Fuel cell health state assessment method

Abstract

The invention discloses a fuel cell health state evaluation method. Improved kernel principal component analysis is used for carrying out feature extraction on working parameters and operating parameters of a fuel cell; the method comprises the following steps: decomposing fuel cell data by adopting a TVF-EMD technology to obtain sub-modes, reducing the non-stationarity of a sequence, reducing the number of the sequence through fuzzy entropy aggregation, and improving the data processing efficiency; a time series prediction model TimeMixer capable of decomposing multi-scale fusion and improved by a self-attention mechanism is provided, multi-scale seasonal and trend components are mixed in two directions from fine to rough and from rough to fine, information in time series data is comprehensively extracted, and prediction robustness is enhanced; and optimizing hyper-parameters of the TimeMixer model by using an improved difference creative search algorithm DCS so as to quickly search out an optimal hyper-parameter combination and realize accurate evaluation of the state of health of the fuel cell. According to the invention, a data-based deep learning technology is adopted to solve the problems of complexity of fuel cell data and health state evaluation.

Description

Technical Field

[0001] The invention belongs to the technical field of renewable energy, and in particular relates to a method for evaluating the health status of a fuel cell. Background Art

[0002] The performance of fuel cells will degrade as the use time increases, affecting their output performance and the power and reliability of the vehicle. Therefore, it is necessary to collect data, monitor signals and diagnose faults in the fuel cell system. Accurate health status assessment (SI) helps to formulate appropriate maintenance plans, reduce unnecessary wear and tear, and extend the service life of fuel cells.

[0003] Existing health status assessment technologies rely on a single parameter such as a voltage signal, which may lead to low accuracy and reliability of the assessment results and inability to take timely fault response measures. In addition, there are some technical problems, such as missing or errors in battery data, difficulty in capturing local characteristics, poor dynamic adaptability of the model, etc., which will directly affect the reliability of the prediction results. Therefore, the health status assessment of fuel cells is inseparable from advanced models and strategies. Data processing methods and full extraction of local characteristics require more advanced strategic support. In addition, accurate assessment of the health status of fuel cells can timely discover potential problems and faults, so as to take preventive measures to avoid unexpected shutdowns and failures and improve the reliability of the entire system. Invention content: In view of the problems mentioned in the background technology, such as poor accuracy of fuel cell data, difficulty in capturing local characteristics, and poor dynamic adaptability of the model, the present invention proposes a fuel cell health status assessment method and system to improve the model prediction accuracy and efficiency and promote the development of the battery industry. Summary of the invention

[0004] Purpose of the invention: The present invention proposes a fuel cell health status assessment method that can more accurately capture the changing trend of the fuel cell health status and improve the accuracy of health status assessment.

[0005] Technical solution: A fuel cell health status assessment method according to the present invention comprises the following steps:

[0006] (1) Obtain fuel cell operation data, environmental parameters, and historical operation data to monitor the status of the fuel cell in real time;

[0007] (2) Using improved kernel principal component analysis (iKPCA) to extract features of fuel cell operating parameters, the data is mapped into a high-dimensional space to reveal the intrinsic structure of the data;

[0008] (3) In view of the random volatility and uncertainty of fuel cell operation data, the TVF-EMD technology is used to decompose the fuel cell data to obtain sub-modes and reduce the non-stationarity of the sequence; and the number of sequences is reduced through fuzzy entropy aggregation to improve data processing efficiency;

[0009] (4) A decomposable multi-scale fusion time series prediction model TimeMixer is improved by self-attention mechanism, which mixes multi-scale seasonal and trend components from fine to coarse and from coarse to fine to fully extract information from time series data;

[0010] (5) The improved differential creative search algorithm DCS is used to optimize the hyperparameters of the TimeMixer model, quickly search for the optimal hyperparameter combination, and achieve accurate assessment of the health status of the fuel cell.

[0011] Furthermore, the fuel cell operating data and load current of step (1) include online monitoring of voltage, current, temperature, pressure and flow; and the environmental parameters include maximum operating temperature, air flow and gas pressure.

[0012] Furthermore, the implementation process of step (2) is as follows:

[0013] Define a time series of input signal as x a , a=1,2,…,m, linearly independent variable y a , a=1,2,…,n, and n≤m; obtain x by singular value decomposition a , then its expression is:

[0014] U=[x1,x2,...,x m ]

[0015] V=[y1,y2,...,y n ]

[0016] U=CV

[0017] Among them, U is the input data matrix, C is the n×n transformation matrix, and V is the output observation matrix; from this matrix, we get:

[0018] R x =BR y B T

[0019] Since the autocorrelation matrix R x is a real symmetric matrix, then the autocorrelation matrix R y Convertible to R x There are m positive real eigenvalues ​​ξ in i , i=1,2,…,m, and B is an n-dimensional vector; that is:

[0020] R y =[E(y i ) 2 ]=[ξ i ]

[0021] ξ1>ξ2>...> m

[0022] In the autocorrelation matrix R y In , m eigenvalues ​​correspond to m eigenvectors, and m eigenvectors constitute m subspaces, and m should satisfy:

[0023]

[0024] From the above formula, we can see that the eigenvalue ξ i The information contained in the original signal is compressed, and the fault mode is classified by eigenvalue and eigenvector;

[0025] Assume a nonlinear mapping function φ n (x), and the original feature set W = [w1,w2,…,w s ] T Transformed into the kernel feature set W' = [φ(w1),φ(w2),…,φ(w s )] T ; where s is the number of samples; then the covariance matrix F of the kernel feature set is expressed as:

[0026]

[0027]

[0028] Among them, ψ(w i ) represents the mapping function, which maps the input data to a high-dimensional space; F is the covariance matrix of the kernel feature set. In KPCA, the F matrix is ​​used to capture the main direction of change of the data; then the eigenvalue λ of the matrix F is used to i and the eigenvector ζ i , we get the expression:

[0029]

[0030]

[0031] Among them, b i Represents the coefficient used to construct the kernel feature vector, which maps each sample to the kernel feature set W' and the feature vector ζ i The dot product of

[0032] For the inner product operation of KPCA in high-dimensional space, Gaussian radial basis function is selected As a kernel function:

[0033]

[0034] in, is the inner product of the two mapped high-dimensional eigenvectors, and then the q eigenvectors corresponding to the q largest eigenvalues ​​are selected as the column vectors of the transformation matrix γ, γ T is the transposed matrix of the transformation matrix, then the result after KPCA dimensionality reduction is:

[0035] R = γ T ·W'

[0036] In order to measure the interval between samples of the same class and samples of different classes, the degree of intra-class aggregation is defined:

[0037]

[0038] Among them, ε is the number of samples in the i-th category, is the standard deviation of the a-th dimension component of the category; at the same time, the discrete degree expression between samples of different categories is:

[0039]

[0040]

[0041] Among them, ε is the number of samples in the i-th category, β aij is the a-th dimension component of the j-th sample in the i-th category, and the degree of separability of sample categories is:

[0042]

[0043] When J a The smaller the value of is, the higher the degree of separability of the sample categories is, and the greater the contribution to sample classification is; the correlation between the extracted feature vector and the sample category vector is:

[0044]

[0045] Among them, τ is the correlation coefficient ζ i is the extracted feature vector, ζ s is the sample category vector in the original feature set; when the absolute value of τ approaches 1, it indicates that the correlation between the two feature vectors is very strong, which means that the two features have similar information in describing the fuel cell data; when the absolute value of τ approaches 0, it indicates that the correlation between the two feature vectors is very weak, which indicates that they provide complementary information in describing the data.

