A multi-working-condition energy storage battery health evaluation system based on unsupervised domain adaptation

Abstract

The application relates to a multi-working-condition energy storage battery health evaluation system based on unsupervised domain adaptation, which comprises a scale perception knowledge query coding module, which extracts battery health state related features in respective fields from original data and target data; a bidirectional cross-attention field mixing module, which utilizes a multi-head cross-attention mechanism to reduce the gap between the battery health state related features in respective fields, and extracts independent features of the battery health state in respective fields again; and a regression module, which predicts the battery health state in the original data field and the target data field according to the independent features of the battery health state in respective fields processed by the multi-head cross-attention mechanism.

Description

Technical Field

[0001] This invention relates to the field of power supply technology, specifically to a multi-condition energy storage battery health assessment system based on unsupervised domain adaptive design. Background Technology

[0002] Long-term fossil fuel development and utilization have led to energy depletion and increasingly prominent climate risks. To mitigate global climate change and promote sustainable growth, the world is undertaking energy transition and advancing green and low-carbon development. In recent years, new energy power generation technologies such as wind and solar power have continuously developed, but their instability and difficulty in grid connection have remained bottlenecks restricting their development. Energy storage batteries, with their strong controllability, smooth transition, peak shaving and valley filling, and frequency and voltage regulation functions, can increase the flexibility of the power system and improve the stability of wind and solar power generation. Because energy storage systems are costly, and failures can cause serious safety problems, timely detection and evaluation of energy storage systems are of great importance.

[0003] This invention is a health assessment system for energy storage batteries based on unsupervised domain adaptive multi-condition operation. It uses a data-driven model to intelligently assess the energy storage system. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a multi-condition energy storage battery health assessment system based on unsupervised domain adaptive design.

[0005] The technical solution of the present invention is as follows: On the one hand, this invention proposes a multi-condition energy storage battery health assessment system based on unsupervised domain adaptive methods, including... The scale-aware knowledge query encoding module extracts battery health status-related features from the raw data and target data respectively, in their respective domains. The bidirectional cross-attention domain hybrid module utilizes a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in various domains and extracts independent features of battery health status in each domain again. The regression module predicts the battery health status in the original data domain and the target data domain based on the independent features of the battery health status in each domain after processing by the multi-head cross-attention mechanism.

[0006] In a preferred embodiment, the scale-aware knowledge query encoding module includes several L-shaped stacked CNN bottleneck layers and CNN pooling layers, and a scale-aware knowledge query layer; the CNN bottleneck layers are used to extract multi-scale battery health status-related features from the original data or target data; the CNN pooling layers are used to transform the dimension of the output features of the CNN bottleneck layers; and the scale-aware knowledge query layer is used to query and output state knowledge from different domains.

[0007] As a preferred implementation, in order to query battery health status related features of different domains in the scale-aware knowledge query layer, the Transformer model is used to attach the label status of the battery health status related features of the multi-scale domains.

[0008] As a preferred embodiment, the step of using a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in different domains specifically includes: Battery health status-related features with domain labels are input into multi-head self-attention transformers to extract independent features from the original domain and the target domain, respectively. Then, a multi-head cross-attention mechanism is used to make the independent features of each domain more similar to those of another domain. The specific calculation formula for the multi-head cross-attention mechanism is as follows:

[0009] In the formula, Q , K and V Represents the query, key, and value vector matrix of the input sequence; . and These are learnable parameters; S and F represent the state knowledge of the original domain and the target domain, respectively.

[0010] As a preferred embodiment, the step of predicting the battery health status in the original data domain and the target data domain based on the independent features of the battery health status in each domain after processing by the multi-head cross-attention mechanism is specifically as follows: The independent features of each domain, calculated by multi-head cross-attention, are substituted into the regression head model for regression calculation; The loss function used in the regression head model is the mean squared error loss function, specifically:

[0011] In the formula, s and t represent the original domain marker and the target domain marker, respectively; n s and n t This represents the number of battery health state-related features in the original domain and the number of battery health state-related features in the target domain. This represents the mean squared error loss function; Predict the battery health status in the original data domain and the target data domain based on the calculation results.

