candidate health indices that meet the conditions can effectively eliminate indices with low correlation to lifespan decay and high redundancy, reduce invalid feature inputs, and improve feature representation capabilities. Furthermore, by obtaining the optimal input dimension and optimal combination of health indices based on candidate health indices and multi-objective optimization functions, a balance can be achieved between prediction accuracy, model complexity, and computational efficiency, avoiding overfitting or prediction performance degradation caused by excessive dimensionality or improper feature selection. At the same time, by constructing a heterogeneous prediction model for fuel cells that includes a multi-task LSTM network and a heterogeneous feature extraction network, the temporal evolution relationship between multidimensional health indices and the coupling relationship between heterogeneous features can be simultaneously explored, more fully representing the nonlinearity, dynamics, and multi-source correlation in the fuel cell lifespan decay process, thereby improving the accuracy and stability of fuel cell lifespan prediction.
[0075] Based on the above method, this application discloses a fuel cell lifetime prediction system based on multi-source feature fusion, referring to... Figure 4 The fuel cell life prediction system 1 includes a data acquisition module 11, a data processing module 12, and a life prediction module 13, wherein... The data acquisition module 11 is used to collect multidimensional health indicator data during the operation of the fuel cell and to perform data preprocessing on the multidimensional health indicator data to obtain multidimensional health time series data. The data processing module 12 is used to verify the correlation of health indicators in the multidimensional health time series data, and to determine the health indicator data that meet the correlation indicator conditions as candidate health indicators. Based on each candidate health indicator and the multi-objective optimization function, the optimal input dimension and the optimal combination of health indicators are obtained. Based on the optimal input dimension and the optimal combination of health indicators, a fuel cell heterogeneous prediction model is constructed. The fuel cell heterogeneous prediction model includes a multi-task LSTM network and a heterogeneous feature extraction network. The life prediction module 13 is used to perform global optimization of the fuel cell heterogeneous prediction model according to the particle swarm optimization algorithm, and input the multi-dimensional health index data to be predicted into the optimized fuel cell heterogeneous prediction model to obtain the fuel cell life degradation prediction results.
[0076] In one example, the data acquisition module 11 performs Gaussian filtering on each health indicator in the multidimensional health indicator data to obtain standard health indicator data after removing outliers; compensates the standard health indicator data based on cubic interpolation to obtain a multidimensional time series dataset; and normalizes the health indicator data corresponding to each dimension in the multidimensional time series dataset to obtain multidimensional health time series data.
[0077] In one example, the data processing module 12 is used to calculate the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient between each health indicator data and the voltage change in the multidimensional health time series data, using the voltage change of the fuel cell as a benchmark; if the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is greater than the preset correlation index value, the health indicator data corresponding to the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is determined as a candidate health indicator.
[0078] In one example, the data processing module 12 is used to construct a multi-objective optimization function by taking the prediction target accuracy function, the optimal permutation and combination function of health indicators, and the model computational complexity function as optimization content; and to optimize and solve each candidate health indicator combination under different input dimensions to obtain the optimal input dimension and the optimal health indicator combination corresponding to the optimal input dimension.
[0079] In one example, a multi-task LSTM network includes a multi-index assignment layer, a multi-branch feature extraction layer, a shared fully connected layer, and a transition output layer connected in sequence. The multi-indicator allocation layer is used to receive multi-dimensional health time-series data determined by the optimal combination of health indicators, and allocate each health indicator data in the multi-dimensional health time-series data to the corresponding task branch according to the health indicator category. The multi-branch feature extraction layer includes multiple parallel LSTM task branches. Each LSTM task branch is used to model the temporal dependency relationship and extract aging features from the corresponding health indicator data, and outputs the corresponding aging feature vector. The shared fully connected layer is used to score the importance of the aging feature vectors output by each LSTM task branch, and the Softmax function is used to convert the importance score results into normalized weights corresponding to each LSTM task branch. Based on the normalized weights corresponding to each LSTM task branch, the aging feature vectors are weighted and aggregated to generate a multi-dimensional aging feature vector. The transition output layer is used to output the multidimensional aging feature vector to the heterogeneous feature extraction network.
