Method for predicting life of fuel cell based on long short-term memory network

Abstract

The application relates to the field of battery life prediction, in particular to a multi-feature fuel cell life prediction method based on a long short-term memory network; the method comprises the following steps: collecting voltage attenuation data and collecting running state data to construct an original input data set; analyzing the running state data, calculating a correlation coefficient according to the voltage attenuation data and each running state data, analyzing the correlation coefficient, and obtaining an input feature set; processing the input feature set and the voltage attenuation data to construct a neural network training data structure; dividing the training data structure into a training set and a test set; constructing a network architecture solution space to be searched, analyzing the network architecture solution space, constructing a fitness function, and determining an optimal network architecture parameter combination based on the fitness function; and inputting the training set into the optimal network architecture parameter combination for training to generate a target detection prediction result. The application can improve the prediction accuracy of the voltage attenuation trend.

Description

Technical Field

[0001] This invention relates to the field of battery life prediction, and more specifically to a multi-feature fuel cell life prediction method based on long short-term memory networks. Background Technology

[0002] Fuel cells, as efficient and clean energy conversion devices, have been widely used in new energy vehicles, distributed power generation, and backup power supplies. Their lifespan prediction is crucial for system safety, operation and maintenance planning, and lifecycle cost control. With the development of artificial intelligence, more and more research is attempting to use deep learning models for data-driven modeling and prediction of fuel cell lifespan. In particular, Long Short-Term Memory (LSTM) networks, due to their advantages in time series modeling, are widely used in the task of modeling fuel cell voltage decay curves.

[0003] In existing technologies, the feature selection process relies on empirical rules and lacks a data-driven key feature recognition mechanism; the structural parameters of LSTM models are optimized by manual setting or grid search, resulting in low search efficiency and an inability to achieve joint adaptive adjustment of structural levels, number of neurons, training hyperparameters, etc.; a closed-loop network architecture search-training-validation mechanism has not been introduced, making it impossible to dynamically backtrack prediction performance and adjust the structure; and an effective fitness function has not been constructed to address the dual constraints of prediction error and curve structure consistency, resulting in a small numerical error but a large deviation in lifetime trend, which limits its engineering applications. These are the problems we need to solve. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the background technology by proposing a multi-feature fuel cell lifetime prediction method based on long short-term memory networks.

[0005] The technical solution of this invention: a multi-feature fuel cell lifetime prediction method based on long short-term memory networks, comprising the following steps:

[0006] S1. Collect voltage decay data of the target detection battery, and simultaneously collect the operating status data of the battery corresponding to each voltage decay data during the image acquisition process. Synchronously align the voltage decay data and the operating status data to construct the original input dataset.

[0007] S2. Identify and process outliers in the operating status data of the original input dataset, perform correlation analysis between voltage decay data and each operating status data, calculate the correlation coefficient between each voltage decay data and the operating status data, and analyze the correlation coefficient to obtain the input feature set.

[0008] S3. Normalize the input feature set and the corresponding voltage decay data, and construct a neural network training data structure based on the normalized voltage decay data and the input feature set; divide the training data structure into a training set and a test set according to a preset ratio.

[0009] S4. Construct a solution space for the network architecture to be searched based on a long short-term memory neural network, analyze the solution space, construct a fitness function, and determine the optimal combination of network architecture parameters based on the fitness function; input the training set into the optimal combination of network architecture parameters for training, and generate object detection prediction results.

[0010] Preferably, the process of collecting voltage decay data of the target detection battery and simultaneously collecting the battery's operating status data corresponding to each voltage decay data point during image acquisition, synchronizing and aligning the voltage decay data and operating status data to construct the original input dataset includes:

[0011] The system collects voltage decay data of the target detection battery during continuous operation. The voltage decay data includes the output voltage value of a single cell, the terminal voltage value of the battery pack, and the corresponding sampling timestamp. While collecting the voltage decay data, the system also collects the operating status data of the fuel cell corresponding to each sampling timestamp. The operating status data includes the hydrogen inlet flow rate, hydrogen outlet flow rate, air inlet flow rate, air outlet flow rate, cooling water flow rate, and cooling water temperature.