[0012] On the other hand, this invention proposes a multi-condition energy storage battery health assessment method based on unsupervised domain adaptive methods, specifically including the following steps: Extract battery health status-related features from the original data and target data respectively, for their respective domains; By utilizing a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in different domains, the independent features of battery health status in each domain are extracted again. Based on the independent features of battery health status in each domain after processing by the multi-head cross-attention mechanism, predict the battery health status in the original data domain and the target data domain.

[0013] In a preferred embodiment, the scale-aware knowledge query encoding module includes several L-shaped stacked CNN bottleneck layers and CNN pooling layers, and a scale-aware knowledge query layer; the CNN bottleneck layers are used to extract multi-scale battery health status-related features from the original data or target data; the CNN pooling layers are used to transform the dimension of the output features of the CNN bottleneck layers; and the scale-aware knowledge query layer is used to query and output state knowledge from different domains.

[0014] As a preferred implementation, in order to query battery health status related features of different domains in the scale-aware knowledge query layer, the Transformer model is used to attach the label status of the battery health status related features of the multi-scale domains.

[0015] As a preferred embodiment, the step of using a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in different domains specifically includes: Battery health status-related features with domain labels are input into multi-head self-attention transformers to extract independent features from the original domain and the target domain, respectively. Then, a multi-head cross-attention mechanism is used to make the independent features of each domain more similar to those of another domain. The specific calculation formula for the multi-head cross-attention mechanism is as follows:

[0016] In the formula, Q , K and V Represents the query, key, and value vector matrix of the input sequence; . and These are learnable parameters; S and F represent the state knowledge of the original domain and the target domain, respectively.

[0017] As a preferred embodiment, the step of predicting the battery health status in the original data domain and the target data domain based on the independent features of the battery health status in each domain after processing by the multi-head cross-attention mechanism is specifically as follows: The independent features of each domain, calculated by multi-head cross-attention, are substituted into the regression head model for regression calculation; The loss function used in the regression head model is the mean squared error loss function, specifically:

[0018]

[0019] In the formula, s and t represent the original domain marker and the target domain marker, respectively; n s and n t This represents the number of battery health state-related features in the original domain and the number of battery health state-related features in the target domain. This represents the mean squared error loss function; Predict the battery health status in the original data domain and the target data domain based on the calculation results.

[0020] The present invention has the following beneficial effects: 1. This invention can use raw battery operating data under different operating conditions to evaluate the health status of the battery under different operating conditions, making the most efficient use of the energy storage system and fully utilizing the energy storage system while ensuring its safety.

[0021] 2. This invention utilizes the characteristics of convolutional neural networks to achieve complete extraction of all features from both the original domain data and the target domain data, thereby improving the accuracy of the prediction results.

[0022] 3. This invention utilizes the multi-head cross-attention mechanism to learn the domain-invariant characteristic, thereby reducing the differences between different domains.

[0023] 4. This invention utilizes the mean squared error loss function to minimize the regression error of all samples, making the final prediction result of battery health status sufficiently accurate. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is an overall architecture diagram of a hybrid network in the knowledge query domain according to an embodiment of the present invention; Figure 3 A schematic diagram of the scale-aware knowledge query encoder; Figure 4 This is a diagram of the bottleneck layer structure of a CNN. Figure 5 This is a diagram of the pooling layer structure in a CNN. Figure 6 This is a schematic diagram of the Transformer structure; Figure 7 This is a schematic diagram of the bidirectional cross-attention domain mixer. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0027] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0029] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0030] Example 1: See Figure 1 A health assessment system for multi-condition energy storage batteries based on unsupervised domain adaptive methods includes... The scale-aware knowledge query encoding module extracts battery health status-related features from the raw data and target data respectively, in their respective domains. The bidirectional cross-attention domain hybrid module utilizes a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in various domains and extracts independent features of battery health status in each domain again. The regression module predicts the battery health status in the original data domain and the target data domain based on the independent features of the battery health status in each domain after processing by the multi-head cross-attention mechanism.