[0046] Furthermore, the implementation process of step (3) is as follows:

[0047] By using adaptive decomposition, the intrinsic modal component IMF is replaced by a local narrowband signal, and the formula is as follows:

[0048]

[0049] Among them, IMF i (t) represents the i-th IMF, x(t) represents the given fuel cell data signal, and r(t) represents the residual; the time-varying filter empirical mode distribution formula is as follows:

[0050]

[0051] Among them, u i represents the maximum value of x(t), i=1, 2, 3, ...; represents the second-order resolution, ρ is the preset threshold of the frequency change between two consecutive maximum values; find the local cutoff frequency Interpolate between the peaks to get Adjustment Improve modal aliasing; refactor Get the new signal h(t):

[0052]

[0053] Calculate the criterion value θ(t) to determine whether the residual signal meets the cutoff standard:

[0054]

[0055] Where θ(t) represents the calculation criterion value, cos(·) represents the cosine function, and B Loughlin (t) represents the Loughlin instantaneous bandwidth, represents the weighted average instantaneous frequency;

[0056] The number of sequences is reduced by fuzzy entropy aggregation, the formula is as follows:

[0057] H(D,B)=-∑(μ(x)*log(μ(x)))

[0058] Among them, H(D, B) represents the fuzzy entropy of sets D and B, μ(x) represents the membership of the fuzzy set, and log(·) represents the logarithmic function.

[0059] Furthermore, the decomposable multi-scale fusion time series prediction model TimeMixer in step (4) is based on a fully connected multi-layer perceptron MLP to capture and learn complex patterns in data; and adopts past decomposable hybrid PDM and future multi-predictor hybrid FMM, uses multi-scale historical data, and integrates sequence data of different time scales to predict the future; the implementation process is as follows:

[0060] First, the historical data of fuel cell operating parameters, operating parameters, environmental parameters, etc. are used as observation data for the model. Downsample and finally obtain a multi-scale time series set through average pooling in m∈{0, 1, ..., M}, C represents the number of variables, P is the length of the time series, i.e., the total number of observed data points, the lowest level sequence j0=j is the input sequence, containing the finest temporal variations, and the highest level sequence j M Represents macroscopic changes; projects multi-scale sequences into the embedding layer to obtain deep features The formula is:

[0061]

[0062] Among them, Embed is the embedding process, which means that the data is transformed from its original form or space to a different feature space;

[0063] Using stacked PDM blocks to mix past information of different scales, for the Lth layer, the input is The process of PDM is formulated as:

[0064]

[0065] Where L is the total number of layers, Indicates mixed past expressions, with d model aisle;

[0066] For the future prediction stage, the FMM block is used to integrate the extracted multi-scale past information And generate future predictions, namely:

[0067]

[0068] in, represents the final prediction and F is the number of predicted time steps, i.e. the total number of future time points that the model will predict.

[0069] Furthermore, the past decomposable hybrid PDM implementation process is as follows:

[0070]

[0071]

[0072] in, For the seasonal part, For the trend part, FeedForward is a feedforward network structure, which includes two linear layers with a GELU activation function in the middle for information interaction between channels. S-Mix(·) and T-Mix(·) represent seasonal and trend mixing, respectively;

[0073] Seasonal mixing uses a bottom-up mixing layer to achieve bottom-up interaction of seasonal information in a residual manner from the mth scale:

[0074]

[0075] Among them, Bottom-Up-Mixing (·) is two linear layers with activation functions in the middle. Along the time dimension, its input dimension and output dimension are both

[0076] For trend mixing, a top-down mixing layer is used to achieve top-down trend information interaction in a residual manner at the mth scale:

[0077]

[0078] In the formula, Top-Down-Mixing (·) is two linear layers with an intermediate GELU activation function, and its input dimension is The output dimension is Empowered by a mixture of seasonality and trends, PDM gradually aggregates seasonal data from micro-scales, while combining macro-trends with historical experience, ultimately achieving multi-scale fusion in the process of extracting historical information.

[0079] Furthermore, the future multi-predictor hybrid FMM implementation process is as follows:

[0080]

[0081] in, represents the future prediction from the m-th scale sequence, and the final output is Predictor m (·) indicates the mth scale; the predictor first uses a single-layer linear transformation to directly predict the future time step of length F; from the length The deep representation is then projected onto C variables; FMM is a collection of different predictors based on past information from different scales, enabling FMM to integrate the complementary predictive capabilities of mixed multi-scale sequences.

[0082] Furthermore, the fully connected multilayer perceptron MLP organically connects the PDM module and the FMM module, and uses the self-attention mechanism to improve the MLP to increase the connection efficiency and stability; the calculation formula of the self-attention mechanism SELF is:

[0083] U l =SELF(LN(Z l-1 ))+Z l-1 , i=1,…,l

[0084] Where LN represents layer normalization, l represents the number of encoding layers in the encoder, and U l Represents the self-attention output value corresponding to the l-th encoding layer;

[0085] Feature Set For each element, the self-attention mechanism calculates a weight vector W i , indicating the importance of this element to other elements in the sequence:

[0086]

[0087] Among them, Q i ,K i ,V i are the query, key, and value vectors obtained by calculation, d k is the dimension of the attention head, and the softmax function ensures the normalization of the attention weights; finally, through the value vector V of each element i Multiply it by its corresponding attention weight W i And sum them up to get the final output:

[0088]

[0089] Furthermore, the improved difference creative search algorithm DCS implementation process in step (5) is as follows:

[0090] Population initialization: In DCS, the optimization process starts with a set of candidate solutions x, which are randomly generated between the upper limit U and the lower limit L of the optimization problem. The initialization formula is as follows:

[0091] x i,d =L d +U(0,1)×(U d -L d )