[0012] Based on the sampling timestamp, the voltage decay data and the operating status data are time-aligned. Data with inconsistent sampling frequencies are synchronized and corrected using time window matching or interpolation. Data segments that cannot establish a correspondence are removed. The time-aligned voltage decay data and the corresponding operating status data are combined according to the sampling timestamp correspondence to generate the original input dataset indexed by the time series.

[0013] Preferably, the process of identifying and processing outliers in the operating status data of the original input dataset, performing correlation analysis between voltage decay data and each operating status data, calculating the correlation coefficient between each voltage decay data and the operating status data, and analyzing the correlation coefficient to obtain the input feature set includes:

[0014] Anomalies are identified in the running status data corresponding to each sampling timestamp in the original input dataset. Anomalies include missing values ​​in the running status data, abrupt changes in values ​​that exceed the preset change range between adjacent sampling times, and data points that are inconsistent with the sampling timestamps of the voltage decay data. The identified anomalies are processed by deletion, interpolation correction, or smoothing based on time windows according to their corresponding types.

[0015] After anomaly handling is completed, voltage decay data is used as the first variable and each operating state data is used as the second variable. Correlation analysis is performed on the two under a unified sampling time series to calculate the correlation coefficient between voltage decay data and each operating state data. Threshold analysis is performed on the correlation coefficients corresponding to each operating state data, and operating state data with correlation coefficients not lower than the preset threshold are selected as input features to generate an input feature set.

[0016] Preferably, the input feature set and the corresponding voltage decay data are normalized, and a neural network training data structure is constructed based on the normalized voltage decay data and the input feature set; the process of dividing the training data structure into a training set and a test set according to a preset ratio is as follows:

[0017] For each dimension of the operating state data and the voltage decay data of the corresponding sampling timestamp in the input feature set, normalization processing is performed within the sequence of complete data. The normalization method adopts a linear interval mapping approach.

[0018] After normalization, the normalized input feature vectors are paired with the corresponding normalized voltage decay data one by one according to the sampling timestamps to construct data pairs and integrate them into the training data structure required for neural network input. The training data structure includes input tensors and target tensors. The training data structure is sorted according to the time series and the data is allocated according to the set ratio of training set to test set.

[0019] Preferably, the process of constructing the solution space of the network architecture to be searched based on a long short-term memory neural network includes:

[0020] A solution space for the network architecture to be searched is constructed based on a Long Short-Term Memory (LSTM) neural network. The solution space consists of multiple structural control parameters, including the number of hidden layer neurons, the number of feature channels, the network hierarchical connections, the dropout probability, and the number of training batches. The number of hidden layer neurons controls the network's deep feature representation capability; the number of feature channels defines the number of multi-dimensional channels in the input and output of each layer; the network hierarchical connections define the forward propagation structure in the LSTM network; the dropout probability prevents overfitting; and the number of training batches controls the sample batch size for each gradient update.

[0021] Preferably, the process of analyzing the network architecture solution space and constructing the fitness function includes:

[0022] Based on the artificial hummingbird algorithm, a structure search optimization of the network architecture solution space is performed. A three-stage operation—initial population generation, fitness calculation, and iterative structure update—is employed to construct an adaptive optimization mechanism for neural network architectures under the dual constraints of prediction error and structural consistency. In the initial stage, multiple initial network architecture candidates are generated according to the defined range of structural control parameters, forming a network architecture population. Each candidate is traversed within the population. The training set is input into the network model for training, and the test set is input into the trained model to generate prediction results. Prediction error and structural consistency indices are calculated. The prediction error indices include root mean square error (RMSE) and mean absolute error (MAE), while the structural consistency indices include the coefficient of determination (COD). The two indices are normalized and weighted to form a fitness function. The fitness function is a weighted sum of the error indices between the predicted results and the actual voltage decay data, and the curve shape similarity indices. The error indices include RMSE, MAE, or other regression error indices, while the curve shape similarity indices include the COD.

