Battery system health state evaluation method and system based on multi-dimensional feature fusion

Abstract

The invention relates to the field of battery fault detection, in particular to a battery system health state evaluation method and system based on multi-dimensional feature fusion, and the method comprises the steps: S1, obtaining original operation data of a battery system, and carrying out the preprocessing of the original operation data to obtain a standardized data matrix; s2, based on the standardized data matrix, extracting health state features of a plurality of preset dimensions to form an original feature set; s3, constructing an incidence relation model among the features in the original feature set, and performing nonlinear fusion processing based on the incidence relation model to generate a low-dimensional fusion feature vector; s4, performing principal component analysis based on the low-dimensional fusion feature vector to extract a principal component, calculating a T2 statistical magnitude and an SPE statistical magnitude, and constructing a comprehensive health index; and S5, when the comprehensive health index exceeds a preset threshold value, analyzing the contribution degree of each feature in the original feature vector to the comprehensive health index, and positioning the fault single battery according to the contribution degree. The problems of insensitive fault symptoms, inaccurate fault positioning and high false report and missing report rate are solved.

Description

Technical Field

[0001] The present invention relates to the field of battery fault detection, and in particular to a battery system health status assessment method and system based on multi-dimensional feature fusion. Background Art

[0002] With the increasing popularity of electric vehicles and large-scale energy storage power stations, the safety and reliability of lithium-ion battery systems have become a focus of public attention. One of the core functions of a battery management system (BMS) is to provide real-time assessment of the battery's state of health (SOH) and fault warnings to prevent serious safety incidents such as thermal runaway.

[0003] Currently, existing battery health status assessment methods are mainly divided into three categories: the first category is the threshold-based judgment method, which sets an alarm by setting a fixed threshold for a single parameter such as voltage or temperature. Although this method is simple, it has high false alarm and missed alarm rates, cannot cope with complex and changeable actual working conditions, and cannot locate the source of the fault. The second category is the physical model-based method, which relies on an accurate electrochemical model, is complex to calculate, and the model parameters will drift as the battery ages, making it less practical. The third category is the data-driven method, which has become the mainstream of current research. The existing technology generally has the following technical defects: 1. Traditional dimensionality reduction methods assume that data is linearly separable and cannot effectively handle the nonlinear coupling relationship between multidimensional battery features, resulting in insufficient feature fusion and significant information loss. 2. Existing methods analyze each feature as an independent entity, ignoring the inherent physical and statistical connections between features and failing to build an evaluation model that can reflect the mutual influence between features. 3. Existing solutions are insensitive to early fault signs caused by the coupling of subtle changes in multiple features, and the timeliness and accuracy of early warnings need to be improved. Summary of the Invention

[0004] The present invention aims to provide a battery system health status assessment method and system based on multi-dimensional feature fusion to solve the problems of insufficient processing of nonlinear coupling relationships of multi-dimensional features and insufficient mining of correlations between features in existing battery health status assessment, resulting in insensitivity to early fault signs, inaccurate fault location, and high false alarm and missed alarm rates.

[0005] To achieve the above objectives, the present invention provides a technical solution: a battery system health status assessment method based on multi-dimensional feature fusion, comprising: S1: Obtain the original operating data of the battery system and preprocess the original operating data to obtain a standardized data matrix; S2: Based on the standardized data matrix, extract health status features of multiple preset dimensions to form an original feature set. The health status features include voltage standard deviation, voltage Z score, voltage change rate, voltage-to-state-of-charge differential, temperature maximum, temperature average, temperature change rate, temperature-to-state-of-charge differential, and internal short-circuit resistance. S3: Construct a correlation model between the features in the original feature set, and perform nonlinear fusion processing on the original feature set based on the correlation model to generate a low-dimensional fusion feature vector; S4: Perform principal component analysis based on the low-dimensional fusion feature vector, extract the principal component, and calculate T 2 Statistics and SPE statistics, according to T 2 Statistics and SPE statistics are used to construct comprehensive health indicators; S5: When the comprehensive health index exceeds a preset threshold, the contribution of each feature in the original feature vector to the comprehensive health index is analyzed, and the faulty battery cell is located according to the feature with the largest contribution.

[0006] Preferably, the association relationship model is a graph neural network model, and step S3 includes S3-1) Define each feature in the original feature vector as a graph node and construct a feature graph based on the preset adjacency relationship; S3-2) Use the graph neural network model to process the feature graph, aggregate node neighbor information through the message passing mechanism, update the node representation, and generate a low-dimensional fused feature vector through graph readout operations.

