A vehicle-mounted lithium battery state of health prediction method based on historical degradation interaction reconstruction network

Abstract

The application discloses a vehicle-mounted lithium battery health state prediction method based on a historical degradation interaction reconstruction network, comprising the following steps: acquiring multi-cycle period multi-channel time sequence data of a vehicle-mounted lithium battery and constructing a health state label; constructing original time sequence input and relative degradation change input for a target cycle; selecting a plurality of labeled reference samples from historical cycles to construct a historical reference set; inputting the target and historical samples into a double-branch encoder to extract original state and relative degradation features, and dynamically fusing the features through a bidirectional cross-branch interaction module; constructing a historical memory based on the fusion features, label embedding and position coding of the historical samples, performing historical enhancement decoding with the target features as a query, and outputting health state prediction of a subsequent cycle; and simultaneously introducing historical label reconstruction loss and transfer learning to improve the historical representation quality and small sample cross-domain adaptation capability. The application utilizes historical degradation context to improve the accuracy and stability of vehicle-mounted lithium battery health state prediction.

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted lithium battery health management and degradation time series prediction technology, specifically to a method for predicting the health status of vehicle-mounted lithium batteries based on a historical degradation interaction reconstruction network. Background Technology

[0002] As a core energy storage component of electric vehicles, the health status of on-board lithium batteries directly affects the vehicle's range, power output performance, operational safety, and secondary utilization value. With long-term vehicle operation, on-board lithium batteries are affected by factors such as cycle aging, calendar aging, temperature changes, charge / discharge rate fluctuations, and complex driving conditions, leading to a gradual decrease in usable capacity, increased internal resistance, and reduced power output. Therefore, accurately predicting the State of Health (SOH) of on-board lithium batteries is crucial for on-board battery management systems in terms of lifespan assessment, safety warnings, maintenance decisions, and energy management. Existing methods for predicting the SOH of on-board lithium batteries mainly include mechanistic model-based methods and data-driven methods. Mechanism-based methods typically rely on electrochemical reaction processes, equivalent circuit models, or semi-empirical degradation models to estimate the health status by establishing a mapping relationship between the battery's internal state and external observation signals. However, the battery degradation process is influenced by various factors such as material system, temperature, rate, state of charge range, and operating conditions, and its internal reaction mechanisms are complex with parameters dynamically changing with the aging stage. Furthermore, real-world operational data may contain issues such as uneven sampling, local fluctuations, measurement noise, missing features, and a limited number of available labeled samples, making battery health state prediction a challenging task. Existing methods still have room for improvement in modeling complex degradation processes, adapting to different operational scenarios, and providing stable predictions under limited sample conditions.

[0003] In summary, existing lithium battery life prediction methods do not pay enough attention to data interaction and cannot accurately capture the relationship between degradation state and capacity decay, resulting in insufficient prediction accuracy and difficulty in predicting the early life of lithium-ion batteries. Summary of the Invention

[0004] Purpose of the Invention: The purpose of this invention is to provide a method for predicting the health status of on-board lithium batteries based on a historical degradation interaction reconstruction network. This method constructs a dual-branch input form containing both the original temporal input and the relative degradation change input. Combined with a dual-branch cross-channel interaction module, it achieves dynamic alignment and fusion of intra-cycle state features and inter-cycle degradation change features. Simultaneously, it utilizes multiple historical reference samples labeled with health status to construct a historical memory representation, and retrieves degradation patterns related to the target cycle from the historical memory through historical enhancement decoding. Furthermore, it improves the quality of the historical memory representation through historical label reconstruction constraints, and enhances the model's cross-domain generalization ability under conditions of a small number of labeled samples by combining source domain pre-training and target domain fine-tuning. This addresses the problems of insufficient utilization of historical degradation information, lack of effective interaction between the original temporal state and relative degradation change, unstable representation of historical reference information, and insufficient small-sample adaptation ability across on-board lithium battery datasets in existing on-board lithium battery health status prediction methods.

[0005] Technical solution: The present invention provides a method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks, comprising the following steps:

[0006] Step 1: Acquire multi-channel time-series observation data of the vehicle lithium battery under multiple cycle periods, and construct a health status label based on the available capacity corresponding to each cycle period;

[0007] Step 2: For the target cycle period, construct the original time series input based on the multi-channel time series observation data, and construct the relative degradation change input based on the original time series input of the reference cycle period. The relative degradation change input is used to characterize the degree of degradation shift of the target cycle period relative to the reference cycle period.

[0008] Step 3: Select historical reference samples with health status labels from multiple cycles earlier than the target cycle to construct a historical reference set;

[0009] Step 4: Input the original time series input and relative degradation change input corresponding to the target cycle, as well as each historical reference sample in the historical reference set, into the dual-branch encoder to extract the original state features and relative degradation features;

[0010] Step 5: Through the dual-branch cross-channel interaction module, the original state features and relative degradation features are bidirectionally interacted and dynamically fused to obtain the target cycle period and the fused degradation features corresponding to each historical reference sample.

[0011] Step 6: Construct a historical memory representation based on the fusion degradation features of each historical reference sample, the corresponding health status label embedding, and the location encoding;

[0012] Step 7: Using the fusion degradation features of the target cycle as a query, perform cross-attention decoding on the historical memory representation to obtain the enhanced degradation representation of the target cycle, and output the health status prediction results for one or more cycles after the target cycle based on the enhanced degradation representation.

[0013] Furthermore, in step 1, the health status label is the ratio of the available capacity to the reference capacity corresponding to the target cycle; the reference capacity is the rated capacity, initial capacity, or maximum capacity observed in the data set of the on-board lithium battery.

[0014] Furthermore, in step 2, the original timing input consists of one or more timing observation signals of voltage, current, temperature, capacity, state of charge, charging time, or discharging time within the target cycle; the relative degradation change input is obtained by differential calculation of the original timing input of the target cycle and the original timing input of the reference cycle.

[0015] Furthermore, in step 3, historical reference samples are selected from cycles preceding the target cycle. The selection methods include selecting according to time sequence, selecting according to preset intervals, selecting according to degradation stage, selecting according to health state similarity, or a combination of the above methods. The historical reference set is used to provide degradation reference information with known health state labels for the target cycle.

[0016] Furthermore, in step 4, the dual-branch encoder includes an original temporal branch and a relative degradation branch; the original temporal branch is used to extract original state features, and the relative degradation branch is used to extract relative degradation features, with the parameters of the two branches being independent of each other; both the original temporal branch and the relative degradation branch include a multi-scale convolutional module and a dense connection module. The multi-scale convolutional module extracts local degradation features under different temporal receptive fields through multiple one-dimensional convolutional units with different kernel sizes, and the dense connection module enhances the ability to express deep degradation features through inter-layer feature reuse.

