Offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries

Abstract

This invention discloses an offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries, relating to the field of electric vehicle battery health monitoring technology. This invention combines online and offline data streams, collecting parameters such as battery voltage, temperature, and charge / discharge status in real time. The online component handles real-time monitoring and data acquisition, while the offline component performs deep learning and model training based on historical degradation data. Online data and the offline model work collaboratively through a closed-loop system, updating the battery health status in real time and adjusting the prediction results in each cycle. Through this closed-loop prediction system combining online real-time data and offline historical data, this invention can provide a more accurate health status assessment under limited historical degradation data conditions. This system is particularly suitable for battery management in practical applications, improving the accuracy and stability of battery health monitoring and remaining life prediction.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle battery health monitoring technology, specifically to an offline multi-scale training method for online prediction of the remaining life of electric vehicle lithium batteries. Background Technology

[0002] With the accelerated development of global transportation electrification, lithium-ion batteries (LIBs), with their advantages of high energy density, long cycle life, and low self-discharge rate, have become the core energy storage component for electric vehicles (EVs), hybrid vehicles, and various portable electronic devices. However, during the long-term driving and charge-discharge cycles of electric vehicles, irreversible physical and chemical reactions occur inside the power battery, such as structural degradation of electrode materials, electrolyte decomposition, and increased internal resistance. These aging phenomena directly manifest as a gradual decrease in battery capacity and a decline in power performance, affecting not only the vehicle's driving range but also potentially causing serious safety risks. Therefore, accurate and real-time online prediction of the battery's State of Health (SOH) and Remaining Useful Life (RUL) is crucial for ensuring driving safety, optimizing battery maintenance strategies, and reducing total life-cycle costs.

[0003] Current battery life prediction methods are mainly divided into two categories: physical model-based methods and data-driven methods.

[0004] The first category is based on physical models, mainly including electrochemical models and equivalent circuit models (ECMs). These methods attempt to simulate the electrochemical reaction mechanisms or circuit characteristics inside the battery through mathematical equations. For example, some studies have proposed electrothermal coupling models that combine cycle aging with differences between batteries, or use discrete capacity loss models combined with particle filtering algorithms for state estimation. Although physical models have strong interpretability in controlled laboratory environments, in real-world automotive applications, batteries often operate under varying temperatures, complex current loads, and different aging paths, making the assumptions of the physical models prone to failure. Furthermore, these methods heavily rely on accurate identification of mechanistic parameters, and the model construction is complex and extremely sensitive to external disturbances. This leads to significant inconsistencies between offline model construction and online application, making it difficult to support stable real-time predictions in computationally limited automotive battery management systems (BMS).

[0005] The second category is data-driven methods, particularly the emerging deep learning (DL) techniques. These methods do not require in-depth exploration of the complex electrochemical mechanisms within the battery; instead, they make predictions by mining nonlinear patterns in massive amounts of historical monitoring data. For example, methods such as using variational mode decomposition (VMD) combined with machine learning, or using sparse autoencoders combined with long short-term memory networks (LSTM), have achieved some success in battery RUL prediction tasks. Data-driven methods have proven excellent in handling the complex nonlinear relationships in battery degradation processes, and are particularly suitable for scenarios where it is difficult to build accurate physical models.

[0006] Despite significant progress in battery life prediction using data-driven models, they still face serious challenges in practical applications for the complex dynamic conditions of electric vehicles. First, existing research generally lacks a unified verification mechanism connecting offline training and online deployment, making it difficult to guarantee consistent prediction performance when migrating models from limited historical data to varied real-world vehicle conditions. Second, the degradation process of power batteries is highly nonlinear, combining long-term cumulative decay caused by physical aging with short-term capacity fluctuations caused by driving behavior. Existing single network structures struggle to simultaneously achieve effective modeling of long-range dependencies and detailed characterization of local multi-scale features. Finally, limited by sensor noise, missing samples, and distribution bias issues commonly found in real-world vehicle data, existing methods often exhibit insufficient generalization ability and weak robustness when facing complex scenarios across different temperatures, current rates, and individual batteries.

[0007] To this end, this invention proposes an offline multi-scale training method for online prediction of the remaining life of electric vehicle lithium batteries. Summary of the Invention

[0008] The purpose of this invention is to provide an offline multi-scale training method for online prediction of the remaining life of electric vehicle lithium batteries, so as to solve the problems mentioned in the background art.