[0080] In one example, the heterogeneous feature extraction network comprises, in sequence, an input mapping layer, a position encoding layer, a multi-head self-attention feature extraction layer, a residual normalization layer, and a prediction output layer, wherein, The input mapping layer is used to receive the multidimensional aging features output by the multi-task LSTM network and perform dimensional mapping on the multidimensional aging features to generate a sequence of input features with uniform dimensions. The positional coding layer is used to introduce temporal positional information into the input feature sequence to represent the sequential relationship between features at different times. The multi-head self-attention feature extraction layer is used to perform multi-head parallel attention computation on the input feature sequence after introducing temporal location information, so as to extract global correlation features and coupled aging features between different health indicators and between different time points; The residual normalization layer is used to perform residual connection and normalization processing on the global correlation features and coupled aging features output by the multi-head self-attention feature extraction layer; The prediction output layer is used to output the fuel cell lifetime degradation prediction results.
[0081] In one example, the particle swarm optimization algorithm performs secondary optimization on the heterogeneous feature extraction network in the fuel cell heterogeneous prediction model based on the aging characteristics of the multi-task LSTM network in the fixed fuel cell heterogeneous prediction model, in order to balance the prediction accuracy of different aging indices in the fuel cell heterogeneous prediction model.
[0082] Please see Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 2 may include: at least one processor 21, at least one network interface 24, user interface 23, memory 25, and at least one communication bus 22.
[0083] The communication bus 22 is used to enable communication between these components.
[0084] The user interface 23 may include a display screen and a camera. Optionally, the user interface 23 may also include a standard wired interface and a wireless interface.
[0085] The network interface 24 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).
[0086] The processor 21 may include one or more processing cores. The processor 21 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 25, and by calling data stored in the memory 25. Optionally, the processor 21 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 21.
[0087] The memory 25 may include random access memory (RAM) or read-only memory. Optionally, the memory 25 may include non-transitory computer-readable storage medium. The memory 25 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 25 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 25 may also be at least one storage device located remotely from the aforementioned processor 21. Figure 5As shown, the memory 25, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for predicting fuel cell lifetime based on multi-source feature fusion.
[0088] exist Figure 5 In the electronic device 2 shown, the user interface 23 is mainly used to provide an input interface for the user and obtain the user input data; while the processor 21 can be used to call an application program stored in the memory 25 that is a fuel cell lifetime prediction method based on multi-source feature fusion. When executed by one or more processors, the electronic device executes one or more methods as described in the above embodiments.
[0089] A non-transitory computer-readable storage medium stores instructions that, when executed by one or more processors, cause a computer to perform one or more methods as described in the above embodiments.
[0090] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
[0091] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0092] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.
[0093] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0094] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0095] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.
[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A fuel cell lifetime prediction method based on multi-source feature fusion, characterized in that, The method includes: Collect multidimensional health indicator data during the operation of the fuel cell, and perform data preprocessing on the multidimensional health indicator data to obtain multidimensional health time series data; The correlation of health indicators in the multidimensional health time series data is verified, and the health indicator data that meet the correlation indicator conditions are determined as candidate health indicators. Based on each candidate health indicator and a multi-objective optimization function, the optimal input dimension and the optimal combination of health indicators are obtained. Based on the optimal input dimension and the optimal combination of health indicators, a heterogeneous prediction model for fuel cells is constructed. The heterogeneous prediction model for fuel cells includes a multi-task LSTM network and a heterogeneous feature extraction network. The fuel cell heterogeneous prediction model is globally optimized using the particle swarm optimization algorithm. The multidimensional health index data to be predicted is input into the optimized fuel cell heterogeneous prediction model to obtain the fuel cell life degradation prediction results.
2. The fuel cell lifetime prediction method based on multi-source feature fusion as described in claim 1, characterized in that, The process of preprocessing the multidimensional health indicator data to obtain multidimensional health time-series data specifically includes: Gaussian filtering is applied to each health indicator in the multidimensional health indicator data to obtain standard health indicator data after removing outliers. The standard health indicator data are compensated using cubic interpolation to obtain a multidimensional time series dataset. The health indicator data corresponding to each dimension in the multidimensional time series dataset are normalized to obtain the multidimensional health time series data.
3. The fuel cell lifetime prediction method based on multi-source feature fusion as described in claim 1, characterized in that, The step of verifying the correlation between health indicators in the multidimensional health time-series data and determining the health indicator data that meets the correlation indicator conditions as candidate health indicators specifically includes: Using the voltage change of the fuel cell as a benchmark, the linear Pearson correlation coefficient and the nonlinear maximum mutual information coefficient between each health indicator data in the multidimensional health time series data and the voltage change are calculated respectively. If the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is greater than a preset correlation index value, the health index data corresponding to the linear Pearson correlation coefficient or the nonlinear maximum mutual information coefficient is determined as a candidate health index.