[0092] Among them, x i,d Indicates that the i-th candidate solution is in the dimension of d, L d represents the next term of the dth dimension specified by the optimization problem, U drepresents the upper bound of the d-th dimension specified by the optimization problem, and U(0,1) represents the uniform distribution on the interval (0, 1);

[0093] Based on differentiated knowledge acquisition, DKA focuses on the speed of acquiring new knowledge. The formula is as follows:

[0094]

[0095] Among them, η i,t The parameter represents the quantitative knowledge acquisition rate of the individual at the tth iteration; is the individual's variable at the tth iteration The coefficient value is calculated as follows:

[0096]

[0097] Among them, R i,t is the rank of the i-th individual at the beginning of the t-th iteration, and NP represents the number of individuals in the population; is a quantitative determinant of the degree of imperfection of an individual's knowledge. The coefficient varies with the size of the knowledge gap; higher values ​​mean larger knowledge gaps, The smaller the value, the less imperfect the knowledge;

[0098] DKA uses the following formula to achieve differentiation for each individual X:

[0099] j rand =randint(1,D)

[0100]

[0101] In the formula, y i,d Represents the test member y i,t The element on the dth dimension, y i,t is the experimental member at the tth iteration; j rand is an integer randomly selected from 1 to D and generated once for each individual sample i;

[0102] Convergence strategy: Based on the knowledge reserve of the top performers and incorporating the random contributions from the two team members introduced by the current individual, the formula is as follows:

[0103] y i,d =w×x best,d +λ t ×(x r2,d -x i,d )×ω i,t ×(x r1,d -x i,d )

[0104] Among them, x i,d Drawing inspiration from convergent thinking, by integrating team members x r1,d and x r2,d Provide information to distill the knowledge of team leaders; x best,d is the dth position of the best performing individual in the current iteration, ω represents the cognitive weight of the best performance; ω i,t is the ω coefficient value of the individual at the tth iteration; x r1,d is the dth position of a randomly selected individual in {1,2,…,NP}, x r1,d ≠x i,d ≠x best,d best represents three different individuals; t is the λ coefficient value at the tth iteration, x r2,d is the dth position of a randomly selected individual in {ng+1,…,NP}, x r2,d ≠x r1,d ≠x i,d ≠x best,d ; The parameter ngs indicates the amount of high performance, determined from 100p%;

[0105]

[0106] Among them, NFE t Indicates the number of current function evaluations at time t, NFE max Indicates the maximum number of function evaluations;

[0107] Divergent thinking: Introduce the optimization speed and resource utilization efficiency of the triangle walk strategy optimization algorithm to complete the optimization iteration process faster. The formula is as follows:

[0108] y i,d =x r1,d +Lk(α,σ)

[0109] L1=pos b (t-pos c (t)

[0110]

[0111] Where Lk(α,σ) represents a Linnik distribution random number generator with control parameters α and σ, y i,d is the test vector, x r1,d is existing knowledge, team member x r1,d and x r2,d The knowledge difference is L1, the iteration step of the population is L2, and “→” represents the knowledge difference according to x i,d The position of the individual in the team is used to calculate the moving direction of each individual, pos c(t) is the position of the individual in the team at time t, pos b (t) is the nearest and most important neighbor I ii location;

[0112] Boundary processing:

[0113]

[0114] Among them, y i,d Indicates the test member v i,d The element at the dth position, v i,d is the trial member vi at the tth iteration; L d and U d Respectively represent the lower and upper bounds of the d-th dimension;

[0115] The diversification of team members and the generation of new members follow the following formula:

[0116] y NP =L+U(0,1)×(UL)

[0117] Among them, y NP represents the npth vector; U(0,1) represents the uniform distribution on the interval (0,1), L and U are the lower and upper bounds of the optimization problem;

[0118] Retrospective evaluation: By analyzing past performance data to identify trends and areas for improvement, the efficiency of the selection mechanism and tracking of best performance can be improved, which can significantly improve the algorithm's optimization speed and optimization effect:

[0119]

[0120] Furthermore, the improved difference creative search algorithm DCS in step (5) optimizes the hyperparameters of the TimeMixer model as follows:

[0121] The learning rate and GELU activation function of the PDM module of the TimeMixer model are optimized. The objective function formula is as follows:

[0122] f(lr,β)=GELU(TimeMixer(lr,β))

[0123] Among them, lr represents the step size in the model training process, β represents the GELU activation function parameter, which controls the smoothness of the GELU activation function, and f(·) represents the objective function.

[0124] Compared with the prior art, the present invention has the following beneficial effects:

[0125] 1. Using improved KPCA to extract features from fuel cell data and map the data to a high-dimensional space helps reveal the inherent structure and potential nonlinear relationships of the data, improves the prediction model's ability to identify complex data patterns, and thus improves prediction accuracy;

[0126] 2. Use the improved differential creative search algorithm (DCS) to optimize the hyperparameters of the TimeMixer model improved by the self-attention mechanism. This method can quickly search for the optimal hyperparameter combination, thereby improving the prediction accuracy of the model. Use TVF-EMD technology to decompose the fuel cell data and decompose the complex time series into multiple sub-modals, which helps to reduce the non-stationarity of the data and enables the prediction model to more accurately capture the changing trend of the health status of the fuel cell; reduce the number of sequences through fuzzy entropy aggregation, which not only improves the efficiency of data processing, but also helps to reduce the impact of noise and outliers, making the prediction model more robust and improving the accuracy of health status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0127] Figure 1 is a flow chart of the present invention;

[0128] Figure 2 The network topology diagram of the TimeMixer model optimized for the self-attention mechanism;

[0129] Figure 3 This is the network topology diagram of the S-DCS algorithm. DETAILED DESCRIPTION

[0130] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0131] like Figure 1 As shown, the present invention proposes a method for evaluating the health status of a fuel cell, comprising the following steps:

[0132] Step 1: Obtain fuel cell operating data and historical operating data, including fuel cell load current, maximum operating temperature, air flow, gas pressure, etc.; monitor operating parameters such as voltage, current, temperature, pressure and flow online to monitor the status of the fuel cell in real time.

[0133] Step 2: Use improved Kernel Principal Component Analysis (iKPCA) to extract features from fuel cell data, including 20 types of data features such as voltage, current, temperature, pressure, flow, and air flow. Mapping the data to a high-dimensional space can better reveal the intrinsic structure of the data.

[0134] Fuel cell parameter data often contain nonlinear relationships and complex patterns, which makes traditional linear methods (such as principal component analysis PCA) may not be able to fully capture all the important features in the data. iKPCA can more accurately reveal the intrinsic structure of the data by mapping the data into a high-dimensional space using a kernel method.