[0009] According to a first aspect of the present invention, in order to achieve the above-mentioned objective, the present invention provides the following technical solution: an offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries, comprising the following steps: Obtain the full lifecycle dataset of source domain batteries, which contains the capacity degradation sequence of multiple batteries from the initial state to the end of their lifespan and their corresponding actual remaining lifespan values. A deep neural network prediction model for end-to-end remaining lifetime prediction is constructed. The prediction model includes a cascaded local fluctuation perception module, a global temporal context aggregation module, and a bidirectional degradation trend inference module. Using the source domain dataset, we perform offline, fully parameter-supervised learning on a deep neural network prediction model. By minimizing the loss function between the predicted remaining lifetime and the actual remaining lifetime, we train and solidify all network parameters of the model to obtain an offline-trained prediction model. The offline-trained prediction model is deployed to the battery management system of the target vehicle, and all its network parameters are frozen. During the online prediction phase, the current capacity degradation sequence of the target battery is collected in real time and input into the prediction model with frozen parameters. The remaining life prediction value of the target battery is directly output through one forward inference.

[0010] Furthermore, the local fluctuation sensing module in the deep neural network prediction model is configured to detect the input battery original capacity degradation sequence. Performing sliding window convolution operations extracts geometric features reflecting capacity regeneration and local mutations from short temporal segments, and then applying nonlinear activation... With downsampling operation Filtering high-frequency measurement noise to generate highly robust local feature vectors Its computational logic is defined as follows: in This represents the convolution operation. and These are the learnable parameters of the convolutional layer; Through the local fluctuation sensing module, the original battery signal is transformed into a local feature stream with high robustness.

[0011] Furthermore, the global temporal context aggregation module in the deep neural network prediction model is configured to receive the local feature vector. Long-range dependency modeling is based on a self-attention mechanism. By calculating the semantic association weights between any two time points in the sequence, the degradation patterns throughout the entire lifecycle are dynamically aggregated to generate feature representations containing global contextual information. The mapping relationship is defined as follows: in Includes projected weights used to calculate the query, key, and value matrices.

[0012] Furthermore, the local fluctuation perception module and the global temporal context aggregation module employ the following compact mathematical model for the local feature extraction and global dependency aggregation process: (41) For the input sequence First, through convolution operations and downsampling functions... Obtaining local features: in This represents the convolution operation. For activation functions; (42) Construct a multi-head attention module to directly calculate global features. Unlike separate calculations of query, key, and value matrices, this method involves concatenation. A unified linear mapping is performed on the results of multiple parallel attention heads: Among them, any attention head The calculation formulas are combined as follows: In the formula, The first Projection weight matrix of the head, The output transformation matrix is ​​used to implement feature dimensionality upscaling from local texture to global semantics.

[0013] Furthermore, the bidirectional degradation trend inference module in the deep neural network prediction model is configured to receive the global feature representation. By utilizing a bidirectional recurrent neural network structure, the cumulative decay trend of battery capacity is fitted along the forward path of the time axis, and the feedback constraints of future states on the current state are captured along the reverse path. Finally, a deep temporal feature incorporating bidirectional evolutionary logic is output. Its state update process is characterized as follows: (3) in and These represent the hidden states at the last moment in the forward and reverse directions, respectively. This indicates a feature fusion operation.

[0014] Furthermore, during the implementation of the bidirectional degradation trend inference module, the bidirectional dynamic evolution process adopts a vectorized gating update mechanism: Define the state update function of the LSTM cell as follows: It will come from The Current input of the step Compared to the previous hidden state As input; Forward evolution: by time series The deduction, which updates the cell state through a composite gating mechanism of forget gate, input gate, and output gate, is described in the following computational process: In the formula, Indicates the first The positive hidden state at each time step. Indicates the first The positive cell state at each time step. Indicates the first The positive hidden state at each time step. Indicates the first The positive cell state at each time step; Weights of the forward network The parameter set includes the forget gate, input gate, output gate, and the weight matrix and bias terms corresponding to the candidate memory units; For the reverse evolution process, according to the time series Reverse reasoning, applying the current input features The reverse hidden state at the next time step And the reverse cell state at the next moment. By inputting the same data into the inverse LSTM unit, the inverse hidden state at the current time step is obtained. and reverse cell state The calculation relationship is as follows: In the formula, Indicates the first The reverse hidden state at each time step Indicates the first Reverse cell state at each time step Indicates the first The reverse hidden state at each time step Indicates the first Reverse cell state at each time step; This represents the set of parameters for a reverse LSTM network. During bidirectional fusion, the positive hidden states corresponding to the same time step are... With reverse hidden state Feature fusion is performed to obtain bidirectional time-step features. , is represented as: .

