New energy automobile power battery performance test method and system

Abstract

The invention discloses a new energy automobile power battery performance test method and system, and relates to the technical field of battery detection, and the method comprises the steps: obtaining a voltage sequence, a current sequence and a temperature sequence, carrying out the embedded alignment of data from different sample domains based on a structure contrast learning model, and generating a unified high-dimensional semantic feature space; constructing a feature sequence in the unified high-dimensional semantic feature space into a time sequence tensor, and inputting the time sequence tensor into a multi-scale convolution residual network for reconstruction to generate a reconstruction tensor; constructing a disturbance characteristic sample based on the reconstructed tensor, forming a structural consistency training data set in combination with the original sample, and performing graph representation embedding on the original sample and the disturbance characteristic sample through a graph neural network; inputting the structural consistency training data set into the double-branch neural network to generate a health degree predicted value; according to the method, high-precision, high-robustness and generalizable SOH prediction of the new energy automobile power battery health state in a label-free real operation environment can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery detection, and in particular to a new energy automobile power battery performance test method and system. BACKGROUND

[0002] As a core component of new energy vehicles, the performance degradation process of power batteries has high nonlinearity, environment dependence and individual difference, which leads to significant challenges in the actual deployment of traditional SOH (State of Health) prediction methods that rely on unified model training. Existing battery health modeling methods based on supervised learning usually rely on large-scale labeled data collected in standard test environments for model training. However, in real road running scenarios, the collected data samples often lack clear health labels, and due to the variability of working conditions, the battery operation data has obvious domain differences in statistical distribution and semantic structure. This cross-sample domain distribution inconsistency makes it difficult for existing models to generalize and apply, severely restricting the landing effect of SOH prediction ability in engineering practice.

[0003] In addition, battery operation data is represented as a multi-dimensional time series that changes over time, including state variables such as voltage, current, and temperature. During the life evolution process, it not only presents complex short-term fluctuations, but also contains long-term stable trends. Traditional methods often directly input sequence features into shallow networks or regression models for modeling, without effectively capturing the structural relevance between context semantics, thus limiting the model's ability to express non-linear degradation patterns. At the same time, in the actual running process, due to the influence of sensor noise, environmental disturbances or local abnormal discharge behavior, the battery state data has a certain degree of structural disturbance, and existing models generally lack the mechanism to model and correct such weak changes, making it difficult to ensure the robustness of SOH prediction in complex scenarios.

[0004] Therefore, it is urgent to build a battery SOH prediction method that has cross-domain semantic alignment capability, time structure understanding capability, input disturbance robustness and multi-feature fusion modeling capability to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide a new energy automobile power battery performance test method and system to solve the problems in the background art.

[0006] To achieve the above purpose, the present application provides the following technical solutions: In a first aspect, the present application provides a new energy automobile power battery performance test method, comprising: The voltage sequence, the current sequence and the temperature sequence in the historical operation data and the current to-be-tested data of the target battery are obtained, data from different sample domains are embedded and aligned based on a structure contrast learning model, and a unified high-dimensional semantic feature space is generated; Feature sequences in the unified high-dimensional semantic feature space are constructed into a time sequence tensor in a fixed window, and input into a multi-scale convolution residual network for reconstruction to generate a reconstructed tensor containing local life factors and global trend features; Based on the reconstructed tensor, a perturbation feature sample is constructed, combined with the original sample to form a structure consistency training data set, and the original sample and the perturbation feature sample are embedded into a graph atlas representation through a graph neural network. The structure consistency training data set is input into a double-branch neural network to generate a health degree prediction value.

[0007] Further, the construction method of the structure contrast learning model comprises: A first encoder and a second encoder for source domain samples and target domain samples are constructed, the first encoder and the second encoder both adopt the same one-dimensional convolutional neural network structure, the one-dimensional convolutional neural network comprises a first convolutional layer with a convolution kernel size of 5, a second convolutional layer with a convolution kernel size of 3 and a fully connected embedding layer in sequence, and the output channel numbers are 64 and 128 in sequence; The first encoder and the second encoder are configured to share the main network structure parameters, and normalization layers are respectively set to adapt to the distribution difference between the source domain samples and the target domain samples; Based on the state health degree labels of the source domain samples, a positive and negative sample pair is constructed, wherein the positive sample pair is composed of source domain samples with a label difference less than a first set threshold, and the negative sample pair includes source domain sample pairs with a label difference greater than a second set threshold and unlabeled target domain samples; A pseudo-label estimation operation is performed on the target domain samples to construct a cross-domain sample pair; A structure contrast loss function is constructed, the loss function is composed of a category center constraint term and a mutual information structure separation term, the category center constraint term adopts a Center Loss form for same-class aggregation constraint, and the mutual information structure separation term adopts an InfoNCE form for different-class sample structure contrast; The structure contrast loss function is used as an objective function, and the network parameters of the first encoder and the second encoder are iteratively updated through a back propagation and an optimization algorithm to complete the training of the structure contrast learning model.

[0008] Further, the pseudo-label estimation operation performed on the target domain samples comprises: The embedding vectors obtained after inputting the target domain sample structure into the contrast learning model are subjected to a soft clustering operation, which constructs a response probability function between the sample and each class cluster center based on a student t distribution, and the specific response probability function is as follows: ; In the formula: is the i th target domain embedding vector, is the j th class cluster center, is a degree of freedom parameter, usually set to 1, is the response probability, indicating the semantic similarity of the sample to the cluster center ; A temperature adjustment operation is performed on the response probability, which includes power scaling and normalization processing of each response probability value, for adjusting the distribution amplitude of the response probability, and the calculation formula is as follows: ; Where T is a temperature factor for adjusting the distribution amplitude of the sample to multiple categories; The response probability after temperature adjustment is matched with the source domain sample label distribution, and the class matching includes calculating the KL divergence between the response distribution of the target domain sample and the prior probability distribution of the source domain label, and minimizing the divergence to optimize the pseudo label prediction probability to obtain the class matching probability, which is as follows: ; Wherein, is the label proportion of the j th class in the source domain, is the sum of the clustering probability of the pseudo label belonging to the j th class in the target domain, is the label matching weight coefficient; The class with the maximum class matching probability in each target domain sample is taken as the pseudo label to form a pseudo label sample set, and the positive class sample or the different class sample corresponding to the pseudo label is selected from the source domain sample set to form a cross-domain positive sample pair or a cross-domain negative sample pair between the target domain sample and the source domain sample, and a cross-domain sample pair is obtained.