[0007] Further preferably, the strategy for processing the feature graph using the graph neural network model in step S3-2 includes: Input node features and adjacency matrix into the graph convolutional network layer; Inter-layer propagation is performed using the following formula:

[0008] in, represents the node feature matrix of the lth layer, represents the normalized adjacency matrix, represents the trainable weight matrix of layer l, Represents a nonlinear activation function, wherein the activation function adopts Sigmoid; After at least two layers of graph convolution operations, the final feature representations of all nodes are aggregated using a global pooling function to generate a low-dimensional fused feature vector.

[0009] Preferably, the association relationship model is a graph attention network based on the attention mechanism, and step S3 includes S3-1) Calculate a dynamic weight coefficient for each feature in the original feature set. The weight coefficient is used to represent the importance of the feature relative to other features in the current state. S3-2) Multiply each feature by its corresponding dynamic weight coefficient to obtain a weighted feature set; S3-3) Perform dimensionality reduction processing on the weighted feature set to generate a low-dimensional fused feature vector.

[0010] Preferably, the calculation method of the comprehensive health index is as follows:

[0011] in, is the SPE statistic, T 2 Statistics, is the control limit of the SPE statistic, T 2 The control limits for the statistic.

[0012] Further preferably, in S5, the control limit of the SPE statistic is calculated as follows:

[0013] in, (r=1,2,3), , is the sum of the eigenvalues ​​of the residual subspace, is the sum of squared eigenvalues ​​of the residual subspace, is the sum of the cubed eigenvalues ​​of the residual subspace, is the confidence interval of the standard normal distribution; T 2 The control limits of the statistic are calculated as follows:

[0014] in, is the number of principal components after dimensionality reduction, n is the number of training samples, is the significance level, is the F distribution function.

[0015] In another preferred step, in S5, the contribution calculation method includes a. Calculate T according to the following formula 2 Statistical contribution:

[0016] Among them, r is the number of retained principal components, t i is the score of the i-th pivot, is the eigenvalue of the i-th principal component, are the elements of the loading matrix, is the current measured value of the jth original variable, is the mean of the jth original variable; b. Calculate the SPE statistic contribution according to the following formula:

[0017] in, is the prediction error of the j-th variable; c. According to T 2 Statistical contribution, SPE statistical contribution calculation comprehensive contribution:

[0018] in, T 2 Statistical contribution, is the contribution to the SPE statistic, is the control limit of the SPE statistic, T 2 The control limits for the statistic.

[0019] Preferably, in S2, there are sixteen preset dimensions, namely, voltage standard deviation, maximum voltage Z score, second largest voltage Z score, third largest voltage Z score, average voltage Z score, maximum voltage change rate, second largest voltage change rate, third largest voltage change rate, average voltage change rate, standard deviation of voltage change rate, maximum value of differential value of voltage to state of charge, maximum temperature, average temperature, maximum value of temperature change rate, maximum value of differential value of temperature to state of charge, and minimum value of internal short circuit resistance.

[0020] The present invention provides another technical solution: A battery system health status assessment system based on multi-dimensional feature fusion for running the above method, comprising: Data preprocessing module, used to obtain original operation data and perform preprocessing; Feature extraction module, used to calculate health status features of multiple dimensions from preprocessed data and generate the original feature set; The fusion module builds an association relationship model based on the association relationship between each feature in the original feature set, and uses the association relationship model to perform nonlinear fusion and output a low-dimensional fusion feature vector; Health status monitoring module, used to perform principal component analysis on low-dimensional fusion feature vectors and calculate T 2 statistics, SPE statistics and comprehensive health indicators; The fault tracing and positioning module is used to calculate the contribution of each feature and locate the faulty battery cell when the comprehensive health indicators are abnormal.

[0021] Preferably, the association relationship model is a graph neural network model or a graph attention network based on an attention mechanism.

[0022] The principles and advantages of the present invention are: 1. The nonlinear coupling relationship between multidimensional battery features is processed through the correlation model, breaking the linear assumption of traditional dimensionality reduction methods, fully integrating features, reducing information loss, and at the same time mining the intrinsic physical and statistical correlations between features. The constructed evaluation model is more in line with the complex characteristics of the battery system, solving the problems of existing battery health status assessment that insufficiently handle the nonlinear coupling relationship of multidimensional features and insufficient mining of correlations between features, resulting in inaccurate fault location and high false alarm and missed alarm rates.