[0017] Furthermore, in step 5, the dual-branch cross-channel interaction module includes a first interaction direction, a second interaction direction, and a gating fusion unit. In the first interaction direction, the original state feature is used as the query, and the relative degradation feature is used as the key and value to calculate the first interaction feature. In the second interaction direction, the relative degradation feature is used as the query, and the original state feature is used as the key and value to calculate the second interaction feature. The gating fusion unit generates gating weights based on one or more of the original state feature, relative degradation feature, first interaction feature, or second interaction feature, and adjusts the injection intensity of cross-branch interaction information based on the gating weights. The original state feature and the relative degradation feature after interaction are multiplied element-wise, and the difference features of the original state feature and the relative degradation feature after interaction are concatenated. The concatenation result is then input into the feature mapping layer to obtain the fused degradation feature.

[0018] Furthermore, in step 6, the historical memory representation is obtained by adding or concatenating the fusion degradation features of each historical reference sample, the corresponding health status label embedding, and the position encoding; wherein, the health status label embedding is used to introduce known degradation status information into the historical memory representation, and the position encoding is used to characterize the temporal order relationship or relative cycle distance between each historical reference sample and the target cycle.

[0019] Furthermore, in step 7, the historical enhancement decoding is as follows: the fused degradation features of the target cycle are mapped to a query vector, and the historical memory representation is mapped to a key vector and a value vector; one or more layers of cross-attention calculation are performed based on the query vector, key vector, and value vector to obtain historical context features related to the degradation state of the target cycle; the historical context features are fused with the fused degradation features of the target cycle to obtain the enhanced degradation representation of the target cycle; wherein each layer of cross-attention decoding structure includes a cross-attention calculation unit, a residual connection unit, a normalization unit, and a feedforward mapping unit.

[0020] Furthermore, the method also includes, during model training, constructing a prediction loss based on the health status prediction results and the true health status labels, and reconstructing the health status labels of historical reference samples based on historical memory representation to construct a historical label reconstruction loss. The prediction loss and the historical label reconstruction loss are weighted and combined into a joint training loss to optimize the historical degradation interaction reconstruction network. The historical degradation interaction reconstruction network is pre-trained using source domain battery data and then fine-tuned using a small number of labeled samples in the target domain to achieve target domain vehicle lithium battery health status prediction. Among these, the prediction loss is used to constrain the error between the health status prediction results and the corresponding health status labels in one or more cycles after the target cycle; the historical label reconstruction loss is used to constrain the error between the historical health status labels reconstructed from the historical memory representation and the true historical health status labels; and the joint training loss is a weighted sum of the prediction loss and the historical label reconstruction loss.

[0021] Furthermore, transfer learning includes a source domain pre-training stage and a target domain fine-tuning stage. In the source domain pre-training stage, the historical degradation interaction reconstruction network is trained using source domain battery data to learn a general degradation representation. In the target domain fine-tuning stage, the parameters of the pre-trained network are updated using a small amount of labeled battery data from the target domain to adapt to the degradation distribution of the target vehicle lithium battery. In the target domain fine-tuning stage, some parameters in the dual-branch encoder, dual-branch cross-channel interaction module, historical enhancement decoder, or prediction output layer are frozen and the remaining parameters are updated, or all parameters are fine-tuned.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0023] (1) This invention constructs a historical reference set containing multiple historical reference samples for the target cycle, enabling the model to utilize historical degradation information with known health status labels when predicting the health status of the current target cycle. Compared with prediction methods that rely solely on the current cycle or a fixed continuous window, this invention can more fully capture the cumulative and phased nature of the lithium-ion battery degradation process, improving the model's ability to identify the degradation stage of the target cycle.

[0024] (2) This invention simultaneously constructs the original temporal input and the relative degradation change input, and extracts the original state features and the relative degradation features respectively through a dual-branch encoder. The original temporal input is used to characterize the state evolution process within the cycle, while the relative degradation change input is used to characterize the degree of degradation offset of the current cycle relative to the reference cycle. Dual-branch modeling avoids the problem of insufficient information representation caused by a single input form and improves the completeness of the degradation feature representation.

[0025] (3) This invention achieves bidirectional interaction and dynamic fusion between original state features and relatively degraded features through a dual-branch cross-channel interaction module. This module can perform cross-attention calculation between the two branches and adjust the cross-branch information injection intensity through a gating mechanism, thereby enhancing the degraded feature response related to health state prediction, suppressing redundant or irrelevant information, and improving prediction stability.

[0026] (4) This invention constructs a historical memory representation based on the fusion degradation features of historical reference samples, the embedding of health status labels, and location encoding, and retrieves historical degradation patterns related to the target cycle from the historical memory through historical enhancement decoding. Compared with simple pooling or fixed-weight aggregation of historical information, this invention can adaptively select historical samples with reference value according to the degradation state of the target cycle, thereby improving the targeting of historical information utilization.

[0027] (5) This invention introduces historical label reconstruction loss during model training, enabling the historical memory representation to not only serve the prediction of the target health status, but also to reconstruct the health status labels of historical reference samples. This reconstruction constraint can enhance the ability of the historical memory representation to retain information related to the deterioration state and reduce the impact of noise or redundant information in historical features on the prediction results.

[0028] (6) This invention employs a transfer learning strategy of source domain pre-training and target domain fine-tuning, enabling the model to learn general degradation representations from source domain battery data and quickly adapt to the target battery data distribution using a small number of labeled samples in the target domain. This strategy can alleviate the problem of insufficient labeled samples in the target domain and improve the model's generalization ability across battery, operating conditions, and datasets. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention;

[0030] Figure 2 This is a flowchart illustrating the construction and input generation process of historical reference samples for this invention.

[0031] Figure 3 This is a flowchart illustrating the overall structure of the historical degradation interactive reconstruction network of the present invention.

[0032] Figure 4 This is a flowchart of the dual-branch cross-channel interaction module of the present invention;

[0033] Figure 5 This is a flowchart of the historical memory construction and historical enhancement decoding process of the present invention;

[0034] Figure 6 This is a graph showing the overall prediction performance of the present invention on different lithium-ion vehicle battery datasets; wherein, Figure 6 (a) in Figure 6 In the diagram, (c) represents the SOH prediction curves for three representative vehicles in dataset I. Figure 6 In this context, (d) represents the absolute error distribution of all test samples in dataset I; Figure 6 (e) in Figure 6 In the dataset, (g) represents the SOH prediction curves of three representative vehicles in dataset II. Figure 6 In this context, (h) represents the absolute error distribution of all test samples in dataset II;

[0035] Figure 7 This is a comparison chart of the absolute error distributions of different models of the present invention, used to demonstrate the prediction error distribution of the method of the present invention compared with LSTM, TCN, Transformer, and other comparative models on different vehicle lithium battery datasets; wherein Figure 7 (a) in the text represents dataset I; Figure 7 (b) in the dataset is dataset II. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the health status of an on-board lithium battery based on a historical degradation interaction reconstruction network. This embodiment uses multi-channel time-series observation data of an on-board lithium battery across multiple cycles as a foundation. For the target cycle to be predicted, both the original time-series input and the relative degradation change input are constructed simultaneously. Furthermore, multiple historical reference samples are selected from before the target cycle to construct a historical reference set with health status labels. Then, a dual-branch encoder, a dual-branch cross-channel interaction module, a historical memory construction module, and a historical enhancement decoding module are used to complete the health status prediction. The model's prediction stability and cross-dataset adaptability are improved through historical label reconstruction constraints and transfer learning strategies. Specifically, the following steps are included:

