Real vehicle power battery capacity attenuation trajectory prediction method based on de-noising diffusion model

Abstract

The invention discloses a real vehicle power battery capacity attenuation trajectory prediction method based on a de-noising diffusion model. According to the method, a diffusion generation model is applied to the field of battery capacity prediction for the first time, multi-modal features are constructed for charging and discharging data under a real vehicle working condition, feature expressions are extracted through Transform and CNN, and the features are integrated into high-dimensional embedded vectors. And then, a DDPM model based on a ContextUnet architecture is utilized to learn a dynamic change rule of capacity attenuation in a noise environment through a noise diffusion and reverse denoising process under the guidance of multi-modal features, so that high-precision capacity prediction is realized. The training process is further combined with an implicit denoising mechanism of DDIM, and the reasoning speed is increased while the prediction quality is improved. Finally, a capacity attenuation track confidence interval is obtained by performing multiple times of prediction sampling calculation on a predicted capacity attenuation result, and the capacity attenuation track confidence interval can be used for deducing key indexes such as the state of health (SOH) and the remaining service life (RUL) of the battery. The method has the characteristics of high prediction precision and strong adaptability, obviously enhances the intelligent level of the battery management system, and has important practical value for health assessment and life management of the new energy automobile battery.

Description

Technical Field

[0001] The present invention proposes a novel battery capacity attenuation trajectory prediction technology, which aims to predict the capacity attenuation trajectory of new energy vehicle power batteries by using the denoising diffusion model in the field of graph generation. It belongs to the technical field of new energy vehicle battery management and battery capacity attenuation prediction. Background Art

[0002] Battery capacity decay is a key metric of widespread interest in electric vehicle purchases, vehicle evaluations, insurance, and policy subsidy decisions. In particular, in real-world driving environments, electric vehicle battery performance inevitably degrades over time. This often leads to "battery lifespan anxiety"—the fear that battery capacity will decline too quickly, impacting range and driving experience. Therefore, accurately predicting battery capacity decay and visualizing its degradation trends can help users intuitively understand their vehicle's real-time performance, alleviate anxiety, and optimize their vehicle usage habits from multiple perspectives, including economic and safety considerations. Specifically, users can flexibly adjust their charging strategies based on battery capacity decay trends, rationally scheduling charging schedules to avoid overcharging, frequent charging, and charging in high temperatures, effectively extending the lifespan of their power batteries. Furthermore, this trend can guide users in improving their driving and maintenance habits, including regular vehicle inspections and maintenance, to avoid potential safety hazards caused by battery performance degradation. By properly scheduling battery maintenance and replacement, users can extend the vehicle's lifespan and ultimately reduce vehicle costs. Vehicle manufacturers can optimize the design of power batteries and improve battery durability and endurance by analyzing the accurately predicted battery capacity decay trajectory. They can also use this data to improve the battery management system (BMS), optimize battery charging and discharging strategies through algorithms, and minimize the rate of battery decay. At the same time, vehicle manufacturers can adjust after-sales service strategies based on battery decay trends, provide customized battery health management solutions and capacity decay warnings, and remind users to perform battery maintenance or replacement in a timely manner, thereby increasing the brand's market competitiveness. In general, the accurate prediction of battery capacity decay trajectory provides a win-win solution to the core issues of electric vehicle users and electric vehicle manufacturers. This technology is an important tool for improving the user experience of car owners and alleviating their battery life anxiety. It is also a key means for vehicle manufacturers to improve their product quality and user satisfaction and enhance their competitiveness, laying a solid foundation for the sustainable development of the electric vehicle industry.

[0003] Current methods for predicting battery capacity degradation fall into the following categories. Model-based approaches primarily describe battery performance degradation through physical degradation modeling or empirical formulas. Common methods include semi-empirical models, electrochemical models, and equivalent circuit models. Semi-empirical models describe the relationship between capacity loss and cycle number by fitting experimental data; electrochemical models and equivalent circuit models, combined with filtering methods, predict RUL by identifying degradation parameters. These methods can be applied under real-world vehicle conditions, but their reliance on assumptions and parameter calculations makes them difficult to address the complexity of battery capacity degradation, limiting their feasibility in real-world applications. The rapid development of artificial intelligence (AI) technology in recent years has brought new breakthroughs in many fields, and this trend has also had a profound impact on the field of battery life prediction. Data-driven AI-based approaches use statistical or machine learning techniques to construct battery degradation prediction models directly from data, bypassing the complex physical modeling process. The basic steps of data-driven modeling include extracting characteristic data related to battery capacity degradation and selecting algorithms suitable for specific application scenarios, resulting in high adaptability. This paper, using a data-driven approach, applies the diffusion model to the field of battery capacity degradation trajectory prediction for the first time. It proposes a capacity degradation trajectory prediction method based on the multimodal features of a denoised diffusion model and generative learning. Compared to traditional models, this method's greatest advantage lies in its ability to effectively adapt to battery data under complex operating conditions. In particular, it demonstrates greater robustness and generalization capabilities when dealing with the noise and non-stationary characteristics commonly found in real-world vehicle operating conditions, generating highly accurate predictions that better reflect the actual degradation trends of battery capacity. Summary of the Invention

[0004] With the rapid development of the new energy vehicle industry, battery capacity decay trajectory prediction is becoming an important technical direction. Accurate capacity decay trajectory prediction can help users optimize battery use and maintenance, extend battery life, reduce vehicle costs, and thus improve vehicle economy and reliability. This paper proposes a method for predicting power battery capacity decay trajectory based on a denoising diffusion model. First, the on-board battery management system regularly collects and uploads battery charge and discharge data to the cloud; secondly, the collected data is preprocessed and the capacity label value is calculated. Deep learning models such as Transformer and CNN are used to extract and fuse the processed data to generate multimodal features as input for subsequent network training; then, the ContextUNet denoising diffusion probability model architecture is used, and the hyperparameters required by the model are initialized, including the number of diffusion steps and the noise variation range; after that, the capacity label of each sample is perturbed by random forward diffusion noise, and the training network is modulated by multimodal features to iteratively recover the capacity data distribution of the original sample from the noise, thereby learning the potential relationship between the input features and the capacity decay trajectory; finally, after training is completed, the optimized model is used to predict the battery capacity decay trajectory, and its accuracy and reliability under actual vehicle operating conditions are verified to ensure that periodically updated battery capacity decay trajectory predictions can be provided in the battery management system.

[0005] 1. To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the capacity decay trajectory of a power battery based on a denoising diffusion model, comprising the following steps:

[0006] A. Vehicle charging and discharging data collection, preprocessing and label acquisition;

[0007] B. Extract and filter multimodal features based on the data processed in step A and embed them into ContextUnet in C as context information.

[0008] C. Build a capacity decay trajectory prediction model based on DDPM (denoising diffusion model) with ContextUnet as the core;

[0009] D. Based on the label values and multimodal feature embedding values, the DDPM-improved DDIM denoising strategy is used to predict the diffusion capacity decay trajectory. Finally, the confidence interval of the capacity decay trajectory prediction is constructed through multiple sampling.

[0010] The specific steps of vehicle charging and discharging data collection, preprocessing and label acquisition in step A are as follows:

[0011] A1. Utilize the vehicle's onboard BMS system to regularly collect and upload electric vehicle charging and discharging data, including time, total voltage, current, SOC (battery state of charge), maximum cell voltage, minimum cell voltage, maximum cell temperature, and minimum cell temperature. The collection interval is weekly. All of this data is recorded and stored for subsequent analysis and processing.

