CN-121995239-A - Method, system and storage medium for predicting aging trend and residual life of proton exchange membrane fuel cell

CN121995239ACN 121995239 ACN121995239 ACN 121995239ACN-121995239-A

Abstract

The application relates to the technical field of artificial intelligence, in particular to a method, a system and a storage medium for predicting aging trend and residual life of a proton exchange membrane fuel cell. The method comprises the steps of utilizing a variable-scale convolution kernel to adaptively slide along a time dimension, carrying out downsampling through an adaptive pooling operation, capturing a short-term aging time mode with discriminant, introducing an aging characteristic attention mask, shielding attention calculation of a redundancy stable section, simultaneously optimizing a head number distribution strategy of multi-head self-attention, adapting aging dependency characteristics of a proton exchange membrane fuel cell in different time scales, constructing an adaptive function for optimizing super-parameters and aging trend characteristics of the proton exchange membrane fuel cell by using a prediction error minimization and feature redundancy minimization as a common target, and providing an adaptive genetic operation operator strategy to select the aging trend characteristics related to the aging trend in time sequence data of the proton exchange membrane fuel cell. The method aims at solving the problem of how to improve the prediction accuracy of the proton exchange membrane fuel cell.

Inventors

YANG BO
Mao Yunjiang
LI PEIRUI
JIANG LIN

Assignees

昆明理工大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A method for predicting aging trend and remaining life of a proton exchange membrane fuel cell, the method comprising the steps of: s10, inputting the acquired time sequence data of the proton exchange membrane fuel cell to be tested into a pre-training prediction model; Wherein the pre-trained predictive model comprises: The feature extraction layer adopts a variable-scale one-dimensional convolution kernel function to determine an adaptive sliding window corresponding to time-varying features in time sequence data of the proton exchange membrane fuel cell, and the local features of the time sequence data are sampled in a pooling manner in the adaptive sliding window; The encoding layer adopts an attention calculating process of a redundant segment in the time sequence data, which is shielded by an aging characteristic attention mask, distributes attention heads positively correlated with the time scale of the non-redundant segment to the non-redundant segment for aging characteristic capture, and determines the weight matrix size of the feedforward network according to the activation strength of the aging characteristic in the time sequence data; The pre-training prediction model optimizes the super-parameters and aging trend characteristics of the proton exchange membrane fuel cell by constructing an fitness function with the minimization of prediction errors and the minimization of characteristic redundancy as common targets in the training process, and selects the aging trend characteristics related to the aging trend from time sequence data of the proton exchange membrane fuel cell based on a self-adaptive genetic operation operator strategy; and S20, obtaining a voltage prediction result output by the pre-training prediction model, and determining a residual life prediction value of the proton exchange membrane fuel cell to be tested according to the voltage prediction result.
2. The method for predicting aging trend and residual life of a proton exchange membrane fuel cell as claimed in claim 1, wherein the expression of the variable-scale one-dimensional convolution kernel function is: ; in the formula, Representing the convolution layer at time steps Linear pre-activation output of (a); representing time series data at a time point Is a value of (2); Is convolution kernel at the first Weight parameters for the individual locations; Is a bias term; the size of the convolution kernel is positively correlated with the local fluctuation degree of time sequence data; representing the final output of the ReLU function after activation, and representing the activation strength of the local aging characteristic of the proton exchange membrane fuel cell; The mathematical expression of the pooling downsampling is as follows: ; in the formula, Indicating that the pooling layer is at the first S is the self-adaptive pooling step length, and is positively correlated with the activation intensity of the convolution output characteristic; Refers to the time step value in the corresponding adaptive window in the ReLU output feature map.
3. The method for predicting aging trend and remaining life of a pem fuel cell of claim 1 wherein said encoding layer uses an aging feature attention mask to mask the attention calculation process of redundant segments in said time series data, and assigns the number of attention points positively correlated to the time scale to non-redundant segments for aging feature capture, comprising: the multi-headed self-attention output equation with the aging feature attention mask is set as: ; Wherein a single attention head The mathematical expression of (2) is: ; in the formula, M is an aging characteristic attention mask, wherein 0 is assigned to redundant segments in time sequence data to shield attention calculation, and 1 is assigned to non-redundant segments to reserve attention calculation; in the formula, To pay attention to the head count, the low head count captures short-time scale aging dependence, and the high head count captures long-time scale aging dependence; , , Respectively the first Head-by-head query, key, value weight matrix; Is that Is a dimension of (2); inputting dimensions for a transducer; The matrix of output weights for multi-head attention, ⊗ is an element-wise multiplication operation.