[0038] S101, acquire multi-channel time-series observation data of the lithium-ion battery over multiple cycle periods, and construct a health status prediction label. In this embodiment, the multi-channel time-series observation data can be one or more of the following data collected during the cyclic charge-discharge process of the lithium-ion battery: voltage, current, temperature, capacity, state of charge, charging time, and discharging time. For any cycle period, the multi-channel observation signals arranged according to the sampling time within that cycle period can be used as the time-series observation data for that cycle period. The health status prediction label can be constructed based on the available capacity corresponding to the cycle period. Preferably, the first... The health status label corresponding to each cycle is represented as follows:

[0039]

[0040] in, This represents the available capacity corresponding to the i-th cycle. This represents the reference capacity. The reference capacity can be the battery's rated capacity, initial capacity, or the maximum capacity observed in the dataset. This method transforms the capacity degradation process into a normalized sequence of health states, facilitating subsequent model training and prediction.

[0041] S102, construct the original time series input for the target cycle and construct the relative degradation change input based on the reference cycle. After determining the target cycle, extract the multi-channel time series observation sequence within the target cycle to construct the original time series input. The original time series input is used to reflect the state evolution characteristics within the target cycle, such as voltage changes, current changes, temperature changes, and capacity changes during charging or discharging. Simultaneously, a reference cycle is selected, and the original time series input of the target cycle is differentially calculated with the original time series input of the reference cycle to obtain the relative degradation change input. The reference cycle can be the initial cycle, the cycle with the highest health state, a preset benchmark cycle, or other cycle that meets preset conditions. The relative degradation change input is used to characterize the degree of degradation offset of the target cycle relative to the reference cycle, thereby supplementing the cross-cycle degradation change information that is difficult to directly reflect with the original time series input.

[0042] S103, Select multiple historical reference samples from before the target cycle to construct a historical reference set. In this embodiment, multiple historical cycles before the target cycle are selected as historical reference cycles. For each historical reference cycle, construct the corresponding original time-series input, relative degradation change input, and health status label to form a historical reference sample. The historical reference samples can be selected according to time order, preset interval, degradation stage, health status similarity, or a combination of the above methods. The historical reference set composed of multiple historical reference samples can provide a historical degradation reference of known health status for the target cycle, enabling the model to utilize trend and stage information in the historical degradation process when predicting the subsequent health status of the target cycle.

[0043] S104, the target sample and historical reference samples are input into the dual-branch encoder to extract the original state features and relative degradation features. After constructing the target sample and historical reference sets, the original temporal input and relative degradation change input are respectively input into the dual-branch encoder. The dual-branch encoder includes an original branch encoder and a relative degradation branch encoder. The original branch encoder is used to extract features from the original temporal input to obtain the original state features; the relative degradation branch encoder is used to extract features from the relative degradation change input to obtain the relative degradation features. The two branches can use the same or different network structures, and their parameters are independent of each other. Through dual-branch encoding, intra-cycle state evolution information and cross-cycle degradation change information can be extracted separately, thus providing a foundation for subsequent feature interaction.

[0044] S105, the two types of features are bidirectionally interacted and dynamically fused through a dual-branch cross-channel interaction module to obtain fused degenerate features. After obtaining the original state features and relative degenerate features, the two types of features are input into the dual-branch cross-channel interaction module. This module is used to establish a bidirectional information interaction relationship between the two branches. Specifically, on the one hand, the original state features can be used as the query, and the relative degenerate features as the key and value to calculate the first-direction interaction features; on the other hand, the relative degenerate features can be used as the query, and the original state features as the key and value to calculate the second-direction interaction features. The intensity of the interaction information injection is further adjusted by a gating fusion unit, and the two types of features after interaction are concatenated and mapped to obtain the fused degenerate features. The fused degenerate features simultaneously contain original state information, relative degenerate change information, and the complementary and differential relationships between the two, which can enhance the model's ability to represent complex degenerate patterns.

[0045] S106. A historical memory representation is constructed based on the fusion degradation features, health status labels, and positional encoding of historical reference samples. For each historical reference sample in the historical reference set, a corresponding fusion degradation feature is obtained after dual-branch encoding and cross-channel interaction. This fusion degradation feature is further combined with the corresponding health status label embedding and positional encoding to construct the historical memory representation. The health status label embedding is used to introduce known health status information into the historical memory representation; the positional encoding is used to characterize the temporal order relationship between historical reference samples and the relative cycle distance between historical reference samples and the target cycle. Through this historical memory representation, the model can preserve historical degradation context information related to the prediction of the target cycle.

[0046] S107, using the fusion degradation feature of the target cycle as a query, historical enhancement decoding is performed and the health status prediction result is output. In this embodiment, the original temporal input and relative degradation change input corresponding to the target cycle are processed by dual-branch encoding and cross-channel interaction to obtain the target fusion degradation feature. This target fusion degradation feature is used as a query, and the historical memory representation is used as the key and value, and input into the historical enhancement decoder. The historical enhancement decoder retrieves historical context information related to the degradation state of the target cycle from the historical memory representation through one or more layers of cross-attention computation, and obtains the enhanced degradation representation of the target cycle. Subsequently, the enhanced degradation representation is mapped to the health status prediction result of one or more cycles after the target cycle through the prediction output header. The health status prediction result can be the SOH of the next cycle, or it can be the SOH sequence of multiple future cycles.

[0047] S108, Construct prediction loss and historical label reconstruction loss, and optimize the network using a joint training loss. During model training, firstly, a prediction loss is constructed based on the error between the health state prediction result obtained from the prediction output head and the corresponding true health state label. This prediction loss is used to constrain the model's health state prediction accuracy. Simultaneously, a reconstruction head is set based on the historical memory representation to reconstruct the health state labels corresponding to historical reference samples, obtaining historical label reconstruction results. A historical label reconstruction loss is constructed based on the error between the historical label reconstruction results and the true historical health state labels. This reconstruction loss enables the historical memory representation to retain effective information related to battery degradation state, reducing the interference of irrelevant historical information or noise information on the prediction process. Further, the prediction loss and historical label reconstruction loss are weighted and combined to obtain a joint training loss, and the parameters of the historical degradation interaction reconstruction network are optimized using the joint training loss.

