A capacity recovery-aware power battery degradation prediction method and system for vehicle

Abstract

The present application relates to a kind of capacity recovery perceived vehicle power battery degradation prediction method and system, belong to vehicle power battery health state evaluation and life prediction technical field, solve capacity recovery and long-term degradation trend coupling, the problem such as insufficient degradation stage perception ability.First, the cycling operation data of vehicle power battery of source domain and target domain are collected and preprocessed and sample division are carried out;Capacity recovery perceived vehicle power battery degradation prediction model is constructed;Multi-loss joint pre-training is carried out, and source domain pre-training model is obtained;The source domain pre-training model is transferred fine-tuning, and target domain vehicle power battery capacity prediction model is obtained;After deployment, it is carried out online prediction by Monte Carlo Dropout uncertainty perception inference;Finally, grading capacity degradation early warning is carried out.The present application can effectively decouple long-term degradation trend and short-term capacity recovery, enhance degradation stage perception and inflection point identification ability, improve the prediction accuracy and reliability in the cross-batch, small sample scene.

Description

Technical Field

[0001] This invention relates to the field of vehicle power battery health status assessment and life prediction technology, specifically to a method and system for predicting vehicle power battery degradation based on capacity recovery sensing. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the health of vehicle power batteries, as the core energy storage component of vehicles, directly affects the vehicle's range, power performance, safety, and operation and maintenance costs. During long-term cyclic charging and discharging, battery capacity gradually decreases. When the capacity drops below the preset failure threshold, it may lead to insufficient range, reduced power output, or abnormal overheating, thereby causing safety risks and economic losses.

[0003] The capacity degradation of automotive power batteries is affected by various factors such as charge / discharge rate, ambient temperature, operating conditions, battery consistency, and manufacturing batch. The degradation process exhibits nonlinearity, stages, and significant individual differences. In actual operation, the capacity curve does not strictly decrease monotonically; short-term capacity rebounds or local fluctuations may occur. Although these capacity rebounds are small, they are often related to the recovery of internal polarization, reversible capacity release, or changes in operating conditions. If these are not distinguished during the prediction process, it may lead to misjudging the true rebound as noise and losing important state information, or it may interfere with the accurate estimation of the long-term degradation trend.

[0004] Existing methods for predicting the capacity of automotive power batteries mainly include electrochemical mechanism models, equivalent circuit models, empirical degradation models, and data-driven models. Electrochemical mechanism models can describe the internal chemical reaction processes of the battery, but their parameters are complex, modeling costs are high, and they are difficult to adapt to online applications under multiple operating conditions. Equivalent circuit models have relatively simple structures, simulating battery performance through circuit elements, but their ability to characterize long-term capacity decay and complex aging mechanisms is limited. Empirical degradation models typically rely on pre-defined function forms and are insufficiently adaptable to nonlinear, phased, and significantly individual-varying capacity sequences. In recent years, data-driven models, especially those based on deep learning, have been widely used for capacity prediction and remaining lifetime estimation. These methods can automatically learn the temporal characteristics in cyclic data, improving prediction accuracy to some extent, but they still suffer from insufficient accuracy and stability in multi-operating condition, small sample, and cross-batch application scenarios.

[0005] While existing methods can predict the capacity degradation of automotive power batteries to some extent, they still have several limitations, most notably: Capacity recovery and long-term degradation trends are often coupled, and existing methods struggle to effectively separate short-term recovery from long-term degradation, potentially leading to prediction curves that do not conform to battery degradation patterns on a macroscopic scale; the prediction models lack sufficient perception of degradation stages, failing to accurately identify whether the battery is in an early, slow degradation phase, a stable degradation phase, or a rapid degradation phase, especially exhibiting insufficient fitting ability in the rapid degradation phase; furthermore, degradation inflection points are not adequately modeled, making it difficult for the model to identify the critical point where the battery transitions from stable to rapid degradation in advance, thus limiting risk warning capabilities. In addition, existing transfer learning strategies often struggle to balance the preservation of general knowledge from the source domain with adaptation to the target domain in cross-batch or small-sample target domain applications, easily leading to overfitting or prediction bias. Regarding prediction output, the lack of physical constraints and uncertainty metrics can cause abnormal increases or drastic fluctuations in capacity predictions, and the inability to provide confidence information further impacts the risk assessment and graded early warning capabilities of the battery management system's online operation.