[0012] A2. Data cleaning is performed on the collected vehicle charging and discharging data. First, a threshold range is used to filter out abnormal data, such as unreasonable and drastic fluctuations in voltage, current, or temperature. Then, the K-nearest neighbor interpolation method is used to fill in missing values to ensure the integrity of the time series data. Finally, a Gaussian filter is used to smooth the noisy data to ensure data stability and reliability.

[0013] A3. Normalize the cleaned charge and discharge data, use the median for standardization, align the time, and ensure that the feature data is in the same time and space range through interpolation and filling.

[0014] A4. Based on the actual vehicle charge and discharge data processed above, the battery capacity of each charge and discharge cycle is calculated using the following variable ampere integral formula: in, is the sampling interval, is the charging current, The battery charge state is set to a fixed range (such as 30%-80%) to avoid extreme The value affects the capacity evaluation, ensuring that the battery is calculated under a relatively stable working state, and improving the accuracy and reliability of capacity evaluation. In real-world vehicle operating conditions, estimation errors and data noise can cause significant fluctuations in the capacity value calculated from a single calculation. Therefore, we perform statistical processing (using the mean or median) on the capacity values calculated from all charge and discharge cycles over a certain period of time (such as a week or a month) to obtain the calibrated capacity for that period. This approach effectively reduces the impact of errors and produces more stable capacity data as labels.

[0015] In step B, multiple key multimodal features are extracted based on the pre-processed raw vehicle charge and discharge data. These features are used to train a diffusion model to predict the battery capacity decay trajectory. The specific steps are:

[0016] B1. We extracted statistical features calculated based on weekly charge and discharge cycle data. On this basis, we extracted the mechanism characteristics of battery packs and battery cells, and constructed graph features based on the mechanism characteristics of battery cells. We combined three different types of features to construct multimodal features as model input to improve the model's predictive capabilities.

[0017] B2. Statistical features are key indicators extracted from battery data to describe data distribution and variation patterns. Because battery capacity decay is a slow, long-term process, performing statistical calculations on the data (such as mean, variance, and skewness) on a weekly or monthly basis can capture the temporal characteristics of battery data, reduce noise interference on the model, and reveal patterns of capacity decay. The extracted statistical features are as follows.

[0018] (1) Mileage Increase (MI): This refers to the cumulative mileage of an electric vehicle during driving. As the mileage increases, the battery undergoes more charge and discharge cycles, its capacity gradually decays, its internal impedance increases, and its performance gradually deteriorates.

[0019] (2) Average Maximum Cell Temperature (AMCT): refers to the highest average temperature reached by a battery cell during operation. High temperatures will accelerate battery capacity degradation and increase safety risks.

[0020] (3) Average State of Charge (ASOC): This indicates the average state of charge of the battery during use, reflecting the battery's charge and discharge trajectory. A long period of high or low SOC will accelerate battery capacity degradation.

[0021] (4) Average Pack Voltage (APV): This refers to the average voltage of the entire battery pack, which can reflect the health status, charge and discharge conditions, and balance of the battery pack. Excessively high or low average voltage will affect the battery's capacity retention rate.

[0022] (5) Average Cell Voltage Difference (ACVD): describes the voltage difference between battery cells. A larger voltage difference indicates an increased imbalance in the battery pack, causing some cells to be overcharged or over-discharged, exacerbating capacity degradation.

[0023] (6) Impedance Increase (II): This indicates that the internal impedance of the battery increases over time and with use. Impedance increase usually indicates battery aging, resulting in power loss and increased heat generation.

[0024] (7) Cycling Time (CCT): Indicates the number of complete charge and discharge cycles a battery undergoes. The battery life is usually related to the number of cycles. The more cycles, the more obvious the battery capacity decay.

[0025] (8) Calendar Time (CAT): refers to the natural time from the production of the battery to the present. Even if the battery is not used, its performance will still decline due to calendar aging.

[0026] (9) Charging Accumulated Ah (CAAh): It indicates the amount of electricity (ampere-hours) charged to the battery during the entire use process. It is used to evaluate the impact of the charging trajectory on the battery life. A high cumulative charging ampere-hour usually means that the battery has experienced more cycles, resulting in capacity decay.

[0027] (10) Discharging Accumulated Ah (DAAh): refers to the cumulative amount of electricity discharged by the battery (ampere-hours), reflecting the intensity of battery use and its impact on battery life.

[0028] (11) Regenerated Accumulated Ah (RAAh): This refers to the amount of energy recovered by the battery through regenerative braking and other methods. Frequent energy recovery processes may cause the battery to be in a high instantaneous charging current state, increase polarization, affect the uniform insertion of lithium ions, and thus accelerate battery aging.

[0029] (12) Charging Accumulated Energy (CAE): Indicates the total energy charged into the battery, usually calculated in kilowatt-hours (kWh), which helps analyze charging efficiency and attenuation trends.

[0030] (13) Discharging Accumulated Energy (DAE): refers to the total energy released by the battery (kWh), reflecting the overall energy output of the battery. High-rate output during discharge leads to increased polarization, lithium deposition or local temperature increase, thereby accelerating capacity decay.

[0031] (14) Regenerated Accumulated Energy (RAE): refers to the total energy recovered by the battery through regenerative braking and other methods. High-frequency energy recovery can cause transient high-current charging, leading to uneven deposition of lithium ions, increasing the risk of lithium plating and affecting battery life.

[0032] B3. To improve model efficiency and reduce redundant features, we screened several extracted statistical features and used XGBoost to evaluate feature importance, eliminating features that contributed less to the capacity decay trajectory. During XGBoost training, the importance of each feature can be calculated using the gain, which represents the contribution of the feature to the optimization of the loss function during the decision-making process. The calculation formula is as follows: in, It is a feature Importance score; is the total number of all decision trees; and It is a decision tree The accumulated values of gradient and second-order derivative in ; Is a regularization term to prevent overfitting. XGBoost calculates the importance score of each feature , set the feature importance threshold , features below the threshold will be eliminated, so we can effectively remove redundant information, retain the key statistical features that are most valuable for capacity decay prediction, and improve the robustness and generalization ability of the model.

[0033] B4. Mechanism characteristics are key variables extracted based on the physical and chemical mechanisms within the battery, explaining the root causes of capacity decay. These characteristics are usually closely related to the internal reactions, material decay, and kinetic properties of the battery. Compared with statistical characteristics, mechanism characteristics can more directly explain the health status of the battery and its aging mechanism, and have stronger interpretability. For example, by extracting features from the voltage and current curves during the battery's charge and discharge process within a week or a month, it is possible to analyze the changes in internal materials and reveal the specific trajectory of capacity decay. For battery packs, the battery mechanism characteristics we extracted are as follows.

[0034] (1) IC Sequence: IC (Incremental Capacity) analysis is a method based on the dQ / dV-V curve. For the curve obtained by multiple cycles within a week or a month, the average dQ / dV value corresponding to each voltage horizontal coordinate point is calculated to reduce data fluctuations and the impact of data loss. By analyzing the peak position and height changes of the IC curve, the battery capacity attenuation can be obtained. As the battery ages, the peak of the IC curve will drift and the peak height will gradually decrease, reflecting the attenuation of the battery capacity.

[0035] (2) Resistance and Capacitance Model Parameter Sequence: The parameter sequence of the battery internal resistance and capacitance model (such as ohmic internal resistance Rint, polarization resistance Rp, and polarization capacitance Cp) is defined based on the Thevenin equivalent circuit model and is obtained by calculating the voltage and current data during the discharge process through the recursive least squares method (RLS). To reduce noise interference, the parameter values corresponding to each time point within a week or a month can be averaged or median. The internal resistance and capacitance sequence is highly sensitive to battery aging and can effectively capture the overall aging trend of the battery system according to the law of time and number of cycles, reflecting the changing law of the capacity attenuation trajectory.