[0135] Define a time series of input signal as x a (a=1,2,…,m), linearly independent variable y a , (a=1,2,…,n) and n≤m. The singular value decomposition can be used to obtain x a , then its expression is:

[0136] U=[x1,x2,...,x m ]

[0137] V=[y1,y2,...,y n ]

[0138] U=CV

[0139] Among them, U is the input data matrix, C is the n×n transformation matrix, and V is the output observation matrix. From this matrix, we can get:

[0140] R x =BR y B T

[0141] Since the autocorrelation matrix R x is a real symmetric matrix, then the autocorrelation matrix R y Convertible to R x There are m positive real eigenvalues ​​ζ in i (i=1,2,…,m), B is an n-dimensional vector. That is:

[0142]

[0143] ζ1>ξ2>...>ξ m

[0144] In the autocorrelation matrix R y In , m eigenvalues ​​correspond to m eigenvectors, and m eigenvectors constitute m subspaces. In order to preserve the original information in the m eigenvectors, m should satisfy:

[0145]

[0146] From the above formula, we can see that the eigenvalue ζ i The information contained in the original signal is compressed, so the fault mode can be classified by eigenvalues ​​and eigenvectors.

[0147] Assume a nonlinear mapping function φ n (x), and the original feature set W = [w1,w2,…,w s ] T Transformed into the kernel feature set W' = [φ(w1),φ(w2),…,φ(w s )] T ; where s is the number of samples. Then the covariance matrix F of the kernel feature set is expressed as:

[0148]

[0149]

[0150] Among them, ψ(w i ) represents the mapping function, which maps the input data to a high-dimensional space. F is the covariance matrix of the kernel feature set. In KPCA, the F matrix is ​​used to capture the main direction of change of the data, and then the eigenvalue λ of the matrix F is used to i and the eigenvector ζλ i , we can get the expression:

[0151]

[0152]

[0153] Among them, b i Represents the coefficient used to construct the kernel feature vector, which maps each sample to the kernel feature set W' and the feature vector ζ i The dot product of .

[0154] For the inner product operation of KPCA in high-dimensional space, Gaussian radial basis function is selected As a kernel function:

[0155]

[0156] in, is the inner product of the two mapped high-dimensional eigenvectors, and then the q eigenvectors corresponding to the q largest eigenvalues ​​are selected as the column vectors of the transformation matrix γ, γ T is the transposed matrix of the transformation matrix, then the result after KPCA dimensionality reduction is:

[0157] R = γ T ·W'

[0158] In order to measure the interval between samples of the same class and samples of different classes, the degree of intra-class aggregation is defined, and its expression is:

[0159]

[0160] Among them, ε is the number of samples in the i-th category, is the standard deviation of the a-th dimension component of the category; at the same time, the discrete degree expression between samples of different categories is:

[0161]

[0162]

[0163] Among them, ε is the number of samples in the i-th category, β aij is the a-th dimension component of the j-th sample in the i-th category, and the degree of separability of sample categories is:

[0164]

[0165] When J a The smaller the value of is, the higher the degree of separability of the sample categories is, and the greater the contribution to sample classification is. In order to illustrate the correlation between the extracted feature vector and the sample category vector, the expression is defined as:

[0166]

[0167] Among them, τ is the correlation coefficient ζ i is the extracted feature vector, ζ s is the sample category vector in the original feature set; when the absolute value of τ approaches 1, it indicates that the correlation between the two feature vectors is very strong, which means that the two features have similar information in describing the fuel cell data. When the absolute value of τ approaches 0, it indicates that the correlation between the two feature vectors is very weak, which indicates that they provide complementary information in describing the data. Based on the value of τ, we can select feature vectors that contribute highly to the classification of fuel cell data. These feature vectors can better capture the key information of the data, thus helping to improve the accuracy of classification.

[0168] Step 3: In view of the random volatility and uncertainty of fuel cell data, the TVF-EMD technology is used to decompose the fuel cell data to obtain each sub-mode, reduce the non-stationarity of the sequence, and reduce the number of sequences through fuzzy entropy aggregation to improve data processing efficiency.

[0169] The traditional data deconstruction technology, empirical mode decomposition (EMD), may encounter some problems of incomplete or over-decomposition in certain circumstances. By using TVF-EMD, the dynamic changes of the signal can be captured more accurately, avoiding the problems of incomplete or over-decomposition that may occur in the traditional EMD method, better retaining the original characteristics of the signal, while reducing the feature mixing phenomenon, and improving the accuracy and reliability of data decomposition. This screening is achieved through time-varying filtered empirical mode distribution. TVF-EMD uses adaptive decomposition to replace the intrinsic mode component (IMF) with a local narrowband signal. The formula is as follows:

[0170]

[0171] Among them, IMF i (t) represents the i-th IMF, x(t) represents the given fuel cell data signal, and r(t) represents the residual. The formula for the time-varying filter empirical mode distribution is as follows:

[0172]

[0173] Among them, u i Indicates the maximum value of x(t), expressed as u i , i=1,2,3,... represents the second-order resolution, and ρ is the preset threshold for the frequency change between two consecutive maxima. Find the local cutoff frequency Interpolate between the peaks to get Adjustment Improved modal aliasing. )Reconstruction Get the new signal h(t):

[0174]

[0175] Calculate the criterion value θ(t) to determine whether the residual signal meets the cutoff standard.

[0176]

[0177] Where θ(t) represents the calculation criterion value, cos(·) represents the cosine function, and B Loughlin (t) represents the Loughlin instantaneous bandwidth, represents the weighted average instantaneous frequency.

[0178] TVF-EMD technology decomposes fuel cell data to obtain various sub-modes. The increase in modes is not conducive to model training data. Fuzzy entropy aggregation is needed to reduce the number of sequences and improve data processing efficiency. The formula is as follows:

[0179] H(D,B)=-∑(μ(x)*log(μ(x)))

[0180] Among them, H(D,B) represents the fuzzy entropy of sets D and B, μ(x) represents the membership of the fuzzy set, and log(·) represents the logarithmic function.

[0181] The use of TVF-EMD can solve the problems that when data is over-decomposed, the key features of the original data may be lost, thus affecting the accuracy of data deconstruction; the decomposition process is not carried out thoroughly, which may lead to the mixing of data features, making it difficult to accurately describe the data features. Fuzzy entropy aggregation reduces the number of sequences and improves the efficiency of the model in data processing.

[0182] Step 4: A decomposable multi-scale fusion time series prediction model (TimeMixer) improved by the self-attention mechanism is proposed. The model can mix multi-scale seasonal and trend components from fine to coarse and from coarse to fine, comprehensively extract information from time series data, and enhance the robustness of prediction.

