A method for evaluating state of health of a storage battery based on multi-dimensional parameters

Abstract

The present application belongs to the technical field of battery state monitoring, and particularly relates to a battery health state evaluation method based on multi-dimensional parameters, comprising: obtaining multi-dimensional measurement data of charge-discharge cycles; solving the optimal wavelet base function with minimum energy spectrum overlap degree through a differentiable programming, and extracting a feature vector, and using the feature vector to decompose the measurement data to obtain a time-frequency domain feature sequence; extracting a health index to form a time domain feature sequence, combining a self-attention mechanism with a period correlation adjustment matrix to generate a weighted time domain feature; inputting the weighted time domain feature into a double network channel to extract local time sequence correlation and global dependence relationship, and performing weighted fusion with the wavelet base feature vector as a static condition to obtain a fusion time domain feature; fusing the time-frequency and time domain features in multiple modes, generating a state fusion matrix through wavelet base feature modulation, and outputting an evaluation value by a fully connected network. The present application can improve the accuracy and stability of the battery health state evaluation result under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of battery condition monitoring technology. More specifically, this invention relates to a method for assessing battery health status based on multi-dimensional parameters. Background Technology

[0002] Batteries are widely used in new energy vehicles, smart grid energy storage systems, and backup power supplies. Their health status is an important basis for battery management systems to conduct safety warnings, range estimations, and maintenance decisions. As the number of charge-discharge cycles increases, internal side reactions gradually accumulate, typically manifesting as capacity decay, increased internal resistance, and changes in thermal characteristics. The degradation process is nonlinear, time-sequential, and involves multiple factors, including voltage, current, and temperature.

[0003] Currently, data-driven health status assessment methods are widely used due to their ability to avoid complex electrochemical mechanism modeling and their strong adaptability. Among them, the technical approach of combining time-frequency feature fusion with temporal neural networks has attracted attention. For example, Chinese patent document CN120802105B discloses a SOH assessment method based on time-frequency feature fusion and LSTM. This method extracts time-domain health indicators such as constant current charging time and constant voltage charging time from three operating conditions: full charge-discharge cycle, partial charge cycle, and partial charge-discharge cycle. It obtains the energy of each sub-frequency band as frequency-domain features through three-layer wavelet packet decomposition, uses Pearson correlation coefficient to screen highly correlated time-domain indicators, and then inputs them into an LSTM network model to achieve battery health status assessment. This method takes into account both time-domain evolution information and frequency-domain energy distribution characteristics under standard test conditions, solving the problem of health indicator failure under certain operating conditions.

[0004] However, the above-mentioned technical solutions still have some shortcomings: First, they use fixed wavelet packet basis functions for signal decomposition, which cannot adaptively adjust the filter coefficients according to the signal characteristics of different battery types or different aging stages. This results in a high degree of overlap in time-frequency domain features under different health states, making it difficult to effectively distinguish subtle degradation differences in batteries. Second, using only a single LSTM network for time series modeling makes it difficult to simultaneously consider the continuity of local degradation between adjacent cycles and the long-range degradation dependency across cycles, resulting in insufficient modeling of degradation stage differences caused by cycle evolution. Furthermore, the time-domain features and frequency-domain features are simply input into the network through concatenation, failing to fully explore the nonlinear interaction between the two modal features, which easily leads to information redundancy and affects the accuracy of the evaluation results.

[0005] In addition to the above-mentioned schemes, other existing health status assessment methods also generally have similar limitations: some methods rely too much on artificially constructed single time-domain indicators or conventional statistical features, making it difficult to fully characterize the latent frequency-domain micro-changes during long-term cycling; some schemes that introduce attention mechanisms do not combine the physical characteristics of battery cycle time to adjust the weights, resulting in attention allocation that does not match the actual degradation pattern; multi-source feature fusion methods usually use shallow network processing, which cannot effectively capture the deep coupling relationship between different modal features, and the assessment results are prone to fluctuations under complex working conditions.

[0006] Therefore, extracting more representative degradation features from multidimensional operational data and effectively fusing time-domain evolution information with time-frequency energy distribution information has become crucial for improving the performance of battery health status assessment. This field requires a battery health status assessment method that can adaptively select and adapt wavelet bases, extract latent time-frequency features, and combine cyclic periodic information, multi-path time-series modeling, and deep feature fusion mechanisms to improve the accuracy, stability, and reliability of the assessment results. Summary of the Invention

[0007] To address the problems of inaccurate assessment results under complex operating conditions, difficulty in distinguishing subtle degradation differences, and insufficient characterization of degradation stages in existing battery health status assessment methods, this invention provides a battery health status assessment method based on multidimensional parameters.

[0008] A method for assessing the health status of a battery based on multidimensional parameters includes: S1, acquiring multidimensional measurement data of battery charge-discharge cycles; based on a benchmark battery degradation dataset, solving for the optimal wavelet basis function that minimizes energy spectrum overlap using differentiable programming, extracting corresponding filter coefficients and converting them into optimal wavelet basis feature vectors, and decomposing the multidimensional measurement data using the optimal wavelet basis function to obtain a time-frequency domain feature sequence; S2, extracting health indicators from the multidimensional measurement data to form a time-domain feature sequence, processing the time-domain feature sequence using a self-attention mechanism and a periodic correlation adjustment matrix related to the cycle period to generate a weighted average. Temporal features; the weighted temporal features are input into the first network path and the second network path to obtain the first path output reflecting local temporal correlation and the second path output reflecting global dependency; the fusion weight tensor is calculated using the optimal wavelet basis feature vector as a static condition, and the two outputs are weighted and fused to obtain the fused temporal features; S3, the time-frequency domain feature sequence and the fused temporal features are fused using low-rank bilinear pooling operation, and feature modulation is performed in combination with the optimal wavelet basis feature vector to generate a state fusion matrix; based on the state fusion matrix, the health status assessment value of the battery is output through a fully connected network.

[0009] This invention first extracts highly discriminative time-frequency domain features by adaptively solving for the optimal wavelet basis with minimum energy spectrum overlap using differentiable programming. Second, it extracts time-domain health indicators, adjusts temporal attention weights based on cyclic periods, and utilizes dual-network paths in parallel to capture local temporal continuity and global long-range dependencies. Simultaneously, it dynamically fuses the dual-path outputs based on wavelet basis features. Finally, it employs low-rank bilinear pooling to achieve deep multimodal interaction between the time-frequency and temporal domains, and outputs the evaluation results after modulation using wavelet basis features. This scheme overcomes the limitations of existing methods that rely on fixed wavelet bases and single-temporal modeling, effectively integrating temporal evolution and frequency domain micro-changes, significantly improving the accuracy and reliability of health status assessment.