[0009] Further, the multi-scale convolution residual network includes: At least three parallel time convolution channels, the convolution kernel size is set to 3, 5 and 7, and the output channel number is set to 64; A channel splicing module is used to splice three groups of channel outputs in the channel dimension to form a fusion tensor; A residual connection module is used to map the original input to a fusion channel dimension and add it to the fusion tensor element by element to form a structure reconstruction tensor; An output layer outputs a reconstruction tensor with the same length as the input window.

[0010] Further, the model structure of the graph neural network comprises: a graph input module, configured to receive all sample nodes V and their connection edges E in a structure-consistent training data set, wherein the nodes include original sample nodes and perturbed sample nodes, and the edges include explicit perturbed edges and adjacency edges established based on feature similarity; two-layer graph convolution units, configured to perform feature aggregation on the nodes and their neighbor nodes; an activation function layer and a normalization layer, configured to improve the nonlinear modeling capability and numerical stability of graph embedding; an output layer, configured to finally map each node to a semantic embedding vector of a unified dimension.

[0011] Further, the construction method of the graph neural network is as follows: each original sample and its corresponding perturbed sample in the structure-consistent training data set are taken as node inputs to construct a structure-consistent graph, wherein the structure-consistent graph comprises a node set V and an edge set E, the node set V includes all original sample nodes and perturbed sample nodes, and the edge set E includes explicit perturbed edges and feature similarity edges; In the structure-consistent graph, an explicit perturbed edge is established between each pair of original sample nodes and perturbed sample nodes, and the feature similarity is calculated between any two nodes, and when the similarity is greater than or equal to a set threshold, a feature similarity edge is established to express the semantic adjacent relationship; the reconstructed tensor of each node is taken as an initial input feature and set as a node embedding vector, which is input into a graph neural network comprising at least two layers of graph convolution units, and each layer of graph convolution unit updates the embedding vector of the target node based on the feature aggregation operation of the adjacent nodes, and the specific calculation is as follows: ; In the formula: denotes the embedding vector of node i at the l+1 layer, denotes the adjacency edge aggregation coefficient of node j to node i, denotes the transformation matrix of the l layer, is an activation function, such as a ReLU function or a LeakyReLU function, denotes the input embedding vector of the neighbor node j in the l layer of the graph neural network; denotes the set of adjacent nodes of node i; In the graph neural network, a batch normalization layer and a residual connection structure are adopted to improve the training stability, and the node embedding represented by the final one-layer output is denoted as The trainable parameters of each graph convolution unit in the graph neural network are optimized by back propagation and gradient descent method based on the structural consistency contrast loss until training convergence to obtain the graph neural network. The structural consistency contrast loss function is constructed based on the cosine similarity between the embedding vectors of the original sample nodes and the corresponding perturbed sample nodes, and is specifically as follows: ; In the formula: are the graph embedding vectors of the i-th original sample node and the perturbed sample node, is a cosine similarity function.

[0012] Further, the double-branch neural network includes a trend modeling branch which is a transformer neural network and an anomaly modeling branch which is a time series auto-encoding network, the trend modeling branch is used to extract trend features, and the anomaly modeling branch is used to extract anomaly features, and the outputs of the two are fused through an attention gate mechanism to generate a health degree prediction value. The trend modeling branch adopts a multi-layer Transformer encoder structure; each layer of the Transformer encoder includes a multi-head self-attention module and a feedforward network module, and residual connections and layer normalization structures are introduced between the modules; the number of heads of the multi-head self-attention module is set to 8, and the representation dimension of each head is 64, and the feedforward network adopts a two-layer fully connected structure to project the input features to 512 dimensions and 256 dimensions, respectively. The anomaly detection branch adopts a time series auto-encoding network structure, specifically including an encoder and a decoder; the encoder is composed of two one-dimensional convolutional layers, which are used to extract local time series structures in the input sequence, and the convolution kernel size is 5 and 3 in turn, the output channel number is 64 and 128, the step is 1, and the padding method is same-padding.

[0013] Further, the attention gate mechanism is used to fuse the 256-dimensional feature vector output by the trend modeling branch and the 128-dimensional feature vector output by the anomaly modeling branch, and the fusion process includes the following operations: Splicing ; The fusion weight of the trend feature is calculated through an attention network , and the calculation formula is as follows: ; In the formula: is an attention weight matrix, ​The bias term can be used to control the normalization weight; The channel-level fusion coefficient is calculated through the gating network The calculation method is as follows: In the formula, is a gating weight matrix, is a gating bias term, is used to compress the weight to the interval [0, 1] and has a smooth control capability; The two branch outputs are fused according to the attention weight and the gating coefficient, and the fusion calculation method is as follows: In the formula, denotes element-wise multiplication, denotes a fusion vector.

[0014] Further, the training process of the double-branch neural network comprises the following steps: Obtain the original sample embedding vector and the corresponding perturbed sample embedding vector of each training sample, wherein the embedding vectors are generated by a graph neural network module and are used to represent the semantic representation of the original sample and the perturbed sample in the structure-consistent graph; Splice the structure embedding vector of the original sample and the structure embedding vector of the corresponding perturbed sample to obtain a joint input vector; Input the joint input vector into the trend modeling branch and the anomaly modeling branch, respectively; Input the feature vectors output by the two branches into an attention gate fusion module to generate a fusion vector by calculating an attention response weight and a gating coefficient; Input the fusion vector into a prediction output layer, and after processing by a fully connected layer and a Sigmoid function, generate a normalized health degree prediction value; In the training process, the mean square error between the health degree prediction value and the corresponding true SOH label is used as a supervision loss function, and the learnable parameters of the double-branch neural network are iteratively updated through a back propagation algorithm and a gradient descent method until the loss converges, thereby obtaining the double-branch neural network.