[0036] B5. The extracted mechanism sequence characteristics of the battery pack above can characterize the physical and chemical changes inside the battery and have good interpretability. However, battery aging is a process of interaction of multi-dimensional factors. The sequence characteristics of the battery pack can only describe the data trend of a single dimension, and it is difficult to fully capture the interaction between different mechanism characteristics. Moreover, the characteristics extracted at the battery pack level are the result of the combined effect of multiple battery cells inside it. It is difficult to measure the specific aging of these different cells, and the abnormal degradation of individual cells may be masked. In contrast, the mechanism characteristics of battery cells can provide more detailed information, which can not only reflect the changes in the state of the cells themselves, but also can be used to analyze the consistency and differences between different cells in the battery pack. Therefore, we further uniformly sampled 100 times a week or a month and extracted the mechanism characteristics of each battery cell during the charging process by time alignment and interpolation filling, and constructed an n×m two-dimensional matrix as follows:

[0037] (1) Cell Voltage (Vc): {V}_{c}=\left [ {{V}_{11},{V}_{12}\cdot \cdot \cdot ,{V}_{1m},{V}_{21},{V}_{22},\cdot \cdot \cdot {V}_{2m},\cdot \cdot \cdot ,{V}_{n1},{V}_{n2},\cdot \cdot \cdot {V}_{nm}{}^{}} \right ] in Indicates the Unit No. The target voltage range is set to 3.9V-4.0V, and 100 samples are uniformly sampled in this voltage range. Time alignment and interpolation filling are used to avoid data loss. Therefore, different charging conditions and different charging start conditions under actual vehicle working conditions are achieved. The battery voltage data are controlled to be of similar length to facilitate subsequent analysis.

[0038] (2) Cell Temperature (Tc): {T}_{c}=\left [ {{T}_{11},{T}_{12}\cdot \cdot \cdot ,{T}_{1m},{T}_{21},{T}_{22},\cdot \cdot \cdot {T}_{2m},\cdot \cdot \cdot ,{T}_{n1},{T}_{n2},\cdot \cdot \cdot {T}_{nm}{}^{}} \right ] in Indicates the Unit No. The target temperature range is set to 25°C-45°C, and 100 samples are evenly sampled within this temperature range. To ensure sufficient valid data at the evenly sampled points, the temperature data is also processed using time alignment and interpolation.

[0039] B6. and To further integrate voltage and temperature information based on the matrix, a cell feature map (CFM) was constructed. The CFM is constructed by concatenating the cell voltage and temperature matrices to generate a two-channel, two-dimensional matrix. The matrix dimensions are kept consistent, ensuring that voltage and temperature information can be processed in parallel. This fusion approach preserves the characteristics of battery cells at different time steps, captures the correlation between voltage and temperature, and enhances the ability to characterize cell capacity changes, making it suitable for subsequent analysis using 2D convolutional neural networks (2D CNNs).

[0040] B7. After the above feature extraction, in order to fully explore the different information in the battery health status, we use CNN (convolutional neural network) to process the single feature map, Transformer to process the sequence features, and use the fully connected layer to fuse the statistical features (point features) in the final stage to form a complete prediction framework. For the single feature map, we first use several two-dimensional convolution kernels to extract the local spatial relationship features and output the feature map; then perform pooling on the feature map to obtain a reduced-size feature map; after multiple layers of stacked convolution and pooling operations, we flatten the final feature map and convert it into a one-dimensional vector As a high-dimensional embedding representation of the two-dimensional graph features of the battery cell, It fully captures the complex correlation pattern between battery voltage and temperature, and will serve as the input of the subsequent battery capacity attenuation evaluation and prediction module; for time series features, we use Transformer to process them. Transformer uses a multi-head self-attention mechanism to calculate the spatiotemporal dependencies between different features, while introducing position encoding to retain time series information. This fusion method not only retains the macroscopic laws of statistical features, but also combines the microscopic characteristics of mechanism features, thereby significantly improving the model's ability to characterize the battery degradation process, laying a solid foundation for the subsequent prediction of battery capacity attenuation trajectory, and finally high-dimensional embedding representation Finally, in the face of statistical features without temporal features, we directly input the independent final fully connected layer (FC) for fusion, where we integrate the graph features extracted by CNN , Sequence features extracted by Transformer And point features , forming the final comprehensive feature representation , Z=FC([{Z}_{G},{Z}_{S},{X}_{P}]) . This multimodal feature fusion method not only preserves the macroscopic laws of statistical features but also combines the physical interpretability of mechanism features with the high-dimensional spatial expression capability of graph features as the model prediction input for battery capacity decay trajectories. This combination of features enables the model to take into account both the overall picture and the details, providing comprehensive data support for subsequent accurate prediction of battery capacity decay trajectories.

[0041] In step C, after label value acquisition and multimodal feature organization, the input of the capacity decay trajectory prediction model has been prepared. Then, the capacity decay trajectory prediction model is built around the ContextUnet core. The specific steps are as follows:

[0042] C1. Before using the denoising diffusion model (DDPM) to predict capacity decay trajectories, you need to initialize the diffusion hyperparameters to define the model's noise diffusion process, including the number of noise diffusion steps and the noise variation range. These parameters control the noise evolution during training.

[0043] C2. We use ContextUnet (a multimodal feature modulator) as the primary model architecture for the Diffusion Model. Based on the classic U-Net network structure, it combines temporal and feature embedding mechanisms, enabling it to better perceive the rich information of battery changes and model the complex evolution of battery health status. The various components of the model architecture function as follows: the residual convolution module performs feature extraction, improving feature representation and preventing gradient vanishing; the downsampling module reduces data dimensionality through convolution to obtain high-level features; the upsampling module restores data dimensions through deconvolution and strengthens feature representation; the position encoder encodes the relative position of the input data in the sequence, enabling the model to identify temporal relationships between features when processing time-dependent data, enhancing its ability to model data with long temporal dependencies; the temporal encoder ensures the effective utilization of temporal information in the diffusion process, introducing independent representations at different time steps. This allows the model to accurately track data evolution during denoising and improves its dynamic modeling capabilities; the multi-layer perceptron (MLP) embeds contextual information (features) into the model, and finally, the fully connected layer generates the final prediction output, ensuring accurate predictions after multimodal feature fusion.

[0044] In step D, the established capacity decay trajectory prediction model is used to embed multimodal features as context information to predict noise, thereby restoring the capacity decay trajectory predicted by the data output. The specific steps are as follows:

[0045] D1. Before training, the standardized features and labels are divided into training sets, validation sets, and test sets, and loaded in batches through DataLoader.

[0046] D2. During training, since the noise diffusion parameters have been initialized before, noise is gradually added to the loaded training data. The core formula of noise diffusion is: in, is the original data, i.e., the extracted battery features; is the data after adding noise; is the noise that conforms to the standard normal distribution; is the noise scheduling factor, which determines the How much noise to add. It decreases as the time step increases, making the final data become pure noise.