[0048] S109. Pre-training is performed using source domain data, followed by fine-tuning using a small number of labeled samples from the target domain. During training, the learning rate is preferably between 1e-4 and 5e-4, and the batch size is set to 16-64 to balance training stability and model generalization ability. To improve the model's adaptability to different on-board lithium battery datasets, different vehicles, or different operating conditions, this embodiment employs a transfer learning strategy. First, the historical degradation interaction reconstruction network is pre-trained using source domain battery data, enabling the model to learn general battery degradation time-series representations and historical memory modeling capabilities. Then, the pre-trained model is fine-tuned using a small number of labeled samples from the target domain to adapt the model to the degradation distribution of batteries in the target domain. During fine-tuning, some parameters in the dual-branch encoder, dual-branch cross-channel interaction module, historical enhancement decoder, or prediction output head can be frozen, and only the remaining parameters can be updated; alternatively, all network parameters can be fine-tuned. Through these methods, the predictive stability and generalization ability of the model can be improved even with limited labeled samples in the target domain. In the target domain fine-tuning stage, to balance source domain feature transfer and target domain adaptability, the first 2-3 layers of convolutional modules of the dual-branch encoder can be frozen, and only the high-level fusion module, BCCI module, historical enhancement decoder and prediction output layer are updated. The freezing process involves turning off gradient calculation of the frozen layer parameters before fine-tuning and updating the gradients of the unfrozen layers. Alternatively, all layers can be fine-tuned according to task requirements to obtain greater flexibility.

[0049] S110, output the SOH prediction result of the target vehicle lithium battery. After completing model training or target domain adaptation, the target vehicle lithium battery's to-be-predicted cycle data is input into the trained historical degradation interactive reconstruction network to obtain the SOH prediction results for one or more cycles after the target cycle. The SOH prediction results can be used in applications such as health status monitoring, life assessment, maintenance decision-making, and safety early warning in vehicle battery management systems. In summary, as... Figure 1 As shown, this embodiment achieves the prediction of the health status of lithium-ion batteries through a complete process of "multi-channel time-series data acquisition—dual-input sample construction—historical reference set construction—dual-branch encoding—cross-channel interactive fusion—historical memory construction—historical enhancement decoding—joint training—transfer adaptation". This process can make full use of historical degradation information before the target cycle and improve the stability and adaptability of the model through reconstruction learning and transfer learning.

[0050] like Figure 2 As shown in the illustration, this embodiment further provides the process for constructing historical reference samples and generating inputs. This process is used to construct the original temporal input, relative degradation change input, and historical reference set of the target samples after the target cycle period is determined, providing input data for the subsequent historical degradation interactive reconstruction network. Specifically, it includes the following steps.

[0051] S201, acquire multi-channel time-series observation data of the lithium-ion battery over multiple cycle periods. In this embodiment, the operational observation data of the vehicle-mounted lithium battery over multiple cycle periods is first acquired. The operational observation data can be derived from laboratory cycle test data, data collected by the vehicle-mounted battery management system, or historical data during vehicle operation. The multi-channel time-series observation data can include one or more of the following: voltage, current, temperature, capacity, state of charge, charging time, and discharging time.

[0052] For the A cycle can be used to represent the multi-channel observation signals arranged according to the sampling time within that cycle as follows:

[0053]

[0054] in, This indicates the number of sampling time steps within this cycle. This indicates the number of observation channels. The multi-channel time-series observation data is used to describe the state evolution of a lithium-ion battery within a single cycle.

[0055] S202, Determine the target cycle period T. After obtaining time-series observation data for multiple cycle periods, determine the target cycle period to be predicted. The target cycle period can be determined according to the actual prediction requirements. For example, in an online health status prediction scenario, the latest observed cycle period can be used as the target cycle period; in an offline training or validation scenario, any cycle period that meets the sample construction conditions can be selected from the existing cycle sequences as the target cycle period. To avoid confusion with the sampling time step, the target cycle period can also be denoted as the th cycle period in the following text. The target cycle is used to construct the target sample and serves as the basis for the model to predict the health status of one or more subsequent cycles.

[0056] S203, Extract the multi-channel time series sequence within the target cycle to construct the original time series input. For the target cycle, extract the multi-channel time series observation sequence during its cycle to construct the original time series input. The original time series input can be expressed as:

[0057]

[0058] in, Indicates the first The cycle period is in the _ ... A multi-channel observation vector at each sampling time. This input is used to retain dynamic changes in voltage, current, temperature, capacity, or state of charge within the target's cycle.

[0059] In practice, the time step lengths of different cycle periods can be uniformly processed, for example by using interpolation, resampling, truncation or padding, so that the inputs of different cycle periods have a consistent time step length, which facilitates batch processing by neural networks.

[0060] S204, Select a reference cycle period and perform differential calculation with the target cycle period to construct the relative degradation change input. After constructing the original timing input, further select a reference cycle period. The reference cycle period can be the initial cycle period, the cycle period corresponding to the rated capacity, the cycle period with the highest health status, the preset benchmark cycle period, or any cycle period that meets the preset health status conditions. Let the original timing input corresponding to the reference cycle period be... The original timing input corresponding to the target cycle period is Then the relative degradation change input corresponding to the target cycle period can be expressed as:

[0061]

[0062] in, This method is used to characterize the degree of variation of the target cycle period relative to the reference cycle period in the multi-channel time-series observation space. By constructing this difference, the absolute amplitude differences between different cells or different cycle periods can be weakened, while highlighting the gradual accumulation of variation information during the degradation process. In other embodiments, the relative degradation variation input can also be constructed using proportional difference, normalized difference, standardized residual, or other calculation methods that can characterize the degree of variation of the target cycle period relative to the reference cycle period.

[0063] S205, Select multiple historical cycles preceding the target cycle. After the target cycle is determined, select multiple historical cycles from the cycle sequence earlier than the target cycle as candidate historical reference cycles. These historical cycles all precede the target cycle, so their corresponding health status labels can be used as known historical information during training or inference. The selection methods for historical cycles can include selection according to time order, selection according to fixed intervals, selection according to a combination of recent and long-term cycles, selection according to degradation stage, selection according to health status similarity, or a combination of the above methods. By selecting multiple historical cycles, degradation reference information at different time scales can be provided for the target cycle. In one embodiment, a portion of recent historical cycles closer to the target cycle can be selected to characterize local degradation trends; simultaneously, a portion of long-term historical cycles covering a longer time range can be selected to characterize the global degradation trajectory. Thus, the historical reference set can contain both short-term change information and long-term degradation trend information. The number of historical reference samples K is preferably 4 to 32, and the optimal K value can be determined through validation set grid search or cross-validation. When selecting historical reference samples, both recent and long-term cycles can be considered to capture both local degradation trends and global degradation trajectories.

[0064] S206, construct the original time series input, relative degradation change input and health status label for each historical cycle.