[0015] In a second aspect, the present application provides a new energy automobile power battery performance test system, which is realized based on the new energy automobile power battery performance test method described above, and comprises: An acquisition module is configured to acquire voltage sequences, current sequences and temperature sequences in historical operation data and current to-be-tested data of a target battery, embed and align data from different sample domains based on a structure comparison learning model, and generate a unified high-dimensional semantic feature space; ​​The reconstruction module is configured to construct a feature sequence in a unified high-dimensional semantic feature space into a time sequence tensor in a fixed window, input the time sequence tensor into a multi-scale convolution residual network for reconstruction, and generate a reconstructed tensor containing a local life factor and a global trend feature; The aggregation module is configured to construct a perturbation feature sample based on the reconstructed tensor, combine the original sample to form a structure-consistent training data set, and perform graph representation embedding on the original sample and the perturbation feature sample through a graph neural network. The test module is configured to input the structure-consistent training data set into a double-branch neural network to generate a health degree prediction value.

[0016] In the above technical solution, the present application provides technical effects and advantages: The present application introduces a structure comparison learning model to construct a unified semantic embedding space for source domain and target domain samples, effectively solving the problems of label missing and sample distribution inconsistency in battery health modeling. By constructing positive and negative sample pairs in the source domain and combining the pseudo-label estimation mechanism of the target domain sample, the semantic consistency of the pseudo-label generation is improved by using t-distribution soft clustering, temperature adjustment and KL divergence optimization. On this basis, cross-domain positive and negative sample pairs are constructed, so that the model still has training ability in the actual operating environment lacking of health labels, greatly improving the generalization performance and cross-platform adaptability of SOH modeling.

[0017] In the feature modeling aspect, the present application designs a multi-scale convolution residual network to reconstruct the time structure of the unified semantic feature sequence, realizes the joint extraction of short-term local life factors and long-term trend changes, and extracts context semantics in parallel through multi-scale receptive fields. The residual structure is combined to maintain the continuity of the original input, so that the output features have both local sensitivity and global robustness. This structure effectively improves the modeling ability of the battery nonlinear degradation mode, makes up for the shortcomings of existing models in long-term and short-term feature unified modeling, and enhances the expression integrity and prediction accuracy of the model under complex working conditions.

[0018] Further, the present application constructs a structure-consistent graph based on a perturbation sample, and introduces a graph neural network to model the structure-dependent relationship between the original and perturbation samples, improving the robustness of the model to slight perturbations and abnormal behaviors. On this basis, a double-branch neural network architecture that integrates trend modeling and anomaly modeling is proposed, which realizes dynamic fusion of multi-source semantics through an attention gate mechanism. This structure not only improves the expression ability of multiple degradation behaviors, but also enhances the discriminant stability and explanation ability of the SOH prediction result, providing a more reliable technical path for actual battery health evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0020] Figure 1 This is a flow chart of a new energy vehicle power battery performance testing method of the present invention; Figure 2 This is a framework diagram of a new energy vehicle power battery performance testing system according to the present invention. DETAILED DESCRIPTION

[0021] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be more comprehensive and complete, and will fully convey the concepts of the example embodiments to those skilled in the art. The accompanying drawings are merely schematic illustrations of the disclosure and are not necessarily drawn to scale. Identical reference numerals in the figures indicate identical or similar parts, and thus any repetitive description thereof will be omitted.

[0022] In addition, the described features, structures or characteristics can be combined in one or more example embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the example embodiments disclosed in this application. However, those skilled in the art will appreciate that the technical solutions disclosed in this application can be practiced while omitting one or more of the specific details, or other methods, components, steps, etc. can be adopted. In other cases, well-known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring the various aspects disclosed in this application.

