Hydrogen fuel cell state monitoring method based on multi-stack series-parallel connection

Abstract

The invention relates to the field of hydrogen fuel cells, and discloses a hydrogen fuel cell state monitoring method based on multi-stack series-parallel connection, and the method comprises the following steps: 1, collecting actual operation data and simulation operation data of a hydrogen fuel cell; step 2, building a CNN-LSTM model-based feature extractor, and performing deep feature extraction on the database by the feature extractor; 3, building a multi-task neural network model; 4, sharing the deep feature matrix into a multi-task neural network model; 5, respectively calculating output indexes of all models in the multi-task neural network model; step 6, forming a comprehensive performance index by a plurality of output indexes in the step 5; and step 7, inputting the test set into the multi-task neural network model, and outputting each performance index of the hydrogen fuel cell. Therefore, high distinguishability, good generalization, good mobility, high precision and high operability of fault diagnosis characteristics are ensured.

Description

Technical Field

[0001] The present invention relates to the field of hydrogen fuel cells, and in particular to a method for monitoring the status of hydrogen fuel cells based on multiple stacks connected in series and parallel. Background Art

[0002] Hydrogen energy is an efficient and clean energy with advantages such as high calorific value, renewable, zero pollution, abundant resources and storability. At present, the application of hydrogen energy is mainly concentrated in hydrogen fuel cell vehicles, hydrogen fuel engine, hydrogen-powered aviation, hydrogen metallurgy and other fields. In order to realize the "green" vision of the Paris Agreement, many countries around the world plan to use hydrogen fuel cells as one of the new energy sources to replace fossil energy and internal combustion engines. On this basis, hydrogen fuel cell vehicles fueled by hydrogen energy have been further promoted and developed worldwide with the advantages of low noise, zero pollution, low temperature cold start and good fuel economy. In the context of my country's vigorous development of hydrogen fuel cells, how to monitor the status and analyze the data of hydrogen fuel cell systems is a hot topic for many scholars. Due to the complex working environment and changeable working conditions of hydrogen fuel cell power generation systems, the risk of failure increases greatly with the increase of operating time. Once a hydrogen fuel cell fails, the performance of the control system will decline, which will bring huge economic losses. Therefore, it is necessary to monitor the operating status of hydrogen fuel cells, including identifying their operating conditions, monitoring their health status and predicting their operating performance, preventing their control performance from deteriorating, and further improving the safety and reliability of hydrogen fuel cells. In recent years, a large amount of research work has been devoted to solving the problem of hydrogen fuel cell fault diagnosis at home and abroad, and remarkable results have been achieved. The fault diagnosis problem of hydrogen fuel cells comes from the following three aspects: (1) the characteristics of different fault diagnoses are weakly distinguishable, and the dynamic response of different membrane electrode voltage acquisition is inconsistent; (2) there are distribution differences between single cell voltage and stack voltage, and the category space is asymmetric; (3) intrusive current distribution detection technology interferes with the operating state of hydrogen fuel cells.

[0003] Therefore, hydrogen fuel cell fault diagnosis has several difficulties, such as weak distinguishability of fault diagnosis features, poor generalization, poor transferability, low accuracy, and low operability. Summary of the invention

[0004] The present invention aims to provide a hydrogen fuel cell state monitoring method based on multiple stacks connected in series and parallel, with fault diagnosis features having high distinguishability, good generalization, good transferability, high accuracy and high operability.

[0005] To achieve the above object, the present invention adopts the following technical solution: a method for monitoring the state of a hydrogen fuel cell based on multiple stacks connected in series and parallel, comprising the following steps:

[0006] Step 1: Collect actual operation data and simulation operation data of hydrogen fuel cells to form a database, which is divided into a training set and a test set;

[0007] Step 2: Build a feature extractor based on the CNN-LSTM model to perform deep feature extraction on the database;

[0008] The deep feature extraction method is: use the CNN model to extract spatial features from the database, use the LSTM model to extract temporal features, and combine spatial features and temporal features to obtain a deep feature matrix;

[0009] Step 3: Build a multi-task neural network model;

[0010] Step 4: Share the deep feature matrix obtained in step 2 to the multi-task neural network model established in step 3. Multiple models in the multi-task neural network model jointly learn and share from the deep feature matrix and optimize their own established models respectively.