[0023] 2. The health indicator constructed by comprehensive statistics can keenly capture early fault signs caused by the coupling of subtle changes in multiple features, reduce false alarm rates and missed alarm rates, and provide early and more accurate warnings of potential battery system failures.

[0024] 3. The fault location method based on feature contribution can accurately locate the faulty battery cell, providing clear guidance for subsequent inspection and maintenance, improving the operation and maintenance efficiency of the battery system, and ensuring the safety and reliability of the battery system in application scenarios such as electric vehicles and energy storage power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flow chart of the method of the present invention; Figure 2 This is a flow chart of fault tracing of the present invention; Figure 3 This is a logic block diagram of the system of the present invention. DETAILED DESCRIPTION

[0026] The following is further described in detail through specific implementation methods: Example 1 See also Figures 1 to 2 , a battery system health status assessment method based on multi-dimensional feature fusion, including: S1: Obtain the original operating data of the battery system and preprocess the original operating data to obtain a standardized data matrix; Downloading GBT32960 standard operating data from the electric vehicle cloud platform; Preprocessing of operational data to identify, correct, or delete inaccurate, incomplete, duplicate, or inapplicable data in the dataset. Preprocessing includes: a) Outlier cleaning Delete the entire data that is abnormal, such as: '-', control, 'abnormal', 'INVALID', 'ERROR', '0-1'; b) Range cleaning Temperature cleaning: Retain data with a temperature less than 80°C to avoid data distortion caused by acquisition failures or data anomalies during data network transmission; Voltage cleaning: Retain data with 3V < cell (voltage) <= 4.5V to ensure the operating range of the battery cell under normal conditions and exclude abnormal data points; Soc cleaning: Retain 0 <= Soc <= 100 to ensure that the SOC value is within a reasonable range and exclude potentially abnormal or invalid values; Insulation resistance value cleaning: Retain 10 < insulation resistance value < 50000. Insulation resistance values outside this range have no calculation value. To ensure the good insulation of the vehicle electrical system, eliminate potential faults or unqualified data; c) Sampling fault cleaning Identify voltage jumps caused by sensor failures and replace them with the average value of adjacent cells. Specifically, first calculate the maximum value of the voltage difference at the current moment. If it is greater than the threshold, check if the voltage difference between adjacent cells is greater than 95%. If both conditions are met, continue to determine if the voltages of adjacent cells are symmetrically distributed. If so, a sampling fault has occurred, and the value of the sampling fault is replaced with the average value.

[0027] d) Standardization processing Perform Z-score standardization on the cleaned data to eliminate the influence of dimensions.

[0028] It also includes current cleaning, converting the current to a floating-point number for numerical accuracy and unified data format.

[0029] S2: Based on the standardized data matrix, extract health state features of multiple preset dimensions to form an original feature set. The health state features include voltage standard deviation, voltage Z-score, voltage change rate, voltage derivative with respect to state of charge, maximum temperature, average temperature, temperature change rate, temperature derivative with respect to state of charge, internal short-circuit resistance; Specifically, in S2, there are a total of sixteen preset dimensions, namely the maximum value of the voltage standard deviation, the second-largest value of the voltage Z-score, the third-largest value of the voltage Z-score, the average value of the voltage Z-score, the maximum value of the voltage change rate, the second-largest value of the voltage change rate, the third-largest value of the voltage change rate, the average value of the voltage change rate, the standard deviation of the voltage change rate, the maximum value of the voltage derivative with respect to state of charge, the maximum temperature, the average temperature, the maximum value of the temperature change rate, the maximum value of the temperature derivative with respect to state of charge, and the minimum value of the internal short-circuit resistance.

[0030] S3: Construct a correlation relationship model between the features in the original feature set and perform non-linear fusion processing on the original feature set based on the correlation relationship model to generate a low-dimensional fusion feature vector; The association relationship model is a graph neural network model, and step S3 includes S3-1) Define each feature in the original feature vector as a graph node and construct a feature graph based on the preset adjacency relationship; Specifically, each feature in the original feature vector (such as voltage standard deviation, voltage Z score, etc.) is defined as a graph node. In this embodiment, a total of 16 features are extracted, so there are a total of 16 nodes in the graph. The initial feature of each node is a scalar, that is, the specific value of the feature after S2 calculation. The initial features of all nodes together form the node feature matrix; Define the connection relationships between nodes based on prior knowledge (mechanistic relationships) or data statistical relationships (such as correlation coefficients) in the battery field. For example, if the "voltage" related features and the "temperature" related features are closely coupled physically, connections are established between their corresponding nodes. These connection relationships are represented by an adjacency matrix. Normalize the adjacency matrix to stabilize model training; S3-2) Use the graph neural network model to process the feature graph, aggregate node neighbor information through the message passing mechanism, update the node representation, and generate a low-dimensional fused feature vector through graph readout operations.