[0183] like Figure 2 As shown in the figure, the TimeMixer model designs a multi-scale hybrid architecture to deal with the complex time dynamic problems encountered in time series forecasting. This model is mainly based on a fully connected multi-layer perceptron (MLP) structure and consists of two core parts: past decomposable mixture (PDM) and future multi-predictor mixture (FMM). This design enables the model to efficiently integrate sequence data of different time scales to improve the accuracy of prediction.

[0184] First, the historical data of fuel cell operating parameters, operating parameters, environmental parameters, etc. are used as observation data for the model. Downsample and finally obtain a multi-scale time series set through average pooling in m∈{0,1,…,M}, C represents the number of variables, P is the length of the time series, that is, the total number of observed data points, the lowest level sequence j0=j is the input sequence, containing the most detailed time changes, and the highest level sequence j M Represents macroscopic changes. These multi-scale sequences are then projected into the embedding layer to obtain deep features The formula is:

[0185]

[0186] Among them, Embed is the embedding process, which means that the data is transformed from its original form or space to a different feature space.

[0187] Using stacked past decomposable mixture (PDM) blocks to mix past information of different scales, for the Lth layer, the input is The process through PDM can be formulated as:

[0188]

[0189] Where L is the total number of layers, Indicates mixed past expressions, with d model For the future prediction stage, a future multi-predictor mixture (FMM) block is used to integrate the extracted multi-scale past information And generate future predictions, namely:

[0190]

[0191] in, represents the final prediction, and F is the number of predicted time steps, i.e., the total number of future time points that the model will predict. Through PDM and FMM, TimeMixer successfully captures key historical information from multi-scale observational data and predicts the future based on it, making full use of the advantages of multi-scale historical data.

[0192] The PDM module is responsible for extracting past information and mixing seasonal and trend components at different scales. Specifically, for each PDM block, the multi-scale time series It is decomposed into seasonal and trend parts, and the mixing operation is applied to seasonality and trend to interact with information from multiple scales respectively. The specific formula is as follows:

[0193]

[0194]

[0195] in, For the seasonal part, For the trend part, FeedForward is a feedforward network structure, which includes two linear layers with a GELU activation function in the middle for information interaction between channels. S-Mix(·) and T-Mix(·) represent seasonal and trend mixing, respectively.

[0196] Seasonal mixing: Using a bottom-up approach, integrating detailed time series data from lower levels helps provide additional details for high-level seasonal patterns. A bottom-up mixing layer is used to achieve bottom-up interaction of seasonal information in a residual manner from the mth scale, and the formula is expressed as:

[0197]

[0198] Among them, Bottom-Up-Mixing (·) is two linear layers with activation functions in the middle. Along the time dimension, its input dimension and output dimension are both

[0199] Trend mixing: For the trend part, the detailed changes of the trend items may introduce noise and affect the capture of the macro trend. The coarse-scale time series at a higher level can more easily provide clear macro information. Therefore, a top-down approach is adopted, and a top-down mixing layer is used to realize the top-down trend information interaction in a residual manner at the mth scale. The formula is as follows:

[0200]

[0201] Among them, Top-Down-Mixing (·) is two linear layers with an intermediate GELU activation function, and its input dimension is The output dimension is Empowered by a mixture of seasonality and trends, PDM gradually aggregates seasonal data from micro-scales, while combining macro-trends with historical experience, ultimately achieving multi-scale fusion in the process of extracting historical information.

[0202] Given that sequences at each scale show different dominant changes and their specific advantages in prediction, the FMM module aggregates the prediction capabilities of multi-scale sequences and presents this through the future multi-predictor hybrid block. The specific execution formula is as follows:

[0203]

[0204] in, represents the future prediction from the m-th scale sequence, and the final output is Predictor m (·) indicates that at the mth scale, the predictor first uses a single-layer linear transformation to directly predict the future time steps of length F. The deep representation is then projected onto C variables. Note that FMM is an ensemble where different predictors are based on past information from different scales, enabling FMM to integrate complementary prediction capabilities of mixed multi-scale sequences.

[0205] The TimeMixer model organically connects the PDM module and the FMM module based on a fully connected multi-layer perceptron (MLP) structure, and uses the self-attention mechanism to improve the MLP to increase the connection efficiency and stability. The calculation formula of the self-attention mechanism SELF is:

[0206] U l =SELF(LN(Z l-1 ))+Z l-1 , i=1,…,l

[0207] Where LN represents layer normalization, l represents the number of encoding layers in the encoder, and U lRepresents the self-attention output value corresponding to the l-th encoding layer;

[0208] Feature Set For each element, the self-attention mechanism calculates a weight vector W i , indicating the importance of this element to other elements in the sequence:

[0209]

[0210] Among them, Q i ,K i ,V i are the query, key, and value vectors obtained by calculation, d k is the dimension of the attention head, and the softmax function ensures the normalization of the attention weights; finally, through the value vector V of each element i Multiply it by its corresponding attention weight W i And sum them up to get the final output:

[0211]

[0212] Step 5: Figure 3 As shown in the figure, the improved differential creative search algorithm (S-DCS) is applied to optimize the hyperparameters of the TimeMixer model, including learning rate, activation function, etc., to quickly search for the optimal hyperparameter combination and improve the prediction accuracy of the model.

[0213] Differentiated Creative Search (DCS) uses differentiated knowledge acquisition and creative realism strategies to solve complex optimization problems. DCS effectively optimizes in a broad search space through mathematical modeling and implementation. It has strong evolutionary capabilities, fast search speed, and excellent optimization capabilities in solving complex optimization problems.

[0214] Population initialization: In DCS, the optimization process starts with a set of candidate solutions x, which are randomly generated between the upper limit U and the lower limit L of the optimization problem. The initialization formula is as follows:

[0215] x i,d =L d +U(0,1)×(U d -L d )

[0216] Among them, x i,d Indicates that the i-th candidate solution is in the dimension of d, L d represents the next term of the dth dimension specified by the optimization problem, U d represents the upper bound of the d-th dimension specified by the optimization problem, and U(0,1) represents the uniform distribution on the interval (0, 1).