[0015] Furthermore, during the offline training of the deep neural network prediction model, the predicted values ​​are used. Compared to actual lifespan The mean squared error between them is used as the objective function: In the formula, This represents a deep neural network.

[0016] Furthermore, offline, fully parameter-supervised learning of the deep neural network prediction model is performed using the source domain dataset, as detailed below: (81) The network is analyzed using the backpropagation algorithm. Perform full-parameter training to enable the model to learn a universal degradation feature representation across batteries, until the model converges and locks all weight parameters. Generate a fixed reasoning model, with the optimization objective being: In the formula, The optimal parameter set obtained through supervised learning; The network parameters to be optimized; This represents the number of training samples; For sample index; The loss function is used to predict the output and the true label. For deep neural networks to the first The mapping output of each input sample; For the first One source domain input sequence; This corresponds to the actual remaining lifespan value; (82) Online zero-sample cross-condition prediction: For target batteries with operating conditions different from those in the source domain, the parameters of the deep neural network prediction model are forced to remain frozen, and only the real-time observation sequence of the target battery is input. The remaining lifetime prediction is directly mapped and output through forward inference. To achieve lossless transfer of knowledge from the source domain to the target domain: In the formula, The predicted remaining lifespan of the target battery; This is a real-time observation sequence for the target battery; For the solidified inference model deployed in the target battery management system; This is the optimal parameter set corresponding to the fixed inference model, and it remains unchanged during the online prediction phase.

[0017] According to a second aspect of the present invention, the present invention provides an offline multi-scale training system for online remaining life prediction of electric vehicle lithium batteries, used in the offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries described in the first aspect, comprising: The dataset acquisition module is used to acquire the full life cycle dataset of the source domain battery. The dataset contains the capacity degradation sequence of multiple batteries from the initial state to the end of their life and their corresponding real remaining life values. The model building module is used to build a deep neural network prediction model for end-to-end remaining lifetime prediction. The prediction model includes a cascaded local fluctuation perception module, a global temporal context aggregation module, and a bidirectional degradation trend inference module. The model training module is used to perform offline, fully parameter-supervised learning of a deep neural network prediction model using the source domain dataset. By minimizing the loss function between the predicted remaining lifetime and the actual remaining lifetime, all network parameters of the model are trained and solidified to obtain the offline-trained prediction model. The prediction output module is used to deploy the offline-trained prediction model to the battery management system of the target vehicle and freeze all its network parameters. In the online prediction stage, the current capacity degradation sequence of the target battery is collected in real time and input into the prediction model with frozen parameters. The remaining life prediction value of the target battery is directly output through one forward inference.

[0018] According to a third aspect of the present invention, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein when the processor loads and executes the computer program, it employs the offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries described in the first aspect.

[0019] This invention has at least the following beneficial effects: (1) This invention constructs a closed-loop prediction architecture that combines offline deep training with online frozen inference. In the offline stage, the model performs full-parameter supervised learning based on multi-source historical data, fully mining and solidifying degradation knowledge common to all batteries; in the online stage, the model parameters are completely frozen and deployed to the vehicle battery management system, and the remaining lifetime prediction value can be directly output without any online fine-tuning or gradient updates. This design eliminates the computational overhead and storage burden caused by online backpropagation, significantly improves the real-time response capability and engineering deployment feasibility of the model on embedded platforms with limited computing power, and effectively avoids the model drift problem caused by online data distribution offset.

[0020] (2) This invention designs a multi-level temporal feature decoupling and fusion network, which can comprehensively characterize the multi-scale evolution law in the battery degradation process. By performing sliding convolution and downsampling on the original capacity sequence through the local fluctuation perception module, the model can keenly capture short-term local fluctuations caused by capacity regeneration, load mutation, etc., and effectively suppress sensor noise interference. By introducing a multi-head self-attention mechanism through the global temporal context aggregation module, the model can establish long-range dependencies throughout the entire life cycle, dynamically aggregate degradation patterns at different time points, and achieve accurate modeling of the global evolution trend. This multi-scale feature extraction mechanism significantly enhances the model's adaptability to complex non-stationary degradation sequences.

[0021] (3) This invention introduces a bidirectional degradation trend inference module in the time series modeling layer. It uses a bidirectional recurrent neural network to fit the historical cumulative decay process along the forward time axis and capture the feedback constraints of future trends on the current health status along the reverse time axis. This bidirectional evolution modeling method breaks the limitation of traditional unidirectional time series models that can only rely on historical information. It enables the model to learn more complete and consistent dynamic evolution laws from the overall degradation trajectory, thereby improving the stability and convergence efficiency of remaining lifetime prediction. It is especially suitable for long-sequence, highly nonlinear degradation scenarios.