[0047] D3, then denoising is performed. The denoising process is the core of the Diffsion Model, and the goal is to restore the original data from the noisy data. Noisy data , we need to restore the original data by denoising The key here is noise prediction, which is to recover the noise from the noisy input through the encoder-decoder learning of the ContextUnet model. The process is as follows: in is in the time step Noisy data (i.e., the battery at The noise-contaminated data of the first step is the input of the model; is the noise predicted by the ContextUnet model; is the denoised data, indicating that at time step At the same time, the loss function Mean Square Error (MSE) is set to minimize the noise prediction error, and the Adam optimizer adaptively adjusts the learning rate of the parameters to help the model converge stably. The calculation formulas of the loss function and optimizer are as follows: in, is the actual noise added; in, is the learning rate, are the current model parameters, is the gradient momentum, weighted average of squared gradients, is a small constant that prevents division by zero and predicts the added noise. Through the noise diffusion and denoising training process, the model can restore battery data under complex working conditions to characteristics that conform to the battery degradation law. In addition, to improve the denoising efficiency and accelerate the sampling process, we introduced DDIM. Compared with the traditional DDPM denoising method mentioned above, DDIM adopts an implicit denoising strategy, which allows the noise to be retained or some noise to be added to the data during the denoising process, avoiding the over-denoising phenomenon that may occur in DDPM, maintaining data diversity and improving generation quality. The implicit denoising strategy formula is as follows: in, is the recovered original data (predicted value); and is the noise scheduling factor, is the noise predicted by the ContextUnet model; compared with the above DDPM, DDIM introduces a front time step The noise scheduling factor , in addition, an additional term related to the noise of the previous step is added This allows DDIM to retain more noise information during the denoising process, allowing for more flexible denoising operations while reducing the amount of computation required at each time step. Through this approach, we can accurately recover the true battery capacity from noisy data, reducing the impact of noise on the estimation results and laying the foundation for accurately predicting battery capacity decay trajectories.

[0048] D4. In order to further improve the stability of capacity decay trajectory prediction, after using DDIM to generate battery capacity decay trajectory samples, the present invention introduces a capacity decay trajectory confidence interval estimation mechanism based on multiple sample statistics. For each set of input features, we repeat the DDIM denoising process based on different random noise initial conditions to generate a total of 20 sets of battery capacity decay prediction trajectory samples. Subsequently, the capacity values of these 20 groups of prediction samples at each time step are statistically processed, and their mean, standard deviation and variance are calculated to capture the prediction fluctuations due to randomness. Afterwards, we adopt the normal distribution assumption and use the 95% confidence interval calculation formula: in, is the mean of 20 prediction values, is the standard deviation of the corresponding time step, and 1.96 is the critical value of the 95% confidence interval under the standard normal distribution. This method can effectively quantify the uncertainty of the prediction results, thereby constructing the upper and lower confidence intervals of the battery capacity decay trajectory, providing solid data support for subsequent battery state of health (SOH) and remaining useful life (RUL) assessments.

[0049] The capacity decay trajectory predicted by the combined diffusion model of D5, DDPM, and DDIM not only provides detailed trends in battery capacity decay but also lays the foundation for further deriving key indicators such as the battery's state of health (SOH) and remaining useful life (RUL). The SOH can be calculated as the ratio of the predicted mean capacity to the initial capacity, while the confidence interval provides estimates of the upper and lower limits of the SOH. This extends the SOH from a single value to a dynamic health indicator in the form of a credible interval. As the battery capacity decays, the SOH value can be used to monitor battery health in real time, providing a basis for maintenance decisions. Traditional methods typically use the time point when the capacity trajectory first drops to the EOL threshold (e.g., 80%) as the endpoint for remaining useful life (RUL). However, in this invention, the earliest possible failure time (upper bound reaching EOL) and the latest possible failure time (lower bound reaching EOL) are calculated using the upper and lower bounds of the predicted trajectory's confidence interval, thereby deriving a predicted RUL interval. This approach significantly improves the robustness and reliability of RUL assessments, allowing users to plan battery replacement or maintenance within a more flexible timeframe, thereby avoiding the safety and economic losses caused by unexpected failures. Through these predictive indicators, users and electric vehicle manufacturers can more accurately optimize vehicle usage, extend the service life of power batteries, reduce downtime and maintenance costs, and reduce hidden risks of safety issues.

[0050] Figure 1 Flowchart for battery capacity attenuation trajectory prediction;

[0051] Figure 2 Schematic diagram of the process for building and screening multimodal features for batteries;

[0052] Figure 3 Schematic diagram of modeling the denoised diffusion model for battery capacity estimation. DETAILED DESCRIPTION

[0053] In order to make the purpose and technical solution of the present invention clearer, the following is a detailed description with reference to specific examples. It should be understood that the specific implementation examples described herein are only used to explain the present invention, but are not limited to the present invention.

[0054] like Figure 1 As shown, the present invention mainly includes 4 steps.

[0055] A. Collection and preprocessing of charge and discharge data of actual vehicle power batteries;

[0056] B. Based on the preprocessed data, extract point features, mechanism features, and graph features, and construct multimodal features;

[0057] C. Build a denoising diffusion model with ContextUnet as the core, aiming to achieve capacity decay trajectory prediction;

[0058] D. Based on the label values and multimodal feature embedding values, the DDPM-improved DDIM denoising strategy is used to predict the diffusion capacity decay trajectory. Finally, the confidence interval of the capacity decay trajectory prediction is constructed through multiple sampling.

[0059] A1. The vehicle's BMS system collects battery operation data, including time, voltage, current, SOC (battery state of charge), maximum and minimum voltage of a single cell, maximum and minimum temperature of a single cell, etc., records the data and uploads it to the cloud.

[0060] A2. Clean the collected data and use cleaning strategies such as outlier removal, missing value filling (K-nearest neighbor interpolation), and Gaussian filtering to smooth the data to ensure data stability and reliability.

[0061] A3. Normalize the data and unify the time scale so that each feature data remains in the same time and space range, thereby improving the accuracy of subsequent feature extraction and modeling.

[0062] A4. Finally, calculate the capacity label value based on the cleaned data and use the variable ampere integral method to calculate the battery capacity of each charge and discharge cycle. The specific formula is as follows: in, is the sampling interval, is the charging current, The battery charge state is set to a fixed range (such as 30%-80%) to avoid extreme The capacity evaluation is affected by the error value. Then, the capacity values calculated for all charge and discharge cycles within a certain period of time (such as a week) are statistically processed (using the mean or median) to obtain the calibrated capacity for that period. This method can effectively reduce the impact of errors and obtain more stable capacity data as labels.

[0063] B1、 Figure 2 As shown in the figure, statistical features and mechanism features are first extracted from the charge and discharge data. The statistical features include the mileage growth of the single battery, cycle time, and the extreme values of the single battery voltage and temperature.

[0064] B2. To improve model efficiency and reduce redundant features, we screened several extracted statistical features and used XGBoost to evaluate feature importance, eliminating features that contributed less to the capacity decay trajectory. During XGBoost training, the importance of each feature can be calculated using the gain, which represents the contribution of the feature to the optimization of the loss function during the decision-making process. The calculation formula is as follows: in, It is a feature Importance score; is the total number of all decision trees; and It is a decision tree The accumulated values of gradient and second-order derivative in ; Is a regularization term to prevent overfitting. XGBoost calculates the importance score of each feature , set the feature importance threshold , features below the threshold will be eliminated, so we can effectively remove redundant information, retain the key statistical features that are most valuable for capacity decay prediction, and improve the robustness and generalization ability of the model.