[0065] S207, Forming a historical reference set. Multiple historical reference samples are combined according to chronological order, relative distance order, or a preset arrangement rule to form a historical reference set. This historical reference set can be represented as:

[0066]

[0067] in, Indicates the number of historical reference samples. Indicates the first There are 1 historical reference sample. Each historical reference sample can be represented as:

[0068]

[0069] in, This indicates that the cycle period corresponding to the historical reference sample is earlier than the target cycle period. The historical reference set is used to construct a historical memory representation in subsequent models, providing a historical degradation reference with a known health status label for the prediction of the target cycle period.

[0070] S208, Output target samples and historical reference set. After the above steps, output target samples and historical reference set. The target samples include the original time-series input of the target cycle. and relative degradation change input The historical reference set includes multiple historical reference samples, each of which includes the original time-series input, relative degradation change input, and health status label for the corresponding cycle.

[0071] like Figure 3 As shown, this embodiment further presents the overall structure and flow of the historical degradation interactive reconstruction network. This network takes target samples and historical reference samples as input, and achieves on-board lithium battery health status prediction through dual-branch encoding, dual-branch cross-channel interaction, historical memory construction, historical enhancement decoding, and a dual-output structure of prediction and reconstruction. Specifically, it includes the following steps: S301, input target samples and historical reference samples. In... Figure 2 Following the sample construction process shown, the target sample and historical reference samples are input into the historical degradation interactive reconstruction network. The target sample includes the original temporal input and relative degradation change input corresponding to the target cycle; the historical reference sample includes the original temporal input, relative degradation change input, and health status label corresponding to the historical cycle. For the target sample, the model needs to predict the health status of one or more cycles after the target cycle; for the historical reference sample, the model uses its known health status label to construct a historical memory representation and assists in the prediction of the target sample.

[0072] S302, the original timing input enters the original branch encoder, and the relative degradation change input enters the relative degradation branch encoder. In this embodiment, the historical degradation interactive reconstruction network adopts a dual-branch coding structure. The original timing input enters the original branch encoder, and the relative degradation change input enters the relative degradation branch encoder. The two branches can be executed in parallel; the diagram is only used to indicate that different inputs enter the corresponding coding branches.

[0073] The original branch encoder is used to extract the state evolution features within the cycle; the relative degradation branch encoder is used to extract the degradation change features relative to the reference cycle. Both branches can employ one-dimensional convolution, temporal convolution, recurrent neural networks, Transformer encoders, or other coding structures capable of processing temporal data. Preferably, the two branches employ a structurally similar but parameter-independent multi-scale temporal coding structure to learn the feature representations of the two types of inputs respectively.

[0074] S303, extract the original state features and relative degradation features. After passing through the dual-branch encoder, the original state features and relative degradation features are obtained respectively. Let the original state features be represented as... The relative degradation characteristics are expressed as .in, It mainly characterizes the multi-channel temporal state changes within the target cycle or historical cycle. This primarily characterizes the degradation offset information of the current cycle relative to the reference cycle. For each historical reference sample in the historical reference set, the same dual-branch encoding method is used to obtain the corresponding original state features and relative degradation features. Thus, both the target sample and the historical reference samples are mapped to a unified feature representation space. During model training, the kernel size of the multi-scale convolutional modules of the original branch encoder and the relative degradation branch encoder is preferably a combination of {1, 3, 5, 7}, and the number of network layers can be set to 3~6.

[0075] S304, interactive fusion is performed through a dual-branch cross-channel interaction module to obtain fused degenerate features. After obtaining the original state features and relative degenerate features, both are input into the dual-branch cross-channel interaction module. This module is used for bidirectional information interaction between the original state features and the relative degenerate features, enabling the two types of features to complement each other. Specifically, the dual-branch cross-channel interaction module can achieve feature alignment through a cross-attention mechanism and control the injection intensity of interactive information through a gated fusion mechanism. After processing by this module, the fused degenerate features are obtained, denoted as... The fused degradation features simultaneously include original state information, relative degradation change information, and the correlation information between the two. For historical reference samples, corresponding historical fused degradation features are also obtained through the dual-branch cross-channel interaction module. The fused degradation features of the target sample are used for subsequent queries, while the fused degradation features of the historical reference samples are used to construct a historical memory representation.

[0076] S305, a historical memory representation is constructed based on the fused degradation features, label embedding, and position encoding of historical samples. In this embodiment, for each historical reference sample, its fused degradation features, health status label embedding, and position encoding are fused to construct a historical memory representation. The fused degradation features describe the degradation state of the historical reference sample; the health status label embedding incorporates the known SOH information of the historical reference sample into the memory representation; and the position encoding represents the temporal order of the historical reference sample in the cyclic sequence or its distance relative to the target cyclic period. By combining these three types of information, a historical memory representation containing historical degradation features, historical health status, and temporal position information can be formed.

[0077] Historical memory can be represented as:

[0078]

[0079] in, Indicates the first The memory vector corresponding to each historical reference sample. This historical memory representation is used as the key and value in the subsequent historical enhancement decoding process.

[0080] S306, using the fusion degradation features of the target sample as a query, historical enhancement decoding is performed to obtain an enhanced degradation representation. The fusion degradation features of the target sample are mapped to a query vector, and the historical memory representation is used as the key and value, and input into the historical enhancement decoding module. The historical enhancement decoding module may include one or more layers of cross-attention decoding structures.

[0081] In cross-attention computation, the target sample automatically retrieves historical reference information related to its current degradation state by querying the historical memory representation. In this way, the model can select more valuable historical information from the historical reference set based on the target sample's own degradation characteristics, and obtain an enhanced degradation representation of the target cycle. .

[0082] S307, the prediction head outputs the SOH prediction result. The enhanced degradation representation output by the historical enhancement decoding module is input into the prediction head. The prediction head can consist of a fully connected layer, a multilayer perceptron, or other regression mapping structure, used to map the enhanced degradation representation to the SOH prediction result for one or more cycles after the target cycle. In one embodiment, the prediction head outputs the SOH prediction value for the next cycle; in another embodiment, the prediction head can also output a sequence of SOH predictions for multiple future cycles. The SOH prediction result output by the prediction head can be used to evaluate the current and future health status of the battery.

[0083] S308, the reconstruction head outputs the historical label reconstruction results. To enhance the ability of the historical memory representation to retain health status information, this embodiment also includes a reconstruction head. The reconstruction head outputs the health status label reconstruction results corresponding to the historical reference sample based on the historical memory representation. By comparing the reconstructed historical labels with the true historical health status labels, the historical label reconstruction loss can be obtained. This reconstruction process can constrain the historical memory representation to retain effective information related to battery degradation status, so that the historical memory not only serves target prediction but also reflects the health status of the historical reference sample itself.

[0084] S309, Joint Training Loss Optimizes Network Parameters. During the training phase, the error between the predicted SOH from the prediction head output and the true SOH label constitutes the prediction loss, and the error between the historical label reconstruction from the reconstruction head output and the true historical SOH label constitutes the reconstruction loss. Furthermore, the prediction loss and reconstruction loss are weighted and combined to form the joint training loss:

[0085]

[0086] in, Indicates the predicted loss. This indicates the loss from rebuilding historical tags. , where represents the weighting coefficients. Optimizing the parameters of the historical degradation interaction reconstruction network using this joint training loss can simultaneously improve the target health status prediction capability and the quality of historical memory representation.