[0031] Preferably, the strategy for processing the feature graph using the graph neural network model in step S3-2 includes: Input node features and adjacency matrix into the graph convolutional network layer; Inter-layer propagation is performed using the following formula:

[0032] in, represents the node feature matrix of the lth layer, represents the normalized adjacency matrix, represents the trainable weight matrix of layer l, Represents a nonlinear activation function, wherein the activation function adopts Sigmoid; After at least two layers of graph convolution operations, the final feature representations of all nodes are aggregated using a global pooling function to generate a low-dimensional fused feature vector.

[0033] S4: Perform principal component analysis based on the low-dimensional fusion feature vector, extract the principal component, and calculate T 2 Statistics and SPE statistics, according to T 2 Statistics and SPE statistics are used to construct comprehensive health indicators; S4-1) A large number of low-dimensional fused feature vectors obtained after processing in steps S1-S3 are used to establish a PCA model and determine the dimensionality reduction parameters and control limits; Specifically, a. Select the low-dimensional fusion feature vector obtained by preprocessing historical normal operation data in step S1, extracting features in step S2, and fusing with the graph neural network in step S3, and organize it into a training dataset X with a dimension of [n, m], where n is the number of training samples and m is the dimension of the low-dimensional fusion feature; b. Perform Z-score normalization again to center the original data, eliminate the influence of the mean and standardize the variance:

[0034] Where, , is the mean of the jth column, , is the standard deviation c. Based on the standardized training data, calculate the covariance matrix C (used to measure the linear correlation between fusion features). The formula is:

[0035] Where m is the number of training samples, X is the standardized training data matrix, is the transpose of matrix X.

[0036] d. Perform eigendecomposition on the covariance matrix C to obtain the eigenvalues and the corresponding eigenvector matrix ; e. Determine the number of retained pivots based on the principle that the cumulative contribution rate is ≥ 90%. For example, if the cumulative contribution rate of the first eight eigenvalues ​​reaches 92%, then set r = 8, and form the first r eigenvectors into a dimensionality reduction transformation matrix P (with dimensions [m, r]). Store P for online dimensionality reduction. The calculation formula for the cumulative contribution rate is as follows:

[0037] Where, is the i-th eigenvalue, k is the number of principal components currently considered, and n is the total number of original features.

[0038] f. Based on the training data and the selected principal components, calculate the control limits of the two types of statistics in advance (as standard thresholds for subsequent health risk judgment).

[0039] S4-2) Calculate control limits T 2 The control limits of the statistic are calculated as follows:

[0040] in, is the number of principal components after dimensionality reduction, n is the number of training samples, is the significance level, is the F distribution function; The control limits for the SPE statistic are calculated as follows:

[0041] in, (r=1,2,3), , is the confidence interval of the standard normal distribution, is the sum of the eigenvalues ​​of the residual subspace, is the sum of squared eigenvalues ​​of the residual subspace, is the sum of the cubed eigenvalues ​​of the residual subspace, is the confidence interval of the standard normal distribution; S4-3) performing PCA projection and statistics calculation on the single low-dimensional fusion feature vector obtained by processing in steps S1-S3 at the current sampling moment; T 2 The calculation formula of the statistic is as follows:

[0042] Where, is the standardized data vector of the i-th sample, for The transpose of , C is the covariance matrix of the principal component scores, is the loading matrix of the first k principal components in principal component analysis, for The transposed matrix of The calculation formula of SPESPE statistic is as follows:

[0043] Where I is the identity matrix, is the loading matrix of the first k principal components in principal component analysis, for The transposed matrix, x is the standardized data, is the square of the L2 norm of the vector; S4-4) According to T 2 Statistics and SPE statistics are used to calculate the comprehensive health index. The calculation method of the comprehensive health index is as follows:

[0044] in, is the SPE statistic, T 2 Statistics, is the control limit of the SPE statistic, T 2 Control limits of statistics; Final storage comprehensive health indicators As the preset threshold value of step S5.