[0217] Differentiated knowledge-acquisition (DKA) is rooted in the concept of "different potentials of human beings to acquire knowledge". It breaks the random mechanism of traditional differential evolution algorithm and pays more attention to the speed of acquiring new knowledge. The formula is as follows:

[0218]

[0219] Among them, η i,t The parameter represents the quantitative knowledge acquisition rate of the individual at the tth iteration. is the individual's variable at the tth iteration The coefficient value is calculated as follows:

[0220]

[0221] Among them, R i,t is the rank of the i-th individual at the beginning of the t-th iteration, and NP represents the number of individuals in the population. The coefficient is a quantitative determinant of the degree to which an individual's knowledge is imperfect, and this approach, when working with fuel cell data, can help identify and highlight areas for improvement, thereby determining learning needs and guiding targeted interventions. The coefficient varies with the size of the knowledge gap; higher values ​​mean a larger knowledge gap, indicating that individuals need to learn, absorb, and integrate new knowledge or experience. The smaller the value, the less imperfect the knowledge is, which means that the individual's knowledge base is more comprehensive and solid. The DCS algorithm's differentiated knowledge acquisition simulates this human skill. The coefficients can highlight areas for improvement, identify learning needs, and even guide targeted interventions to enhance the algorithm’s problem-solving abilities.

[0222] DKA uses the following formula to achieve differentiation for each individual X:

[0223] j rand =randint(1,D)

[0224]

[0225] In the formula, y i,d Represents the test member y i,t The element on the dth dimension, y i,t is the experimental member at the tth iteration. rand is an integer randomly selected from 1 to D and is generated once for each individual sample i.

[0226] Convergent thinking: This strategy is based on the knowledge of the top performers and incorporates the random contributions of the two team members introduced by the current individual. In this way, it enhances diversity and promotes convergence to the optimal solution in the process of solution formation. The formula is as follows:

[0227] y i,d =w×x best,d +λ t ×(x r2,d -x i,d )+ω i,t ×(x r1,d -x i,d )

[0228] Among them, x i,d Drawing inspiration from convergent thinking, by integrating team members x r1,d and x r2,d Provides information that distills the knowledge of team leaders. best,d is the dth position of the best performing individual in the current iteration, ω represents the cognitive weight of the best performance, and the default value is 1. i,t is the ω coefficient value of the individual at the tth iteration. r1,d is the dth position of a randomly selected individual in {1,2,…,NP}, x r1,d ≠x i,d ≠x best,d best represents three different individuals; t is the λ coefficient value at the tth iteration, x r2,d is the dth position of a randomly selected individual in {ng+1,…,NP}, x r2,d ≠x r1,d ≠x i,d ≠x best,d The parameter ngs represents the number of high performances, starting from 100p%. For larger populations, 'p' is set to 0.2, i.e. 20% of the individuals in the population are considered high-performing individuals. For populations with a population size of no more than 30 individuals, the minimum value of NGS is 6 to ensure that the algorithm can effectively utilize high-performing individuals.

[0229]

[0230] Among them, NFE t Indicates the number of current function evaluations at time t, NFE max Indicates the maximum number of function evaluations.

[0231] Divergent thinking: There are two possible transformations of X. First, X may move to a new position in the divergent region under the influence of Xe and Linnik (α, σ). Alternatively, it may be attracted to a point near Xe.

[0232] y i,d =x r1,d +Lk(α,σ)

[0233] Where Lk(α,σ) represents a random number generator with Linnik distribution whose control parameters are α and σ. The new strategy means that the trial vector (y i,d ) is the existing knowledge (x r1,d ) and a mixture of novelty or creative elements Lk(α,σ). Linnik distribution increases the potential to explore less frequent areas in the search space. However, this divergent thinking will increase the algorithm's optimization time and slow down the optimization speed. Therefore, the introduction of the triangle walk strategy optimizes the algorithm's optimization speed and resource utilization efficiency, and completes the optimization iteration process faster. The formula is as follows:

[0234] L1=pos b (t)-pos c (t)

[0235]

[0236] Among them, team members x r1,d and x r2,d The knowledge difference is L1, the iteration step of the population is L2, and “→” represents the knowledge difference according to x i,d The position of the individual in the team is used to calculate the moving direction of each individual, pos c (t) is the position of the individual in the team at time t, pos b (t) is the nearest and most important neighbor I ii location.

[0237] Boundary processing: The role of the boundary processing mechanism is to ensure that the generated solutions remain within the feasible and realistic range. If the position reached by an individual in the dth dimension exceeds the actual allowed limit, the mechanism will adjust the value to the nearest boundary value, whether it is the maximum or minimum value. This processing method is often called boundary interpolation, which avoids the generation of unrealistic or unfeasible ideas. The formula is as follows:

[0238]

[0239] Among them, y i,d Indicates the test member v i,d The element at the dth position (dimension), v i,d is the trial member vi at the tth iteration. d and U d They represent the lower and upper bounds of the d-th dimension, respectively, which are determined by the optimization problem.

[0240] Diversity of team members: The continuous updating of team members can inspire more diverse creativity. Based on this, the algorithm regularly eliminates those members with poor performance and introduces new members to inject fresh blood. The generation of new members follows the following formula:

[0241] y NP =L+U(0,1)×(UL)

[0242] Among them, y NP represents the npth vector. U(0,1) represents the uniform distribution on the interval (0,1), and L and U are the lower and upper bounds of the optimization problem.

[0243] Retrospective evaluation: By analyzing past performance data to identify trends and areas for improvement, the efficiency of the selection mechanism and tracking of best performance has been improved, which significantly improves the algorithm's optimization speed and optimization effect:

[0244]

[0245] The hyperparameters of the PDM module of the TimeMixer model are optimized by applying the differential creative search algorithm (DCS) improved by the triangle walk strategy, including the learning rate and GELU activation function. The objective function formula is as follows:

[0246] f(lr,β)=GELU(TimeMixer(lr,β))

[0247] Among them, lr represents the step size in the model training process, β represents the GELU activation function parameter, which controls the smoothness of the GELU activation function, and f(·) represents the objective function. The differential creative search algorithm (DCS) improved by the triangle walk strategy is applied to optimize the hyperparameters of the PDM module of the TimeMixer model, so that the TimeMixer model can show strong evolutionary ability, fast search speed and excellent optimization ability in solving complex optimization problems.

[0248] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for evaluating the health status of a fuel cell, characterized in that: The following steps are involved: (1) Obtain fuel cell operation data, environmental parameters, and historical operation data to monitor the status of the fuel cell in real time; (2) Using improved kernel principal component analysis (iKPCA) to extract features of fuel cell operating parameters, the data is mapped into a high-dimensional space to reveal the intrinsic structure of the data; (3) In view of the random volatility and uncertainty of fuel cell operation data, the TVF-EMD technology is used to decompose the fuel cell data to obtain sub-modes and reduce the non-stationarity of the sequence; and the number of sequences is reduced through fuzzy entropy aggregation to improve data processing efficiency; (4) A decomposable multi-scale fusion time series prediction model TimeMixer is improved by self-attention mechanism, which mixes multi-scale seasonal and trend components from fine to coarse and from coarse to fine to fully extract information from time series data; (5) The improved differential creative search algorithm DCS is used to optimize the hyperparameters of the TimeMixer model, quickly search for the optimal hyperparameter combination, and achieve accurate assessment of the health status of the fuel cell.