[0065] B3. Mechanism characteristics are key variables extracted based on the physical and chemical mechanisms inside the battery, which explain the root causes of capacity decay. These characteristics are usually closely related to the internal reactions, material decay, and kinetic characteristics of the battery. Compared with statistical characteristics, mechanism characteristics can more directly explain the health status of the battery and its aging mechanism, and have stronger interpretability. For example, by extracting features from the voltage and current curves during the battery charging and discharging process within a week or a month, the changes in internal materials can be analyzed to reveal the specific trajectory of capacity decay. For battery packs, the battery mechanism characteristics we extract include the IC sequence and internal resistance sequence of the battery pack.

[0066] B4. The extracted mechanism sequence characteristics of the battery pack above can characterize the physical and chemical changes inside the battery and have good interpretability. However, battery aging is a process of interaction of multi-dimensional factors. The sequence characteristics of the battery pack can only describe the data trend of a single dimension, and it is difficult to fully capture the interaction between different mechanism characteristics. Moreover, the characteristics extracted at the battery pack level are the result of the combined effect of multiple battery cells inside it. It is difficult to measure the specific aging of these different cells, and the abnormal degradation of individual cells may be masked. In contrast, the mechanism characteristics of battery cells can provide more detailed information, which can not only reflect the changes in the state of the cells themselves, but also be used to analyze the consistency and differences between different cells in the battery pack. Therefore, we further perform 100 uniform samplings on a weekly or monthly basis and extract the mechanism characteristics of each battery cell during the charging process by time alignment and interpolation filling, and construct an n×m two-dimensional matrix as follows:

[0067] (1) Cell Voltage (Vc): {V}_{c}=\left [ {{V}_{11},{V}_{12}\cdot \cdot \cdot ,{V}_{1m},{V}_{21},{V}_{22},\cdot \cdot \cdot {V}_{2m},\cdot \cdot \cdot ,{V}_{n1},{V}_{n2},\cdot \cdot \cdot {V}_{nm}{}^{}} \right ] in Indicates the Unit No. The target voltage range is set to 3.9V-4.0V, and 100 samples are uniformly sampled in this voltage range. Time alignment and interpolation filling are used to avoid data loss. Therefore, different charging conditions and different charging start conditions under actual vehicle working conditions are achieved. The battery voltage data are controlled to be of similar length to facilitate subsequent analysis.

[0068] (2) Cell Temperature (Tc): {T}_{c}=\left [ {{T}_{11},{T}_{12}\cdot \cdot \cdot ,{T}_{1m},{T}_{21},{T}_{22},\cdot \cdot \cdot {T}_{2m},\cdot \cdot \cdot ,{T}_{n1},{T}_{n2},\cdot \cdot \cdot {T}_{nm}{}^{}} \right ] in Indicates the Unit No. The target temperature range is set to 25°C-45°C, and 100 samples are evenly sampled within this temperature range. To ensure sufficient valid data at the evenly sampled points, the temperature data is also processed using time alignment and interpolation.

[0069] B5. and To further integrate voltage and temperature information based on the matrix, a cell feature map (CFM) was constructed. The CFM is constructed by concatenating the cell voltage and temperature matrices to generate a two-channel, two-dimensional matrix. The matrix dimensions are kept consistent, ensuring that voltage and temperature information can be processed in parallel. This fusion approach preserves the characteristics of battery cells at different time steps, captures the correlation between voltage and temperature, and enhances the ability to characterize cell capacity changes, making it suitable for subsequent analysis using 2D convolutional neural networks (2D CNNs).

[0070] B6. After the above feature extraction, in order to fully explore the different information in the battery health status, we use CNN (convolutional neural network) to process the single feature map, Transformer to process the sequence features, and use the fully connected layer to fuse the statistical features (point features) in the final stage to form a complete prediction framework. For the single feature map, we first use several two-dimensional convolution kernels to extract the local spatial relationship features and output the feature map; then perform pooling on the feature map to obtain a reduced-size feature map; after multiple layers of stacked convolution and pooling operations, we flatten the final feature map and convert it into a one-dimensional vector As a high-dimensional embedding representation of the two-dimensional graph features of the battery cell, It fully captures the complex correlation pattern between battery voltage and temperature, and will serve as the input of the subsequent battery capacity attenuation evaluation and prediction module; for time series features, we use Transformer to process them. Transformer uses a multi-head self-attention mechanism to calculate the spatiotemporal dependencies between different features, while introducing position encoding to retain time series information. This fusion method not only retains the macroscopic laws of statistical features, but also combines the microscopic characteristics of mechanism features, thereby significantly improving the model's ability to characterize the battery degradation process, laying a solid foundation for the subsequent prediction of battery capacity attenuation trajectory, and finally high-dimensional embedding representation Finally, in the face of statistical features without temporal features, we directly input the independent final fully connected layer (FC) for fusion, where we integrate the graph features extracted by CNN , Sequence features extracted by Transformer And point features , forming the final comprehensive feature representation , Z=FC([{Z}_{G},{Z}_{S},{X}_{P}]) . This multimodal feature fusion method not only preserves the macroscopic laws of statistical features but also combines the physical interpretability of mechanism features with the high-dimensional spatial expression capability of graph features as the model prediction input for battery capacity decay trajectories. This combination of features enables the model to take into account both the overall picture and the details, providing comprehensive data support for subsequent accurate prediction of battery capacity decay trajectories.

[0071] C1. Before using the denoising diffusion model (DDPM) to predict capacity decay trajectories, you need to initialize the diffusion hyperparameters to define the model's noise diffusion process, including the number of noise diffusion steps and the noise variation range. These parameters control the noise evolution during training.

[0072] C2. We use ContextUnet (Multimodal Feature Modulator) as the primary model architecture for DDPM. Based on the classic U-Net network structure, it combines temporal and feature embedding mechanisms, enabling it to better perceive the rich information of battery changes and model the complex evolution of battery health status. The various components of the model architecture function as follows: the residual convolution module performs feature extraction, improving feature representation and preventing gradient vanishing; the downsampling module reduces data dimensionality through convolution to obtain high-level features; the upsampling module restores data dimensions through deconvolution and strengthens feature representation; the positional encoder encodes the relative position of the input data in the sequence, enabling the model to identify temporal relationships between features when processing time-dependent data, enhancing its ability to model data with long temporal dependencies; the temporal encoder ensures the effective utilization of temporal information during the diffusion process, introducing independent representations at different time steps. This allows the model to accurately track data evolution during denoising and improves dynamic modeling capabilities; the multi-layer perceptron (MLP) embeds contextual information (features) into the model, and finally, the fully connected layer generates the final prediction output, ensuring accurate predictions after multimodal feature fusion.

[0073] D1. Before training, the standardized features and labels are divided into training sets, validation sets, and test sets, and loaded in batches through DataLoader.

[0074] D2. During training, since the noise diffusion parameters have been initialized before, noise is gradually added to the loaded training data. The core formula of noise diffusion is: in, is the original data, i.e., the extracted battery features; is the data after adding noise; is the noise that conforms to the standard normal distribution; is the noise scheduling factor, which determines the How much noise to add. It decreases as the time step increases, making the final data become pure noise.