[0023] Preferably, the process of determining the optimal combination of network architecture parameters based on the fitness function includes:

[0024] During the structure search phase, three foraging operations are performed sequentially based on the artificial hummingbird algorithm, and the candidate structure solutions are updated. These operations include guided foraging, regional foraging, and migratory foraging. All three foraging operations incorporate three types of hummingbird flight skills: axial flight, diagonal flight, and omnidirectional flight.

[0025] When the fitness function reaches the set convergence threshold, the optimal combination of structural control parameters is selected as the optimal network architecture. The target detection neural network model is reconstructed based on the optimal network architecture parameter combination, and the training set data is input into the model for full rounds of training to obtain the optimal network architecture parameter combination.

[0026] Preferably, the process of training the training set using the optimal combination of network architecture parameters to generate object detection prediction results includes:

[0027] The training set is input into the trained model to generate target detection prediction results, and structural evaluation indicators are calculated based on the prediction results. The structural evaluation indicators include the root mean square error, mean absolute error, and structural prediction consistency judgment index of the test set. When the structural evaluation indicators meet the preset performance conditions, the current target detection model is output as the fuel cell lifetime prediction model. If the performance conditions are not met, the process of searching and optimizing the network architecture solution space is returned, and the structural search and training are restarted.

[0028] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: By performing correlation analysis based on fuel cell operating status data and voltage decay data, and dynamically selecting the input feature set, the objectivity and adaptability of the feature selection process are achieved, avoiding the modeling bias problem caused by relying on human experience to select input features, and improving the life prediction model's ability to represent the actual operating state; by constructing an LSTM network architecture solution space containing network structure control parameters, and using the artificial hummingbird algorithm for closed-loop structure search and performance verification, adaptive optimization of the network structure parameter combination is achieved, ensuring that the constructed life prediction model has excellent performance in error control and trend matching, and improving the model's fitting ability and generalization ability for complex voltage decay processes. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of one embodiment of the present invention. Detailed Implementation

[0030] like Figure 1 As shown, the multi-feature fuel cell lifetime prediction method based on long short-term memory networks proposed in this invention includes the following steps:

[0031] S1. Collect voltage decay data of the target detection battery, and simultaneously collect the operating status data of the battery corresponding to each voltage decay data during the image acquisition process. Synchronously align the voltage decay data and the operating status data to construct the original input dataset.

[0032] S2. Identify and process outliers in the operating status data of the original input dataset, perform correlation analysis between voltage decay data and each operating status data, calculate the correlation coefficient between each voltage decay data and the operating status data, and analyze the correlation coefficient to obtain the input feature set.

[0033] S3. Normalize the input feature set and the corresponding voltage decay data, and construct a neural network training data structure based on the normalized voltage decay data and the input feature set; divide the training data structure into a training set and a test set according to a preset ratio.

[0034] S4. Construct a solution space for the network architecture to be searched based on a long short-term memory neural network, analyze the solution space, construct a fitness function, and determine the optimal combination of network architecture parameters based on the fitness function; input the training set into the optimal combination of network architecture parameters for training, and generate object detection prediction results.

[0035] It should be further explained that, in the specific implementation process, the voltage decay data of the target detection battery is collected, and the operating status data of the battery corresponding to each voltage decay data is collected synchronously during image acquisition. The process of synchronizing and aligning the voltage decay data and operating status data to construct the original input dataset is as follows:

[0036] The voltage decay data of the target detection battery during continuous operation is collected. The voltage decay data is used to represent the voltage change trend of the battery during the operation cycle, including the output voltage value of the individual battery, the voltage value of the battery pack terminal, and the corresponding sampling timestamp recorded according to a unified sampling time series.