[0023] Example 1 like Figure 1 As shown, this embodiment discloses a method for testing the performance of a new energy vehicle power battery, including: S101: Obtain the historical operating data of the target battery and the voltage sequence, current sequence, and temperature sequence in the current test data. Based on the structural contrastive learning model, embed and align the data from different sample domains to generate a unified high-dimensional semantic feature space. It should be noted that: the above "historical running data" and "current to be measured data" both refer to data samples collected in different time periods from the same target battery monomer or battery pack, the sampling time span can cover from the initial state of the factory to the current SOH (State of Health) evaluation time point, ensuring that the multi-stage performance evolution information is included; the sampling method is online sensor data recording, and the original data is aligned by time stamp to form a multi-dimensional time sequence sample; It can be understood that: the voltage sequence, the current sequence and the temperature sequence respectively represent the working voltage (unit: volt) of the battery, the size of the current flowing in or out (unit: ampere) and the thermal state (unit: Celsius); in the embodiment, the sampling frequency is uniformly set to 1Hz, that is, one sample is collected every second, which ensures the strict alignment of the three sequences on the time axis, avoiding the influence of time sequence dislocation on subsequent modeling; It should be noted that: after collection, the three types of sequences will be normalized to eliminate the interference of dimensional differences on neural network learning; the normalization method can be min-max normalization or z-score standardization, and after normalization, they are spliced into a three-channel tensor; It should be understood that: the sample domain refers to the source platform to which the sample belongs; in the method, there are two sample domains: a source domain (Source Domain): battery data from a standardized experimental platform, with reliable and complete SOH labels (health labels); a target domain (Target Domain): samples collected from real road running vehicles, which usually do not have explicit SOH labels; Specifically, the method for constructing the structure contrast learning model comprises: A first encoder and a second encoder for source domain samples and target domain samples are constructed, the first encoder and the second encoder both adopt the same one-dimensional convolutional neural network structure, the one-dimensional convolutional neural network comprises a first convolutional layer with a convolution kernel size of 5, a second convolutional layer with a convolution kernel size of 3 and a fully connected embedding layer in sequence, and the output channel numbers are 64 and 128 in sequence; The first encoder and the second encoder are configured to share the main network structure parameters, but are respectively provided with normalization layers to adapt to the distribution difference between the source domain samples and the target domain samples; A positive sample pair and a negative sample pair are constructed based on the state health labels of the source domain samples, wherein the positive sample pair is composed of source domain samples with a label difference less than a first set threshold, and the negative sample pair includes source domain sample pairs with a label difference greater than a second set threshold and unlabeled target domain samples; A pseudo-label estimation operation is performed on the target domain samples to construct a cross-domain sample pair; The pseudo-label estimation operation performed on the target domain samples comprises: The embedding vectors obtained after inputting the target domain sample structure into the contrast learning model are subjected to a soft clustering operation, which constructs a response probability function between the sample and each class cluster center based on a student t distribution, and is used to estimate the similarity probability of each embedding vector and the multiple semantic class centers. The specific response probability function is as follows: ; In the formula: is the i th target domain embedding vector, is the j th class cluster center, is a degree of freedom parameter, usually set to 1, is the response probability, indicating the semantic similarity of the sample to the cluster center ; A temperature adjustment operation is performed on the response probability, which includes power scaling and normalization processing of each response probability value to control the distribution entropy of the target domain sample response to each cluster center. The temperature factor is a fixed constant or a trainable parameter, which is used to adjust the distribution amplitude of the response probability. The calculation formula is as follows: ; Where T is the temperature factor, used to adjust the distribution amplitude of the sample response probability to multiple classes. T can be fixed at 1.5 or used as a learnable parameter for iterative optimization; The temperature-adjusted response probability is matched with the source domain sample label distribution. The class matching includes calculating the KL divergence between the response distribution of the target domain sample and the prior probability distribution of the source domain label, and minimizing the divergence to optimize the pseudo-label prediction probability to obtain the class matching probability. The specific formula is as follows: ; Wherein, is the label proportion of the j th class in the source domain, is the sum of the clustering probabilities of the pseudo-label belonging to the j th class in the target domain, is the label matching weight coefficient; Wherein, the cluster center is dynamically updated by exponential moving average, and the update formula is as follows: ; Wherein, is the momentum coefficient, which is obtained according to experimental data (such as = 0.95), is the average value of the embedding vector of the j th class in the current clustering round. After each round of update, all cluster center vectors are normalized to unit vectors to control the scale consistency; The category with the highest category matching probability in each target domain sample is used as a pseudo label to form a pseudo label sample set, and positive samples or outlier samples corresponding to the pseudo label are selected from the source domain sample set to form a cross-domain positive sample pair or cross-domain negative sample pair between the target domain sample and the source domain sample to obtain a cross-domain sample pair; Construct a structural contrast loss function, which consists of a category center constraint term and a mutual information structure separation term. The category center constraint term uses the Center Loss form to perform similar aggregation constraints, and the mutual information structure separation term uses the InfoNCE form to compare the structures of heterogeneous samples. Among them, Center Loss is used to aggregate similar embeddings, and its formula is: ; Where: For the The center embedding of the class; Among them, InfoNCE loss is used to distinguish heterogeneous samples, and its definition is as follows: ; Where: is the cosine similarity function; Specifically, the structural contrast loss function is a weighted sum of two terms: ; Where: and is a weight parameter greater than zero, , determined according to specific experiments, typical values ​​(such as =0.8, =0.2); Using the structural contrast loss function as the objective function, the network parameters of the first encoder and the second encoder are iteratively updated through back propagation and optimization algorithms to complete the training of the structural contrast learning model; Through a structural contrastive learning model, the semantic embeddings of target and source domain battery data are aligned, breaking through the traditional health modeling's reliance on large-scale labeled data. This step effectively improves the semantic usability of unlabeled samples from real road environments through mechanisms such as constructing positive and negative sample pairs, pseudo-label estimation, and cross-domain contrast constraints. This unified high-dimensional feature representation space not only enhances the model's generalization ability for cross-domain samples but also provides a semantically consistent foundation for subsequent lifespan modeling, improving the practicality and robustness of SOH modeling.

[0024] S102: Constructing the feature sequence in the unified high-dimensional semantic feature space into a time series tensor according to a fixed window, inputting it into a multi-scale convolutional residual network for reconstruction, and generating a reconstructed tensor containing local life factors and global trend features; Specifically, the fixed window is constructed as a time series tensor, including: The obtained unified high-dimensional semantic feature sequence is sliced and constructed as a three-dimensional time series tensor in a fixed time window mode , and the specific mode is: Let the original feature sequence be a vector sequence of length L Each , for example, d=256; Set the window length to w=16 and the step to s=8, perform sliding window cutting, and obtain the time series tensor: ; Wherein, ; It should be noted that: the multi-scale convolution residual network is a deep neural network structure for modeling sequence context semantics, which belongs to a multi-branch residual convolution network for time series feature modeling; the network takes a plurality of feature tensors with fixed window lengths as input and outputs a structure reconstruction tensor with the same dimension, which is used to restore the semantic context continuity; Wherein, the multi-scale convolution residual network comprises: At least three parallel time convolution channels, respectively corresponding to different receptive field sizes, and the convolution kernel size is preferably set to 3, 5, and 7, and the output channel number is set to 64; It can be understood that: the channel with a convolution kernel of 3 is more sensitive to short-term changes and can extract the micro-change behavior of voltage, current and temperature in a short time; the channel with a convolution kernel of 7 covers a longer time interval, which is beneficial to identify the trend evolution characteristics; A channel splicing module is used to splice three groups of channel outputs in the channel dimension to form a fusion tensor; A residual connection module is used to map the original input to the fusion channel dimension and add it to the fusion tensor element by element to form a structure reconstruction tensor: , wherein, is the original input, represents a linear transformation layer; An output layer outputs a reconstruction tensor with the same length as the input window; The multi-scale convolution residual network is trained in an end-to-end manner, the loss function combines the structure reconstruction error (such as L2 loss) and the trend constraint term (such as first derivative information), and the multi-channel parameter update is completed through back propagation, and the training target is to minimize the semantic difference between the reconstructed features and the original input; In this embodiment, in order to extract life representation features with time context semantics from the unified high-dimensional semantic feature space, a multi-scale convolution residual network is used to process the feature time series tensor, thereby generating a reconstruction tensor containing local life factors and global trend features; It can be understood that the feature sequence in the unified high-dimensional semantic feature space is a set of feature vectors encoded by a structural contrast learning model, which has consistent semantic expression across sample domains, but still lacks deep modeling capability for temporal context structure; therefore, it is necessary to further analyze the sequence structure by a convolutional network; It should be noted that the feature sequence is first sliced in a fixed window manner to construct a three-dimensional time series tensor. Let the length of the original feature sequence be L, the dimension of the feature vector at each time step be d, the window length be w, and the step length be S. Then, the sliding window method is used to segment the sequence to generate a set of tensor fragments with a length of ; Next, the time series tensor constructed above is input into a multi-scale convolutional residual network for processing. The basic structure of the network includes three parallel time convolution channels, each using one-dimensional convolution operation with a convolution kernel size of 3, 5, and 7 to extract context semantic features under different time receptive fields; Among them, the channel with a convolution kernel size of 3 mainly perceives local context changes and is suitable for detecting short-term voltage, current and temperature micro-fluctuation behavior. This part of the feature is defined as the local life factor. The channel with a convolution kernel size of 7 is used to cover a larger time window to extract trend features in the battery life evolution process, such as capacity decay curve shape, temperature stability drift, etc. Such features are defined as global trend features. The channel with a convolution kernel of 5 serves as a medium-scale supplement to capture the transition information between local and global, and improve the feature continuity and semantic level perception ability; It is worth noting that the tensors output by the three convolution channels are concatenated in the channel dimension to form a fusion tensor , where C is the output channel number of each channel (such as 64); then, the original input tensor is adjusted to the same channel dimension as the fusion tensor through a linear mapping module, and is added element by element with the fusion tensor to form a residual reconstruction tensor; It should be understood that the residual reconstruction tensor not only retains the low-level semantic information of the original input feature, but also fuses the context structure semantics at multiple time scales, having unified expression ability of local sensitivity and global robustness; For example, assuming that the tensors output by the three convolution branches are (local features extracted by a convolution kernel of 3), (local features extracted by a convolution kernel of 5), and (local features extracted by a convolution kernel of 7); then the fusion tensor is: , the residual reconstruction tensor is: , and the final output contains local life factors (from​ embodied), also contains global trend features (embodied by embodied), realizing double-feature fusion, wherein, represents a channel dimension splicing operation, used to connect multiple tensors in the feature dimension (i.e., the last dimension); By using a multi-scale convolution residual network to reconstruct the time structure of the high-dimensional semantic feature sequence, short-term micro-change behavior and long-term trend evolution can be modeled at the same time, and local sensitivity and global robustness are taken into account. The network combines three types of receptive field convolution channels and uses a residual mechanism to maintain low-level semantic continuity, realizes the fusion and extraction of local life factors and global trend features, and effectively solves the problem of insufficient modeling capability of traditional models for battery nonlinear degradation patterns, and significantly improves the context integrity and sequence modeling accuracy of feature expression.