[0011] Step 5: Calculate the output index of each model in the multi-task neural network model respectively;

[0012] Step 6: Combine the output indicators in step 5 into a comprehensive performance indicator, use the training set to perform multi-task joint training, obtain the optimal comprehensive performance indicator through different weight coefficients of each output indicator, and update the parameters of multiple models in real time;

[0013] Step 7: Input the test set into the multi-task neural network model and output various performance indicators of the hydrogen fuel cell.

[0014] The beneficial effects of this program are:

[0015] 1. The present invention combines the excellent feature extraction capability of convolutional neural networks and the ability of long-short-term memory neural networks to capture long-term dependencies in time series data, and efficiently and accurately extracts spatial structure information and dynamic time series evolution laws in hydrogen fuel cell operation data;

[0016] 2. The present invention designs a feature fusion and sharing mechanism to organically integrate the deep features extracted by the convolutional neural network and the long short-term memory neural network, which not only retains their respective advantages, but also promotes the interaction and complementarity of cross-dimensional information, greatly enhancing the model's adaptability and understanding of the complex operating environment of fuel cells;

[0017] 3. The present invention introduces a multi-task joint learning strategy and establishes a multi-task learning neural network model to achieve correct identification of the operating conditions of hydrogen fuel cells, real-time monitoring of health status, and accurate prediction of operating performance. By learning multiple tasks simultaneously, the accuracy of each task execution is improved, and information sharing and synergy between tasks are promoted, so that the model can learn more generalized feature representations;

[0018] 4. In summary, the fault diagnosis features of the present invention have high distinguishability, good generalization, good transferability, high accuracy, and high operability. They can comprehensively evaluate the overall health status of hydrogen fuel cells, promptly discover potential operating failures, provide early warnings and precise maintenance, effectively reduce economic losses caused by downtime due to faults, extend the service life of fuel cells, and effectively improve the safety, reliability, and stability of hydrogen fuel cell operation.

[0019] Furthermore, in step six, the different weight coefficients of various output indicators are automatically adjusted; the specific method is:

[0020] 6.1. Build time division module and prediction module;

[0021] 6.2. The time division module sets the current time point, which is a variable. The fixed-length time period before the current time point is T0, and the fixed-length time period after the current time point is T1.

[0022] 6.3. The training set is divided into historical data and verification data according to time characteristics. The historical data is the data in the T0 time period, and the verification data is the data in the T1 time period. The historical data is input into the multi-task neural network model, and the various performance indicators of the hydrogen fuel cell of the multi-task neural network model are imported into the prediction module;

[0023] 6.4. The prediction module performs simulation based on various indicators. The prediction module outputs the prediction data within the T1 time period, compares the prediction data with the verification data, extracts the part where the deviation value between the prediction data and the verification data is greater than the design deviation, and forms a deviation set. According to the correlation between the deviation set and each output indicator, the weight coefficient of each output indicator is automatically adjusted.

[0024] Furthermore, in step six, the comprehensive performance index is:

[0025] J=λ p J1+λ m J2+λ n J3

[0026] Among them, λ p Represents the weight coefficient of the operating condition classification performance index, λ m represents the weight coefficient of the health status monitoring performance index, λ nRepresents the weight coefficient of the performance indicator for running performance prediction.

[0027] Furthermore, in step one, the collected operating data of the hydrogen fuel cell includes battery voltage, battery current, battery temperature, external magnetic field information and load.