2. A fuel cell health status assessment method according to claim 1, characterized in that: The fuel cell operating data and load current of step (1) include online monitoring of voltage, current, temperature, pressure and flow; the environmental parameters include maximum operating temperature, air flow and gas pressure.

3. A fuel cell health status assessment method according to claim 1, characterized in that: The implementation process of step (2) is as follows: Define a time series of input signal as x a , a=1,2,…,m, linearly independent variable y a , a=1,2,…,n, and n≤m; obtain x by singular value decomposition a , then its expression is: U=[x1,x2,...,x m ] V=[y1,y2,...,y n ] U=CV Among them, U is the input data matrix, C is the n×n transformation matrix, and V is the output observation matrix; from this matrix, we get: R x =BR y B T Since the autocorrelation matrix R x is a real symmetric matrix, then the autocorrelation matrix R y Convertible to R x There are m positive real eigenvalues ​​ξ in i , i=1,2,…,m, and B is an n-dimensional vector; that is: R y =[E(y i ) 2 ]=[ξ i ] ξ1>ξ2>...>ξ m In the autocorrelation matrix R y In , m eigenvalues ​​correspond to m eigenvectors, and m eigenvectors constitute m subspaces, and m should satisfy: From the above formula, we can see that the eigenvalue ξ i The information contained in the original signal is compressed, and the fault mode is classified by eigenvalue and eigenvector; Assume a nonlinear mapping function φ n (x), and the original feature set W = [w1,w2,…,w s ] T Transformed into the kernel feature set W' = [φ(w1),φ(w2),…,φ(w s )] T ; where s is the number of samples; then the covariance matrix F of the kernel feature set is expressed as: Among them, ψ(w i ) represents the mapping function, which maps the input data to a high-dimensional space; F is the covariance matrix of the kernel feature set. In KPCA, the F matrix is ​​used to capture the main direction of change of the data; then the eigenvalue λ of the matrix F is used to i and the eigenvector ζ i , we get the expression: Among them, b i Represents the coefficient used to construct the kernel feature vector, which maps each sample to the kernel feature set W' and the feature vector ζ i The dot product of For the inner product operation of KPCA in high-dimensional space, Gaussian radial basis function is selected As a kernel function: in, is the inner product of the two mapped high-dimensional eigenvectors, and then the q eigenvectors corresponding to the q largest eigenvalues ​​are selected as the column vectors of the transformation matrix γ, γ T is the transposed matrix of the transformation matrix, then the result after KPCA dimensionality reduction is: R=γ T ·W' In order to measure the interval between samples of the same class and samples of different classes, the degree of intra-class aggregation is defined: Among them, ε is the number of samples in the i-th category, is the standard deviation of the a-th dimension component of the category; at the same time, the discrete degree expression between samples of different categories is: Among them, ε is the number of samples in the i-th category, β aij is the a-th dimension component of the j-th sample in the i-th category, and the degree of separability of sample categories is: When J a The smaller the value of is, the higher the degree of separability of the sample categories is, and the greater the contribution to sample classification is; the correlation between the extracted feature vector and the sample category vector is: Among them, τ is the correlation coefficient ζ i is the extracted feature vector, ζ s is the sample category vector in the original feature set; when the absolute value of τ approaches 1, it indicates that the correlation between the two feature vectors is very strong, which means that the two features have similar information in describing the fuel cell data; when the absolute value of τ approaches 0, it indicates that the correlation between the two feature vectors is very weak, which indicates that they provide complementary information in describing the data.

4. A fuel cell health status assessment method according to claim 1, characterized in that: The implementation process of step (3) is as follows: By using adaptive decomposition, the intrinsic modal component IMF is replaced by a local narrowband signal, and the formula is as follows: Among them, IMF i (t) represents the i-th IMF, x(t) represents the given fuel cell data signal, and r(t) represents the residual; the time-varying filter empirical mode distribution formula is as follows: Among them, u i represents the maximum value of x(t), i=1, 2, 3, ...; represents the second-order resolution, ρ is the preset threshold of the frequency change between two consecutive maximum values; find the local cutoff frequency Interpolate between the peaks to get Adjustment Improve modal aliasing; refactor Get the new signal h(t): Calculate the criterion value θ(t) to determine whether the residual signal meets the cutoff standard: Where θ(t) represents the calculation criterion value, cos(·) represents the cosine function, and B Loughlin (t) represents the Loughlin instantaneous bandwidth, represents the weighted average instantaneous frequency; The number of sequences is reduced by fuzzy entropy aggregation, the formula is as follows: H(D,B)=-∑(μ(x)*log(μ(x))) Among them, H(D,B) represents the fuzzy entropy of sets D and B, μ(x) represents the membership of the fuzzy set, and log(·) represents the logarithmic function.

5. A fuel cell health status assessment method according to claim 1, characterized in that: The decomposable multi-scale fusion time series prediction model TimeMixer in step (4) is based on a fully connected multi-layer perceptron MLP to capture and learn complex patterns in data; and adopts past decomposable hybrid PDM and future multi-predictor hybrid FMM, uses multi-scale historical data, and integrates sequence data of different time scales to predict the future; the implementation process is as follows: First, the historical data of fuel cell operating parameters, operating parameters, environmental parameters, etc. are used as observation data for the model. Downsample and finally obtain a multi-scale time series set through average pooling in C represents the number of variables, P is the length of the time series, that is, the total number of observed data points, the lowest level sequence j0=j is the input sequence, containing the most detailed time changes, and the highest level sequence j M Represents macro changes; Projecting multi-scale sequences into the embedding layer to obtain deep features The formula is: Among them, Embed is the embedding process, which means that the data is transformed from its original form or space to a different feature space; Using stacked PDM blocks to mix past information of different scales, for the Lth layer, the input is The process of PDM is formulated as: Where L is the total number of layers, Indicates mixed past expressions, with d model aisle; For the future prediction stage, the FMM block is used to integrate the extracted multi-scale past information And generate future predictions, namely: in, represents the final prediction and F is the number of predicted time steps, i.e. the total number of future time points that the model will predict.