[0010] Preferably, the multidimensional measurement data includes voltage, current, and temperature measurement data; extracting health indicators from the multidimensional measurement data to form a time-domain feature sequence includes: extracting a first time interval from the voltage measurement data during the discharge process when the terminal voltage experiences a preset voltage drop range; extracting a second time interval from the current measurement data during the constant voltage charging stage when the constant voltage is maintained; calculating the difference between the highest and lowest temperatures in the charge-discharge cycle as a temperature difference; and concatenating the first time interval, the second time interval, and the temperature difference in the order of the acquisition cycle into a feature matrix as the time-domain feature sequence.

[0011] By defining three core measurement parameters—voltage, current, and temperature—and extracting the discharge voltage drop time, constant voltage charging duration, and charge-discharge temperature difference to construct a time-domain feature matrix, we can understand how these three parameters directly correspond to the physical processes of battery capacity decay, internal resistance increase, and thermal characteristic deterioration. This approach controls computational load while fully preserving the temporal evolution of the battery degradation trajectory, providing highly relevant inputs with clear physical meaning for subsequent models.

[0012] Preferably, the step of extracting corresponding filter coefficients by solving the optimal wavelet basis function that minimizes energy spectrum overlap using differentiable programming based on the benchmark battery degradation dataset includes: constructing a differentiable wavelet basis parameterization model with scale filter coefficients and corresponding translation parameters as optimization variables; dividing samples from different capacity degradation intervals in the benchmark battery degradation dataset into multiple continuous aging stages; calculating the average wavelet energy spectrum distribution of samples in each aging stage; calculating the cosine similarity of the average wavelet energy spectrum distributions of any two aging stages; and using the mean of each cosine similarity as a penalty term for the energy spectrum overlap; updating the differentiable wavelet basis parameterization model by minimizing the penalty term; and extracting the converged low-pass filter coefficient sequence and high-pass filter coefficient sequence as the corresponding filter coefficients.

[0013] This approach enables end-to-end adaptive optimization of wavelet basis functions, allowing the generated filter coefficients to maximize the differentiation of time-frequency characteristics of different health states. It solves the problem that traditional fixed wavelet basis functions cannot adapt to battery degradation characteristics and improves the ability of time-frequency domain features to characterize minor degradation states.

[0014] Preferably, the step of extracting the corresponding filter coefficients and converting them into the optimal wavelet basis feature vector includes: concatenating the extracted low-pass filter coefficient sequence and high-pass filter coefficient sequence end to end to construct a one-dimensional filter parameter array; normalizing the one-dimensional filter parameter array using mean-variance standardization; mapping the normalized one-dimensional filter parameter array into a feature vector of a set dimension through a linear mapping layer, and outputting it as the optimal wavelet basis feature vector after processing by an activation function.

[0015] Preferably, the step of processing the temporal feature sequence by combining a self-attention mechanism and utilizing a cycle-related adjustment matrix to generate weighted temporal features includes: generating a query matrix, a key matrix, and a value matrix from the temporal feature sequence through linear transformation; calculating an initial attention score matrix based on the query matrix and the key matrix; obtaining the number of each cycle corresponding to the temporal feature sequence; calculating the quotient of each cycle number to the rated cycle life cycle number; and calculating the natural exponent of the quotient to generate a cycle weight vector; calculating and constructing the cycle-related adjustment matrix based on the cycle weight values ​​corresponding to any two cycle periods and their normalized cycle difference; performing element-wise addition and normalization of the cycle-related adjustment matrix and the initial attention score matrix; and multiplying the normalized result by the value matrix to generate the weighted temporal features.

[0016] By injecting the physical law of nonlinear acceleration of battery degradation rate with life cycle into the model, the model is guided to adaptively increase the attention weight of key nodes in the mid-to-late aging stage. This breaks the limitation of traditional attention relying solely on numerical similarity and improves the rationality and accuracy of long-range time-series feature aggregation.

[0017] Preferably, the first network path is a gated recurrent unit path, and the second network path is a self-attention network path; the output of the first path is a gated sequence tensor, and the output of the second path is an attention sequence tensor.

[0018] Preferably, the fusion weight tensor is calculated using the optimal wavelet basis feature vector as a static condition, and the two outputs are weighted and fused to obtain the fused temporal features. This includes: concatenating the gated sequence tensor and the attention sequence tensor along the feature dimension to form a synthetic sequence tensor; expanding the optimal wavelet basis feature vector in the time dimension to match the size of the synthetic sequence tensor, and concatenating the expanded optimal wavelet basis feature vector with the synthetic sequence tensor to form an overall tensor; mapping the overall tensor through a feedforward neural network and activation function to generate the fusion weight tensor; calculating the difference between the constant 1 and the fusion weight tensor to obtain a complementary weight tensor; and performing element-wise multiplication and summation of the gated sequence tensor, the attention sequence tensor, the fusion weight tensor, and the complementary weight tensor to obtain the fused temporal features.

[0019] Preferably, the step of fusing the time-frequency domain feature sequence and the fused time-domain feature using low-rank bilinear pooling includes: using two sets of preset low-rank projection matrices to perform dimensionality compression transformation on the time-frequency domain feature sequence and the fused time-domain feature after cyclic periodic dimension alignment processing, respectively, to obtain a dimension-matched dimensionality-reduced time-frequency feature sequence and a dimensionality-reduced time-domain feature sequence; and performing Hadamard product calculation on the dimensionality-reduced time-frequency feature sequence and the dimensionality-reduced time-domain feature sequence to obtain an interactive feature sequence.

[0020] While maintaining the consistency of the cycle dimension, the computational complexity of bilinear interaction is significantly reduced, and deep nonlinear coupling between time-domain evolution law and frequency-domain energy distribution characteristics is realized, fully exploring the multimodal potential correlations driving battery degradation.

[0021] Preferably, the process of combining the optimal wavelet basis eigenvector with feature modulation to generate a state fusion matrix includes: concatenating the optimal wavelet basis eigenvector after extending the time dimension with the fused time-domain features, performing feature mapping through a multilayer perceptron network to generate a conditional modulation coefficient sequence; performing element-wise multiplication on the interactive feature sequence and the conditional modulation coefficient sequence to obtain a hybrid feature sequence; sequentially performing a sign square root function operation and L2 norm normalization along the feature dimension on the hybrid feature sequence, and using the processed hybrid feature sequence as the state fusion matrix.