[0006] In summary, existing methods for predicting the capacity of automotive power batteries still have shortcomings in areas such as capacity recovery processing, degradation stage perception, degradation inflection point modeling, cross-batch migration adaptation, and prediction uncertainty measurement. Therefore, there is an urgent need for a new capacity degradation prediction method and system that can achieve coordinated processing of capacity recovery decoupling, degradation stage perception, degradation inflection point assisted identification, migration fine-tuning, and uncertainty prediction under multi-condition, small-sample, and cross-batch application scenarios. This would improve prediction accuracy and provide reliable online health monitoring and graded early warning capabilities for battery management systems. Summary of the Invention

[0007] The main objective of this invention is to address the problems in existing technologies regarding the coupling of capacity recovery signals with long-term degradation trends in automotive power batteries, insufficient perception of degradation stages, lack of proactive identification of degradation inflection points, weak adaptability in cross-batch and small-sample scenarios, lack of uncertainty measurement in prediction results, and insufficient physical rationality of capacity prediction curves. This invention proposes a capacity recovery perception-based method and system for predicting degradation in automotive power batteries. This method effectively decouples long-term degradation trends from short-term capacity recovery, enhances the perception of degradation stages and inflection point identification capabilities, improves prediction accuracy and reliability in cross-batch and small-sample scenarios, and provides reliable hierarchical early warning information for battery management systems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for predicting the degradation of automotive power batteries based on capacity recovery sensing includes the following steps:

[0010] Step 1: Collect the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery, and perform preprocessing and sample division to obtain the source domain training sample set, the source domain validation sample set, the target domain fine-tuning sample set, and the target domain test sample set.

[0011] Step 2: Construct a vehicle power battery degradation prediction model with capacity recovery perception;

[0012] Step 3: Use the source domain training sample set to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model to obtain the source domain pre-trained model;

[0013] Step 4: Use the target domain fine-tuning sample set to perform transfer fine-tuning on the source domain pre-trained model, and introduce hierarchical differential learning rate configuration and maximum mean difference domain alignment loss to obtain the target domain vehicle power battery capacity prediction model.

[0014] Step 5: Deploy the target domain vehicle power battery capacity prediction model in an online application environment, and use the Monte Carlo Dropout uncertainty perception reasoning mechanism to perform online capacity degradation prediction for the vehicle power battery to be predicted, calculate the predicted capacity mean sequence, the predicted confidence interval sequence and the probability of the final degradation inflection point, and determine the estimated remaining service life.

[0015] Step 6: Establish tiered capacity degradation early warning judgment conditions. Based on the predicted mean capacity sequence, predicted confidence interval sequence, probability of the final degradation inflection point, and estimated remaining useful life, tiered capacity degradation early warning is performed. It includes at least one or more of the following: uncertainty warning, inflection point imminent warning, and end of useful life warning. The tiered capacity degradation early warning results are generated.

[0016] Accordingly, this invention also proposes a capacity recovery sensing system for predicting the degradation of automotive power batteries, comprising:

[0017] The data acquisition and preprocessing unit is used to acquire the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery, and to perform preprocessing and sample partitioning to obtain the source domain training sample set, the source domain verification sample set, the target domain fine-tuning sample set, and the target domain test sample set.

[0018] The model building unit is used to build a vehicle power battery degradation prediction model that senses capacity recovery.

[0019] The multi-loss joint pre-training unit is used to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model using the source domain training sample set to obtain the source domain pre-trained model.

[0020] The domain adaptive transfer fine-tuning unit is used to transfer fine-tun the source domain pre-trained model with the target domain fine-tuning sample set, and introduces hierarchical differential learning rate configuration and maximum mean differential domain alignment loss to obtain the target domain vehicle power battery capacity prediction model.