[0028] Furthermore, the specific method of deep feature extraction is:

[0029] Step 1.1: Perform convolution operation: Given an input data X∈R H*W*C (height H, width W, number of channels C), the convolution layer passes through the convolution kernel K∈R k*k*C*F (size k*k, number of input channels C, number of output features F) to generate the output feature Y∈R H *W*F ;

[0030] It is expressed as: Y[i,j,f]=max(0,Y[i,j,f]);

[0031] Among them, Y[i,j,f] represents the value of the output feature map Y at the i-th row, j-th column, and f-th feature channel;

[0032] Downsampling, reducing feature size and introducing translation invariance;

[0033] Step 1.2: Use LSTM model to extract time features;

[0034] Define h t is the hidden state of the LSTM model at time step t, x t is the input of this time step, and the update rule of the LSTM model is:

[0035] Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f (;

[0036] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i );

[0037]

[0038] Update unit status:

[0039] Output gate: o t =σ(W o ·[h t-1 ,xt ]+b o );

[0040] h t =o t *tanh(C t );

[0041] Where σ is the sigmoid function, * represents element-wise multiplication of matrices or vectors, W and b are the weights and biases learned by the network, and [·,·] represents the concatenation of vectors.

[0042] Furthermore, in step three, the steps for building a multi-task neural network model are:

[0043] Step 3.1. Specify clear performance indicators for each task.

[0044] Step 3.2, the input value is the deep feature matrix in step 2;

[0045] Step 3.3: Network layer design. The network layer design includes the design of the shared layer and the design of the specific task layer.

[0046] Shared layer design: The activation function is used for calculation, and the lth layer is defined as the shared layer. The calculation is expressed as:

[0047] 1xfz (l) =W (l) a (l-1) +b (l) ;

[0048] mathbfa (l) =σ(z (l) );

[0049] Among them, a (l-1) is the output of layer l-1, W (l) and b (l) are the weight and bias of the lth layer, σ is the activation function;

[0050] Task-specific layer design: The task-specific layer is designed for each task separately. The task-specific layer generates the performance indicators of the task based on the output of the shared layer.

[0051] Furthermore, after step 3.3, the following steps are also included:

[0052] Step 3.4, model training: define a loss function for each task. The loss function is used to measure the difference between the model prediction result and the true value label. The total loss function is:

[0053]

[0054] Among them, α iis the weight of the i-th task, N is the total number of tasks, and loss is the loss function of each task;

[0055] Optimization algorithm selection: Select the stochastic gradient descent method as the optimization algorithm to update the weights and biases of the model and minimize the total loss function;

[0056] Training process: Use the training set to calculate the prediction results and losses of each task through forward propagation; repeat the training until the stopping conditions are met, including the number of iterations reaching the preset value and the decrease in the loss function value being less than the set value.

[0057] Furthermore, after step 3.4, the following steps are also included:

[0058] Step 3.5: Model evaluation and optimization;

[0059] Evaluate model performance: Use a validation set or a test set to evaluate the performance of the model; calculate performance metrics for each task.

[0060] Optimize model structure: Adjust the model structure based on the evaluation results. Methods for adjusting the model structure include increasing or decreasing the number of shared layers, adjusting the structure of specific task layers, and adjusting hyperparameters.

[0061] Furthermore, in step three, the multi-task neural network model includes an operating condition neural network identification model, a health status neural network monitoring model and an operating performance neural network prediction module.

[0062] Furthermore, in step 5, the output indicators of the operating condition neural network identification model, the health status neural network monitoring model and the operating performance neural network prediction module are the operating condition classification performance indicator, the hydrogen fuel cell health status monitoring performance indicator and the hydrogen fuel cell operating performance prediction performance indicator, respectively, and the calculation formula is as follows:

[0063] Operating condition classification performance indicators:

[0064] Health status monitoring performance indicators:

[0065] Operational performance prediction performance indicators:

[0066] in, Indicates the actual output value of the operating condition. Represents the identification output value of the operating condition at time t; The actual output value representing the health status, Represents the monitoring output value of the health status at time t; Indicates the actual output value of the running performance, Represents the predicted output value of the operating performance at time t.