[0037] While collecting the voltage decay data, fuel cell operating status data corresponding to each sampling timestamp is collected simultaneously. The operating status data includes hydrogen side inlet flow rate, hydrogen side outlet flow rate, air side inlet flow rate, air side outlet flow rate, cooling water flow rate, and cooling water temperature.

[0038] Based on the sampling timestamp, the voltage attenuation data and the operating status data are time-aligned. Data with inconsistent sampling frequencies are synchronized and corrected by time window matching or interpolation. Data segments that cannot establish a corresponding relationship are removed.

[0039] The time-aligned voltage decay data and the corresponding operating status data are combined according to the sampling timestamp correspondence to generate a raw input dataset indexed by time series. This raw input dataset serves as the data input basis for subsequent feature selection and neural network model training.

[0040] It should be further explained that, in the specific implementation process, the process of identifying and processing outliers in the operating status data of the original input dataset, performing correlation analysis between voltage attenuation data and each operating status data, calculating the correlation coefficient between each voltage attenuation data and the operating status data, and analyzing the correlation coefficient to obtain the input feature set is as follows:

[0041] Anomalies are identified in the running status data corresponding to each sampling timestamp in the original input dataset. The anomalies include missing values ​​in the running status data, abrupt changes in values ​​that exceed a preset change range between adjacent sampling times, and data points that are inconsistent with the sampling timestamps of the voltage decay data.

[0042] The identified anomalies are processed according to their corresponding types using deletion, interpolation correction, or time window-based smoothing methods to ensure the continuity and consistency of the operational status data over time.

[0043] After the anomaly handling is completed, voltage decay data is used as the first variable and each operating state data is used as the second variable. Correlation analysis is performed on the two under a unified sampling time series to calculate the correlation coefficient r between voltage decay data and each operating state data. The correlation coefficient is used to characterize the statistical correlation between voltage decay data and corresponding operating state data.

[0044] ;

[0045] Where r represents the correlation coefficient. This represents the covariance of variables X and Y. and The variances of variables X and Y are represented; the correlation analysis can also be replaced by Spearman correlation analysis or other statistical methods that can quantify the correlation between voltage and characteristic parameters;

[0046] Threshold analysis is performed on the correlation coefficients of each operating status data, and operating status data with correlation coefficients not lower than the preset threshold are selected as input features to generate an input feature set.

[0047] It should be further explained that, in the specific implementation process, the input feature set and the corresponding voltage decay data are normalized, and a neural network training data structure is constructed based on the normalized voltage decay data and the input feature set; the process of dividing the training data structure into a training set and a test set according to a preset ratio is as follows:

[0048] For each dimension of the operating state data and the corresponding voltage decay data at the sampling timestamp in the input feature set, normalization is performed within the sequence of the complete data. The normalization method adopts a linear interval mapping approach and is calculated according to the following formula:

[0049] ;

[0050] in, Represents the original data. This represents the minimum value of the data. This represents the maximum value of the data. The data represents normalized data; the normalization process is applicable to all operating status data and the corresponding voltage value sequence, and ensures that all values ​​are uniformly mapped to the range of [0,1]; the normalization method can also be replaced by other transformation methods that maintain the proportional relationship of input variables, including standard deviation normalization, Z-score normalization or unit interval scaling. The alternative method must ensure that the normalized data is comparable across different dimensions;

[0051] After normalization, the normalized input feature vectors are paired with the corresponding normalized voltage decay data in chronological order of sampling timestamps to construct data pairs. These pairs are then integrated into the training data structure required for neural network input. The training data structure includes an input tensor and a target tensor. The input tensor consists of a sequence of running state data, and the target tensor consists of a sequence of voltage decay data.

[0052] The training data structure is sorted according to the time series, and the data is allocated according to the set ratio of training set to test set. The ratio is 6:4, or it can be adjusted within the range of 5:1 to 1:1 according to the total amount of data and the complexity of the task. The partitioning process must ensure that the data of the test set does not appear in the training set, and the data allocation maintains the temporal continuity and distribution consistency of the original data, in order to support the evaluation of the generalization ability of the model in the future.