[0022] By utilizing frequency domain base properties to target and enhance aging-sensitive features, and by using sign square root and L2 normalization to smooth feature amplitude fluctuations, reduce noise interference, and improve the numerical stability of fused features and the cross-condition adaptability of back-end regression prediction, the following approach is adopted.

[0023] Preferably, the health status assessment value of the battery is output through a fully connected network based on the state fusion matrix, including: adjusting the state fusion matrix to a fixed size, flattening it, inputting it into a health status regression network composed of multiple cascaded fully connected layers, and outputting a scalar prediction value representing the current capacity retention rate as the health status assessment value.

[0024] The technical solution of the present invention has the following beneficial technical effects: This invention employs differentiable programming to adaptively solve for the optimal wavelet basis function, combines a dual-pathway network with a periodic correlation matrix to process time-series data, and uses low-rank bilinear pooling to complete multimodal feature interaction. By optimizing and adapting the wavelet basis, the energy spectrum overlap between different health states is reduced, overcoming the limitation of traditional fixed wavelet bases in effectively distinguishing subtle differences in battery degradation. The dual-pathway architecture collaboratively extracts local degradation continuity and long-range dependencies across cycles, and adjusts the time-series correlation weights based on the physical laws of battery cyclic aging, solving the problem of insufficient modeling of stage-specific degradation patterns by a single network. Simultaneously, a deep feature coupling mechanism replaces the traditional simple splicing method, fully exploring the nonlinear interaction between time-frequency and time-domain while avoiding information redundancy, significantly improving the accuracy and stability of battery health status assessment results under complex operating conditions. Attached Figure Description

[0025] Figure 1 This is a flowchart of a battery health status assessment method based on multidimensional parameters; Figure 2 The diagram shows the results of the optimal wavelet basis solution process. Figure 3 The diagram shows the results of temporal feature sequence extraction and periodic weight evolution. Figure 4 Output an error comparison chart for the health status assessment values. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] This invention discloses a method for assessing the health status of a battery based on multi-dimensional parameters, referring to... Figure 1 This includes the following steps: S1. Optimal wavelet basis solution and time-frequency domain feature sequence extraction.

[0029] This step extracts highly discriminative time-frequency domain degradation features from the original running data. By adaptively solving for the optimal wavelet basis to replace the traditional fixed wavelet basis, it addresses the problem of high feature overlap at different aging stages from the source. At the same time, it transforms the wavelet basis structure information into static conditional features that can be utilized by deep learning networks, providing frequency domain prior guidance for subsequent multimodal fusion.

[0030] First, the continuous cyclic test data collected by the battery management system is read and a time-series tensor is constructed. Specifically, three types of parameters—voltage, current, and temperature—are collected within multiple consecutive charge-discharge cycles of the battery. Each charge-discharge cycle serves as a basic processing unit, providing uniformly organized raw multidimensional data for subsequent wavelet basis solving and feature extraction. It should be noted that the benchmark battery degradation dataset used in this embodiment is a measured dataset of the same type of lithium-ion battery, containing 50 sets of complete charge-discharge cycle degradation data. The test conditions are 1C constant current charge-discharge and an ambient temperature of 25℃, fully covering the battery capacity degradation range of 100%-80%.

[0031] When constructing a wavelet parameterization model, the scaling filter coefficients and corresponding translation parameters are used as optimization variables. For example, a single-layer one-dimensional convolutional network is used as the wavelet parameterization network, with a convolution stride of 1 and zero padding to ensure consistent input and output sequence lengths. A baseline battery degradation data sequence is used as input, and a trainable convolutional kernel representing a low-pass scaling filter is used to convolve the input sequence to obtain low-frequency approximation coefficients. Simultaneously, based on the orthogonal conjugate mirror filter bank constraint, the low-pass filter coefficients are reversed overall, and then the even-numbered coefficients are alternately inverted to generate the corresponding high-pass filter coefficients, which are then convolved with the input sequence to obtain high-frequency detail coefficients.

[0032] In one implementation, the convolution kernel length is set to 8, and the convolution bias term is used as a differentiable translation parameter during the training phase for optimization, but is not output separately as filter coefficients required for subsequent discrete wavelet decomposition. After each parameter update, the low-pass filter coefficients are energy normalized so that the low-pass filter coefficients and high-pass filter coefficients together constitute a filter bank that can be used for discrete wavelet decomposition. This ensures that the filter bank satisfies the energy conservation condition of discrete wavelet transform, avoiding signal energy amplification or attenuation and distortion during the decomposition process.

[0033] Next, capacity label values ​​are extracted from the baseline battery degradation dataset, representing the range of capacity decay from 100% to 80%, and divided into multiple consecutive aging stages. In this embodiment, it is divided into five consecutive aging stages: 100% to 96%, 96% to 92%, 92% to 88%, 88% to 84%, and 84% to 80%. During each forward propagation, multi-resolution wavelet decomposition is performed on the charge and discharge samples within each aging stage. In this embodiment, three-level resolution wavelet decomposition is performed to obtain the low-frequency approximation coefficients and high-frequency detail coefficients of each level. The wavelet energy features of the corresponding frequency band are obtained by summing the squares of the wavelet coefficients within each frequency band. Then, the average wavelet energy features of all samples within the same aging stage are calculated to form the average wavelet energy spectrum distribution of that aging stage. Let the first... The average wavelet energy spectrum distribution vector for each aging stage is: .

[0034] To improve the distinguishability of features at each stage, the cosine similarity of the average wavelet energy spectrum distributions of any two aging stages is calculated and averaged, serving as a penalty term for energy spectrum overlap. ; in, This is the energy spectrum overlap penalty term, with a value ranging from 0 to 1, used to characterize the degree of overlap of time-frequency domain features at different aging stages; The total number of aging stages, in this embodiment This was obtained by manually dividing the benchmark dataset into capacity ranges; , The first , The average wavelet energy spectral distribution vector of each aging stage is obtained by taking the mean value of the wavelet energy features of all samples within the corresponding stage; molecule Given the inner product of two vectors, calculate the degree of coincidence of their projections in space; denominator The product of the magnitudes of two vectors is used to normalize the inner product result. This relationship characterizes the degree of overlap of features between different aging stages by calculating the mean cosine similarity of the average wavelet energy spectrum distribution between all pairs of aging stages. Cosine similarity only considers the shape consistency of the energy spectrum distribution and is not affected by the difference in absolute energy amplitude, thus stably characterizing the similarity of the energy spectrum distribution between two aging stages even with inconsistent dimensions. The penalty term takes a value between 0 and 1; the lower the value, the better the separation of the energy spectrum between different aging stages.