4. The method for predicting aging trend and remaining life of a pem fuel cell of claim 1 wherein said encoding layer determines a weight matrix size of a feed-forward network based on an activation strength of aging features in said time series data, comprising: Setting feature enhancement factors As the weight coefficient of the weight matrix of the feedforward network, the characteristic enhancement factor is adjusted according to the activation intensity of the aging characteristic in the time sequence data Thereby adjusting the size of the weight matrix of the feed-forward network; wherein, the expression of the feedforward network is: ; in the formula, , A weight matrix for FFN; , Is a bias term; And X is an input characteristic of FFN, and is output from a multi-head self-attention module at the upper layer and comprises a long-distance dependence characteristic of PEMFC time sequence data.
5. The method for predicting aging tendency and remaining life of a proton exchange membrane fuel cell as recited in claim 1, wherein optimizing is performed with a goal of maximizing said fitness function, wherein said fitness function The expression of (2) is: ; in the formula, A prediction error loss function; a feature redundancy loss function; The weight coefficient is used for balancing the prediction error and the feature redundancy; Wherein: ; ; in the formula, Individual encoded as a subset of super parameters and features; The number of samples for the test set; Predicting a voltage value for a model corresponding to the individual x; Is the actual voltage value; Is characterized by And (3) with And F is the total number of features.
6. The method for predicting aging trend and residual life of a proton exchange membrane fuel cell as claimed in claim 1, wherein the adaptive genetic operator strategy comprises a selection operator, an adaptive crossover operator and an adaptive mutation operator; Wherein the adaptive genetic operator strategy comprises the following steps: S11, screening a selection operator in the test set sample by adopting roulette so as to positively correlate the probability of the test set sample individual being selected with the fitness: ; in the formula, Is the first The probability of selection of the individual test set samples; Is the population scale; Is the first A plurality of test set samples; S12, dynamically adjusting an adaptive crossover operator according to individual fitness to reserve excellent genes by adopting small crossover coefficients for individuals with high fitness and adopting large crossover coefficients for individuals with low fitness to promote evolution, wherein the expression of the adaptive crossover operator is as follows: ; in the formula, In order to adapt the crossover operator, The upper limit and the lower limit of the crossing coefficient are respectively; Fitness for the individual; Maximum and minimum fitness of the contemporary population respectively; is the average fitness of the contemporary population; s13, dynamically adjusting an adaptive mutation operator according to individual fitness, wherein the expression of the adaptive mutation operator is as follows: ; in the formula, Is an adaptive mutation operator; The lower limit and the upper limit of the variation probability are respectively; Is the disturbance variance; Is the mutation probability.
7. The method for predicting aging tendency and remaining life of a proton exchange membrane fuel cell as recited in claim 1, further comprising, prior to step S10: s30, carrying out median filtering with noise detection on the acquired proton exchange membrane fuel cell data to be detected: ; in the formula, Representing the voltage after noise reduction at the time t; for half width of adaptive filtering window, the noise intensity is used for filtering Determining; Is a data sampling interval; is a median function; s40, determining a voltage attenuation rate according to the variation of the noise-reduced voltage, and determining a target sampling sliding window length according to the voltage attenuation rate so as to sample data based on the target sampling sliding window length: ; Wherein: Is the first A plurality of input samples; For the sliding window length, determined by the voltage decay rate; Is the first The expiration time of each sample; S50, introducing aging characteristic contribution degree constraint, and screening aging trend characteristics from the proton exchange membrane fuel cell data to be tested, wherein the expression of the aging characteristic contribution degree constraint is as follows: ; in the formula, Is the first Importance of the individual features; Is the first Sample prediction voltage pair Partial derivatives of the individual features; contributing a degree factor to the aging characteristics.
8. The method for predicting aging trend and remaining life of a proton exchange membrane fuel cell as claimed in claim 1, wherein in S20, determining the predicted remaining life of the proton exchange membrane fuel cell to be measured according to the voltage prediction result comprises: S21, determining the prediction time of the prediction voltage meeting the multi-time window failure verification for the first time according to the voltage prediction result ; S22, according to the predicted time With the actual measured current monitoring time The difference between the two, determining the residual life prediction value RUL: 。
9. A computer system comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program when executed by the processor implementing the steps of the method for predicting aging tendency and remaining life of a proton exchange membrane fuel cell as claimed in any one of claims 1 to 8.
10. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program when executed by a processor implements the steps of the method for predicting aging tendency and remaining life of a proton exchange membrane fuel cell according to any one of claims 1 to 8.