[0031] In this embodiment, the scale-aware knowledge query encoding module, the bidirectional cross-attention domain hybrid module, and the regression module together constitute the knowledge query domain hybrid network, and its overall structural framework is as follows: Figure 2 As shown.

[0032] The scale-aware knowledge query encoding module takes the raw data as input to generate domain-specific health status-related embedded features. Since each domain provides time-series data of energy storage batteries from different feature spaces and aging modes, this invention sets up private feature encoders for the source and target domains, denoted as follows: and Raw data and target data Input a feature encoder to generate domain-specific features. After that, characteristics of specific fields. It is fed into a bidirectional cross-attention domain hybrid module to alleviate the domain feature gap and extract domain-invariant features. Finally, the domain-invariant features. Prediction results are generated using the regression head. The regression head consists of a two-layer fully connected neural network. The algorithm flow of the proposed framework can be represented as follows:

[0033] In the formula: and These represent the functions of the bidirectional cross-attention domain mixer and the regression head, respectively, where n and m represent the domain and knowledge state, respectively.

[0034] In a preferred embodiment of this invention, the scale-aware knowledge query encoding module includes several L-shaped stacked CNN bottleneck layers and CNN pooling layers, and a scale-aware knowledge query layer. The CNN bottleneck layers are used to extract multi-scale battery health status-related features from the original data or target data. The CNN pooling layers are used to transform the dimension of the output features of the CNN bottleneck layers. The scale-aware knowledge query layer is used to query and output state knowledge from different domains.

[0035] In this embodiment, the domain-oriented scale-aware knowledge query encoder aims to learn the feature space of a specific domain. This invention employs two private branches, each used to learn a domain-oriented feature embedding function. and Furthermore, this invention leverages the powerful feature extraction capabilities of convolutional neural networks (CNNs), such as parameter sharing and local information compression, using a one-dimensional CNN as the backbone network for extracting high-order features. Through these methods, this invention can effectively learn rich feature representations from different domains, thereby achieving better domain adaptability. Figure 3 As shown, the encoder consists of L stacked CNN bottleneck layers, CNN pooling layers, and a scale-aware knowledge query layer.

[0036] Formally, given an input This can be either the source domain or the target domain. This invention employs L-stacked CNN bottleneck layers to continuously and alternately extract multi-scale features. For example... Figure 4 As shown, the bottleneck layer of a CNN consists of Conv1D blocks, CBAM blocks, and element-wise convolutional blocks.

[0037] One-dimensional convolution is a technique that performs a sliding window convolution operation on signal, sequence, or time series data. It can effectively extract features from the data, enabling machine learning algorithms to better understand the data and extract features. Specifically, assuming the input feature map... X The size is C in × L convolution kernel W The size is C out × C in × H The size of the output feature map Y is C out × L′ The calculation method for one-dimensional convolution is as follows:

[0038] In the formula: k Indicates the output feature map Y The passage in the middle, c Represents the input feature map X The passage in the middle, H This indicates the length of the convolution kernel.

[0039] Convolutional block attention modules are attention mechanisms in computer vision. To improve the representational power of convolutional neural networks (CNNs), this invention introduces convolutional block attention modules to enhance the robustness and generalization performance of features. Specifically, the channel attention module obtains a global feature vector by globally aggregating the feature maps of each channel. This feature vector is then input into a fully connected network to learn the relationships between channels. The learned weights are applied to the feature maps of each channel and used to perform a weighted average of the features from each channel. The spatial attention module learns features at different scales by applying different convolutional kernels in the spatial dimension. It learns spatial relationships by globally aggregating these features and then applies the learned weights to the features at each spatial location, performing a weighted average of the features at each location.