[0035] The gradient descent method is used to minimize the penalty term to update the insimilarity wavelet basis parameterized model. In one implementation, the Adam optimizer is used for backpropagation training, with an initial learning rate of 0.001, momentum decay coefficients of 0.9 and 0.999 respectively, a batch size of 64, and updates are stopped after approximately 500 iterations when the loss value converges to less than 0.05. After convergence, the 8-length one-dimensional convolutional kernel weights in the wavelet parameterized network are extracted as a low-pass filter coefficient sequence. After energy normalization, a high-pass filter coefficient sequence of the same length (8) is generated according to the orthogonal conjugate mirror filter bank constraint relationship. The two together constitute the analysis filter bank required for subsequent multi-resolution wavelet decomposition.

[0036] The obtained filter coefficients are converted into compact feature representations that can be directly used by subsequent networks, enabling the structural information of the filter bank to participate in dual-path fusion and multi-mode modulation as static conditions. The extracted low-pass and high-pass filter coefficient sequences are concatenated sequentially in a one-dimensional manner to construct a one-dimensional filter parameter array. In one implementation, an 8-length low-pass coefficient sequence and an 8-length high-pass coefficient sequence are concatenated into a 16-dimensional array. This array is then normalized using mean-variance standardization. ; in, This represents the first element in the normalized one-dimensional filter parameter array. One element; The first element in the one-dimensional filter parameter array Each element is obtained by concatenating the coefficients of the low-pass and high-pass filters end to end; The array length is equal to the sum of the lengths of the low-pass and high-pass filter coefficients, as shown in this embodiment. ; , These are the arithmetic mean and standard deviation of the original filter parameter array, respectively. To prevent division by zero of extremely small positive numbers, this embodiment uses 1×10⁻⁶. -8 The coefficients of low-pass and high-pass filters differ in numerical dimensions and extreme value ranges. After standardization, the array elements have a distribution characteristic of zero mean and unit variance, which helps to reduce the impact of differences in dimensions and extreme values ​​on subsequent linear mapping and stabilize the training process.

[0037] Next, a linear mapping layer maps the normalized array into feature vectors of a set dimension. In one implementation, the linear mapping layer consists of a single fully connected layer with an input dimension of 16 and an output dimension of 32. Its parameters include a weight matrix and a bias vector that can be initialized with all zeros. The resulting 32-dimensional high-order feature vector is then input element-wise into a hyperbolic tangent activation function. ; in, This represents the optimal wavelet basis eigenvector, used to characterize the structural features of the optimal wavelet basis function and its filter bank; This is the normalized filter parameter array; , These are the weight matrix and bias vector of the linear mapping layer, respectively; The hyperbolic tangent activation function restricts the output of each feature dimension to... to Within a continuous interval, the optimal wavelet basis eigenvectors are... The numerical distribution is stable, making it convenient to use as a static condition input in subsequent modules. This 32-dimensional vector serves as the optimal wavelet basis eigenvector, used to characterize the structural features of the optimal wavelet basis function and its filter coefficients.

[0038] Finally, the optimal wavelet basis function obtained from the solution is used to perform multi-resolution wavelet decomposition on the measurement data to extract the frequency domain micro-variation information hidden during the charging and discharging process. The discrete wavelet transform algorithm is called, using a single charging and discharging cycle as the basic processing unit, to perform multi-level decomposition on the voltage, current, and temperature time series within each cycle. Taking three-level wavelet decomposition as an example, each signal is decomposed into third-level low-frequency approximation coefficients. Level 3 high-frequency detail coefficient Second-level high-frequency detail coefficients and the first-level high-frequency detail coefficient For each wavelet frequency band of each signal, five statistical parameters are calculated: band energy, mean coefficient, standard deviation of coefficient, peak coefficient, and band energy percentage. These parameters are then concatenated according to a preset order of signal type, wavelet frequency band, and statistical parameter type to obtain a 60-dimensional original cyclic-level time-frequency feature vector corresponding to each charge-discharge cycle. The continuous... The 60-dimensional feature vectors corresponding to each charge-discharge cycle are stacked row by row in cyclic order to obtain... The original time-frequency domain feature sequence is divided into 60 rows and 60 columns; each 60-dimensional vector is further mapped to 64 dimensions through a linear mapping layer to generate... Time-frequency domain feature sequence with 64 rows and 64 columns This is used for subsequent low-rank bilinear pooling fusion processing.

[0039] S2, Temporal feature dual-path modeling and gating fusion.

[0040] The purpose of this step is to extract time-domain health indicators with clear physical meaning from the raw data, modify the temporal attention weights in combination with the physical laws of battery cycle aging, capture the local degradation continuity and global long-range dependency relationship at the same time through a dual-path parallel architecture, and dynamically fuse the dual-path outputs with frequency domain priors as conditions, so as to solve the problem that a single network is insufficient in modeling the phased degradation law, and generate comprehensive and accurate fused time-domain features.

[0041] Low-dimensional health indicators directly related to the degradation mechanism are extracted from multidimensional measurement data to form a concise and representative time-domain feature sequence, eliminating data redundancy caused by the original long sequence. The first time interval during the discharge process, when the terminal voltage experiences a preset voltage drop range, is extracted from the voltage measurement data. In one embodiment, the voltage drop curve of the battery during the 1C constant current discharge stage is monitored in real time, and the timestamps of the first voltage drops to 4V and 3.6V are recorded. The difference between these two times is calculated as the first time interval, which is, for example, approximately 2200 seconds when the battery is in good health. The second time interval during the constant voltage maintenance process of the constant voltage charging stage is extracted from the current measurement data. The start time from when the terminal voltage reaches the 4.2V cutoff voltage and enters the constant voltage charging mode is recorded until the charging current drops to the 0.05C cutoff threshold. The difference between these two times reflects the evolution of electrode polarization, and the example value typically fluctuates between 1500 and 2000 seconds.

[0042] Meanwhile, by using a thermocouple sensor placed at the center of the battery surface, temperature time-series data of a single charge-discharge cycle (including charging, discharging and resting stages) are continuously collected. The highest and lowest operating temperatures within the cycle are selected, and the absolute difference between the two is used as the temperature difference value to characterize the degradation of the battery's internal heat generation power.