[0022] (4) This invention proposes a "full parameter locking" cross-condition knowledge transfer strategy, which achieves lossless transfer of source domain knowledge to the target domain without the need for target domain data to participate in training. This strategy enables the model to maintain high prediction accuracy and good generalization ability when facing target batteries with significantly different conditions from the training set. It effectively overcomes the shortcomings of traditional data-driven methods in terms of weak generalization ability and poor adaptability under complex conditions such as variable temperature and variable rate, and significantly improves the model's cross-condition robustness and practical application value.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0024] Figure 1 This is a flowchart of the training method described in this invention; Figure 2 This is a capacity cycle diagram of the degradation curve in this invention; Figure 3 This is a diagram showing the experimental results of CALCE I in this invention; Figure 4 This is a diagram showing the experimental results of CALCE II in this invention; Figure 5 This is a diagram showing the experimental results from NASA I in this invention; Figure 6 This is a diagram showing the results of the NASA II experiment in this invention; Figure 7 This is a scatter plot showing the error distribution in this invention. Detailed Implementation

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

[0026] Example 1: This embodiment is applicable to on-board power battery systems for pure electric vehicles (BEV), plug-in hybrid electric vehicles (PHEV), and electric heavy trucks. It is specifically designed to enable real-time, high-precision online prediction of the remaining service life (RUL) of the power battery under dynamic driving and complex load conditions.