[0025] S103; based on the reconstructed tensor, construct a perturbation feature sample, form a structure-consistent training data set in combination with the original sample, and embed the original sample and the perturbation feature sample into a graph representation through a graph neural network; It should be noted that the perturbation feature sample is obtained by performing a structure perturbation operation on each sample vector in the reconstructed tensor, and the structure perturbation operation includes Gaussian perturbation, feature masking, or time warping, etc. The structure perturbation operation is performed on the segments in the reconstructed tensor to obtain a perturbation sample with similar semantics to the original tensor but with slight changes. In implementation, a structure-consistent training data set is formed in combination with the original sample, and each original tensor sample is stored in combination with its corresponding perturbation sample to form a consistent learning input set: ; To maintain the relevance of the original perturbation sample, a unified index number can be set, and a consistent global sample ID and a perturbation identification bit are used for one-to-one correspondence of nodes in the subsequent graph structure; through the sample pairing and index consistency setting, it is ensured that the relationship between the original sample and the perturbation sample can be tracked in the structure-consistent learning, and the goal of constructing the structure-consistent training data set is achieved; Specifically, the graph neural network is a neural network model for modeling the structure-consistent relationship between samples, and has node feature aggregation and graph structure perception capabilities, and is suitable for graph representation embedding learning scenarios between original samples and perturbation samples. The model structure of the graph neural network includes: a graph input module, configured to receive all sample nodes V and their connection edges E in a structure consistency training data set, wherein the nodes include original sample nodes and perturbed sample nodes, and the edges include explicit perturbation edges and adjacency edges established based on feature similarity; wherein the explicit perturbation edge is mapped to two nodes in the graph for each original sample and its corresponding perturbed sample, and an explicit connection edge is added therebetween to express the strong structural correlation between the sample and its perturbed version; the edge does not depend on feature similarity calculation and belongs to prior explicit connection; the adjacency edge is calculated for any two nodes to calculate the cosine similarity between their initial structure embedding vectors, and when the similarity is greater than a set threshold, an edge is added between the nodes to express their proximity in the structure semantic space; Through the construction of the above two types of edges, the structure consistency graph not only explicitly retains the mapping relationship between each pair of perturbed samples, but also captures the semantic similarity relationship between different samples, thereby providing a structural basis for subsequent adjacency feature aggregation in the graph neural network; two-layer graph convolution units (Graph Convolutional Layers) for feature aggregation of nodes and their neighbor nodes; an activation function layer and a normalization layer for improving the nonlinear modeling capability and numerical stability of the graph embedding; an output layer for finally mapping each node to a semantic embedding vector of a unified dimension, representing its representation result under the global graph structure; In implementation, the construction method of the graph neural network is as follows: each original sample in the structure consistency training data set and its corresponding perturbed sample are respectively taken as a node input to construct a structure consistency graph, wherein the structure consistency graph includes a node set V and an edge set E, the node set V includes all original sample nodes and perturbed sample nodes, and the edge set E includes explicit perturbation edges and feature similarity edges; In the structure consistency graph, an explicit perturbation edge is established between each pair of original sample nodes and perturbed sample nodes, and a feature similarity edge is established between any two nodes when the feature similarity is greater than or equal to a set threshold to express the semantic proximity relationship; the reconstructed tensor of each node is taken as an initial input feature and set as a node embedding vector, which is input into a graph neural network including at least two layers of graph convolution units, and the embedding vector of the target node is updated based on the feature aggregation operation of the adjacent nodes in each layer of graph convolution units, and the specific calculation is as follows: ; In the formula, denotes the embedding vector of node i at the l+1 layer, denotes the aggregation coefficient of the adjacent edges of node j to node i, denotes the transformation matrix of the l-th layer, is an activation function, such as a ReLU function or a LeakyReLU function, denotes the input embedding vector of the neighbor node j in the l-th layer of the graph neural network; denotes the set of adjacent nodes of node i; In the graph neural network, a batch normalization layer and a residual connection structure are used to improve the training stability, and the node embedding representation of the final layer output is obtained; The trainable parameters of each graph convolution unit in the graph neural network are optimized based on the structure consistency contrast loss through back propagation and gradient descent method until the training converges to obtain the graph neural network; The structure consistency contrast loss function is constructed based on the cosine similarity between the embedding vectors of the original sample node and the corresponding perturbed sample node, and is specifically as follows: ; In the formula: are the graph embedding vectors of the i-th original sample node and the perturbed sample node, respectively, is a cosine similarity function; By constructing a perturbed sample pair and introducing a graph neural network to establish a structure consistent graph, this step can effectively simulate the slight abnormal perturbation of battery data in a real environment, thereby improving the robustness of the model to input perturbations. The structure consistency learning mechanism preserves the structural relationship between the original and perturbed samples through graph-level feature aggregation, making the embedding vector have stronger anti-interference ability and structural expression ability, avoiding distortion of key life features under slight perturbations, and improving the stability and reliability of the SOH prediction model in actual deployment.