[0053] It should be further explained that, in the specific implementation process, the solution space of the network architecture to be searched is constructed based on the Long Short-Term Memory Neural Network, the solution space is analyzed, a fitness function is constructed, and the optimal combination of network architecture parameters is determined based on the fitness function; the process of inputting the training set into the optimal combination of network architecture parameters for training to generate object detection prediction results is as follows:

[0054] A solution space for the network architecture to be searched is constructed based on a Long Short-Term Memory (LSTM) neural network. This solution space consists of multiple structural control parameters, including at least: the number of hidden layer neurons, the number of feature channels, the network hierarchical connectivity, the dropout probability, and the number of training batches. The number of hidden layer neurons controls the network's deep feature representation capability; the number of feature channels defines the number of multi-dimensional channels in the input and output of each layer; the network hierarchical connectivity sets the forward propagation structure in the LSTM network; the dropout probability prevents overfitting; and the number of training batches controls the sample batch size for each gradient update.

[0055] Based on the artificial hummingbird algorithm, a structure search optimization of the network architecture solution space is performed. A three-stage operation—initial population generation, fitness calculation, and iterative structure update—is employed to construct an adaptive optimization mechanism for neural network architecture under the dual constraints of prediction error and structural consistency. Specifically: In the initial stage, multiple initial network architecture candidates are generated according to the defined range of structural control parameters, forming a network architecture population. Each candidate is traversed, and the training set is input into the network model for training. The test set is input into the trained model to generate prediction results, and prediction error and structural consistency indices are calculated. The prediction error indices include root mean square error (RMSE) and mean absolute error (MAE), and the structural consistency indices include the coefficient of determination (R²). These two indices are normalized and weighted to form a fitness function used to evaluate network architecture performance. The fitness function is weighted by the error index between the prediction results and the actual voltage decay data, and the curve shape similarity index. The error index includes root mean square error (RMSE), mean absolute error (MAE), or other regression error indices, and the curve shape similarity index includes the coefficient of determination (R²). The relevant calculation formulas are as follows:

[0056] Root mean square error: ;

[0057] Absolute average error: ;

[0058] Coefficient of determination: ;

[0059] in, This represents the number of voltage samples. Represents actual voltage decay data. This represents the predicted voltage decay value.

[0060] During the structure search phase, three foraging operations are performed sequentially based on the artificial hummingbird algorithm, and the candidate structure solutions are updated. These operations include guided foraging, regional foraging, and migratory foraging. All three foraging operations incorporate three types of hummingbird flight skills: axial flight, diagonal flight, and omnidirectional flight.

[0061] Fitness value of food source:

[0062] ;

[0063] Axial flight, diagonal flight, and omnidirectional flight skills:

[0064] ;

[0065] ;

[0066] ;

[0067] Guiding foraging strategies:

[0068] ;

[0069] ;

[0070] Regional foraging strategies: ;

[0071] Migration and foraging strategies: ;

[0072] in, Represents fitness value; Represents the number of voltage samples; Represents true electrical attenuation data; This represents the predicted voltage attenuation value; Represents flying skills; This indicates the generation of random integers from 1 to d; This indicates the creation of a random permutation of integers from 1 to k; d represents a random number uniformly distributed between [0, 1]; d represents the dimension of the solution space. , , These represent the number of neurons, the dropout probability, and the training batch, respectively. Indicates the first In the next iteration, the position of the i-th candidate food source; The position of the i-th candidate food source in the t-th iteration; This indicates the location of the target food source that the i-th hummingbird will visit; and Indicates a guiding factor that follows a normal distribution; Indicates the location of the food source with the worst nectar replenishment rate in the population; and These represent the upper and lower bounds of the search space, respectively.

[0073] When the fitness function reaches the set convergence threshold, the optimal combination of structural control parameters with the best fitness is selected as the optimal network architecture.