[0075] D3, then denoising is performed. The denoising process is the core of the Diffsion Model, and the goal is to restore the original data from the noisy data. Noisy data , we need to restore the original data by denoising The key here is noise prediction, which is to recover the noise from the noisy input through the encoder-decoder learning of the ContextUnet model. The process is as follows: in is in the time step Noisy data (i.e., the battery at The noise-contaminated data of the first step is the input of the model; is the noise predicted by the ContextUnet model; is the denoised data, indicating that at time step At the same time, the loss function Mean Square Error (MSE) is set to minimize the noise prediction error, and the Adam optimizer adaptively adjusts the learning rate of the parameters to help the model converge stably. The calculation formulas of the loss function and optimizer are as follows: in, is the actual noise added; in, is the learning rate, are the current model parameters, is the gradient momentum, weighted average of squared gradients, is a small constant that prevents division by zero and predicts the added noise. Through the noise diffusion and denoising training process, the model can restore battery data under complex working conditions to characteristics that conform to the battery degradation law. In addition, to improve the denoising efficiency and accelerate the sampling process, we introduced DDIM. Compared with the traditional DDPM denoising method mentioned above, DDIM adopts an implicit denoising strategy, which allows the noise to be retained or some noise to be added to the data during the denoising process, avoiding the over-denoising phenomenon that may occur in DDPM, maintaining data diversity and improving generation quality. The implicit denoising strategy formula is as follows: in, is the recovered original data (predicted value); and is the noise scheduling factor, is the noise predicted by the ContextUnet model; compared with the above DDPM, DDIM introduces a front time step The noise scheduling factor , in addition, an additional term related to the noise of the previous step is added This allows DDIM to retain more noise information during the denoising process, allowing for more flexible denoising operations while reducing the amount of computation required at each time step. Through this approach, we can accurately recover the true battery capacity from noisy data, reducing the impact of noise on the estimation results and laying the foundation for accurately predicting battery capacity decay trajectories.

[0076] D4. In order to further improve the stability of capacity decay trajectory prediction, after using DDIM to generate battery capacity decay trajectory samples, the present invention introduces a capacity decay trajectory confidence interval estimation mechanism based on multiple sample statistics. For each set of input features, we repeat the DDIM denoising process based on different random noise initial conditions to generate a total of 20 sets of battery capacity decay prediction trajectory samples. Subsequently, the capacity values of these 20 groups of prediction samples at each time step are statistically processed, and their mean, standard deviation and variance are calculated to capture the prediction fluctuations due to randomness. Afterwards, we adopt the normal distribution assumption and use the 95% confidence interval calculation formula: in, is the mean of 20 prediction values, is the standard deviation of the corresponding time step, and 1.96 is the critical value of the 95% confidence interval under the standard normal distribution. This method can effectively quantify the uncertainty of the prediction results, thereby constructing the upper and lower confidence intervals of the battery capacity decay trajectory, providing solid data support for subsequent battery state of health (SOH) and remaining useful life (RUL) assessments.

[0077] The capacity decay trajectory predicted by the combined diffusion model of D5, DDPM, and DDIM not only provides detailed trends in battery capacity decay but also lays the foundation for further deriving key indicators such as the battery's state of health (SOH) and remaining useful life (RUL). The SOH can be calculated as the ratio of the predicted mean capacity to the initial capacity, while the confidence interval provides estimates of the upper and lower limits of the SOH. This extends the SOH from a single value to a dynamic health indicator in the form of a credible interval. As the battery capacity decays, the SOH value can be used to monitor battery health in real time, providing a basis for maintenance decisions. Traditional methods typically use the time point when the capacity trajectory first drops to the EOL threshold (e.g., 80%) as the endpoint for remaining useful life (RUL). However, in this invention, the earliest possible failure time (upper bound reaching EOL) and the latest possible failure time (lower bound reaching EOL) are calculated using the upper and lower bounds of the predicted trajectory's confidence interval, thereby deriving a predicted RUL interval. This approach significantly improves the robustness and reliability of RUL assessments, allowing users to plan battery replacement or maintenance within a more flexible timeframe, thereby avoiding the safety and economic losses caused by unexpected failures. Through these predictive indicators, users and electric vehicle manufacturers can more accurately optimize vehicle usage, extend the service life of power batteries, reduce downtime and maintenance costs, and reduce hidden risks of safety issues.

Claims

1. A method for predicting the capacity decay trajectory of a real vehicle power battery based on a denoising diffusion model, comprising the following steps: A. Vehicle charging and discharging data collection, preprocessing and label acquisition; B. Extract and filter multimodal features based on the data processed in step A and embed them into ContextUnet in C as context information. C. Build a capacity decay trajectory prediction model based on DDPM (denoising diffusion model) with ContextUnet as the core; D. Based on the label value and the multimodal feature embedding value, the DDIM denoising strategy improved by DDPM is used to predict the diffusion capacity decay trajectory. Finally, the confidence interval of the capacity decay trajectory prediction is constructed through multiple sampling.

2. The specific steps for vehicle charging and discharging data collection, preprocessing and label acquisition in step A are as follows: A1. Use the onboard BMS system to regularly collect and upload electric vehicle charging and discharging data, including time, total voltage, current, SOC (battery state of charge), single cell maximum voltage, single cell minimum voltage, single cell maximum temperature, single cell minimum temperature, etc. The collection interval is weekly. All of this data will be recorded and stored for subsequent analysis and processing; A2. Data cleaning is performed on the collected vehicle charging and discharging data. First, a threshold range is used to filter out abnormal data, such as unreasonable and sharp fluctuations in voltage, current, or temperature. Then, missing values are filled using the K-nearest neighbor interpolation method to ensure the integrity of the time series data. Finally, a Gaussian filter is used to smooth the noisy data to ensure data stability and reliability. A3. Normalize the cleaned charge and discharge data using the median, align the time, and ensure that the feature data is in the same temporal and spatial range through interpolation and padding. A4. Based on the actual vehicle charge and discharge data processed above, the battery capacity of each charge and discharge cycle is calculated using the following variable ampere integral formula: in, is the sampling interval, is the charging current, The battery charge state is set to a fixed range (such as 30%-80%) to avoid extreme The value affects the capacity evaluation, ensuring that the battery is calculated under a relatively stable working state, and improving the accuracy and reliability of capacity evaluation. In real-world vehicle operating conditions, estimation errors and data noise can cause significant fluctuations in the capacity value calculated from a single calculation. Therefore, we perform statistical processing (using the mean or median) on the capacity values calculated from all charge and discharge cycles over a certain period of time (such as a week or a month) to obtain the calibrated capacity for that period. This approach effectively reduces the impact of errors and produces more stable capacity data as labels.