[0087] like Figure 4 As shown, this embodiment further illustrates the specific process of the dual-branch cross-channel interaction module. This module is used to perform bidirectional interaction and dynamic fusion of the original state features output by the original branch encoder and the relative degradation features output by the relative degradation branch encoder, thereby obtaining fused degradation features. Specifically, it includes the following steps: S401, input the original state features and the relative degradation features. In this embodiment, the original timing input is processed by the original branch encoder to obtain the original state features, denoted as... The relative degradation feature is obtained after the relative degradation change input is processed by the relative degradation branch encoder, and is denoted as . Among them, the original state characteristics mainly reflect the temporal variation characteristics of multi-channel observed signals such as voltage, current, temperature, capacity, or state of charge within the current cycle; the relative degradation characteristics mainly reflect the degradation offset information of the current cycle relative to the reference cycle. In one embodiment, and It can be a one-dimensional temporal feature sequence or a feature matrix after pooling or mapping. To facilitate subsequent interactive computation, both types of features can be mapped to the same feature dimension space.

[0088] S402, using the original state features as the query and the relative degradation features as the key and value, performs a first-direction interaction. In the first-direction interaction, the original state features... As a query, relative degradation features Cross-attention is calculated using the keys and values. This process can be represented as:

[0089]

[0090] in, Obtained by mapping the features of the original state. and Obtained from relative degradation feature mapping, This represents the information injected into the branch of the original state features by the relatively degraded features. Through this directional interaction, the original state features can perceive the feature responses related to the degradation changes, thereby enhancing their ability to express the degradation state.

[0091] S403, using the relative degradation feature as the query and the original state feature as the key and value, performs a second-direction interaction. In this second-direction interaction, the relative degradation feature... As a query, the original state features Cross-attention computation is performed using the key and value.

[0092] Cross-attention calculation can be represented as:

[0093]

[0094] in, Indicates the feature dimension.

[0095] This process can be represented as:

[0096]

[0097] in, Obtained from relative degradation feature mapping, and Obtained by mapping the features of the original state. This represents the information injected into the relative degenerate branch from the original state features. Through this directional interaction, the relative degenerate features can obtain supplementary information on the state evolution within the cycle, thereby avoiding the information loss caused by relying solely on the differential input.

[0098] S404, the intensity of the interaction information injection is adjusted through a gating fusion unit. After obtaining the interaction features in two directions, the gating fusion unit further controls the intensity of the interaction information injection. For the original state branch, gating weights can be generated based on the original state features and the interaction features in the first direction:

[0099]

[0100] in, This represents the Sigmoid activation function. This represents a linear mapping, a multilayer perceptron, or other learnable mapping function. This represents the concatenation operation. Based on this gating weight, the enhanced original state features can be obtained:

[0101]

[0102] in, This indicates element-wise multiplication.

[0103] Similarly, for the relatively degenerate branch, gating weights can be generated based on the relatively degenerate features and the second-direction interaction features:

[0104]

[0105] And the enhanced relative degradation characteristics are obtained:

[0106]

[0107] Through the gated fusion unit, the model can adaptively control the amount of cross-branch information injected according to the feature state of the current sample, avoiding excessive interference of irrelevant or redundant information with the original branch and the relatively degenerate branch.

[0108] S405 concatenates and maps the two types of features after interaction, and outputs the fused degenerate features. This yields the enhanced original state features. and the enhanced relative degradation characteristics Then, the two are fused. Preferably, the enhanced original state features, the enhanced relative degradation features, the element-wise product of the two, and the absolute difference between the two can be concatenated:

[0109]

[0110] Then, the fused degenerate features are obtained through the feature mapping layer:

[0111]

[0112] in, This represents a linear layer, convolutional layer, multilayer perceptron, or other feature mapping unit.

[0113] Through the above process, the dual-branch cross-channel interaction module can establish a bidirectional information exchange relationship between the original state features and the relative degradation features, and obtain a more complete degradation representation through gating mechanisms and fusion mapping. This fused degradation feature can be used as a query feature for the target sample, or as a basic feature for constructing historical memory representations from historical reference samples.

[0114] like Figure 5 As shown, this embodiment further illustrates the historical memory construction and historical enhancement decoding process. This process is used to construct a historical memory representation from the fusion degradation features, health status labels, and location information in the historical reference samples, and to retrieve the historical memory representation using the fusion degradation features of the target sample, thereby obtaining the enhanced degradation representation of the target cycle and the SOH prediction result. Specifically, it includes the following steps: S501, input the fusion degradation features, label embedding, and location encoding of the historical samples. In this embodiment, each historical reference sample in the historical reference set undergoes dual-branch encoding and dual-branch cross-channel interaction to obtain the corresponding fusion degradation features. For the first... A historical reference sample, whose fusion degradation characteristics can be represented as: At the same time, the health status label corresponding to this historical reference sample will be... Mapped to a label embedding vector, denoted as Label embedding is used to explicitly introduce known health status information into historical memory, enabling the model to associate historical degradation features with their corresponding SOH states. Furthermore, positional encoding is introduced to characterize the temporal order of different historical reference samples in the cyclic sequence or their relative distance from the target cyclic cycle. Position coding can employ sine / cosine coding, learnable coding, or relative distance coding, with sine / cosine coding being preferred to preserve the temporal order information of historical reference samples, while ensuring that the sequence order relationship is effectively preserved in the historical enhancement decoding.

[0115] S502, Construct a historical memory representation. For each historical reference sample, fuse its degradation features, label embeddings, and positional encodings, add or concatenate them, and obtain the corresponding historical memory vector through a mapping layer. Preferably, the historical memory vector can be represented as:

[0116]

[0117] in, This represents the fusion mapping function, which can be an addition, a concatenated linear mapping, a multilayer perceptron, or a lightweight encoder.

[0118] Depend on The historical memory representation is composed of the historical memory vectors corresponding to each historical reference sample:

[0119]

[0120] In one embodiment, the historical memory representation can also be input into a lightweight Transformer encoder or other sequence encoder to enable further interaction between different historical memory vectors to obtain a more stable historical context representation.

[0121] S503, the fusion degradation features of the target sample are used as the query. The target sample corresponding to the target cycle is processed through dual-branch encoding and dual-branch cross-channel interaction to obtain the fusion degradation features of the target sample. Map the fused degraded features to a query vector:

[0122]

[0123] in, This represents the query mapping parameters. This query vector is used to retrieve historical information related to the target cycle degradation state from the historical memory representation.

[0124] S504 represents the historical memory as a key and value. Mapped to keys and values:

[0125]

[0126] in, and These represent the key mapping parameters and the value mapping parameters, respectively. The keys in the historical memory are used to calculate the relevance between the target sample query and each historical reference sample, while the values ​​are used to provide historical context information to the target sample based on the relevance weights.