[0043] Following the chronological order of charge-discharge cycles, the first time interval, the second time interval, and the temperature difference value of each cycle are horizontally merged into a single row vector, and then continuously... The data from each iteration is accumulated and then vertically concatenated to obtain a shape of... 3-row feature matrix , as a time-domain feature sequence.

[0044] The temporal correlations between cyclic periods in the temporal feature sequence are reweighted to enhance the weight of mid-to-late-stage cyclic features in attention allocation, thus better characterizing the stage differences in the degradation process. In one implementation, a self-attention network is introduced, comprising three parallel unbiased linear fully connected layers and a scaled dot-product attention module. The hidden layer feature dimension is set to 64, and the dropout coefficient of the scaled dot-product attention module is set to 0.1; the shape is... Time-domain feature sequence matrix with 3 rows and 3 columns Parallel input consists of three unbiased linear layers with weight matrices of 3 rows and 64 columns each, and the generation dimension is also 1. Query matrix with 64 rows and 64 columns Key matrix Sum matrix Number of consecutive cycles The value range is [50, 200]. In this embodiment, we take... The value is 100. Perform matrix multiplication between the query matrix and the transposed key matrix, and divide by the square root of the key vector feature dimension to obtain the initial attention score matrix: ; in, The initial attention score matrix has dimensions N×N, representing the initial correlation between different cycle periods; The query matrix has dimensions N×64 and consists of time-domain feature sequences. Obtained through mapping using the first unbiased linear layer; The key matrix has dimensions N×64 and is composed of temporal feature sequences. Obtained by mapping through a second unbiased linear layer; Key matrix Transpose of; Here, is the key vector dimension, and is a preset hyperparameter for the hidden layer feature dimension. In this embodiment... Dividing by the square root of the key vector dimension, the dot product result is scaled using gradient to prevent the dot product value from being too large, which would cause the attention distribution after Softmax to be too sharp, thus stabilizing the training process.

[0045] Next, obtain the cycle number corresponding to each time-domain feature sequence, and then... Number of cycles With respect to the nominal rated cycle life of the battery The ratio of is used as the normalized periodic value. And calculate the normalized periodic value. Periodic weight values: ; in, For the first The normalized periodic value of each cycle, ranging from 0 to 1; For the first The cycle weight value of each cycle increases exponentially with the number of cycles. For the first The sequence number of each charge-discharge cycle is obtained from the cycle count during data acquisition; This refers to the nominal rated cycle life of the battery. The cycle life is normalized to... This eliminates the dimensional impact of differences in the rated lifespan of different batteries; the natural index is used to give greater weight to cycle periods closer to the rated lifespan, thereby enhancing the focus on mid-to-late cycle characteristics.

[0046] For any two cycles and Based on the periodic weight values ​​of the two , and normalized periodic difference Through shared parameter mapping functions The first in the periodic correlation adjustment matrix Line number Column elements: ; In one implementation, the mapping function employs... ;in, For the periodic correlation adjustment matrix, the first Line number The elements of the column are used to correct the initial attention score; , The first , The periodic weight value corresponding to each cycle; This is the normalized period difference between the two cycles; This is the natural exponential function. This mapping function provides a larger correlation gain when the cycle periods are similar, and the overall correlation strength increases with the degree of degradation. After traversing all cycle pairs, the size is obtained as... Periodic correlation adjustment matrix Its dimensions are the same as the initial attention score matrix. Maintain consistency. Add the two elements-wise to obtain the corrected attention score matrix: ; in, To correct the attention score matrix, the dimension is N×N; The initial attention score matrix has dimensions N×N; The periodic correlation adjustment matrix is ​​of dimension N×N. The periodic correlation adjustment matrix is ​​added element-wise to the initial attention score matrix to perform physical correction on the initial attention score. Then, along each row... Perform Softmax normalization so that the sum of the attention weights for each cycle is 1, and then compare the normalization result with the value matrix. Multiplying them together yields the weighted time-domain feature matrix: ; in, The weighted temporal feature matrix has a dimension of N×64 and is the attention-weighted temporal feature representation. The normalization function along the row direction converts the modified attention score into a weight distribution between 0 and 1, with each row summing to 1; The value matrix has dimensions N×64 and is composed of time-domain feature sequences. The result is obtained after mapping through a third unbiased linear layer. Here, Softmax normalization is used to convert the corrected attention score into a non-negative weight distribution that sums to 1 across rows, thereby aggregating the temporal degradation information of each cycle in a probabilistic weighted manner.

[0047] Weighted time-domain features The parallel input dual-path time information aggregation module has two paths. The first network path is a gated loop unit path, used to extract local temporal correlations between adjacent loops and output a gated sequence tensor reflecting these local temporal correlations. Its dimensions are Rows 64 columns; the second network pathway is a self-attention network pathway, used to extract long-range dependencies across cycles, and outputs an attention sequence tensor reflecting global dependencies. Their dimensions are the same Rows 64 columns. Two parallel processing paths focus on modeling local degradation continuity and long-range dependencies respectively, thus structurally taking into account temporal evolution information at different scales.

[0048] A gated fusion unit is configured to adaptively coordinate the contribution ratios of the two pathways in each cycle and feature dimension based on frequency domain prior information, avoiding information confusion caused by simple equal-weight concatenation. The gated sequence tensor is then used... With attention sequence tensor By concatenating and stitching along the last feature dimension, a feature with dimension [value missing] is constructed. A 128-row synthetic sequence tensor. To inject static wavelet guidance information with frequency prior properties, the optimal wavelet basis eigenvector of length 32 is used. Replication in time batch dimension Next, expanded to Row 32, and with The synthesized sequence tensor with 128 rows and 128 columns is horizontally concatenated along the feature dimension to generate a sequence of size 128. The overall context tensor with 160 rows and 160 columns .

[0049] Then, the overall context tensor The input is a feedforward neural network containing nonlinear hidden layers and a linear output layer. In one implementation, the global context tensor... First, a hidden layer containing 128 neurons is input. ReLU activation is used to extract the combination relationships between gated recurrent unit pathways, self-attention network pathways, and optimal wavelet basis features. Then, the resulting intermediate hidden states are input into a fully connected layer with an output dimension of 64. The output is then constrained to between 0 and 1 using the Sigmoid activation function, generating a layer with a size of [missing information]. 64-row fusion weight tensor : ; in, To integrate the weight tensor, the dimension is N×64, and the value range is 0 to 1, representing the contribution ratio of the gated recurrent unit path; It is a feedforward neural network, containing a ReLU hidden layer with 128 neurons and a linear output layer with 64 neurons; The Sigmoid function maps output values ​​to the range of 0 to 1. The overall context tensor has dimensions N×160. A sigmoid function is used to constrain the output to the (0,1) interval, making... It has a clear meaning of percentage weight, representing the relative contribution of the gated recurrent unit pathway in each cycle and each feature dimension.