6. A fuel cell health status assessment method according to claim 5, characterized in that: The implementation process of the past decomposable hybrid PDM is as follows: in, For the seasonal part, For the trend part, FeedForward is a feedforward network structure, which includes two linear layers with a GELU activation function in the middle for information interaction between channels. S-Mix(·) and T-Mix(·) represent seasonal and trend mixing, respectively; Seasonal mixing uses a bottom-up mixing layer to achieve bottom-up interaction of seasonal information in a residual manner from the mth scale: Among them, Bottom-Up-Mixing (·) is two linear layers with activation functions in the middle. Along the time dimension, its input dimension and output dimension are both For trend mixing, a top-down mixing layer is used to achieve top-down trend information interaction in a residual manner at the mth scale: In the formula, Top-Down-Mixing (·) is two linear layers with an intermediate GELU activation function, and its input dimension is The output dimension is Empowered by a mixture of seasonality and trends, PDM gradually aggregates seasonal data from micro-scales, while combining macro-trends with historical experience, ultimately achieving multi-scale fusion in the process of extracting historical information.

7. A fuel cell health status assessment method according to claim 5, characterized in that: The future multi-predictor hybrid FMM implementation process is as follows: in, represents the future prediction from the m-th scale sequence, and the final output is Predictor m (·) indicates the mth scale; the predictor first uses a single-layer linear transformation to directly predict the future time step of length F; from the length The deep representation is then projected onto C variables; FMM is a collection of different predictors based on past information from different scales, enabling FMM to integrate the complementary prediction capabilities of mixed multi-scale sequences.

8. A fuel cell health status assessment method according to claim 5, characterized in that: The fully connected multi-layer perceptron MLP organically connects the PDM module and the FMM module, and uses the self-attention mechanism to improve the MLP to increase the connection efficiency and stability; the calculation formula of the self-attention mechanism SELF is: U l =SELF(LN(Z l-1 ))+Z l-1 ,i=1,…,l Among them, LN represents layer normalization, l represents the number of encoding layers in the encoder, and U l Represents the self-attention output value corresponding to the l-th encoding layer; Feature Set For each element, the self-attention mechanism calculates a weight vector W i , indicating the importance of this element to other elements in the sequence: Among them, Q i ,K i ,V i are the query, key, and value vectors obtained by calculation, d k is the dimension of the attention head, and the softmax function ensures the normalization of the attention weights; finally, through the value vector V of each element i Multiply it by its corresponding attention weight W i And sum them up to get the final output:

9. A fuel cell health status assessment method according to claim 1, characterized in that: The improved difference creative search algorithm DCS implementation process in step (5) is as follows: Population initialization: In DCS, the optimization process starts with a set of candidate solutions x, which are randomly generated between the upper limit U and the lower limit L of the optimization problem. The initialization formula is as follows: x i,d =L d +U(0,1)×(U d -L d ) Among them, x i,d Indicates that the i-th candidate solution is in the dimension of d, L d represents the next term of the dth dimension specified by the optimization problem, U d represents the upper bound of the d-th dimension specified by the optimization problem, and U(0,1) represents the uniform distribution on the interval (0, 1); Based on differentiated knowledge acquisition, DKA focuses on the speed of acquiring new knowledge. The formula is as follows: Among them, η i,t The parameter represents the quantitative knowledge acquisition rate of the individual at the tth iteration; is the individual's variable at the tth iteration The coefficient value is calculated as follows: Among them, R i,t is the rank of the i-th individual at the beginning of the t-th iteration, and NP represents the number of individuals in the population; is a quantitative determinant of the degree of imperfection of an individual's knowledge. The coefficient varies with the size of the knowledge gap; higher values ​​mean larger knowledge gaps, The smaller the value, the less imperfect the knowledge; DKA uses the following formula to achieve differentiation for each individual X: j rand =randint(1,D) In the formula, y i,d Represents the test member y i,t The element on the dth dimension, y i,t is the experimental member at the tth iteration; j rand is an integer randomly selected from 1 to D and generated once for each individual sample i; Convergence strategy: Based on the knowledge reserve of the top performers and incorporating the random contributions from the two team members introduced by the current individual, the formula is as follows: y i,d =w×x best,d +λ t ×(x r2,d -x i,d )+ω i,t ×(x r1,d -x i,d ) Among them, x i,d Drawing inspiration from convergent thinking, by integrating team members x r1,d and x r2,d Provide information to distill the knowledge of team leaders; x best,d is the dth position of the best performing individual in the current iteration, ω represents the cognitive weight of the best performance; ω i,t is the ω coefficient value of the individual at the tth iteration; x r1,d is the dth position of a randomly selected individual in {1,2,…,NP}, x r1,d ≠x i,d ≠x best,d best represents three different individuals; t is the λ coefficient value at the tth iteration, x r2,d is the dth position of a randomly selected individual in {ng+1,…,NP}, x r2,d ≠x r1,d ≠x i,d ≠x best,d ; The parameter ngs indicates the amount of high performance, determined from 100p%; Among them, NFE t Indicates the number of current function evaluations at time t, NFE max Indicates the maximum number of function evaluations; Divergent thinking: Introduce the optimization speed and resource utilization efficiency of the triangle walk strategy optimization algorithm to complete the optimization iteration process faster. The formula is as follows: y i,d =x r1,d +Lk(a,s) L1=pos b (t)-pos c (t) Where Lk(α,σ) represents a Linnik distribution random number generator with control parameters α and σ, y i,d is the test vector, x r1,d is existing knowledge, team member x r1,d and x r2,d The knowledge difference is L1, the iteration step of the population is L2, and "→" represents the i,d The position of the individual in the team is used to calculate the moving direction of each individual, pos c (t) is the position of the individual in the team at time t, pos b (t) is the nearest and most important neighbor I ii location; Boundary processing: Among them, y i,d Indicates the test member v i,d The element at the dth position, v i,d is the trial member vi at the tth iteration; L d and U d Respectively represent the lower and upper bounds of the d-th dimension; The diversification of team members and the generation of new members follow the following formula: y NP =L+U(0,1)×(UL) Among them, y NP represents the npth vector; U(0,1) represents the uniform distribution on the interval (0,1), L and U are the lower and upper bounds of the optimization problem; Retrospective evaluation: By analyzing past performance data to identify trends and areas for improvement, the efficiency of the selection mechanism and tracking of best performance can be improved, which can significantly improve the algorithm's optimization speed and optimization effect:

10. A fuel cell health status assessment method according to claim 1, characterized in that: The improved difference creative search algorithm DCS in step (5) optimizes the hyperparameters of the TimeMixer model as follows: The learning rate and GELU activation function of the PDM module of the TimeMixer model are optimized. The objective function formula is as follows: f(lr,β)=GELU(TimeMixer(lr,β)) Among them, lr represents the step size in the model training process, β represents the GELU activation function parameter, which controls the smoothness of the GELU activation function, and f(·) represents the objective function.

Images