[0027] Please see Figures 1-7 This invention provides a technical solution: an offline multi-scale training method for online prediction of the remaining life of electric vehicle lithium batteries, comprising the following steps: S1. Obtain the full life cycle dataset of the source domain battery, which contains the capacity degradation sequence of multiple batteries from the initial state to the end of their life and their corresponding real remaining life value. S2. Construct a deep neural network prediction model for end-to-end remaining lifetime prediction. The prediction model includes a cascaded local fluctuation perception module, a global temporal context aggregation module, and a bidirectional degradation trend inference module, as detailed below: This embodiment constructs an end-to-end heterogeneous feature learning architecture, aiming to extract multi-dimensional decay information layer by layer from the original battery capacity degradation sequence X={x1,x2,…,xT} through cascaded computing modules. Its internal logic includes three core dimensions: spatial local perception, global semantic aggregation, and bidirectional temporal evolution. S21. The local fluctuation sensing module is configured primarily to capture capacity regeneration or local abrupt changes exhibited by lithium batteries within short cycles. It utilizes a one-dimensional convolutional kernel to perform a sliding scan along the time axis, extracting geometrical abrupt changes in the degradation trajectory. To enhance the model's immunity to measurement random noise, this invention introduces nonlinear mapping and spatial downsampling processing after the convolutional layer. Its local feature representation... The evolutionary logic is as follows: in, This represents the convolution operation. For activation function, and This is the learnable tensor for this layer. Through this step, the original signal is transformed into a highly robust local feature flow; Through this step, the local fluctuation sensing module transforms the original battery signal into a highly robust local feature stream. S22. Global Temporal Context Aggregation Module Construction: To address the difficulty of traditional models in capturing long-range nonlinear dependencies throughout the entire lifecycle, this layer introduces a self-attention mechanism. This module dynamically aggregates degradation patterns by calculating the association weights between any two time points within the sequence. Unlike the local constraints of traditional convolution, this method achieves a higher-dimensional mapping of global semantics through multi-head projection. Specifically, the system performs parallel computing Each attention head is assigned a value, which is then concatenated and subjected to a uniform linear transformation to calculate the value of each attention head. The formula is: This structure enables a feature transition from local texture to a global macroscopic degradation trend; It should be further explained that the local fluctuation perception module and the global temporal context aggregation module use the following compact mathematical model for local feature extraction and global dependency aggregation: For the input sequence First, through convolution operations and downsampling functions... Obtaining local features: in This represents the convolution operation. For activation functions; Construct a multi-head attention module to directly compute global features. Unlike separate calculations of query, key, and value matrices, this method involves concatenation. A unified linear mapping is performed on the results of multiple parallel attention heads: Among them, any attention head The calculation formulas are combined as follows: In the formula, The first Projection weight matrix of the head, The output transformation matrix is ​​used to achieve feature dimensionality upscaling from local texture to global semantics. S23. Construction of the Bidirectional Degradation Trend Inference Module: After acquiring global features, to further characterize the physical evolution of the battery state over time, this embodiment employs a bidirectional recurrent network with a vectorized gating mechanism. This module fits the cumulative loss of battery capacity along the forward path and captures the feedback constraints of future trends on the current health state along the reverse path: the forward evolution is calculated through the composite control of the forget gate, input gate, and output gate to determine the forward hidden state. Reverse inference is used to process the feature sequence in reverse order, generating reverse hidden states. Final fusion: The model performs feature fusion (Concat) on the bidirectional terminal hidden states to output a deep temporal representation. : ; Specifically, during the implementation of the bidirectional degradation trend inference module, the bidirectional dynamic evolution process adopts a vectorized gating update mechanism: Define the state update function of the LSTM cell as follows: It will come from The Current input of the step Compared to the previous hidden state As input; Forward evolution: by time series The deduction, which updates the cell state through a composite gating mechanism of forget gate, input gate, and output gate, is described in the following computational process: In the formula, Indicates the first The positive hidden state at each time step. Indicates the first The positive cell state at each time step. Indicates the first The positive hidden state at each time step. Indicates the first The positive cell state at each time step; Weights of the forward network The parameter set includes the forget gate, input gate, output gate, and the weight matrix and bias terms corresponding to the candidate memory units; For the reverse evolution process, according to the time series Reverse reasoning, applying the current input features The reverse hidden state at the next time step And the reverse cell state at the next moment. By inputting the same data into the inverse LSTM unit, the inverse hidden state at the current time step is obtained. and reverse cell state The calculation relationship is as follows: In the formula, Indicates the first The reverse hidden state at each time step Indicates the first Reverse cell state at each time step Indicates the first The reverse hidden state at each time step Indicates the first Reverse cell state at each time step; This represents the set of parameters for a reverse LSTM network. During bidirectional fusion, the positive hidden states corresponding to the same time step are... With reverse hidden state Feature fusion is performed to obtain bidirectional time-step features. , is represented as: ; S3. Use the source domain dataset to perform offline full-parameter supervised learning on the deep neural network prediction model. By minimizing the loss function between the predicted remaining lifetime and the actual remaining lifetime, train and solidify all network parameters of the model to obtain the offline trained prediction model. During the offline training of deep neural network prediction models, predicted values ​​are used. Compared to actual lifespan The mean squared error between them is used as the objective function: In the formula, Represents deep neural networks; S4. Deploy the offline-trained prediction model into the battery management system of the target vehicle and freeze all its network parameters; during the online prediction phase, collect the current capacity degradation sequence of the target battery in real time and input it into the prediction model with frozen parameters, and output the remaining life prediction value of the target battery directly through one forward inference. Offline, fully parameter-supervised learning of a deep neural network prediction model is performed using a source domain dataset, as detailed below: S41. Use the backpropagation algorithm to analyze the network. Perform full-parameter training to enable the model to learn a universal degradation feature representation across batteries, until the model converges and locks all weight parameters. Generate a fixed reasoning model, with the optimization objective being: In the formula, The optimal parameter set obtained through supervised learning; The network parameters to be optimized; This represents the number of training samples; For sample index; The loss function is used to predict the output and the true label. For deep neural networks to the first The mapping output of each input sample; For the first One source domain input sequence; This corresponds to the actual remaining lifespan value; S42. Online Zero-Shot Cross-Condition Prediction: For target batteries with operating conditions different from the source domain, the parameters of the deep neural network prediction model are forced to remain frozen, and only the real-time observation sequence of the target battery is input. The remaining lifetime prediction is directly mapped and output through forward inference. To achieve lossless transfer of knowledge from the source domain to the target domain: In the formula, The predicted remaining lifespan of the target battery; This is a real-time observation sequence for the target battery; For the solidified inference model deployed in the target battery management system; This is the optimal parameter set corresponding to the fixed inference model, and it remains unchanged during the online prediction phase.

[0028] This design completely eliminates the computational overhead of online backpropagation, ensuring that the model can still provide stable RUL prediction outputs even in resource-constrained environments.

[0029] The technical solution of the present invention will be further described below with reference to specific embodiments: I. Description of the Battery Capacity Degradation Dataset To evaluate the performance of the offline-online lithium battery lifetime prediction closed-loop framework and its multi-scale time-series fusion algorithm proposed in this invention, this embodiment selects two representative publicly available degradation datasets with significantly different physical characteristics. Comparative experiments using the University of Maryland (CALCE) dataset and the National Aeronautics and Space Administration (NASA) dataset verify the model's generalization adaptability under both stationary degradation and non-stationary perturbations.