[0045] S5: When the comprehensive health index exceeds a preset threshold, the contribution of each feature in the original feature vector to the comprehensive health index is analyzed, and the faulty battery cell is located based on the feature with the largest contribution; S5-1) Monitor the comprehensive detection indicators output by S4 in real time and compare them with the preset thresholds. If the thresholds are exceeded, the battery system status is determined to be abnormal and fault tracing is immediately triggered; S5-2) Calculate contribution Contribution calculation methods include a. Calculate T according to the following formula 2 Statistical contribution:

[0046] Among them, r is the number of retained principal components, t i is the score of the i-th pivot, is the eigenvalue of the i-th principal component, are the elements of the loading matrix, is the current measured value of the jth original variable, is the mean of the jth original variable; b. Calculate the SPE statistic contribution according to the following formula:

[0047] in, The prediction error of the j-th variable reconstructed by the PCA model; c. According to T 2 Statistical contribution, SPE statistical contribution calculation comprehensive contribution:

[0048] in, T 2 Statistical contribution, is the contribution to the SPE statistic, is the control limit of the SPE statistic, T 2 The control limits for the statistic.

[0049] The above calculations are completed for all sixteen features.

[0050] S5-3) Sort contribution and identify key features Sort the contribution values ​​of all 16 features from largest to smallest and identify the one or several features with the largest overall contribution; S5-4) Map the abnormalities of key features to the specific faulty battery cells and trace the cell status information Different mapping measurements are used for different key feature types: Specifically, for features that can be directly traced, the individual scores are calculated according to the following formula:

[0051] Where, is the fault score of the i-th battery cell, is the current comprehensive health index value, is the preset threshold, is the total contribution of the i-th battery.

[0052] For statistical features, other data must be combined to assist in positioning. For example, if the most contributing feature is "voltage standard deviation," it indicates that inconsistency is increasing. The system needs to further check the voltage values ​​of all cells to find the cell or cells that deviate the most from the cluster average voltage and locate them as suspected fault cells.

[0053] Also includes S5-5) Generate diagnostic report The diagnostic report includes: Alarm information: alarm time, value of comprehensive health indicators; Tracing results: the top three contributing features and their contribution percentages; Location result: the serial number of the suspected faulty battery cell (e.g., "cell 23"); Auxiliary information: Current data (such as voltage and temperature) of the suspected faulty cell and its deviation from the average value.

[0054] Example 2 The difference from Example 1 is that the association relationship model is a graph attention network based on the attention mechanism, and step S3 includes S3-1) Calculate a dynamic weight coefficient for each feature in the original feature set. The weight coefficient is used to represent the importance of the feature relative to other features in the current state. Specifically, each feature is mapped to a high-dimensional representation through a shared fully connected layer with a tanh activation function to complete the linear transformation; a learnable context vector and dot product are used to obtain the unnormalized attention score; the Softmax function is used to normalize the scores of all features to the attention weight; S3-2) Multiply each feature by its corresponding dynamic weight coefficient to obtain a weighted feature set; S3-3) Perform dimensionality reduction processing on the weighted feature set to generate a low-dimensional fused feature vector.

[0055] Example 3 See also Figure 3 , a battery system health status assessment system based on multi-dimensional feature fusion for running the above method, comprising Data preprocessing module, used to obtain original operation data and perform preprocessing; Feature extraction module, used to calculate health status features of multiple dimensions from preprocessed data and generate the original feature set; The fusion module builds an association relationship model based on the association relationship between each feature in the original feature set, and uses the association relationship model to perform nonlinear fusion and output a low-dimensional fusion feature vector; Health status monitoring module, used to perform principal component analysis on low-dimensional fusion feature vectors and calculate T 2 statistics, SPE statistics and comprehensive health indicators; The fault tracing and positioning module is used to calculate the contribution of each feature and locate the faulty battery cell when the comprehensive health indicators are abnormal.

[0056] The association relationship model is a graph neural network model or a graph attention network based on an attention mechanism.

[0057] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A battery system health status assessment method based on multi-dimensional feature fusion, characterized in that: include: S1: Obtain the original operating data of the battery system and preprocess the original operating data to obtain a standardized data matrix; S2: Based on the standardized data matrix, extract health status features of multiple preset dimensions to form an original feature set. The health status features include voltage standard deviation, voltage Z score, voltage change rate, voltage-to-state-of-charge differential, temperature maximum, temperature average, temperature change rate, temperature-to-state-of-charge differential, and internal short-circuit resistance. S3: Construct a correlation model between the features in the original feature set, and perform nonlinear fusion processing on the original feature set based on the correlation model to generate a low-dimensional fusion feature vector; S4: Perform principal component analysis based on the low-dimensional fusion feature vector, extract the principal component, and calculate T 2 Statistics and SPE statistics, according to T 2 Statistics and SPE statistics are used to construct comprehensive health indicators; S5: When the comprehensive health index exceeds a preset threshold, the contribution of each feature in the original feature vector to the comprehensive health index is analyzed, and the faulty battery cell is located according to the feature with the largest contribution.

2. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The association relationship model is a graph neural network model, and step S3 includes S3-1) Define each feature in the original feature vector as a graph node and construct a feature graph based on the preset adjacency relationship; S3-2) Use the graph neural network model to process the feature graph, aggregate node neighbor information through the message passing mechanism, update the node representation, and generate a low-dimensional fused feature vector through graph readout operations.

3. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 2 is characterized in that: The strategies for processing the feature graph using the graph neural network model in step S3-2 include: Input node features and adjacency matrix into the graph convolutional network layer; Inter-layer propagation is performed using the following formula: in, represents the node feature matrix of the lth layer, represents the normalized adjacency matrix, represents the trainable weight matrix of layer l, Represents a nonlinear activation function, wherein the activation function adopts Sigmoid; After at least two layers of graph convolution operations, the final feature representations of all nodes are aggregated using a global pooling function to generate a low-dimensional fused feature vector.

4. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The association relationship model is a graph attention network based on the attention mechanism, and step S3 includes S3-1) Calculate a dynamic weight coefficient for each feature in the original feature set. The weight coefficient is used to represent the importance of the feature relative to other features in the current state. S3-2) Multiply each feature by its corresponding dynamic weight coefficient to obtain a weighted feature set; S3-3) Perform dimensionality reduction processing on the weighted feature set to generate a low-dimensional fused feature vector.

5. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The calculation method of comprehensive health index is as follows: in, is the SPE statistic, T 2 Statistics, is the control limit of the SPE statistic, T 2 The control limits for the statistic.

6. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 5 is characterized in that: In S5, the control limits of the SPE statistic are calculated as follows: in, (r=1,2,3), , is the sum of the eigenvalues ​​of the residual subspace, is the sum of squared eigenvalues ​​of the residual subspace, is the sum of the cubed eigenvalues ​​of the residual subspace, is the confidence interval of the standard normal distribution; T 2 The control limits of the statistic are calculated as follows: in, is the number of principal components after dimensionality reduction, n is the number of training samples, is the significance level, is the F distribution function.

7. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 6 is characterized in that: In S5, the contribution calculation method includes a. Calculate T according to the following formula 2 Statistical contribution: Among them, r is the number of retained principal components, t i is the score of the i-th pivot, is the eigenvalue of the i-th principal component, are the elements of the loading matrix, is the current measured value of the jth original variable, is the mean of the jth original variable; b. Calculate the SPE statistic contribution according to the following formula: in, is the prediction error of the j-th variable; c. According to T 2 Statistical contribution, SPE statistical contribution calculation comprehensive contribution: in, T 2 Statistical contribution, is the contribution to the SPE statistic, is the control limit of the SPE statistic, T 2 The control limits for the statistic.

8. The battery system health status assessment method based on multi-dimensional feature fusion according to claim 1 is characterized in that: In S2, there are a total of sixteen preset dimensions, namely, voltage standard deviation, maximum voltage Z score, second largest voltage Z score, third largest voltage Z score, average voltage Z score, maximum voltage change rate, second largest voltage change rate, third largest voltage change rate, average voltage change rate, standard deviation of voltage change rate, maximum value of voltage to state of charge differential, maximum temperature, average temperature, maximum temperature change rate, maximum value of temperature to state of charge differential, and minimum internal short circuit resistance.

9. A battery system health status assessment system based on multi-dimensional feature fusion according to claim 1, characterized in that: include Data preprocessing module, used to obtain original operation data and perform preprocessing; Feature extraction module, used to calculate health status features of multiple dimensions from preprocessed data and generate the original feature set; The fusion module builds an association relationship model based on the association relationship between each feature in the original feature set, and uses the association relationship model to perform nonlinear fusion and output a low-dimensional fusion feature vector; Health status monitoring module, used to perform principal component analysis on low-dimensional fusion feature vectors and calculate T 2 statistics, SPE statistics and comprehensive health indicators; The fault tracing and positioning module is used to calculate the contribution of each feature and locate the faulty battery cell when the comprehensive health indicators are abnormal.

10. The battery system health status assessment system based on multi-dimensional feature fusion according to claim 9, characterized in that: The association relationship model is a graph neural network model or a graph attention network based on an attention mechanism.

Images