[0067] This solution also has the following effects:

[0068] 1. The three tasks learn and share from the input data together, and optimize the three established models respectively. The advantages of neural networks can prevent model overfitting, make the established model more robust, and better adapt to different data set tasks, thereby improving model performance, accelerating model training, and enhancing model interpretability.

[0069] 2. The multi-task joint training method is adopted to obtain the optimal comprehensive performance index by different weight coefficients of each performance index, and the parameters of the three model neural networks are updated in real time, finally realizing high-precision monitoring of hydrogen fuel cells.

[0070] 3. In multi-task learning, it is very important to balance the weights of different tasks. It can help the model achieve good performance on multiple tasks and avoid the model overfitting a single task and ignoring other tasks.

[0071] In this solution, the weight adjustment is achieved by building a time division module and a prediction module. Specifically:

[0072] The historical data before the current time point is used as the basis for prediction. The time length of these historical data can be 10 minutes or 5 minutes. If simulation calculations are performed directly based on these historical data, the calculations will be complicated, consuming too much computing power and reducing the efficiency of training;

[0073] Therefore, in this scheme, the historical data is first input into the multi-task neural network model, and the operating condition classification performance index, the hydrogen fuel cell health status monitoring performance index and the hydrogen fuel cell operating performance prediction performance index are output respectively. For example, the output operating condition classification performance index includes the levels of excellent, good and poor. The "parameters when the operating condition classification performance index is excellent, good and poor" are respectively input into the corresponding simulation model in the prediction module in advance. During training, the corresponding simulation model is called according to the operating condition classification performance index, so as to quickly predict and obtain the prediction data after the current time point;

[0074] Then, the output indicators of each task are adjusted according to the deviation between the prediction data and the verification data; for example, in the prediction data, the battery voltage, battery current and operating conditions are related, so the larger the deviation of the battery voltage and battery current, the higher the weight of the operating condition classification performance indicator. Similarly, the relevant data of the health status detection performance indicator are: battery temperature, battery current; the relevant data of the operating performance prediction performance indicator are: battery voltage, external magnetic field information and load. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of an embodiment;

[0076] Figure 2 Schematic diagram of a multi-task neural network model of an embodiment. DETAILED DESCRIPTION

[0077] The following is further described in detail through specific implementation methods:

[0078] The reference numerals in the drawings of the specification include:

[0079] Example

[0080] The embodiment is basically as follows Figure 1 As shown: A method for monitoring the status of a hydrogen fuel cell based on multiple stacks in series and parallel, the flow chart is as follows Figure 1 As shown, the following steps are included:

[0081] Step 1: Collect actual operation data and simulated operation data of the hydrogen fuel cell to form a database; the collected actual operation data and simulated operation data of the hydrogen fuel cell include battery voltage, battery current, battery temperature, external magnetic field information and load;

[0082] The collected data is normalized. The normalization method is as follows:

[0083]

[0084] Among them, x * represents the data after normalization, x represents the data before normalization, and x max Indicates the maximum value of the parameter processed in the data, x min Indicates the minimum value of the parameter processed in the data;

[0085] Divide the database into training and testing sets.

[0086] Step 2: Build a feature extractor based on the CNN-LSTM model to perform deep feature extraction on the database;

[0087] The method of deep feature extraction is: use the CNN model to extract spatial features from the database, use the LSTM model to extract temporal features, and combine spatial features and temporal features to obtain a deep feature matrix; the CNN model consists of multiple convolutional layers, activation layers, pooling layers, and possible fully connected layers. The convolutional layer is the core, which extracts spatial features through convolution operations;

[0088] The specific method of deep feature extraction is:

[0089] Step 1.1: Convolutional layer performs convolution operation: given an input data X∈R H*W*C (height H, width W, number of channels C), the convolution layer passes through the convolution kernel K∈R k*k*C*F(size k*k, number of input channels C, number of output features F) to generate the output feature Y∈R H*W*F ;

[0090] It is expressed as: Y[i,j,f]=max(0,Y[i,j,f]);