[0030] Step 1: Setting up experimental subjects and constructing samples for multi-source batteries Data source and experimental subject selection CALCE subset: Four prism-shaped LiCoO2 batteries, numbered CS2_35 to CS2_38, were selected. These batteries have a nominal capacity of 1.1 Ah, and the experimental environment was strictly controlled at a constant temperature of 1°C to simulate the standard degradation process of power batteries under low-temperature conditions.

[0031] NASA subset: Four 18650 cylindrical LiCoO2 batteries, numbered B0005, B0006, B0007, and B0018, were selected. Their nominal capacity is 2 Ah, and the experiment was conducted at room temperature (24°C). This subset exhibits significant capacity rebound and non-stationary characteristics, and was used to test the model's robustness to noise.

[0032] Charge / discharge strategy and end-of-life (EOL) determination: Charge / discharge logic: The experiment uniformly adopted a constant current constant voltage (CC-CV) charging strategy and a constant current (CC) discharging strategy. The CALCE battery charging current was 0.55A and the discharging current was 1.1A; the NASA battery charging current was 1.5A and the discharging current was 2A.

[0033] EOL Standard Setting: To ensure the uniformity of the evaluation system, this invention defines the failure threshold as a decrease in the measured capacity of a battery to 70% of its nominal capacity. Specifically, the EOL threshold for CALCE batteries is 0.77 Ah, and the EOL threshold for NASA batteries is 1.4 Ah.

[0034] Data Characteristics and Research Significance: By integrating stationary degradation data from CALCE with fluctuating degradation data from NASA, this invention constructs a multi-dimensional experimental space encompassing different material systems, geometries, ambient temperatures, and load intensities. This sample set comprehensively records the dynamic evolution of voltage, current, and temperature, providing reliable physical support for verifying the generalization ability and robustness of the prediction framework across various operating conditions. Step 2: Summary of Dataset Parameters and Experimental Conditions Standardized comparison criteria: To eliminate dimensional differences between heterogeneous data, this invention performs a standardized data cleaning process on the selected samples to ensure the predictive accuracy of the comparison models under the same evaluation system.

[0035] Experimental conditions comparison: Table 1 summarizes the specific charge and discharge constraint parameters of CALCE and NASA batteries in this embodiment, ensuring the scientific nature of the experimental design.

[0036] Table 1 Battery charging and discharging parameters and experimental conditions for NASA and CALCE datasets. II. Data Preprocessing and Experiment Setup To ensure the reproducibility accuracy of the multi-scale temporal fusion model during offline modeling and online deployment, this embodiment designs a standardized data link and a cross-condition verification process.

[0037] Step 1: Data Preprocessing Flow Feature decoupling processing: For the CALCE dataset, this invention accurately locks the discharge cycle by locating Step_Index = 7 and uses the Coulomb integral method to accurately calculate the cycle capacity.

[0038] Capacity sequence extraction: For the NASA dataset, the capacity evolution trajectory is directly analyzed from the discharge observation records as a characterization sequence of battery health status.

[0039] Trend visualization analysis: Figure 2The degradation curves after treatment are shown using representative batteries CS2-36 and B0006, which intuitively present the physical characteristics of monotonic decline and fluctuating degradation.

[0040] Step 2: Experimental Design and Parameter Configuration Robustness control: Ten independent repeated experiments were conducted and the average was taken to eliminate random errors introduced by neural network initialization.

[0041] Prediction node settings: To verify the model's ability to capture key degradation stages, the CALCE scenario SP was set to cycle 300, and the NASA scenario SP was set to cycle 30.

[0042] Hyperparameter standardization: A uniform 100-round training cycle was set, and the time steps for CALCE and NASA were optimized to 5 and 3 respectively.

[0043] Step 3: Closed-loop verification working condition division Cross-sample generalization design: The “multi-source offline training-single-source online inference” model was adopted to construct the four sets of validation experiments shown in Table 2.

[0044] Table 2 Offline–Online Closed-Loop Verification Settings III. Experimental Validation using the CALCE dataset This section aims to evaluate the model's high-precision tracking performance in standard stationary degradation scenarios.

[0045] Fitting accuracy evaluation: In the EOL critical region of the CS2-35 battery, this model demonstrates excellent dynamic sensing capability. For the CS2-36, which exhibits more significant local fluctuations, the model still maintains a high degree of alignment with the curves. Experimental results are as follows: Figure 3 , Figure 4 As shown.