[0050] Construct a with Constant tensors of the same size and all elements having a value of 1, and The difference is used to obtain the complementary weight tensor The gated sequence tensor is element-wise multiplied with the fusion weight tensor, and the attention sequence tensor is element-wise multiplied with the complementary weight tensor. The two weighted results are then summed element-wise to obtain the fused temporal features. ; in, To integrate temporal features, the dimension is N×64, which is the final temporal feature after dynamic weighting of the dual-path output; The gated sequence tensor has a dimension of N×64; To ensure the sequence tensor is appropriate, its dimension is N×64; This is the Hadamard product. The resulting fused temporal feature size is maintained at [value]. Rows 64 and 64 are used for subsequent multimodal fusion with time-frequency domain features. This weighting method allows the two pathways to complement and diminish each other during fusion, which helps to dynamically adjust the relative importance of local and global information at different lifecycle stages.

[0051] S3, Multimodal Feature Fusion and Health Status Assessment Output.

[0052] The purpose of this step is to achieve deep nonlinear interactive fusion of time-frequency domain features and time-domain features. The computational complexity is controlled by low-rank bilinear pooling, and the fusion process is modulated with the optimal wavelet basis features as a condition. Finally, the final health status assessment value is output through a regression network, which solves the problems of information redundancy and insufficient interaction in the traditional simple splicing method, and improves the accuracy and stability of the assessment results under complex working conditions.

[0053] To achieve nonlinear interactive fusion between time-frequency domain energy distribution features and time-domain evolution features while controlling the computational complexity of high-dimensional interactions, a frequency-domain prior is used to conditionally modulate the fusion process. Two unbiased low-rank projection matrix networks with random Gaussian weights are pre-initialized; both networks are single-layer linear unbiased fully connected layers. The dimensions of the low-rank projection matrices are determined based on the dimensions of the time-frequency domain feature sequence and the fused time-domain features. For example, when the input feature dimension is 64, it is projected to 32 dimensions to balance computational complexity and feature representation capability. The network is then aligned according to the cyclic period dimension and has a size of... Time-frequency domain feature sequence with 64 rows and 64 columns The size is the same after integration with the gating system. Fusion temporal features of row 64 columns Dimensional compression is performed through the first and second low-rank projection layers respectively: ; in, The time-frequency feature sequence is a dimension-reduced sequence with dimensions N×32; The time-domain feature sequence is a dimension-reduced sequence with dimensions N×32; It is a time-frequency domain feature sequence with a dimension of N×64; To fuse temporal features, the dimension is N×64; This is the first low-rank projection matrix, with dimensions of 64×32; The second low-rank projection matrix has a dimension of 64×32; both low-rank projection matrices have a size of 64 rows and 32 columns, and their size can be set to 64 rows and 32 columns to make the dimensionality-reduced time-frequency feature sequence... With time-domain feature sequences All The 32-row column is mapped to a unified latent feature space while preserving the original cyclic dimension, which facilitates subsequent execution of Hadamard product and multimodal interaction fusion, and to some extent reduces the problem of excessive dimensionality caused by conventional bilinear pooling.

[0054] To generate feature activation weights, the optimal wavelet basis eigenvectors of dimension 32 are used. Copy and expand 32-row matrix format spliced ​​along the feature axis The fusion of temporal features in rows 64 and columns forms An auxiliary splicing tensor with 96 rows and columns is used. This splicing tensor is fed into a multilayer perceptron network consisting of three sequentially connected fully connected layers with 64, 64, and 32 neurons respectively. The first two layers use the LeakyReLU activation function, and the output layer uses the Sigmoid function to generate a non-negative conditional modulation coefficient sequence. : ; in, The conditional modulation coefficient sequence has a dimension of N×32 and a value range of 0 to 1. It is used to target and weight the interaction features. This represents the optimal wavelet basis eigenvector matrix after time-dimension expansion. This indicates a concatenation operation along the feature dimension. It is a three-layer fully connected network with a structure of 96→64→64→32. The first two layers are activated by LeakyReLU, and the output layer is activated by Sigmoid.

[0055] The Hadamard product is calculated by combining the dimensionality-reduced time-frequency feature sequence with the dimensionality-reduced time-domain feature sequence to obtain the interaction feature sequence: ' in, The interactive feature sequence has a dimension of N×32 and is the interactive result of element-wise multiplication of the time-frequency domain and the time-domain features; The time-frequency feature sequence is a dimension-reduced sequence with dimensions N×32; The time-domain feature sequence is a dimension-reduced sequence with dimensions N×32; The Hadamard product is then used; next, element-wise dot product modulation is performed between the interactive feature sequence and the conditional modulation coefficient sequence to obtain the mixed feature sequence: ; in, It is a mixed feature sequence with a dimension of N×32; The interaction feature sequence has a dimension of N×32; The conditional modulation coefficient sequence has a dimension of N×32. This modulation operation is based on static wavelet domain features to weight and adjust specific aging-related frequency band information, enabling the fusion process to highlight interaction components that are more relevant to degradation based on frequency domain priors.

[0056] To mitigate the model's sensitivity to abrupt changes in amplitude distribution and anomalous noise points, a signed square root activation operation and L2 norm normalization along the feature dimensions are performed sequentially on the mixed feature sequence: ; in, The feature sequence after sign square root activation has a dimension of N×32 and is used to compress large value responses; The final state fusion matrix has a dimension of N×32 and is the fusion feature after L2 normalization. It is a mixed feature sequence with a dimension of N×32; This is a sign function that outputs the sign of an element: 1 for positive numbers, -1 for negative numbers, and 0 for zero. This represents the L2 norm along the feature dimension, which is the square root of the sum of the squares of the elements in each row. To prevent division by zero of extremely small positive numbers, this embodiment uses 1×10⁻⁶. -8 The reason for performing signed square root processing is to compress the large amplitude responses in the features and mitigate the drastic fluctuations in amplitude distribution; L2 norm normalization eliminates the differences in the overall energy scale between samples, making the features fall within a uniform scale range. The combination of these two methods reduces the impact of feature amplitude fluctuations on the regression results. The processed result... 32-row matrix That is, it serves as the state fusion matrix.