[0091] Among them, Y[i,j,f] represents the value of the output feature map Y at the i-th row, j-th column, and f-th feature channel;

[0092] The pooling layer downsamples to reduce feature size and introduce translation invariance;

[0093] Step 1.2: Use LSTM model to extract time features;

[0094] Define h t is the hidden state of the LSTM model at time step t, x t is the input of this time step, and the update rule of the LSTM model is:

[0095] Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f );

[0096] Input gate: i t =σ(W i ·[h t-1 ,x t ]+h i );

[0097]

[0098] Update unit status:

[0099] Output gate: o t =σ(W o ·[b t-1 ,x t ]+b o );

[0100] h t =o t *tanh(C t );

[0101] Where σ is the sigmoid function, * represents element-wise multiplication of matrices or vectors, W and b are the weights and biases learned by the network, and [·,·] represents the concatenation of vectors.

[0102] Step 3: Build a multi-task neural network model; the schematic diagram of the multi-task neural network model is as follows Figure 2As shown, the multi-task neural network model includes an operating condition neural network identification model, a health status neural network monitoring model and an operating performance neural network prediction module;

[0103] The steps to build a multi-task neural network model are:

[0104] Step 3.1. Specify clear performance indicators for each task.

[0105] Step 3.2, the input value is the deep feature matrix in step 2;

[0106] Step 3.3: Network layer design. The network layer design includes the design of the shared layer and the design of the specific task layer.

[0107] Shared layer design: The activation function is used for calculation, and the lth layer is defined as the shared layer. The calculation is expressed as:

[0108] 1xfz (l) =W (l) a (l-1) +b (l) ;

[0109] mathbfa (l) =σ(z (l) );

[0110] Among them, a (l-1) is the output of layer l-1, W (l) and b (l) are the weight and bias of the lth layer, σ is the activation function;

[0111] Design of specific task layer: The specific task layer is designed for each task separately. The specific task layer generates the performance index of the task based on the output of the shared layer.

[0112] Step 3.4, model training: define a loss function for each task. The loss function is used to measure the difference between the model prediction result and the true value label. The total loss function is:

[0113]

[0114] Among them, α i is the weight of the i-th task, N is the total number of tasks, and loss is the loss function of each task;

[0115] Optimization algorithm selection: Select the stochastic gradient descent method as the optimization algorithm to update the weights and biases of the model and minimize the total loss function;

[0116] Training process: Use the training set to calculate the prediction results and losses of each task through forward propagation; repeat the training until the stopping condition is met, which includes the number of iterations reaching the preset value and the decrease in the loss function value being less than the set value;

[0117] Step 3.5: Model evaluation and optimization;

[0118] Evaluate model performance: Use a validation set or a test set to evaluate the performance of the model; calculate performance metrics for each task.

[0119] Optimize model structure: adjust the model structure according to the evaluation results. The methods of adjusting the model structure include increasing or decreasing the number of shared layers, adjusting the structure of specific task layers, and adjusting hyperparameters.

[0120] Step 4: Share the deep feature matrix obtained in step 2 to the multi-task neural network model established in step 3. Multiple models in the multi-task neural network model jointly learn and share from the deep feature matrix and optimize their own established models respectively.

[0121] Step 5. Calculate the output index of each model in the multi-task neural network model respectively; the output indexes of the operating condition neural network recognition model, the health status neural network monitoring model and the operating performance neural network prediction module are the operating condition classification performance index, the hydrogen fuel cell health status monitoring performance index and the hydrogen fuel cell operating performance prediction performance index respectively, and the calculation formula is as follows:

[0122] Operating condition classification performance indicators:

[0123] Health status monitoring performance indicators:

[0124] Operational performance prediction performance indicators:

[0125] in, Indicates the actual output value of the operating condition. Represents the identification output value of the operating condition at time t; The actual output value representing the health status, Represents the monitoring output value of the health status at time t; Indicates the actual output value of the running performance, It represents the predicted output value of the operating performance at time t;