[0026] S104: Input the structure consistency training data set into the double-branch neural network, extract the trend features through the transformer neural network, extract the abnormal features through the time series auto-encoding network, and fuse the output results of the two branches based on the attention gate mechanism to generate the health degree prediction value; The health degree prediction value is a real number with a value range of [0, 1], which is used to represent the relative health status of the i-th target battery sample, and the closer the value is to 1, the healthier the state, and the closer the value is to 0, the worse the health status. Specifically, the double-branch neural network includes a trend modeling branch which is a transformer neural network and an abnormal modeling branch which is a time series auto-encoding network, the trend modeling branch is used to extract trend features, and the abnormal modeling branch is used to extract abnormal features. The outputs of the two are fused through the attention gate mechanism to generate the health degree prediction value; The trend modeling branch adopts a multi-layer Transformer encoder structure, receives the joint input vector, and models the long-term trend evolution relationship between samples; each layer of the Transformer encoder includes a multi-head self-attention module and a feedforward network module, and residual connections and layer normalization structures are introduced between the modules; the number of heads of the multi-head self-attention module is set to 8, and the representation dimension of each head is 64; the feedforward network adopts a two-layer fully connected structure, which projects the input features to 512 dimensions and 256 dimensions, respectively; the final output dimension of the branch is a trend feature vector of 256 dimensions; The anomaly detection branch adopts a time series auto-encoding network structure, specifically including an encoder and a decoder; the encoder is composed of two one-dimensional convolutional layers, which are used to extract local time series structures in the input sequence, the convolution kernel size is 5 and 3 in turn, the output channel number is 64 and 128, the step is 1, and the padding method is same-padding; after the encoder output is reduced to a 128-dimensional latent representation by a fully connected layer, it is input into the symmetric decoder for sequence reconstruction; the reconstruction error (the error between the original embedding and the reconstruction result is measured by MSE, and the error vector can be used to infer the abnormal sensitivity) is used to indirectly extract abnormal behavior features, and finally an abnormal feature vector of 128 dimensions is obtained; The attention gating mechanism is used to fuse the 256-dimensional feature vector output by the trend modeling branch and the 128-dimensional feature vector output by the anomaly modeling branch The fusion process includes the following operations: and are spliced: ; The fusion weight of the trend feature is calculated by the attention network , and the calculation formula is as follows: ; In the formula, is the attention weight matrix, is the bias term, which can be used to control the normalized weight; The channel-level fusion coefficient is calculated by the gating network , and the calculation method is as follows: ; In the formula, is the gating weight matrix, is the gating bias term, is used to compress the weight to the interval [0, 1], which has a smooth control ability; The outputs of the two branches are fused according to the attention weight and the gating coefficient, and the fusion calculation method is as follows: ​ ; wherein: represents element-wise multiplication, represents a fusion vector; It should be noted that the fusion vector is finally input into an output prediction layer, which includes a fully connected layer and a Sigmoid normalization function, for compressing the output value to the interval [0, 1] to represent the health probability or normalized SOH prediction value of the sample; the dual-branch neural network is trained in a supervised manner, and the training target is to minimize the mean square error loss function between the prediction result and the true SOH label, which is defined as follows: ; wherein, is the prediction value of the i-th sample, is the corresponding SOH label value, is a sample loss weight coefficient for adjusting the contribution intensity of different samples in the training process; In implementation, the training process of the dual-branch neural network includes: obtaining the original sample embedding vector and the corresponding perturbed sample embedding vector of each training sample, which are generated by the graph neural network module and used to represent the semantic representation of the original sample and the perturbed sample in the structure-consistent graph; splicing the structure embedding vector of the original sample and the structure embedding vector of the corresponding perturbed sample to obtain a joint input vector; It should be noted that the input of the dual-branch neural network is the structure-aware vector output by the graph neural network embedding module, including the original sample embedding vector and the perturbed sample embedding vector; Specifically, the embedding vectors of each group of original samples and perturbed samples are spliced to form a joint input vector with a dimension of ; ; wherein, represents the embedding vector output by the graph neural network module, the input is a tensor, and the output is a fixed-length semantic vector, with a dimension of , represents a vector splicing operation, and the output is a dimensional vector; The joint input vector is input into the trend modeling branch and the anomaly modeling branch, respectively; The feature vectors output by the two branches are input into an attention gate fusion module to generate a fusion vector by calculating the attention response weight and the gate coefficient; The fusion vector is input to a prediction output layer, and after being processed by a full connection layer and a Sigmoid function, a normalized health degree prediction value is generated, the health degree prediction value being a real number in the interval between 0 and 1, used to represent the state of health (SOH) of the target battery sample; In the training process, the mean square error between the health degree prediction value and the corresponding real SOH label is taken as a supervision loss function, and the learnable parameters of the double-branch neural network are iteratively updated through a back propagation algorithm and a gradient descent method until the loss converges, to obtain the double-branch neural network; By constructing a double-branch neural network architecture that fuses trend modeling and anomaly modeling, and introducing an attention gate mechanism for dynamic feature fusion, the accuracy and robustness of the SOH prediction result are effectively improved. The trend modeling branch extracts the long-term degradation trend of the battery, and the anomaly modeling branch captures potential sudden behaviors and abnormal changes. The fusion module automatically adjusts the weights of each branch according to the semantic of the input sample, realizing adaptive modeling of multiple degradation modes. This structure overcomes the problem of dependence on a single feature in traditional SOH prediction models, and has higher discrimination accuracy and semantic coverage ability.