[0074] The target detection neural network model is reconstructed based on the optimal combination of network architecture parameters. The training set data is input into the model for full rounds of training to obtain the optimal combination of network architecture parameters.

[0075] The training set is input into the trained model to generate object detection prediction results, and the structure evaluation index is calculated based on the prediction results.

[0076] The structural evaluation metrics include RMSE, MAE, and structural prediction consistency criteria of the test set. When the structural evaluation metrics meet the preset performance conditions, the current target detection model is output as the fuel cell lifetime prediction model; if the performance conditions are not met, the process of searching and optimizing the network architecture solution space is returned, and the structural search and training are restarted.

[0077] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A multi-feature fuel cell lifetime prediction method based on long short-term memory networks, characterized in that, Includes the following steps: S1. Collect voltage decay data of the target detection battery, and simultaneously collect the operating status data of the battery corresponding to each voltage decay data during the image acquisition process. Synchronously align the voltage decay data and the operating status data to construct the original input dataset. S2. Identify and process outliers in the operating status data of the original input dataset, perform correlation analysis between voltage decay data and each operating status data, calculate the correlation coefficient between each voltage decay data and the operating status data, and analyze the correlation coefficient to obtain the input feature set. S3. Normalize the input feature set and the corresponding voltage decay data, and construct a neural network training data structure based on the normalized voltage decay data and the input feature set; divide the training data structure into a training set and a test set according to a preset ratio. S4. Construct a solution space for the network architecture to be searched based on a long short-term memory neural network, analyze the solution space, construct a fitness function, and determine the optimal combination of network architecture parameters based on the fitness function; input the training set into the optimal combination of network architecture parameters for training, and generate object detection prediction results; The process of constructing a solution space for the network architecture to be searched based on a Long Short-Term Memory (LSTM) neural network, analyzing the solution space, and constructing the fitness function includes: A solution space for the network architecture to be searched is constructed based on a Long Short-Term Memory (LSTM) neural network. This solution space consists of multiple structural control parameters, including the number of hidden layer neurons, the number of feature channels, network hierarchical connections, the dropout probability, and the number of training batches. The number of hidden layer neurons controls the network's deep feature representation capability; the number of feature channels defines the number of multi-dimensional channels in the input and output of each layer; the network hierarchical connections define the forward propagation structure in the LSTM network; the dropout probability prevents overfitting; and the number of training batches controls the sample batch size for each gradient update. Based on the artificial hummingbird algorithm, the network architecture solution space is optimized by structure search. The three-stage operation of initial population generation, fitness calculation and structure iterative update is adopted to construct an adaptive optimization mechanism for neural network architecture under the dual constraints of prediction error and structural consistency. In the initial stage, multiple initial network architecture candidates are generated according to the defined range of structural control parameters to form a network architecture population. Each candidate network architecture is traversed within the network architecture population. The training set is input into the network model for training, and the test set is input into the trained model to generate prediction results. Prediction error metrics and structural consistency metrics are calculated. The prediction error metrics include root mean square error and mean absolute error, and the structural consistency metrics include the coefficient of determination. The two metrics are normalized and weighted to form a fitness function. The fitness function is composed of a weighted average of the error metrics between the prediction results and the actual voltage decay data and the curve shape similarity metrics. The error metrics include root mean square error, mean absolute error, or other regression error metrics, and the curve shape similarity metrics include the coefficient of determination.

2. The multi-feature fuel cell lifetime prediction method based on long short-term memory networks according to claim 1, characterized in that, The process of collecting voltage decay data of the target detection battery and simultaneously collecting the battery's operating status data corresponding to each voltage decay data point during image acquisition, synchronizing and aligning the voltage decay data and operating status data to construct the original input dataset includes: The system collects voltage decay data of the target detection battery during continuous operation. The voltage decay data includes the output voltage value of a single cell, the terminal voltage value of the battery pack, and the corresponding sampling timestamp. While collecting the voltage decay data, the system also collects the operating status data of the fuel cell corresponding to each sampling timestamp. The operating status data includes the hydrogen inlet flow rate, hydrogen outlet flow rate, air inlet flow rate, air outlet flow rate, cooling water flow rate, and cooling water temperature. Based on the sampling timestamp, the voltage decay data and the operating status data are time-aligned. Data with inconsistent sampling frequencies are synchronized and corrected using time window matching or interpolation. Data segments that cannot establish a correspondence are removed. The time-aligned voltage decay data and the corresponding operating status data are combined according to the sampling timestamp correspondence to generate the original input dataset indexed by the time series.