[0127] S505, perform one or more layers of cross-attention decoding. After obtaining the query, key, and value, a cross-attention mechanism is used for history enhancement decoding. Through cross-attention calculation, the target sample can adaptively select relevant historical reference information from the historical memory representation based on its own degradation characteristics. In one embodiment, the history enhancement decoder may include one or more layers of cross-attention decoding structure. Each layer may include a cross-attention calculation unit, a residual connection unit, a normalization unit, and a feedforward mapping unit. The multi-layer decoding structure can fuse historical context information layer by layer, improving the expressive power of the target degradation representation.

[0128] S506 yields the enhanced degenerate representation. After one or more layers of cross-attention decoding, the enhanced degenerate representation of the target cyclic period is obtained, denoted as... This enhanced degradation representation includes not only the fusion degradation features of the target cycle itself, but also relevant historical context information retrieved from historical memory.

[0129] Compared to methods that directly predict based on the features of the target sample, enhanced degradation representation can make fuller use of known SOH information and degradation stage information in historical reference samples, thereby improving the ability to predict subsequent SOH of the target cycle.

[0130] S507 outputs the target cell SOH prediction result. This will enhance the degradation representation. The prediction head is input to obtain the SOH prediction results for one or more cycles after the target cycle. The prediction head can employ a fully connected layer, a multilayer perceptron, a regression layer, or other structures capable of mapping the enhanced degradation representation to SOH values. In one embodiment, the prediction head outputs the SOH prediction value for the next cycle:

[0131]

[0132] In another implementation, the prediction head can also output the SOH prediction sequence for multiple future cycles:

[0133]

[0134] in, This indicates the prediction step size.

[0135] The attached diagram is for illustrative purposes only.

[0136] like Figure 1 As shown in Figure 2, lithium-ion battery degradation is cumulative and phased, and a single sample from the target cycle is insufficient to fully reflect the historical degradation state. Therefore, this embodiment first acquires multi-channel time-series observation data for multiple cycle cycles and constructs SOH tags; then, it constructs the original time-series input for the target cycle. Simultaneously, a reference cycle period is selected and differiated from the target cycle period to construct a relative degradation change input. Furthermore, multiple historical cycles prior to the target cycle are selected, and for each historical cycle, the original time-series input, relative degradation change input, and health status label are constructed to form a historical reference set. .

[0137] like Figure 1 As shown in Figure 3, the target sample and historical reference samples are input into the historical degradation interactive reconstruction network. The network first processes the original temporal input and the relative degradation change input through a dual-branch encoder, extracting the original state features and relative degradation features respectively. Then, it performs interactive fusion through the BCCI module to obtain fused degradation features. For historical reference samples, historical memory is constructed based on their fused degradation features, health state label embedding, and position encoding. For target samples, their fused degradation features are used as queries for historical enhancement decoding to obtain enhanced degradation representations, which are then output as SOH prediction results via the prediction head. Simultaneously, the reconstruction head reconstructs historical labels, and this reconstruction, together with the prediction loss, forms a joint training loss. During the training phase, this embodiment uses joint optimization of the prediction loss and historical label reconstruction loss. The prediction loss constrains the target SOH prediction results, and the reconstruction loss constrains the historical memory's representation ability of historical health state labels. Furthermore, pre-training can be performed using source domain battery data, followed by fine-tuning and adaptation using a small number of labeled samples from the target domain. In this way, the model can inherit the degradation representation ability of the source domain even with a small number of target domain samples and adapt to the distribution of target battery data.

[0138] against Figure 4 In this context, the original state features and the relative degradation features have different information focuses. The original state features describe the state evolution within the loop, while the relative degradation features describe the degradation offset relative to the reference loop period. To avoid insufficient interaction due to simple concatenation, this embodiment designs a dual-branch cross-channel interaction module. This module includes two interaction directions: a first-direction interaction using the original state features as the query and the relative degradation features as the key and value; and a second-direction interaction using the relative degradation features as the query and the original state features as the key and value. Then, the intensity of the interaction information injection is adjusted through a gating fusion unit, and finally, the two types of features after interaction are concatenated and mapped to obtain the fused degradation features.

[0139] right Figure 5 In this context, historical reference samples are not simply averages or ordinary time windows, but are constructed as historical memory representations. The fusion degradation features of each historical sample are combined with health status label embeddings and location encodings to form historical memory units. Multiple historical memory units constitute the historical memory representation. The fusion degradation features of the target samples are mapped to the query. Historical memory representations are used as keys and values. Through one or more layers of cross-attention, historical enhancement decoding is performed to retrieve information related to the target degradation state from the historical reference set, resulting in an enhanced degradation representation. Finally, the target SOH prediction result is output.

[0140] Depend on Figure 6 As can be seen, on different datasets and representative vehicles, the predicted curves output by the method of this invention can follow the actual SOH degradation trajectory well, and the prediction results maintain a high degree of consistency with the actual curves. For samples with a relatively smooth degradation trend, the method of this invention can stably depict the long-term trend of SOH gradually decreasing with increasing cycle time; for samples with significant local changes or differences in degradation stages, the method of this invention can still maintain good curve tracking ability. Combined with the absolute error distribution, it can be seen that the prediction errors of most test samples are concentrated in a small range, indicating that the method of this invention has good prediction accuracy and stability on different vehicle-mounted lithium battery datasets.

[0141] In this embodiment, an appropriate number of historical reference samples are selected for historical memory construction, enabling the model to achieve relatively stable predictive performance with limited computational overhead. In other words, the parameter configuration selected in this embodiment achieves a good balance between deployment cost and predictive performance. This setting avoids both insufficient degradation context due to too few historical reference samples and increased computational complexity and redundant information interference due to too many historical reference samples, making it suitable for deployment in vehicle battery management systems or vehicle health status monitoring scenarios.

[0142] Depend on Figure 7As can be seen, on both datasets I and II, the absolute error distribution of the method of this invention is more concentrated overall, with both the median and interquartile range of the error at a low level, indicating that the method of this invention can achieve smaller prediction bias on most test samples. Compared with models such as LSTM, TCN, and Transformer, the method of this invention can better reduce the overall level of absolute error and reduce the dispersion of the error distribution, demonstrating its better prediction stability and generalization ability on different lithium-ion vehicle battery datasets.

[0143] Figure 6 and Figure 7 The effectiveness of the method of this invention has been verified. On the one hand, the method of this invention can accurately track the SOH degradation trend on different lithium-ion vehicle-mounted lithium battery datasets while maintaining a small absolute prediction error. On the other hand, ablation experiments show that the historical reference set, bi-branch degradation feature modeling, bi-branch cross-channel interaction, and historical label reconstruction constraints all contribute to improving the prediction accuracy and stability of the model. Furthermore, an appropriate number of historical reference samples can balance prediction performance and computational cost, making the method of this invention valuable for engineering applications and feasible for deployment.