[0027] Embodiment 2 As shown in Figure 2 the embodiment, the embodiment discloses a new energy vehicle power battery performance test system, which comprises: An acquisition module 201 is configured to acquire voltage sequences, current sequences and temperature sequences in historical running data and current to-be-tested data of a target battery, embed and align data from different sample domains based on a structure comparison learning model, and generate a unified high-dimensional semantic feature space; A reconstruction module 202 is configured to construct feature sequences in the unified high-dimensional semantic feature space into a time sequence tensor according to a fixed window, input the time sequence tensor into a multi-scale convolution residual network for reconstruction, and generate a reconstructed tensor containing local life factors and global trend features; An aggregation module 203 is configured to construct perturbation feature samples based on the reconstructed tensor, combine the original samples to form a structure consistency training data set, and perform graph representation embedding on the original samples and the perturbation feature samples through a graph neural network; A test module 204 is configured to input the structure consistency training data set into a double-branch neural network, and generate a health degree prediction value.

[0028] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation. The preset parameters, weights and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0029] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center through a wired network or a wireless network. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0030] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of the claims of the present application.

Claims

1. A new energy vehicle power battery performance testing method, characterized in that: include: Obtain the target battery's historical operating data and the voltage, current, and temperature sequences from the current test data. Use a structural contrastive learning model to embed and align data from different sample domains to generate a unified high-dimensional semantic feature space. The feature sequence in the unified high-dimensional semantic feature space is constructed as a time series tensor according to a fixed window, and input into a multi-scale convolutional residual network for reconstruction to generate a reconstructed tensor containing local life factors and global trend features; Construct perturbation feature samples based on the reconstructed tensor, combine them with the original samples to form a structural consistency training dataset, and embed the original samples and perturbation feature samples into graph representations through a graph neural network; The structural consistency training dataset is input into a two-branch neural network to generate health prediction values.

2. The new energy vehicle power battery performance testing method according to claim 1, characterized in that: The method for constructing the structural contrast learning model includes: Construct a first encoder and a second encoder for source domain samples and target domain samples, respectively. The first encoder and the second encoder both use the same one-dimensional convolutional neural network structure. The one-dimensional convolutional neural network includes a first convolutional layer with a convolution kernel size of 5, a second convolutional layer with a convolution kernel size of 3, and a fully connected embedding layer. The number of output channels is 64 and 128, respectively. The first encoder and the second encoder are configured to share backbone network structure parameters, and normalization layers are set respectively to adapt to the distribution difference between source domain samples and target domain samples; Construct positive and negative sample pairs based on the state health labels of the source domain samples, where the positive sample pairs consist of source domain samples whose label difference is less than a first set threshold, and the negative sample pairs include source domain sample pairs whose label difference is greater than a second set threshold and unlabeled target domain samples; Performing a pseudo-label estimation operation on the target domain samples to construct cross-domain sample pairs; Construct a structural contrast loss function, which consists of a category center constraint term and a mutual information structure separation term. The category center constraint term uses the Center Loss form to perform similar aggregation constraints, and the mutual information structure separation term uses the InfoNCE form to compare the structures of heterogeneous samples. The structural contrast loss function is used as the objective function, and the network parameters of the first encoder and the second encoder are iteratively updated through back propagation and optimization algorithms to complete the training of the structural contrast learning model.

3. The new energy vehicle power battery performance testing method according to claim 2, characterized in that: Performing a pseudo-label estimation operation on the target domain sample, including: A soft clustering operation is performed on the embedding vectors obtained after inputting the target domain samples into the structural contrast learning model. The soft clustering operation constructs a response probability function between the samples and various cluster centers based on the Student's t distribution. The specific response probability function is as follows: ； Where: is the i-th target domain embedding vector, is the j-th cluster center, is the degree of freedom parameter, usually set to 1, is the response probability, representing the sample Belongs to the cluster center Semantic similarity of A temperature adjustment operation is performed on the response probability. The temperature adjustment operation includes power scaling and normalization of each response probability value to adjust the distribution amplitude of the response probability. The calculation formula is as follows: ； Among them, T is the temperature factor, which is used to adjust the distribution amplitude of the sample's response probability to multiple categories; The temperature-adjusted response probability is then matched against the source domain sample label distribution. This matching involves calculating the KL divergence between the target domain sample response distribution and the source domain label prior probability distribution, and minimizing this divergence to optimize the pseudo-label prediction probability. The resulting matching probability is as follows: ； in, is the label ratio of the jth class in the source domain, is the sum of the clustering probabilities that the pseudo label in the target domain belongs to the jth class, is the label matching weight coefficient; The category with the highest category matching probability in each target domain sample is used as a pseudo label to form a pseudo label sample set, and positive samples or heterogeneous samples corresponding to the pseudo label are selected from the source domain sample set to form a cross-domain positive sample pair or cross-domain negative sample pair between the target domain sample and the source domain sample to obtain a cross-domain sample pair.

4. The new energy vehicle power battery performance testing method according to claim 3, characterized in that: The multi-scale convolutional residual network includes: At least three parallel temporal convolution channels, with kernel sizes set to 3, 5, and 7, and the number of output channels set to 64; A channel concatenation module, which concatenates the three sets of channel outputs in the channel dimension to form a fused tensor; A residual connection module is used to map the original input to the fusion channel dimension and then add it element-by-element to the fusion tensor to form the structure reconstruction tensor; An output layer that outputs a reconstructed tensor of the same length as the input window.