3. In step B, multiple key multimodal features are extracted based on the pre-processed vehicle charge and discharge raw data. These features are used to train the diffusion model to predict the battery capacity decay trajectory. The specific steps are: B1. We extracted statistical features calculated based on weekly charge and discharge cycle data. Based on this, we extracted mechanistic features of battery packs and individual cells. We constructed graph features based on the mechanistic features of individual cells. By integrating these three different types of features, we constructed a multimodal feature set as input for the model to enhance its predictive capabilities. B2. Statistical features are key indicators extracted from battery data to describe data distribution and variation patterns. Since battery capacity decay is a slow, long-term process, performing statistical calculations on data on a weekly or monthly basis (such as calculating mean, variance, skewness, etc.) can capture the temporal characteristics of battery data, reduce noise interference on the model, and reveal capacity decay patterns. The extracted statistical features are as follows: (1) Mileage Increase (MI): This refers to the cumulative mileage of an electric vehicle during driving. As the mileage increases, the battery undergoes more charge and discharge cycles, its capacity gradually decays, its internal impedance increases, and its performance gradually deteriorates. (2) Average Maximum Cell Temperature (AMCT): refers to the highest average temperature reached by a battery cell during operation. High temperature will accelerate battery capacity degradation and increase safety risks; (3) Average State of Charge (ASOC): This indicates the average state of charge of the battery during use, reflecting the battery's charge and discharge trajectory. A long period of high or low SOC will accelerate the battery capacity decay; (4) Average Pack Voltage (APV): refers to the average voltage of the entire battery pack, which can reflect the health status, charge and discharge conditions, and balance of the battery pack. Too high or too low average voltage will affect the battery capacity retention rate; (5) Average Cell Voltage Difference (ACVD): describes the voltage difference between battery cells. A larger voltage difference indicates an increased imbalance in the battery pack, causing some cells to be overcharged or over-discharged, exacerbating capacity degradation. (6) Impedance Increase (II): This indicates that the internal impedance of the battery increases with time and use. Impedance increase usually indicates battery aging, resulting in power loss and increased heat generation. (7) Cycling Time (CCT): Indicates the number of complete charge and discharge cycles that a battery has undergone. The service life of a battery is usually related to the number of cycles. The more cycles, the more obvious the battery capacity decay. (8) Calendar Time (CAT): refers to the natural time from the production of the battery to the present. Even if the battery is not used, its performance will still decline due to calendar aging; (9) Charging Accumulated Ah (CAAh): Indicates the amount of electricity (ampere-hours) charged to the battery during the entire use process. It is used to evaluate the impact of the charging trajectory on the battery life. A high cumulative charging ampere-hour usually means that the battery has experienced more cycles, resulting in capacity decay. (10) Discharging Accumulated Ah (DAAh): refers to the cumulative amount of electricity discharged by the battery (ampere-hours), reflecting the intensity of battery use and its impact on battery life; (11) Regenerated Accumulated Ah (RAAh): represents the amount of energy recovered by the battery through regenerative braking and other methods. Frequent energy recovery processes may cause the battery to be in a high instantaneous charging current state, increase polarization, affect the uniform insertion of lithium ions, and thus accelerate battery aging; (12) Charging Accumulated Energy (CAE): Indicates the total energy charged into the battery, usually calculated in kilowatt-hours (kWh), which helps analyze charging efficiency and decay trends; (13) Discharging Accumulated Energy (DAE): refers to the total energy released by the battery (kWh), reflecting the overall energy output of the battery. High-rate output during discharge leads to increased polarization, lithium deposition, or local temperature increase, thereby accelerating capacity decay. (14) Regenerated Accumulated Energy (RAE): refers to the total energy recovered by the battery through regenerative braking and other methods. High-frequency energy recovery will cause transient high-current charging, resulting in uneven deposition of lithium ions, increasing the risk of lithium plating and affecting battery life. B3. To improve model efficiency and reduce redundant features, we screened several extracted statistical features and used XGBoost to evaluate feature importance, eliminating features that contributed less to the capacity decay trajectory. During XGBoost training, the importance of each feature can be calculated using the gain, which represents the contribution of the feature to the optimization of the loss function during the decision-making process. The calculation formula is as follows: in, It is a feature Importance score; is the total number of all decision trees; and It is a decision tree The accumulated values of gradient and second-order derivative in ; Is a regularization term to prevent overfitting. XGBoost calculates the importance score of each feature , set the feature importance threshold , features below the threshold will be eliminated, so we can effectively remove redundant information, retain the key statistical features that are most valuable for capacity decay prediction, and improve the robustness and generalization ability of the model; B4. Mechanism characteristics are key variables extracted based on the physical and chemical mechanisms inside the battery, explaining the root causes of capacity decay. These characteristics are usually closely related to the internal reactions, material decay, and kinetic characteristics of the battery. Compared with statistical characteristics, mechanism characteristics can more directly explain the health status of the battery and its aging mechanism, and have stronger interpretability. For example, by extracting features from the voltage and current curves during the battery charging and discharging process within a week or a month, the changes in internal materials can be analyzed to reveal the specific trajectory of capacity decay. For battery packs, the battery mechanism characteristics we extracted are as follows; (1) IC Sequence: IC (Incremental Capacity) analysis is a method based on the dQ / dV-V curve. For the curve obtained by multiple cycles within a week or a month, the average dQ / dV value corresponding to each voltage horizontal coordinate point is calculated to reduce data fluctuations and the impact of data loss. By analyzing the peak position and height changes of the IC curve, the battery capacity attenuation can be obtained. As the battery ages, the peak of the IC curve will drift and the peak height will gradually decrease, reflecting the attenuation of the battery capacity; (2) Resistance and Capacitance Model Parameter Sequence: The parameter sequence of the battery internal resistance and capacitance model (such as ohmic internal resistance Rint, polarization resistance Rp, and polarization capacitance Cp) is defined based on the Thevenin equivalent circuit model and is obtained by calculating the voltage and current data during the discharge process through the recursive least squares method (RLS). To reduce noise interference, the parameter values corresponding to each time point within a week or a month can be averaged or median. The internal resistance and capacitance sequence is highly sensitive to battery aging and can effectively capture the overall aging trend of the battery system according to the law of time and number of cycles, reflecting the changing law of the capacity attenuation trajectory; B5. The extracted mechanism sequence characteristics of the battery pack above can characterize the physical and chemical changes inside the battery and have good interpretability. However, battery aging is a process of interaction of multi-dimensional factors. The sequence characteristics of the battery pack can only describe the data trend of a single dimension, and it is difficult to fully capture the interaction between different mechanism characteristics. Moreover, the characteristics extracted at the battery pack level are the result of the combined effect of multiple battery cells inside it. It is difficult to measure the specific aging of these different cells, and the abnormal degradation of individual cells may be masked. In contrast, the mechanism characteristics of battery cells can provide more detailed information, which can not only reflect the changes in the state of the cells themselves, but also can be used to analyze the consistency and differences between different cells in the battery pack. Therefore, we further uniformly sampled 100 times a week or a month and extracted the mechanism characteristics of each battery cell during the charging process by time alignment and interpolation filling, and constructed an n×m two-dimensional matrix as follows: (1) Cell Voltage (Vc): in Indicates the Unit No. The target voltage range is set to 3.9V-4.0V, and 100 samples are uniformly sampled in this voltage range. Time alignment and interpolation filling are used to avoid data loss. Therefore, different charging conditions and different charging start conditions under actual vehicle working conditions are achieved. The battery voltage data are controlled to be of similar length to facilitate subsequent analysis; (2) Cell Temperature (Tc): in Indicates the Unit No. The target temperature range is set to 25℃-45℃, and 100 samples are evenly sampled within this temperature range. To ensure sufficient valid data at the evenly sampled points, the temperature data is also processed using time alignment and interpolation. B6. and To further integrate voltage and temperature information based on the matrix, a cell feature map (CFM) was constructed. The CFM is constructed by concatenating the cell voltage and temperature matrices to generate a two-channel, two-dimensional matrix. The matrix dimensions are kept consistent, ensuring that voltage and temperature information can be processed in parallel. This fusion approach preserves the characteristics of battery cells at different time steps, captures the correlation between voltage and temperature, and enhances the ability to characterize cell capacity changes, making it suitable for subsequent analysis using a 2D convolutional neural network (2D CNN). B7. After the above feature extraction, in order to fully explore the different information in the battery health status, we use CNN (convolutional neural network) to process the single feature map, Transformer to process the sequence features, and use the fully connected layer to fuse the statistical features (point features) in the final stage to form a complete prediction framework. For the single feature map, we first use several two-dimensional convolution kernels to extract the local spatial relationship features and output the feature map; then perform pooling on the feature map to obtain a reduced-size feature map; after multiple layers of stacked convolution and pooling operations, we flatten the final feature map and convert it into a one-dimensional vector As a high-dimensional embedding representation of the two-dimensional graph features of the battery cell, It fully captures the complex correlation pattern between battery voltage and temperature, and will serve as the input of the subsequent battery capacity attenuation evaluation and prediction module; for time series features, we use Transformer to process them. Transformer uses a multi-head self-attention mechanism to calculate the spatiotemporal dependencies between different features, while introducing position encoding to retain time series information. This fusion method not only retains the macroscopic laws of statistical features, but also combines the microscopic characteristics of mechanism features, thereby significantly improving the model's ability to characterize the battery degradation process, laying a solid foundation for the subsequent prediction of battery capacity attenuation trajectory, and finally high-dimensional embedding representation Finally, in the face of statistical features without temporal features, we directly input the independent final fully connected layer (FC) for fusion, where we integrate the graph features extracted by CNN , Sequence features extracted by Transformer And point features , forming the final comprehensive feature representation , . This multimodal feature fusion method not only preserves the macroscopic laws of statistical features but also combines the physical interpretability of mechanism features with the high-dimensional spatial expression capability of graph features as the model prediction input for battery capacity decay trajectories. This combination of features enables the model to take into account both the overall picture and the details, providing comprehensive data support for subsequent accurate prediction of battery capacity decay trajectories.