[0144] In summary, this invention, through its overall approach of "constructing multi-channel time-series observation data for vehicles, jointly representing the original time-series input and the relative degradation change input, constructing a historical reference sample set, dual-branch cross-channel interactive fusion, historical memory enhancement decoding, historical label reconstruction constraints, and source domain pre-training—target domain fine-tuning adaptation," can fully utilize the intra-cycle state information, inter-cycle degradation change information, and historical reference information with health status labels during the degradation process of vehicle-mounted lithium batteries to achieve accurate and stable prediction of the state of health (SOH) of the target vehicle-mounted lithium battery. Experimental results show that the prediction curves of this invention on different vehicle-mounted lithium battery datasets can track the real SOH degradation trend well, and have a lower and more concentrated absolute error distribution compared to comparative models such as LSTM, TCN, and Transformer. Furthermore, this invention achieves a good balance between prediction performance and deployment cost, and has high engineering application value.

Claims

1. A method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks, characterized in that, Includes the following steps: Step 1: Acquire multi-channel time-series observation data of the vehicle lithium battery under multiple cycle periods, and construct a health status label based on the available capacity corresponding to each cycle period; Step 2: For the target cycle period, construct the original time series input based on the multi-channel time series observation data, and construct the relative degradation change input based on the original time series input of the reference cycle period. The relative degradation change input is used to characterize the degree of degradation shift of the target cycle period relative to the reference cycle period. Step 3: Select historical reference samples with health status labels from multiple cycles earlier than the target cycle to construct a historical reference set; Step 4: Input the original time series input and relative degradation change input corresponding to the target cycle, as well as each historical reference sample in the historical reference set, into the dual-branch encoder to extract the original state features and relative degradation features; Step 5: The original state features and relative degradation features are bidirectionally interacted and dynamically fused through the dual-branch cross-channel interaction module to obtain the target cycle period and the fused degradation features corresponding to each historical reference sample. The dual-branch cross-channel interaction module includes a first interaction direction, a second interaction direction, and a gated fusion unit. In the first interaction direction, the original state features are used as the query, and the relative degradation features are used as the key and value to calculate the first interaction feature. In the second interaction direction, the relative degradation features are used as the query, and the original state features are used as the key and value to calculate the second interaction feature. The gated fusion unit generates gate weights based on one or more of the original state features, relative degradation features, first interaction features, or second interaction features, and adjusts the injection intensity of cross-branch interaction information based on the gate weights. The original state features and relative degradation features after interaction are multiplied element-wise, and the difference features of the original state features and relative degradation features after interaction are concatenated. The concatenated result is then input into the feature mapping layer to obtain the fused degradation features. Step 6: Construct a historical memory representation based on the fusion degradation features of each historical reference sample, the corresponding health status label embedding, and the location encoding; Step 7: Using the fusion degradation features of the target cycle as a query, perform cross-attention decoding on the historical memory representation to obtain the enhanced degradation representation of the target cycle, and output the health status prediction results for one or more cycles after the target cycle based on the enhanced degradation representation.

2. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, In step 1, the health status label is the ratio of the available capacity to the reference capacity corresponding to the target cycle; the reference capacity is the rated capacity, initial capacity, or maximum capacity observed in the data set of the on-board lithium battery.

3. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, In step 2, the original timing input consists of one or more timing observation signals of voltage, current, temperature, capacity, state of charge, charging time or discharging time within the target cycle; the relative degradation change input is obtained by differential calculation of the original timing input of the target cycle and the original timing input of the reference cycle.

4. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, In step 3, historical reference samples are selected from cycles preceding the target cycle. The selection methods include selecting according to time order, selecting according to preset intervals, selecting according to degradation stage, selecting according to health state similarity, or a combination of the above methods. The historical reference set is used to provide degradation reference information with known health state labels for the target cycle.

5. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, In step 4, the dual-branch encoder includes a raw temporal branch and a relative degradation branch. The raw temporal branch is used to extract raw state features, and the relative degradation branch is used to extract relative degradation features. The parameters of the two branches are independent of each other. Both the raw temporal branch and the relative degradation branch include a multi-scale convolutional module and a dense connection module. The multi-scale convolutional module extracts local degradation features under different temporal receptive fields through multiple one-dimensional convolutional units with different kernel sizes. The dense connection module enhances the expression ability of deep degradation features through inter-layer feature reuse.

6. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, In step 6, the historical memory representation is obtained by adding or concatenating the fusion degradation features of each historical reference sample, the corresponding health status label embedding, and the position encoding; wherein, the health status label embedding is used to introduce known degradation status information into the historical memory representation, and the position encoding is used to characterize the temporal order relationship or relative cycle distance between each historical reference sample and the target cycle.

7. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, In step 7, the history enhancement decoding is as follows: the fused degradation features of the target cycle are mapped to a query vector, and the historical memory representation is mapped to a key vector and a value vector; one or more layers of cross-attention calculation are performed based on the query vector, key vector, and value vector to obtain historical context features related to the degradation state of the target cycle; the historical context features are fused with the fused degradation features of the target cycle to obtain the enhanced degradation representation of the target cycle; wherein each layer of cross-attention decoding structure includes a cross-attention calculation unit, a residual connection unit, a normalization unit, and a feedforward mapping unit.

8. The method for predicting the health status of on-board lithium batteries based on historical degradation interaction reconstruction networks according to claim 1, characterized in that, The method further includes: during model training, constructing a prediction loss based on the health status prediction results and the true health status labels, and reconstructing the health status labels of historical reference samples based on historical memory representation to construct a historical label reconstruction loss; and weighting the prediction loss and the historical label reconstruction loss together to form a joint training loss; pre-training the historical degradation interaction reconstruction network using source domain battery data, and then fine-tuning it using a small number of labeled samples in the target domain to achieve target domain vehicle lithium battery health status prediction; wherein, the prediction loss is used to constrain the error between the health status prediction results and the corresponding health status labels in one or more cycles after the target cycle; the historical label reconstruction loss is used to constrain the error between the historical health status labels reconstructed from the historical memory representation and the true historical health status labels; and the joint training loss is a weighted sum of the prediction loss and the historical label reconstruction loss.

9. The method for predicting the health status of an on-board lithium battery based on a historical degradation interaction reconstruction network according to claim 8, characterized in that, Transfer learning includes a source domain pre-training stage and a target domain fine-tuning stage. In the source domain pre-training stage, the historical degradation interaction reconstruction network is trained using source domain battery data to learn a general degradation representation. In the target domain fine-tuning stage, the parameters of the pre-trained network are updated using a small amount of labeled battery data from the target domain to adapt to the degradation distribution of the target vehicle lithium battery. In the target domain fine-tuning stage, some parameters in the dual-branch encoder, dual-branch cross-channel interaction module, historical enhancement decoder, or prediction output layer are frozen and the remaining parameters are updated, or all parameters are fine-tuned.

Images