[0040] Element-wise convolution has a kernel size of 1. It is used for feature fusion to aggregate feature maps at different levels and adjust the number of channels in the input feature map. W The size is C out × Cin The calculation method for element-wise convolution is as follows:

[0041] To accelerate convergence and improve generalization ability, this invention utilizes batch normalization in a one-dimensional convolutional neural network. Batch normalization normalizes each feature channel of each mini-batch of input data, making the mean of the feature channels close to 0 and the variance close to 1, thereby accelerating the model's convergence speed. Simultaneously, to enhance the network's nonlinear mapping capability, this invention uses the rectified linear unit (ReLU) as the activation function for all convolutional layers, achieving faster convergence compared to other activation functions.

[0042] To ensure that the feature map size of the CNN encoder is consistent with the feature map size of the scale-aware knowledge query block, this invention utilizes CNN pooling layers to transform the dimensionality of the output features from the CNN bottleneck layer. For example... Figure 5 As shown, this invention combines one-dimensional adaptive average pooling and element-wise convolution to adaptively change the length of the input tensor and reduce the number of channels. These two-dimensional vectors are then flattened into one-dimensional vectors along the channels. After these steps, this invention extracts multi-scale feature vectors from the original time series data of lithium-ion batteries. , where L represents the total number of time scales.

[0043] As a preferred embodiment of this example, in order to query battery health status related features of different domains in the scale-aware knowledge query layer, the Transformer model is used to attach the label status of the battery health status related features of the multi-scale domains.

[0044] In this embodiment, the Transformer, a popular neural network architecture, was initially proposed by Google in 2017 to address sequence modeling problems in Natural Language Processing (NLP). For example... Figure 4 As shown, the Transformer consists of multi-head self-attention, layer normalization, and a feedforward network. Based on its self-attention mechanism, the Transformer can learn long-term dependencies in time series, which is crucial for time series prediction tasks. Traditional recurrent neural networks suffer from vanishing or exploding gradients when learning long-term dependencies, while the Transformer avoids these problems and better captures long-term dependencies in time series. The multi-head self-attention mechanism in the Transformer can extract features from the sequence from multiple different perspectives, capturing information at different scales. This is very useful for predicting time series because it can capture both long-term and short-term patterns simultaneously. Suppose we are given an input sequence... X It contains n vectors , where the dimension of each vector is d .like Figure 6 As shown, the calculation method of Transformer is as follows:

[0045] Multi-head self-attention can produce more accurate sequence representations by capturing the relationships between different positions within a time series. For each vector x i Multi-head self-attention calculates its similarity to other vectors using scaled dot products, and the formula can be expressed as:

[0046] In the formula: Q , K and V This represents a query, key, and value vector matrix representing the input sequence. . and These are learnable parameters, typically dv equal dk .

[0047] Another prominent feature of Transformer is its ability to flexibly attach learnable classification labels for different purposes. This can be achieved by enhancing the multi-scale feature vectors... By collecting global and diverse information, this invention sets up a set of learnable tags. As a knowledge query, this invention concatenates these knowledge queries with multi-scale features. The input is fed into a Transformer, and then the output features are generated corresponding to the knowledge query. Used as domain-specific state knowledge, where (·) represents the original data domain or the target data domain.

[0048] In a preferred embodiment of this invention, the step of using a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in different domains specifically includes: Battery health status-related features with domain labels are input into multi-head self-attention transformers to extract independent features from the original domain and the target domain, respectively. Then, a multi-head cross-attention mechanism is used to make the independent features of each domain more similar to those of another domain. The specific calculation formula for the multi-head cross-attention mechanism is as follows:

[0049] In the formula, Q , K and V Represents the query, key, and value vector matrix of the input sequence; . and These are learnable parameters; S and F represent the state knowledge of the original domain and the target domain, respectively.