[0057] Before inputting the fully connected network, the state fusion matrix is ​​limited to a preset cyclic window length. Corresponding fixed size, preset loop window length The value is determined based on the battery degradation rate and data acquisition frequency, ranging from 50 to 200 charge-discharge cycles: when the actual number of cycles is greater than... At that time, extract the most recent The state fusion matrix corresponding to each cycle; when the actual number of cycles is less than At that time, zero-padding is used at the beginning of the sequence to fill in the blanks. This is done to maintain the causality of the time series and to unify the feature length of the input to the fully connected network, thereby adapting to the fixed input dimension of the regression network.

[0058] The fixed-size state fusion matrix (whose size specification is...) is processed in row-major order. Flattened to a length of A one-dimensional state fusion vector is input to a healthy state regression network consisting of three cascaded fully connected layers. In one implementation, the first fully connected layer will have a length of... The input vector is mapped to 256-dimensional latent features. The second fully connected layer maps the 256-dimensional latent features to 64-dimensional latent features. The third fully connected layer maps the 64-dimensional latent features to a 1-dimensional health status output value. The first two fully connected layers are followed by ReLU activation functions. The output layer is used to generate a scalar prediction value representing the current capacity retention rate, which serves as the health status assessment value of the battery.

[0059] During model training, the ratio of the current available capacity to the rated capacity in the baseline battery degradation dataset is used as the true health status label, and the mean squared error loss function is used to calculate the error between the predicted value and the true label: ; in, This is the mean squared error loss function value, used to measure the degree of deviation between the predicted value and the true label; For the first The predicted health status value for each sample. For the corresponding real tags, The sample size is denoted as . The network parameters are updated via backpropagation using the Adam optimization algorithm until the training loss converges or the preset number of iterations is reached.

[0060] To verify the effectiveness of each functional module, a control experiment was set up based on a benchmark battery degradation dataset. The benchmark group used a complete model including a differentially modulo-parameterized model, a gated fusion unit, and a low-rank bilinear pooling multimodal fusion module. Three parallel control ablation groups were also set up: Ablation group 1 removed the differentially modulo-parameterized model and replaced it with a traditional standard wavelet with fixed-scale filter coefficients; Ablation group 2 removed the gated fusion unit and concatenated the output tensor of the gated recurrent unit and the self-attention network with equal weights; Ablation group 3 removed the low-rank bilinear pooling module and used a regular fully connected layer to perform linear dimensionality reduction and feature merging on the time-frequency domain feature sequence and the fused time-domain features. All experiments used the same data partitioning method, Adam optimizer, initial learning rate of 0.001, batch size of 64, and were iterated 500 times or trained until the loss converged under the same training epochs.

[0061] On the test set, the root mean square error (RMSE) of the complete model was 0.0125, and the mean absolute error (MAO) was 0.0086. The RMSE of ablation group 1, using a fixed wavelet basis, was 0.0192, and the MAO was 0.0148. The RMSE of ablation group 2, using dual-path equal-weighted splicing, was 0.0165, and the MAO was 0.0122. The RMSE of ablation group 3, using a common fully connected layer to replace the low-rank bilinear pooling module, was 0.0178, and the MAO was 0.0135. These results show that, under the same experimental conditions, the error index of the complete model is lower than that of each ablation group, indicating that each functional module contributes positively to the battery health status assessment performance.

[0062] Specifically, the complete model reduces the root mean square error by approximately 34.9% compared to ablation group 1, indicating that solving the wavelet basis with low energy spectrum overlap through differentiable programming can extract more discriminative time-frequency domain features under different aging stages to a certain extent. The complete model achieves a lower error compared to ablation group 2, indicating that using the optimal wavelet basis eigenvector as a static condition input and allocating dual-path fusion weights through a gated fusion unit helps coordinate the modeling ability of the gated recurrent unit for local sequence temporal correlations with the modeling ability of the self-attention network for long-range global correlations. The reduced error of the complete model compared to ablation group 3 indicates that the low-rank bilinear pooling operation modulated by the wavelet basis vector can alleviate the redundancy and computational complexity problems caused by direct fusion of high-dimensional features to a certain extent, and helps to characterize the nonlinear interaction between temporal evolution features and frequency domain energy fluctuation features.

[0063] like Figure 2 As shown in the figure, the energy spectrum overlap penalty term changes with the number of iterations during the optimal wavelet basis solution stage. The curve in the figure clearly shows the decay trend of the penalty term decreasing rapidly as the model training progresses.

[0064] like Figure 3 As shown in the figure, this diagram illustrates the physical evolution of the temporal characteristics and attention weighting mechanism in multidimensional battery measurement data as a function of cycle time. Different lines in the figure depict the trends of three characteristic variables. The first time interval, representing the voltage drop at the discharge terminal, gradually shortens with increasing cycle count, while the second time interval, representing the constant-voltage charging stage, exhibits a gradually lengthening characteristic with increasing cycle count. Another curve shows that the cycle weight value increases exponentially with the number of cycle times. This demonstrates that the method can adaptively enhance the attention weight of key aging nodes in the mid-to-late stages by incorporating the physical law of the nonlinear acceleration of battery degradation rate with its lifespan.

[0065] like Figure 4 As shown in the figure, this diagram illustrates the error performance of the proposed health status assessment model in the control experiment. It is clearly visible from the figure that the heights of the two error indicator bars corresponding to the complete model are the lowest among all model groups. This fully demonstrates the significant advantage of the complete technical solution in assessment accuracy, proving that the collaborative work of each functional module effectively improves the stability of battery health status assessment results under complex operating conditions.

[0066] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assessing the health status of a battery based on multidimensional parameters, characterized in that, include: S1. Acquire multi-dimensional measurement data of battery charge-discharge cycles; Based on a benchmark battery degradation dataset, the optimal wavelet basis function that minimizes the energy spectrum overlap is solved by differentiable programming. The corresponding filter coefficients are extracted and converted into the optimal wavelet basis feature vector. The multidimensional measurement data is then decomposed using the optimal wavelet basis function to obtain the time-frequency domain feature sequence. S2. Extract health indicators from the multidimensional measurement data to form a time-domain feature sequence. Combine a self-attention mechanism and use a periodic correlation adjustment matrix related to the cycle cycle to process the time-domain feature sequence and generate weighted time-domain features. Input the weighted time-domain features into the first network path and the second network path to obtain the first path output reflecting local temporal correlation and the second path output reflecting global dependency. Calculate the fusion weight tensor using the optimal wavelet basis feature vector as a static condition, and then perform weighted fusion of the two outputs to obtain the fused time-domain features. S3. The time-frequency domain feature sequence and the fused time-domain feature are fused using low-rank bilinear pooling operation, and feature modulation is performed by combining the optimal wavelet basis feature vector to generate a state fusion matrix; based on the state fusion matrix, the health status assessment value of the battery is output through a fully connected network.