[0046] Quantitative Indicator Statistics: Table 3 shows the error assessment results.

[0047] Table 3 Prediction Error Metrics for the CALCE Dataset IV. Experimental Verification using NASA Datasets This section focuses on examining the model's noise suppression and zero-shot inference capabilities in complex, non-stationary environments.

[0048] Step 1: Response Analysis under Dynamic Disturbance Noise immunity verification: such as Figure 5 , Figure 6As shown, in the rapid decline range of B0005 and the frequent abrupt change points of B0006, the predicted values ​​of this model consistently maintain a high degree of agreement with the measured values, demonstrating excellent filter characteristics and trend capture capabilities.

[0049] Error index statistics: Table 4 summarizes the performance under complex working conditions.

[0050] Table 4 Prediction Error Metrics for NASA Datasets Step 2 Scatter Plot To further characterize the distribution and fluctuation patterns of the prediction error of the model on the time axis under different NASA operating conditions, Figure 7 shows the scatter distribution of predicted and actual capacities under NASA I and NASA II operating conditions. It can be seen that in both operating conditions, the scatter points are generally closely distributed near the ideal diagonal, the slope of the fitted line is close to 1, and most points fall within a relatively narrow confidence band. This indicates that the proposed method does not exhibit significant systematic bias under both NASA I and NASA II conditions, and the error fluctuation remains within a small range. Combined with the aforementioned time series prediction results and evaluation indicators, this further confirms the model's trend fitting ability and error control level in non-stationary degradation scenarios.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on," "mounted on," "fixed to," or "set on" another element, it may be directly on the other element or there may be an intermediate element present. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element present. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible embodiments.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0054] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. An offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries, characterized in that, Includes the following steps: Obtain the full lifecycle dataset of source domain batteries, which contains the capacity degradation sequence of multiple batteries from the initial state to the end of their life and their corresponding actual remaining life values. A deep neural network prediction model for end-to-end remaining lifetime prediction is constructed. The prediction model includes a cascaded local fluctuation perception module, a global temporal context aggregation module, and a bidirectional degradation trend inference module. Using the source domain dataset, we perform offline, fully parameter-supervised learning on a deep neural network prediction model. By minimizing the loss function between the predicted remaining lifetime and the actual remaining lifetime, we train and solidify all network parameters of the model to obtain an offline-trained prediction model. The offline-trained prediction model is deployed to the battery management system of the target vehicle, and all its network parameters are frozen. During the online prediction phase, the current capacity degradation sequence of the target battery is collected in real time and input into the prediction model with frozen parameters. The remaining life prediction value of the target battery is directly output through one forward inference.

2. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 1, characterized in that: The local fluctuation sensing module in the deep neural network prediction model is configured to process the input battery original capacity degradation sequence. Performing sliding window convolution operations extracts geometric features reflecting capacity regeneration and local mutations from short temporal segments, and then applying nonlinear activation... Downsampling operation Filtering high-frequency measurement noise to generate highly robust local feature vectors Its computational logic is defined as follows: in This represents the convolution operation. and These are the learnable parameters of the convolutional layer; Through the local fluctuation sensing module, the original battery signal is transformed into a local feature stream with high robustness.

3. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 2, characterized in that: The global temporal context aggregation module in the deep neural network prediction model is configured to receive the local feature vectors. Long-range dependency modeling is based on a self-attention mechanism. By calculating the semantic association weights between any two time points in the sequence, the degradation patterns throughout the entire lifecycle are dynamically aggregated to generate feature representations containing global contextual information. The mapping relationship is defined as follows: in Includes projected weights used to calculate the query, key, and value matrices.

4. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 3, characterized in that: The local fluctuation sensing module and the global temporal context aggregation module use the following compact mathematical model for local feature extraction and global dependency aggregation: (41) For the input sequence First, through convolution operations and downsampling functions... Obtaining local features: in This represents the convolution operation. For activation functions; (42) Construct a multi-head attention module to directly calculate global features. Unlike separate calculations of query, key, and value matrices, this method involves concatenation. A unified linear mapping is performed on the results of multiple parallel attention heads: Among them, any attention head The calculation formulas are combined as follows: In the formula, The first Projection weight matrix of the head, The output transformation matrix is ​​used to implement feature dimensionality upscaling from local texture to global semantics.

5. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 3, characterized in that: The bidirectional degradation trend inference module in the deep neural network prediction model is configured to receive the global feature representation. By utilizing a bidirectional recurrent neural network structure, the cumulative decay trend of battery capacity is fitted along the forward path of the time axis, and the feedback constraints of future states on the current state are captured along the reverse path. Finally, a deep temporal feature incorporating bidirectional evolutionary logic is output. Its state update process is characterized as follows: (3) in and These represent the hidden states at the last moment in the forward and reverse directions, respectively. This indicates a feature fusion operation.

6. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 1, characterized in that: During the implementation of the bidirectional degradation trend inference module, the bidirectional dynamic evolution process adopts a vectorized gating update mechanism: Define the state update function of the LSTM cell as follows: It will come from The Current input of the step Compared to the previous hidden state As input; Forward evolution: by time series The deduction, which updates the cell state through a composite gating mechanism of forget gate, input gate, and output gate, is described in the following computational process: In the formula, Indicates the first The positive hidden state at each time step. Indicates the first The positive cell state at each time step. Indicates the first The positive hidden state at each time step. Indicates the first The positive cell state at each time step; Weights of the forward network The parameter set includes the forget gate, input gate, output gate, and the weight matrix and bias terms corresponding to the candidate memory units; For the reverse evolution process, according to the time series Reverse reasoning, applying the current input features The reverse hidden state at the next time step And the reverse cell state at the next moment. By inputting the same data into the inverse LSTM unit, the inverse hidden state at the current time step is obtained. and reverse cell state The calculation relationship is as follows: In the formula, Indicates the first The reverse hidden state at each time step Indicates the first Reverse cell state at each time step Indicates the first The reverse hidden state at each time step Indicates the first Reverse cell state at each time step; This represents the set of parameters for a reverse LSTM network. During bidirectional fusion, the positive hidden states corresponding to the same time step are... With reverse hidden state Feature fusion is performed to obtain bidirectional time-step features. , represented as: 。 7. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 6, characterized in that: During the offline training of deep neural network prediction models, predicted values ​​are used. Compared to actual lifespan The mean square error between them is used as the objective function: In the formula, This represents a deep neural network.

8. The offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries according to claim 7, characterized in that: Offline, fully parameter-supervised learning of a deep neural network prediction model is performed using a source domain dataset, as detailed below: (81) The network is analyzed using the backpropagation algorithm. Perform full-parameter training to enable the model to learn a universal degradation feature representation across batteries, until the model converges and locks all weight parameters. Generate a fixed reasoning model, with the optimization objective being: In the formula, The optimal parameter set obtained through supervised learning; The network parameters to be optimized; This represents the number of training samples; For sample index; The loss function is used to predict the output and the true label. For deep neural networks to the first The mapping output of each input sample; For the first One source domain input sequence; This corresponds to the actual remaining lifespan value; (82) Online zero-sample cross-condition prediction: For target batteries with operating conditions different from those in the source domain, the parameters of the deep neural network prediction model are forced to remain frozen, and only the real-time observation sequence of the target battery is input. The remaining lifetime prediction is directly mapped and output through forward inference. To achieve lossless transfer of knowledge from the source domain to the target domain: In the formula, The predicted remaining lifespan of the target battery; This is a real-time observation sequence for the target battery; For the solidified inference model deployed in the target battery management system; This is the optimal parameter set corresponding to the fixed inference model, and it remains unchanged during the online prediction phase.

9. An offline multi-scale training system for online remaining life prediction of electric vehicle lithium batteries, used to implement the offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries as described in claims 1 to 8, characterized in that, include: The dataset acquisition module is used to acquire the full life cycle dataset of the source domain battery. The dataset contains the capacity degradation sequence of multiple batteries from the initial state to the end of their life and their corresponding real remaining life values. The model building module is used to build a deep neural network prediction model for end-to-end remaining lifetime prediction. The prediction model includes a cascaded local fluctuation perception module, a global temporal context aggregation module, and a bidirectional degradation trend inference module. The model training module is used to perform offline, fully parameter-supervised learning of a deep neural network prediction model using the source domain dataset. By minimizing the loss function between the predicted remaining lifetime and the actual remaining lifetime, all network parameters of the model are trained and solidified to obtain the offline-trained prediction model. The prediction output module is used to deploy the offline-trained prediction model to the battery management system of the target vehicle and freeze all its network parameters. In the online prediction stage, the current capacity degradation sequence of the target battery is collected in real time and input into the prediction model with frozen parameters. The remaining life prediction value of the target battery is directly output through one forward inference.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor loads and executes the computer program, it employs the offline multi-scale training method for online remaining life prediction of electric vehicle lithium batteries as described in claims 1 to 8.

Images