3. The multi-feature fuel cell lifetime prediction method based on long short-term memory networks according to claim 2, characterized in that, The process of identifying and processing outliers in the operating status data of the original input dataset, performing correlation analysis between voltage decay data and each operating status data, calculating the correlation coefficient between each voltage decay data and the operating status data, and analyzing the correlation coefficient to obtain the input feature set includes: Anomalies are identified in the running status data corresponding to each sampling timestamp in the original input dataset. Anomalies include missing values ​​in the running status data, abrupt changes in values ​​that exceed the preset change range between adjacent sampling times, and data points that are inconsistent with the sampling timestamps of the voltage decay data. The identified anomalies are processed by deletion, interpolation correction, or smoothing based on time windows according to their corresponding types. After anomaly handling is completed, voltage decay data is used as the first variable and each operating state data is used as the second variable. Correlation analysis is performed on the two under a unified sampling time series to calculate the correlation coefficient between voltage decay data and each operating state data. Threshold analysis is performed on the correlation coefficients corresponding to each operating state data, and operating state data with correlation coefficients not lower than the preset threshold are selected as input features to generate an input feature set.

4. The multi-feature fuel cell lifetime prediction method based on long short-term memory networks according to claim 3, characterized in that, The input feature set and the corresponding voltage decay data are normalized, and a neural network training data structure is constructed based on the normalized voltage decay data and the input feature set. The process of dividing the training data structure into a training set and a test set according to a preset ratio is as follows: For each dimension of the operating state data and the voltage decay data of the corresponding sampling timestamp in the input feature set, normalization processing is performed within the sequence of complete data. The normalization method adopts a linear interval mapping approach. After normalization, the normalized input feature vectors are paired with the corresponding normalized voltage decay data one by one according to the sampling timestamps to construct data pairs and integrate them into the training data structure required for neural network input. The training data structure includes input tensors and target tensors. The training data structure is sorted according to the time series and the data is allocated according to the set ratio of training set to test set.

5. The multi-feature fuel cell lifetime prediction method based on long short-term memory networks according to claim 1, characterized in that, The process of determining the optimal combination of network architecture parameters based on the fitness function includes: During the structure search phase, three foraging operations are performed sequentially based on the artificial hummingbird algorithm, and the candidate structure solutions are updated. These operations include guided foraging, regional foraging, and migratory foraging. All three foraging operations incorporate three types of hummingbird flight skills: axial flight, diagonal flight, and omnidirectional flight. When the fitness function reaches the set convergence threshold, the optimal combination of structural control parameters is selected as the optimal network architecture. The target detection neural network model is reconstructed based on the optimal network architecture parameter combination, and the training set data is input into the model for full rounds of training to obtain the optimal network architecture parameter combination.

6. The multi-feature fuel cell lifetime prediction method based on long short-term memory networks according to claim 5, characterized in that, The process of training the training set with the optimal combination of network architecture parameters to generate object detection prediction results includes: The training set is input into the trained model to generate target detection prediction results, and structural evaluation indicators are calculated based on the prediction results. The structural evaluation indicators include the root mean square error, mean absolute error, and structural prediction consistency judgment index of the test set. When the structural evaluation indicators meet the preset performance conditions, the current target detection model is output as the fuel cell lifetime prediction model. If the performance conditions are not met, the process of searching and optimizing the network architecture solution space is returned, and the structural search and training are restarted.

Images