4. In step C, after label value acquisition and multimodal feature organization, the input of the capacity decay trajectory prediction model has been prepared. Then, the capacity decay trajectory prediction model is built around the ContextUnet core. The specific steps are as follows: C1. Before using the DDPM (denoising diffusion model) to predict capacity decay trajectories, you need to initialize the diffusion hyperparameters to define the model's noise diffusion process, including the number of noise diffusion steps and the noise variation range. These parameters control the noise evolution during training. C2. We use ContextUnet (a multimodal feature modulator) as the primary model architecture for the Diffusion Model. Based on the classic U-Net network structure, it combines temporal and feature embedding mechanisms, enabling it to better perceive the rich information of battery changes and model the complex evolution of battery health status. The various components of the model architecture function as follows: the residual convolution module performs feature extraction, improving feature representation and preventing gradient vanishing; the downsampling module reduces data dimensionality through convolution to obtain high-level features; the upsampling module restores data dimensions through deconvolution and strengthens feature representation; the position encoder encodes the relative position of the input data in the sequence, enabling the model to identify temporal relationships between features when processing time-dependent data, enhancing its ability to model data with long temporal dependencies; the temporal encoder ensures the effective utilization of temporal information in the diffusion process, introducing independent representations at different time steps. This allows the model to accurately track data evolution during denoising and improves its dynamic modeling capabilities; the multi-layer perceptron (MLP) embeds contextual information (features) into the model, and finally, the fully connected layer generates the final prediction output, ensuring accurate predictions after multimodal feature fusion.

5. In step D, the established capacity decay trajectory prediction model is used to embed multimodal features as context information to predict noise, thereby restoring the capacity decay trajectory predicted by the data output. The specific steps are as follows: D1. Before training, the standardized features and labels are divided into training sets, validation sets, and test sets, and batch loaded through DataLoader; During training, since the noise diffusion parameters have been initialized before, noise is added to the loaded training data step by step. The core formula of noise diffusion is: in, is the original data, i.e., the extracted battery features; is the data after adding noise; is the noise that conforms to the standard normal distribution; is the noise scheduling factor, which determines the How much noise to add. It decreases as the time step increases, making the final data become pure noise; D3, then denoising is performed. The denoising process is the core of the Diffsion Model, and the goal is to restore the original data from the noisy data. Noisy data , we need to restore the original data by denoising The key here is noise prediction, which is to recover the noise from the noisy input through the encoder-decoder learning of the ContextUnet model. The process is as follows: in is in the time step Noisy data (i.e., the battery at The noise-contaminated data of the first step is the input of the model; is the noise predicted by the ContextUnet model; is the denoised data, indicating that at time step At the same time, the loss function Mean Square Error (MSE) is set to minimize the noise prediction error, and the Adam optimizer adaptively adjusts the learning rate of the parameters to help the model converge stably. The calculation formulas of the loss function and optimizer are as follows: in, is the actual noise added; in, is the learning rate, are the current model parameters, is the gradient momentum, weighted average of squared gradients, is a small constant that prevents division by zero and predicts the added noise. Through the noise diffusion and denoising training process, the model can restore battery data under complex working conditions to characteristics that conform to the battery degradation law. In addition, to improve the denoising efficiency and accelerate the sampling process, we introduced DDIM. Compared with the traditional DDPM denoising method mentioned above, DDIM adopts an implicit denoising strategy, which allows the noise to be retained or some noise to be added to the data during the denoising process, avoiding the over-denoising phenomenon that may occur in DDPM, maintaining data diversity and improving generation quality. The implicit denoising strategy formula is as follows: in, is the recovered original data (predicted value); and is the noise scheduling factor, is the noise predicted by the ContextUnet model; compared with the above DDPM, DDIM introduces a front time step The noise scheduling factor , in addition, an additional term related to the noise of the previous step is added This allows DDIM to retain more noise information during the denoising process, allowing for more flexible denoising operations while reducing the amount of computation required at each time step. Through this approach, we can accurately recover the true battery capacity from noisy data, reducing the impact of noise on the estimation results and laying the foundation for accurately predicting battery capacity decay trajectories. D4. In order to further improve the stability of capacity decay trajectory prediction, after using DDIM to generate battery capacity decay trajectory samples, the present invention introduces a capacity decay trajectory confidence interval estimation mechanism based on multiple sample statistics. For each set of input features, we repeat the DDIM denoising process based on different random noise initial conditions to generate a total of 20 sets of battery capacity decay prediction trajectory samples. Subsequently, the capacity values of these 20 groups of prediction samples at each time step are statistically processed, and their mean, standard deviation and variance are calculated to capture the prediction fluctuations due to randomness. Afterwards, we adopt the normal distribution assumption and use the 95% confidence interval calculation formula: in, is the mean of 20 prediction values, is the standard deviation of the corresponding time step, and 1.96 is the critical value of the 95% confidence interval under the standard normal distribution. This method can effectively quantify the uncertainty of the prediction results, thereby constructing the upper and lower confidence intervals of the battery capacity decay trajectory, providing solid data support for subsequent battery health state (SOH) and remaining life (RUL) assessments; The capacity decay trajectory predicted by the combined diffusion model of D5, DDPM, and DDIM not only provides detailed trends in battery capacity decay but also lays the foundation for further deriving key indicators such as the battery's state of health (SOH) and remaining useful life (RUL). The SOH can be calculated as the ratio of the predicted mean capacity to the initial capacity, while the confidence interval provides estimates of the upper and lower limits of the SOH. This extends the SOH from a single value to a dynamic health indicator in the form of a credible interval. As the battery capacity decays, the SOH value can be used to monitor battery health in real time, providing a basis for maintenance decisions. Traditional methods typically use the time point when the capacity trajectory first drops to the EOL threshold (e.g., 80%) as the endpoint for remaining useful life (RUL). However, in this invention, the earliest possible failure time (upper bound reaching EOL) and the latest possible failure time (lower bound reaching EOL) are calculated using the upper and lower bounds of the predicted trajectory's confidence interval, thereby deriving a predicted RUL interval. This approach significantly improves the robustness and reliability of RUL assessments, allowing users to plan battery replacement or maintenance within a more flexible timeframe, thereby avoiding the safety and economic losses caused by unexpected failures. Through these predictive indicators, users and electric vehicle manufacturers can more accurately optimize vehicle usage, extend the service life of power batteries, reduce downtime and maintenance costs, and reduce hidden risks of safety issues.

Images