5. The new energy vehicle power battery performance testing method according to claim 1, characterized in that: The model structure of the graph neural network includes: A graph input module is used to receive all sample nodes V and their connecting edges E in the structural consistency training dataset, where the nodes include original sample nodes and perturbed sample nodes, and the edges include explicit perturbation edges and adjacent edges established based on feature similarity; Two-layer graph convolution unit, used to aggregate features of a node and its neighboring nodes; Activation function layer and normalization layer are used to improve the nonlinear modeling ability and numerical stability of graph embedding; The output layer is used to finally map each node into a semantic embedding vector of uniform dimension.

6. The new energy vehicle power battery performance testing method according to claim 1, characterized in that: The method for constructing the graph neural network is as follows: Each original sample in the structural consistency training data set and its corresponding perturbation sample are respectively input as nodes to construct a structural consistency graph, wherein the structural consistency graph includes a node set V and an edge set E, the node set V includes all original sample nodes and perturbation sample nodes, and the edge set E includes explicit perturbation edges and feature similarity edges; In the structural consistency graph, an explicit perturbation edge is established between each pair of original sample nodes and perturbation sample nodes, and feature similarity is calculated between any two nodes. When the similarity is greater than or equal to a set threshold, a feature similarity edge is established to express the semantic proximity relationship; The reconstructed tensor of each node is used as the initial input feature, set as the node embedding vector, and input into a graph neural network containing at least two layers of graph convolution units. Each layer of graph convolution units updates the embedding vector of the target node based on the feature aggregation operation of adjacent nodes. The specific calculation is: ； Where: represents the embedding vector of node i in the l+1 layer, represents the aggregation coefficient of the adjacent edge of node j to node i, Represented as the transformation matrix of the lth layer, is an activation function, such as the ReLU function or the LeakyReLU function, Represents the input embedding vector of the neighbor node j in the lth layer of the graph neural network; represents the set of adjacent nodes of node i; In the graph neural network, batch normalization layers and residual connection structures are used to improve training stability, and the nodes output by the final layer are embedded into the representation set; Based on the structural consistency contrast loss, the trainable parameters of each graph convolution unit in the graph neural network are optimized by back propagation and gradient descent method until the training converges to obtain a graph neural network; The structural consistency contrast loss function is constructed based on the cosine similarity between the embedding vectors of the original sample node and the corresponding perturbation sample node, as follows: ； Where: are the graph embedding vectors of the i-th original sample node and the perturbation sample node, is the cosine similarity function.

7. The new energy vehicle power battery performance testing method according to claim 6, characterized in that: The dual-branch neural network includes a trend modeling branch, which is a transformer neural network, and an anomaly modeling branch, which is a temporal autoencoder network. The trend modeling branch is used to extract trend features, and the anomaly modeling branch is used to extract anomaly features. The outputs of the two are fused through an attention gating mechanism to generate a health prediction value. The trend modeling branch adopts a multi-layer Transformer encoder structure; each layer of the Transformer encoder includes a multi-head self-attention module and a feedforward network module, and introduces residual connections and layer normalization structures between the modules; the number of heads in the multi-head self-attention module is set to 8, and the representation dimension of each head is 64. The feedforward network adopts a two-layer fully connected structure, projecting the input features to 512 dimensions and 256 dimensions respectively; Among them, the anomaly detection branch adopts a temporal autoencoder network structure, which specifically includes an encoder and a decoder. The encoder consists of two one-dimensional convolutional layers, which are used to extract the local temporal structure in the input sequence. The convolution kernel sizes are 5 and 3 respectively, the number of output channels is 64 and 128, the step size is 1, and the padding method is same-padding.

8. The new energy vehicle power battery performance testing method according to claim 7, characterized in that: The attention gating mechanism is used to fuse the 256-dimensional feature vector output by the trend modeling branch and the 128-dimensional feature vector output by the anomaly modeling branch , the fusion process includes the following operations: right and To splice: ； Calculate the fusion weight of trend features through attention network , which is calculated as follows: ； Where; is the attention weight matrix, is a bias term that can be used to control the normalization weight; Calculate channel-level fusion coefficients through gating network , calculated as: ； Where; is the gating weight matrix, is the gate bias term, Used to compress weights to the [0,1] interval, with smooth control capabilities; The two branch outputs are fused according to the attention weight and the gating coefficient. The fusion calculation method is: ； Where: represents element-wise multiplication, Represents the fusion vector.

9. The new energy vehicle power battery performance testing method according to claim 8, characterized in that: The training process of the dual-branch neural network includes: Obtain the original sample embedding vector and the corresponding perturbation sample embedding vector of each training sample. The embedding vector is generated by the graph neural network module and is used to represent the semantic representation of the original sample and the perturbation sample in the structural consistency graph; Concatenate the structural embedding vector of the original sample and the structural embedding vector of its corresponding perturbation sample to obtain a joint input vector; Input the joint input vector into the trend modeling branch and the anomaly modeling branch respectively; The feature vectors output by the two branches are input into the attention gating fusion module, and the fusion vector is generated by calculating the attention response weight and the gating coefficient; The fusion vector is input to the prediction output layer, and after being processed by the fully connected layer and the Sigmoid function, a normalized health prediction value is generated; During the training process, the mean square error between the health prediction value and the corresponding true SOH label is used as the supervised loss function, and the learnable parameters of the two-branch neural network are iteratively updated through the back propagation algorithm and the gradient descent method until the loss converges to obtain a two-branch neural network.

10. A new energy vehicle power battery performance testing system, implemented based on the new energy vehicle power battery performance testing method according to any one of claims 1 to 8, characterized in that: include: The acquisition module is used to obtain the historical operating data of the target battery and the voltage, current, and temperature series in the current test data. It embeds and aligns the data from different sample domains based on the structural contrast learning model to generate a unified high-dimensional semantic feature space. The reconstruction module is used to construct the feature sequence in the unified high-dimensional semantic feature space into a time series tensor according to a fixed window, input it into the multi-scale convolutional residual network for reconstruction, and generate a reconstructed tensor containing local life factors and global trend features; The aggregation module is used to construct perturbation feature samples based on the reconstructed tensor, combine them with the original samples to form a structural consistency training dataset, and embed the original samples and perturbation feature samples into graph representations through a graph neural network; The testing module is used to input the structural consistency training data set into the two-branch neural network to generate a health prediction value.

Images