[0050] In this embodiment, although self-attention can effectively utilize domain-specific knowledge, the model still cannot explore and learn domain-invariant features. Multi-head cross-attention and self-attention are both attention mechanisms, but their application scenarios and computational methods differ slightly. Self-attention is mainly used to handle the internal dependencies of a single input sequence or tensor, while cross-attention is mainly used to handle the interactions between multiple input sequences or tensors. The cross-attention mechanism can transfer information between different inputs and model the relationships between different inputs, thereby improving the accuracy of information transfer. Therefore, multi-head cross-attention is highly suitable for reducing domain differences and learning domain-invariant representations. Specifically, multi-head cross-attention is calculated by modeling the interaction between two input sequences or tensors as an attention distribution, and its calculation formula can be expressed as:

[0051] In the formula: S and F It can be state knowledge of the source domain or the target domain.

[0052] Here, the invention is not merely about maintaining Unchanged, let near Instead of directly aligning domains, this invention achieves domain alignment through a bidirectional approach to transfer knowledge from the source domain. It employs multi-head cross-attention as a bridge connecting different domains, enabling the mixing of feature spaces from different domains. For example... Figure 7 As shown, this invention employs four weight-sharing transformers to achieve bidirectional neighborhood alignment. Specifically, this invention integrates state knowledge... and The input is fed into multi-head self-attention transformers to extract feature representations independently. and Conversely, this invention employs transformers with multi-head cross-attention to make the feature representations of each domain more closely resemble those of another domain. These computational processes can be described as follows:

[0053] In the formula: l represents the layer index of the transformers. It is worth noting that this invention will... and Initialize to state knowledge ,Will and Initialize to state knowledge .

[0054] During the training phase, this invention utilizes combined features. and To reduce the differences in feature spaces across different domains, this invention employs a distributional dissimilarity index to minimize the domain difference between each pair of source and target domains without introducing an additional domain discriminator. The maximum mean dissimilarity (MMD) is a representative distributional dissimilarity index widely used in domain adaptation, but it only calculates the first moment of the inter-domain distance. Therefore, this invention uses the correlated alignment (CORAL) loss, which measures the difference between the second-order statistics (covariance matrices) of the source and target domains.

[0055] In a preferred embodiment of this example, the step of predicting the battery health status in the original data domain and the target data domain based on the independent features of the battery health status in each domain after processing by the multi-head cross-attention mechanism is specifically as follows: The independent features of each domain, calculated by multi-head cross-attention, are substituted into the regression head model for regression calculation; The loss function used in the regression head model is the mean squared error loss function, specifically:

[0056]

[0057] In the formula, s and t represent the original domain marker and the target domain marker, respectively; n s and n t This represents the number of battery health state-related features in the original domain and the number of battery health state-related features in the target domain. This represents the mean squared error loss function; Predict the battery health status in the original data domain and the target data domain based on the calculation results.

[0058] In this embodiment, following the above only and The data are selected and fed into the regression head for final regression. To fully extract the supervisory information from each domain, this invention employs the mean squared error (MSE) loss function to minimize the regression error of all labeled samples.

[0059]

[0060]

[0061]

[0062] In the formula: This represents the mean squared error loss function.

[0063] Example 2: A multi-condition health assessment method for energy storage batteries based on unsupervised domain adaptive methods includes the following steps: Extract battery health status-related features from the original data and target data respectively, for their respective domains; By utilizing a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in different domains, the independent features of battery health status in each domain are extracted again. Based on the independent features of battery health status in each domain after processing by the multi-head cross-attention mechanism, predict the battery health status in the original data domain and the target data domain.

[0064] In a preferred embodiment of this invention, the scale-aware knowledge query encoding module includes several L-shaped stacked CNN bottleneck layers and CNN pooling layers, and a scale-aware knowledge query layer. The CNN bottleneck layers are used to extract multi-scale battery health status-related features from the original data or target data. The CNN pooling layers are used to transform the dimension of the output features of the CNN bottleneck layers. The scale-aware knowledge query layer is used to query and output state knowledge from different domains.

[0065] As a preferred embodiment of this example, in order to query battery health status related features of different domains in the scale-aware knowledge query layer, the Transformer model is used to attach the label status of the battery health status related features of the multi-scale domains.