[0021] The uncertainty-aware online prediction unit is used to deploy the target domain vehicle power battery capacity prediction model in an online application environment. Through the Monte Carlo Dropout uncertainty-aware inference mechanism, it performs online capacity degradation prediction of the vehicle power battery to be predicted, calculates the predicted capacity mean sequence, the predicted confidence interval sequence and the probability of the final degradation inflection point, and determines the estimated remaining service life.

[0022] The graded capacity degradation early warning unit is used to establish graded capacity degradation early warning judgment conditions. It performs graded capacity degradation early warning based on the predicted mean capacity sequence, predicted confidence interval sequence, probability of the final degradation inflection point, and estimated remaining useful life. It includes at least one or more of the following: uncertainty early warning, inflection point imminent warning, and end of useful life warning, and generates graded capacity degradation early warning results.

[0023] This invention proposes a method and system for predicting the degradation of automotive power batteries based on capacity recovery sensing. By uniformly processing the cyclic operation data of automotive power batteries and combining capacity recovery sensing decoupled modeling, degradation degree conditional location encoding, degradation inflection point assisted identification, domain adaptive migration fine-tuning, and uncertainty-aware reasoning, it achieves comprehensive prediction of capacity degradation trends, remaining service life estimation, and degradation risk status. This improves the online health status assessment and graded early warning capabilities under multiple operating conditions, small sample sizes, and cross-batch scenarios. The beneficial effects of this invention include:

[0024] This invention improves the structural consistency and temporal continuity of capacity degradation data under different battery individuals, batches, and operating conditions by uniformly collecting, cleaning, regularizing, calculating degradation degree, and generating sliding window samples for the cyclic operation data of source domain vehicle power batteries and target domain vehicle power batteries.

[0025] This invention uses a capacity recovery sensing dual-branch variational autoencoder to decouple and reconstruct a subnetwork, structurally decoupling the long-term degradation trend component and the short-term capacity recovery component, thus avoiding the real capacity recovery being mistakenly treated as random noise smoothing or mistakenly transmitted as a long-term capacity upward trend to the subsequent prediction network.

[0026] This invention introduces the normalized degradation degree into the Transformer prediction backbone subnetwork through a degradation degree conditional location encoding module, enabling the model to simultaneously perceive the cycle time position and the actual degradation stage, thereby improving the capacity prediction adaptability under different degradation stages.

[0027] This invention uses an inflection point detection auxiliary subnetwork to enable the model to output the probability of degradation inflection points while outputting the future capacity prediction sequence, thereby improving the risk identification capability and early warning capability before the transition from the stable degradation stage to the rapid degradation stage.

[0028] This invention employs a multi-loss joint pre-training strategy to enable the model to simultaneously possess the capabilities of capacity sequence reconstruction, capacity recovery decoupling, long-term degradation trend prediction, physical constraints, and degradation inflection point identification, thereby avoiding functional fragmentation caused by isolated training of each module.

[0029] This invention employs a transfer fine-tuning strategy that combines hierarchical differentiated learning rate configuration with maximum mean difference domain alignment. This enables the source domain pre-trained model to adapt to the differences in target domain data distribution while retaining general capacity degradation knowledge, thereby reducing the risk of overfitting under small sample conditions in the target domain.

[0030] This invention uses the Monte Carlo Dropout uncertainty-aware reasoning mechanism to output the predicted mean capacity sequence, the predicted confidence interval sequence, the probability of the final degradation inflection point, and the estimated remaining lifespan, enabling the battery management system to obtain the predicted capacity value, prediction confidence, remaining lifespan information, and degradation risk information.

[0031] This invention transforms the predicted mean capacity sequence, predicted confidence interval sequence, probability of the final degradation inflection point, and estimated remaining service life into health management information that can be used by the battery management system through a graded capacity degradation early warning method, providing a basis for capacity degradation monitoring, maintenance plan formulation, and safety risk control. Attached Figure Description

[0032] Figure 1 This is a flowchart of a method for predicting the degradation of automotive power batteries based on capacity recovery sensing, as described in an embodiment of the present invention.