Description

Method, system and storage medium for predicting aging trend and residual life of proton exchange membrane fuel cell Technical Field The application relates to the technical field of artificial intelligence, in particular to a method, a system and a storage medium for predicting aging trend and residual life of a proton exchange membrane fuel cell. Background With the popularization of hydrogen energy application, proton exchange membrane fuel cells (Proton Exchange Membrane Fuel Cell, PEMFCs) are used as clean energy conversion devices, and the prediction of the residual life (REMAINING USEFUL LIFE, RUL) is crucial to the reliability of maintenance systems. The conventional common battery prediction method such as CNN-transducer combined feature extraction and genetic algorithm parameter optimization has some technical combination difficulties when transferring to the scene of aging trend of PEMFC and RUL prediction: On one hand, the PEMFC time sequence data has the characteristics of non-stability, strong noise and characteristic time variation, the aging key local characteristics are easy to lose, the fixed convolution kernel and pooling strategy of the traditional CNN cannot be adapted to the PEMFC time sequence data, on the other hand, the self-attention mechanism of the traditional transducer has the defects of computational redundancy and low long-distance dependence capturing efficiency when the PEMFC long time sequence data are processed, on the other hand, the traditional genetic algorithm is easy to be trapped into local optimum when the model super-parameters and the characteristic sets are optimized, and the fixed parameters of the crossover operator and the mutation operator cannot be adapted to the heterogeneous characteristics of the PEMFC multi-characteristics, so that the optimization precision and the efficiency are insufficient. In view of the above, the present application provides a new method for predicting aging trend and remaining life of a proton exchange membrane fuel cell, which aims to overcome the above-mentioned drawbacks and improve the prediction accuracy of the proton exchange membrane fuel cell. Disclosure of Invention The application mainly aims to provide a method for predicting the aging trend and the residual life of a proton exchange membrane fuel cell, which aims to solve the problem of how to improve the prediction accuracy of the proton exchange membrane fuel cell. In order to achieve the above object, the present application provides a method for predicting aging trend and remaining life of a proton exchange membrane fuel cell, the method comprising the steps of: s10, inputting the acquired time sequence data of the proton exchange membrane fuel cell to be tested into a pre-training prediction model; Wherein the pre-trained predictive model comprises: The feature extraction layer adopts a variable-scale one-dimensional convolution kernel function to determine an adaptive sliding window corresponding to time-varying features in time sequence data of the proton exchange membrane fuel cell, and the local features of the time sequence data are sampled in a pooling manner in the adaptive sliding window; The encoding layer adopts an attention calculating process of a redundant segment in the time sequence data, which is shielded by an aging characteristic attention mask, distributes attention heads positively correlated with the time scale of the non-redundant segment to the non-redundant segment for aging characteristic capture, and determines the weight matrix size of the feedforward network according to the activation strength of the aging characteristic in the time sequence data; The pre-training prediction model optimizes the super-parameters and aging trend characteristics of the proton exchange membrane fuel cell by constructing an fitness function with the minimization of prediction errors and the minimization of characteristic redundancy as common targets in the training process, and selects the aging trend characteristics related to the aging trend from time sequence data of the proton exchange membrane fuel cell based on a self-adaptive genetic operation operator strategy; and S20, obtaining a voltage prediction result output by the pre-training prediction model, and determining a residual life prediction value of the proton exchange membrane fuel cell to be tested according to the voltage prediction result. Optionally, the expression of the variable-scale one-dimensional convolution kernel is: in the formula, Representing the convolution layer at time stepsLinear pre-activation output of (a); representing time series data at a time point Is a value of (2); Is convolution kernel at the first Weight parameters for the individual locations; Is a bias term; the size of the convolution kernel is positively correlated with the local fluctuation degree of time sequence data; representing the final output of the ReLU function after activation, and representing the