[0066] In a preferred embodiment of this invention, the step of using a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in different domains specifically includes: Battery health status-related features with domain labels are input into multi-head self-attention transformers to extract independent features from the original domain and the target domain, respectively. Then, a multi-head cross-attention mechanism is used to make the independent features of each domain more similar to those of another domain. The specific calculation formula for the multi-head cross-attention mechanism is as follows:

[0067] In the formula, Q , K and V Represents the query, key, and value vector matrix of the input sequence; . and These are learnable parameters; S and F represent the state knowledge of the original domain and the target domain, respectively.

[0068] In a preferred embodiment of this example, the step of predicting the battery health status in the original data domain and the target data domain based on the independent features of the battery health status in each domain after processing by the multi-head cross-attention mechanism is specifically as follows: The independent features of each domain, calculated by multi-head cross-attention, are substituted into the regression head model for regression calculation; The loss function used in the regression head model is the mean squared error loss function, specifically:

[0069] In the formula, s and t represent the original domain marker and the target domain marker, respectively; n s and n t This represents the number of battery health state-related features in the original domain and the number of battery health state-related features in the target domain. This represents the mean squared error loss function; Predict the battery health status in the original data domain and the target data domain based on the calculation results.

[0070] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A health assessment system for multi-condition energy storage batteries based on unsupervised domain adaptive methods, characterized in that, include The scale-aware knowledge query encoding module extracts battery health status-related features from the original data domain and the target data domain, respectively. The scale-aware knowledge query encoding module includes L stacked CNN bottleneck layers, CNN pooling layers, and a scale-aware knowledge query layer; the CNN bottleneck layers are used to extract multi-scale battery health status-related features from the original data or target data; the CNN pooling layers are used to transform the dimensionality of the output features of the CNN bottleneck layers. The scale-aware knowledge query layer is used to query and output state knowledge from different domains. In order to query battery health status related features in different domains at the scale-aware knowledge query layer, the Transformer model is used to attach the domain-specific label status to the multi-scale battery health status related features. The bidirectional cross-attention domain hybrid module utilizes a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in the original data domain and the target data domain, and then extracts independent features of battery health status in the original data domain and the target data domain again. The specific steps for using a multi-head cross-attention mechanism to reduce the gap between battery health status-related features in the original data domain and the target data domain are as follows: Battery health status-related features with domain labels are input into a multi-head self-attention Transformer model to extract independent features from the original data domain and the target data domain, respectively. Then, a multi-head cross-attention mechanism is used to make the independent features of each domain more similar to those of another domain. The specific calculation formula for the multi-head cross-attention mechanism is as follows: In the formula, Q , K and V Represents the query, key, and value vector matrix of the input sequence; ; and These are learnable parameters; S and F represent the state knowledge of the original data domain and the target data domain, respectively; The regression module predicts the battery health status in the original data domain and the target data domain based on the independent characteristics of the battery health status in the original data domain and the target data domain after processing by the multi-head cross-attention mechanism. The specific steps for predicting the battery health status in the original data domain and the target data domain based on the independent characteristics of the battery health status in the original data domain and the target data domain after processing by the multi-head cross-attention mechanism are as follows: The independent features of battery health status in the original data domain and the target data domain, which have been computed through multi-head cross-attention, are substituted into the regression head model for regression calculation; The loss function used in the regression head model is the mean squared error loss function, specifically: In the formula, s and t represent the original data domain label and the target data domain label, respectively; n s and n t This represents the number of battery health status-related features in the original data domain and the number of battery health status-related features in the target data domain. E represents the mean squared error loss function; Predict the battery health status in the original data domain and the target data domain based on the calculation results.

2. A method for assessing the health of energy storage batteries under multiple operating conditions based on unsupervised domain adaptive multi-condition assessment, based on the system described in claim 1, characterized in that: The specific steps include: Extract battery health status-related features from the original data and target data respectively, for their respective domains; We utilize a multi-head cross-attention mechanism to reduce the gap between battery health status-related features across different domains, and then extract the independent features of battery health status in each domain. Based on the independent features of battery health status in each domain after processing by the multi-head cross-attention mechanism, predict the battery health status in the original data domain and the target data domain.

Images