4. The fuel cell lifetime prediction method based on multi-source feature fusion as described in claim 1, characterized in that, The process of obtaining the optimal input dimension and optimal combination of health indicators based on each candidate health indicator and a multi-objective optimization function specifically includes: A multi-objective optimization function is constructed by taking the prediction target accuracy function, the optimal permutation and combination function of health indicators, and the model computational complexity function as optimization content. The optimal input dimension and the optimal health indicator combination corresponding to the optimal input dimension are obtained by optimizing each candidate health indicator combination under different input dimensions.
5. The fuel cell lifetime prediction method based on multi-source feature fusion as described in claim 1, characterized in that, The multi-task LSTM network comprises a multi-index allocation layer, a multi-branch feature extraction layer, a shared fully connected layer, and a transition output layer connected in sequence. The multi-indicator allocation layer is used to receive multi-dimensional health time-series data determined by the optimal combination of health indicators, and allocate each health indicator data in the multi-dimensional health time-series data to the corresponding task branch according to the health indicator category. The multi-branch feature extraction layer includes multiple parallel LSTM task branches. Each LSTM task branch is used to perform time-series dependency modeling and aging feature extraction on the corresponding health indicator data, and outputs the corresponding aging feature vector. The shared fully connected layer is used to score the importance of the aging feature vectors output by each LSTM task branch, and converts the importance score results into normalized weights corresponding to each LSTM task branch through the Softmax function. Based on the normalized weights corresponding to each LSTM task branch, the aging feature vectors are weighted and aggregated to generate a multidimensional aging feature vector. The transition output layer is used to output the multidimensional aging feature vector to the heterogeneous feature extraction network.
6. The fuel cell lifetime prediction method based on multi-source feature fusion as described in claim 1, characterized in that, The heterogeneous feature extraction network comprises, in sequence, an input mapping layer, a position encoding layer, a multi-head self-attention feature extraction layer, a residual normalization layer, and a prediction output layer, wherein, The input mapping layer is used to receive the multidimensional aging features output by the multi-task LSTM network and perform dimensional mapping on the multidimensional aging features to generate a unified dimension input feature sequence. The position encoding layer is used to introduce temporal position information into the input feature sequence to characterize the sequential relationship between features at each time point; The multi-head self-attention feature extraction layer is used to perform multi-head parallel attention calculation on the input feature sequence after introducing temporal location information, so as to extract global correlation features and coupled aging features between different health indicators and between different time periods; The residual normalization layer is used to perform residual connection and normalization processing on the global correlation features and the coupled aging features output by the multi-head self-attention feature extraction layer; The prediction output layer is used to output the fuel cell lifespan degradation prediction results.
7. The fuel cell lifetime prediction method based on multi-source feature fusion as described in claim 1, characterized in that, The particle swarm optimization algorithm performs secondary optimization on the heterogeneous feature extraction network in the fuel cell heterogeneous prediction model based on the aging characteristics of the multi-task LSTM network in the fixed fuel cell heterogeneous prediction model, so as to balance the prediction accuracy of different aging indicators in the fuel cell heterogeneous prediction model.
8. A fuel cell lifetime prediction system based on multi-source feature fusion, characterized in that, The fuel cell life prediction system (1) includes a data acquisition module (11), a data processing module (12), and a life prediction module (13), wherein, The data acquisition module (11) is used to collect multidimensional health indicator data during the operation of the fuel cell and to perform data preprocessing on the multidimensional health indicator data to obtain multidimensional health time series data. The data processing module (12) is used to verify the correlation of health indicators in the multidimensional health time series data, and to determine the health indicator data that meets the correlation indicator conditions as candidate health indicators. Based on each candidate health indicator and the multi-objective optimization function, the optimal input dimension and the optimal combination of health indicators are obtained, and a fuel cell heterogeneous prediction model is constructed according to the optimal input dimension and the optimal combination of health indicators. The fuel cell heterogeneous prediction model includes a multi-task LSTM network and a heterogeneous feature extraction network. The life prediction module (13) is used to perform global optimization of the fuel cell heterogeneous prediction model according to the particle swarm optimization algorithm, and input the multidimensional health index data to be predicted into the optimized fuel cell heterogeneous prediction model to obtain the fuel cell life decay prediction result.
9. An electronic device, characterized in that, The device includes a processor (21), a memory (25), a user interface (23), and a network interface (24). The memory (25) is used to store instructions. The user interface (23) and the network interface (24) are used to communicate with other devices. The processor (21) is used to execute the instructions stored in the memory (25) to cause the electronic device (2) to perform the method as described in any one of claims 1-7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.