[0126] Step 6: Combine the output indicators in step 5 into a comprehensive performance indicator, use the training set to conduct multi-task joint training, obtain the optimal comprehensive performance indicator through different weight coefficients of each output indicator, and update the parameters of multiple models in real time; the comprehensive performance indicator is:

[0127] J=λ p J1+λ m J2+λ n J3

[0128] Among them, λ p Represents the weight coefficient of the operating condition classification performance index, λ m represents the weight coefficient of the health status monitoring performance index, λ n Represents the weight coefficient of the performance index for predicting the running performance;

[0129] During the training process, the different weight coefficients of each output indicator are automatically adjusted; the specific method is:

[0130] 6.1. Build time division module and prediction module;

[0131] 6.2. The time division module sets the current time point, which is a variable. The fixed-length time period before the current time point is T0, and the fixed-length time period after the current time point is T1.

[0132] 6.3. The training set is divided into historical data and verification data according to time characteristics. The historical data is the data in the T0 time period, and the verification data is the data in the T1 time period. The historical data is input into the multi-task neural network model, and the various performance indicators of the hydrogen fuel cell of the multi-task neural network model are imported into the prediction module;

[0133] 6.4. The prediction module performs simulation based on various indicators. The prediction module outputs the prediction data within the T1 time period, compares the prediction data with the verification data, extracts the part where the deviation value between the prediction data and the verification data is greater than the design deviation, and forms a deviation set. According to the correlation between the deviation set and each output indicator, the weight coefficient of each output indicator is automatically adjusted.

[0134] Step 7: Input the test set into the multi-task neural network model and output various performance indicators of the hydrogen fuel cell.

[0135] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel, characterized in that , including the following steps: Step 1: Collect actual operation data and simulation operation data of hydrogen fuel cells to form a database, which is divided into a training set and a test set; Step 2: Build a feature extractor based on the CNN-LSTM model to perform deep feature extraction on the database; The deep feature extraction method is: use the CNN model to extract spatial features from the database, use the LSTM model to extract temporal features, and combine spatial features and temporal features to obtain a deep feature matrix; Step 3: Build a multi-task neural network model; Step 4: Share the deep feature matrix obtained in step 2 to the multi-task neural network model established in step 3. Multiple models in the multi-task neural network model jointly learn and share from the deep feature matrix and optimize their own established models respectively. Step 5: Calculate the output index of each model in the multi-task neural network model respectively; Step 6: Combine the output indicators in step 5 into a comprehensive performance indicator, use the training set to perform multi-task joint training, obtain the optimal comprehensive performance indicator through different weight coefficients of each output indicator, and update the parameters of multiple models in real time; Step 7: Input the test set into the multi-task neural network model and output various performance indicators of the hydrogen fuel cell.

2. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 1, characterized in that: In step six, the different weight coefficients of each output indicator are automatically adjusted; the specific method is: 6.

1. Build time division module and prediction module; 6.

2. The time division module sets the current time point, which is a variable. The fixed-length time period before the current time point is T0, and the fixed-length time period after the current time point is T1. 6.

3. The training set is divided into historical data and verification data according to time characteristics. The historical data is the data in the T0 time period, and the verification data is the data in the T1 time period. The historical data is input into the multi-task neural network model, and the various performance indicators of the hydrogen fuel cell of the multi-task neural network model are imported into the prediction module; 6.

4. The prediction module performs simulation based on various indicators. The prediction module outputs the prediction data within the T1 time period, compares the prediction data with the verification data, extracts the part where the deviation value between the prediction data and the verification data is greater than the design deviation, and forms a deviation set. According to the correlation between the deviation set and each output indicator, the weight coefficient of each output indicator is automatically adjusted.

3. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to any one of claim 1 or claim 2, characterized in that: In step six, the comprehensive performance indicators are: J=λ p J1+λ m J2+λ n J3; Among them, λ p Represents the weight coefficient of the operating condition classification performance index, λ m represents the weight coefficient of the health status monitoring performance index, λ n Represents the weight coefficient of the performance indicator for running performance prediction.

4. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 3, characterized in that: In step one, the collected operating data of the hydrogen fuel cell include battery voltage, battery current, battery temperature, external magnetic field information and load.

5. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 4, characterized in that: The specific method of deep feature extraction is: Step 1.1: Perform convolution operation: Given an input data X∈R H*W*C (height H, width W, number of channels C), the convolution layer passes through the convolution kernel K∈R k*k*C*F (size k*k, number of input channels C, number of output features F) to generate the output feature Y∈R H*W*F ; It is expressed as: Y[i,j,f]=max(0,Y[i,j,f]); Among them, Y[i,j,f] represents the value of the output feature map Y at the i-th row, j-th column, and f-th feature channel; Downsampling, reducing feature size and introducing translation invariance; Step 1.2: Use LSTM model to extract time features; Define h t is the hidden state of the LSTM model at time step t, x t is the input of this time step, and the update rule of the LSTM model is: Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f ); Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ); Update unit status: Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o ); h t =o t *tanh(C t ); Where σ is the sigmoid function, * represents element-wise multiplication of matrices or vectors, W and b are the weights and biases learned by the network, and [·,·] represents the concatenation of vectors.

6. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 5, characterized in that: In step 3, the steps to build a multi-task neural network model are: Step 3.

1. Specify clear performance indicators for each task. Step 3.2, the input value is the deep feature matrix in step 2; Step 3.3: Network layer design. The network layer design includes the design of the shared layer and the design of the specific task layer. Shared layer design: The activation function is used for calculation, and the lth layer is defined as the shared layer. The calculation is expressed as: mathbfz (l) =W (l) a (l-1) +b (l) ; math (l) =σ(z (l) ); Among them, a (l-1) is the output of layer l-1, W (l) and b (l) are the weight and bias of the lth layer, σ is the activation function; Task-specific layer design: The task-specific layer is designed for each task separately. The task-specific layer generates the performance indicators of the task based on the output of the shared layer.

7. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 6, characterized in that: After step 3.3, the following steps are also included: Step 3.4, model training: define a loss function for each task. The loss function is used to measure the difference between the model prediction result and the true value label. The total loss function is: Among them, α i is the weight of the i-th task, N is the total number of tasks, and loss is the loss function of each task; Optimization algorithm selection: Select the stochastic gradient descent method as the optimization algorithm, update the weights and biases of the model, and minimize the total loss function; Training process: Use the training set to calculate the prediction results and losses of each task through forward propagation; repeat the training until the stopping conditions are met, including the number of iterations reaching the preset value and the decrease in the loss function value being less than the set value.

8. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 7, characterized in that: After step 3.4, the following steps are also included: Step 3.5: Model evaluation and optimization; Evaluate model performance: Use a validation set or a test set to evaluate the performance of the model; calculate performance metrics for each task. Optimize model structure: Adjust the model structure based on the evaluation results. Methods for adjusting the model structure include increasing or decreasing the number of shared layers, adjusting the structure of specific task layers, and adjusting hyperparameters.

9. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 8, characterized in that: In step three, the multi-task neural network model includes an operating condition neural network identification model, a health status neural network monitoring model and an operating performance neural network prediction module.

10. A method for monitoring the status of a hydrogen fuel cell based on multiple stacks connected in series and parallel according to claim 9, characterized in that: In step 5, the output indicators of the operating condition neural network identification model, the health status neural network monitoring model and the operating performance neural network prediction module are the operating condition classification performance indicator, the hydrogen fuel cell health status monitoring performance indicator and the hydrogen fuel cell operating performance prediction performance indicator, respectively, and the calculation formula is as follows: Operating condition classification performance indicators: Health status monitoring performance indicators: Operational performance prediction performance indicators: in, Indicates the actual output value of the operating condition. Represents the identification output value of the operating condition at time t; The actual output value representing the health status, Represents the monitoring output value of the health status at time t; Indicates the actual output value of the running performance, Represents the predicted output value of the operating performance at time t.

Images