2. The battery health status assessment method based on multi-dimensional parameters according to claim 1, characterized in that, The multidimensional measurement data includes voltage, current, and temperature measurement data; the health indicators extracted from the multidimensional measurement data constitute a time-domain feature sequence, including: extracting the first time interval during which the terminal voltage experiences a preset voltage drop range from the voltage measurement data; Extract the second time interval experienced during the constant voltage maintenance process in the constant voltage charging phase from the current measurement data; The difference between the highest and lowest temperatures during the charge-discharge cycle is calculated as the temperature difference. The first time interval, the second time interval, and the temperature difference are concatenated into a feature matrix in the order of the acquisition cycle, which serves as the time-domain feature sequence.

3. The battery health status assessment method based on multi-dimensional parameters according to claim 1, characterized in that, The method, based on a benchmark battery degradation dataset, employs differentiable programming to solve for the optimal wavelet basis function that minimizes energy spectrum overlap, and extracts corresponding filter coefficients. This includes: constructing a differentiable wavelet basis parameterization model with scaling filter coefficients and corresponding translation parameters as optimization variables; dividing samples from different capacity degradation intervals in the benchmark battery degradation dataset into multiple continuous aging stages; calculating the average wavelet energy spectrum distribution of samples within each aging stage; calculating the cosine similarity between the average wavelet energy spectrum distributions of any two aging stages; and using the mean of each cosine similarity as a penalty term for the energy spectrum overlap; updating the differentiable wavelet basis parameterization model by minimizing the penalty term; and extracting the converged low-pass filter coefficient sequence and high-pass filter coefficient sequence as the corresponding filter coefficients.

4. The battery health status assessment method based on multidimensional parameters according to claim 3, characterized in that, The step of extracting the corresponding filter coefficients and converting them into the optimal wavelet basis feature vector includes: concatenating the extracted low-pass filter coefficient sequence and high-pass filter coefficient sequence end to end to construct a one-dimensional filter parameter array; The one-dimensional filter parameter array is normalized using mean-variance standardization. The normalized one-dimensional filter parameter array is mapped to a feature vector of a set dimension through a linear mapping layer, and the output after being processed by an activation function is the optimal wavelet basis feature vector.

5. The battery health status assessment method based on multi-dimensional parameters according to claim 1, characterized in that, The step of processing the temporal feature sequence by combining a self-attention mechanism and utilizing a cycle-related adjustment matrix to generate weighted temporal features includes: generating a query matrix, a key matrix, and a value matrix from the temporal feature sequence through linear transformation; calculating an initial attention score matrix based on the query matrix and the key matrix; obtaining the number of each cycle corresponding to the temporal feature sequence; calculating the quotient of each cycle number to the rated cycle life cycle number; and calculating the natural exponent of the quotient to generate a cycle weight vector; calculating and constructing the cycle-related adjustment matrix based on the cycle weight values ​​corresponding to any two cycle periods and their normalized cycle difference; performing element-wise addition and normalization of the cycle-related adjustment matrix and the initial attention score matrix; and multiplying the normalized result by the value matrix to generate the weighted temporal features.

6. The battery health status assessment method based on multi-dimensional parameters according to claim 1, characterized in that, The first network path is a gated recurrent unit path, and the second network path is a self-attention network path; the output of the first path is a gated sequence tensor, and the output of the second path is an attention sequence tensor.

7. The battery health status assessment method based on multidimensional parameters according to claim 6, characterized in that, The fusion weight tensor is calculated using the optimal wavelet basis feature vector as a static condition. Based on this, the two outputs are weighted and fused to obtain the fused temporal features. This includes: concatenating the gated sequence tensor and the attention sequence tensor along the feature dimension to form a synthetic sequence tensor; expanding the optimal wavelet basis feature vector in the time dimension to match the size of the synthetic sequence tensor, and concatenating the expanded optimal wavelet basis feature vector with the synthetic sequence tensor to form a whole tensor; mapping the whole tensor through a feedforward neural network and activation function to generate the fusion weight tensor; calculating the difference between the constant 1 and the fusion weight tensor to obtain a complementary weight tensor; and performing element-wise multiplication of the gated sequence tensor and the attention sequence tensor with the fusion weight tensor and the complementary weight tensor, respectively, and then adding them to obtain the fused temporal features.

8. The battery health status assessment method based on multi-dimensional parameters according to claim 1, characterized in that, The method of fusing the time-frequency domain feature sequence and the fused time-domain feature using low-rank bilinear pooling operation includes: using two sets of preset low-rank projection matrices to perform dimensionality compression transformation on the time-frequency domain feature sequence and the fused time-domain feature after cyclic period dimension alignment processing, respectively, to obtain dimension-matched dimensionality-reduced time-frequency feature sequence and dimensionality-reduced time-domain feature sequence. The Hadamard product is calculated by performing a Hadamard product on the reduced time-frequency feature sequence and the reduced time-domain feature sequence to obtain the interaction feature sequence.

9. The battery health status assessment method based on multidimensional parameters according to claim 8, characterized in that, The process of combining the optimal wavelet basis eigenvector with the fused temporal features to generate a state fusion matrix includes: concatenating the optimal wavelet basis eigenvector after extending the time dimension with the fused temporal features, performing feature mapping through a multilayer perceptron network, and generating a conditional modulation coefficient sequence. The interactive feature sequence and the conditional modulation coefficient sequence are multiplied element-wise to obtain a hybrid feature sequence. The mixed feature sequence is subjected to a sign square root function operation and L2 norm normalization along the feature dimension in sequence, and the processed mixed feature sequence is used as the state fusion matrix.

10. The battery health status assessment method based on multi-dimensional parameters according to claim 1, characterized in that, Based on the state fusion matrix, the battery health status assessment value is output through a fully connected network, including: adjusting the state fusion matrix to a fixed size, flattening it, inputting it into a health status regression network composed of multiple cascaded fully connected layers, and outputting a scalar prediction value representing the current capacity retention rate as the health status assessment value.

Images