[0033] Figure 2 A schematic diagram of the structure of a vehicle power battery degradation prediction model for capacity recovery sensing;

[0034] Figure 3 This is a schematic diagram of the overall structure of a vehicle power battery degradation prediction system with capacity recovery sensing, as described in another embodiment of the present invention. Detailed Implementation

[0035] To make the technical problems, technical solutions, and beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] like Figure 1As shown, this embodiment provides a capacity recovery sensing method for predicting the degradation of automotive power batteries. Based on the cyclic operation data collection and preprocessing of source domain and target domain automotive power batteries, this method sequentially completes the following processes: construction of a capacity recovery sensing model for predicting automotive power battery degradation, multi-loss joint pre-training, domain adaptive transfer fine-tuning, uncertainty-aware online prediction, and graded capacity degradation early warning. It is applicable to the prediction of capacity decay trends, estimation of remaining capacity, estimation of remaining service life, and early warning of degradation risks for individual automotive power battery cells, modules, and battery packs under multiple operating conditions, small sample sizes, and cross-batch scenarios. The method specifically includes the following steps:

[0037] Step 1: Collect the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery, and perform preprocessing, including data cleaning, regularization, degradation degree calculation, sliding window sample generation, and degradation inflection point label generation, to obtain preprocessed samples. Then, divide the preprocessed samples to obtain the source domain training sample set, the source domain validation sample set, the target domain fine-tuning sample set, and the target domain test sample set.

[0038] Step 1 specifically includes the following steps:

[0039] Step 1.1: Collect the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery.

[0040] This step is used to obtain the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery, providing a data foundation for subsequent capacity recovery decoupling, degradation degree calculation, degradation inflection point identification and capacity degradation prediction.

[0041] The source domain automotive power battery refers to an automotive power battery with relatively complete cycle life data that can reflect general capacity degradation patterns. It is used to provide relatively complete capacity degradation samples to support the model in learning general degradation patterns. The target domain automotive power battery refers to an automotive power battery to be predicted, adapted, or monitored online in the target application scenario. It is used for migration adaptation, online prediction, and performance evaluation in the target application scenario. The source domain automotive power battery and the target domain automotive power battery can be any one of the following: automotive power battery cell, automotive power battery module, or automotive power battery pack. They can be battery objects at the same level or battery objects at different levels. When the two are at different levels, a unified input feature can be constructed through cell capacity characteristics, module statistical characteristics, battery pack operating condition characteristics, equivalent capacity characteristics, or a combination of the above features.

[0042] The cyclic operation data collected in this step includes the number of cycles, cycle start time, discharge capacity, charging capacity, charge / discharge voltage, charge / discharge current, battery surface temperature, resting time, and operating condition parameters related to the operating status of the vehicle power battery. Among them, capacity parameters are used to reflect the battery's capacity retention capability and degradation trend; voltage, current, temperature, resting time, and operating condition parameters are used to characterize operating conditions and help determine the relationship between local capacity recovery and changes in operating conditions.

[0043] As an optional implementation, when using publicly available single-cell battery cycle data for model validation, the single-cell data is used to characterize the capacity decay law of the basic degradation unit. When applied to automotive power battery modules or battery packs, single-cell capacity characteristics, module statistical characteristics, battery pack operating condition characteristics, or combinations of the above characteristics can be input into the subsequent automotive power battery degradation prediction model, thereby achieving adaptation from single-cell degradation laws to module and battery pack application scenarios. When a complete capacity sequence cannot be directly obtained for an automotive power battery module or battery pack, input features can be constructed based on single-cell voltage, current, temperature, SOC, SOH, capacity estimates, consistency indicators, and total voltage, total current, and equivalent capacity estimates of the module or battery pack collected by the battery management system. The input to the automotive power battery degradation prediction model is not limited to single-cell capacity sequences, but also includes equivalent capacity characteristics or health status characteristics that can reflect the capacity degradation state.

[0044] In this embodiment, the preprocessing step is not simply a smoothing of the capacity sequence. Instead, based on data cleaning, regularization, degradation degree calculation, and sliding window sample generation, it retains local capacity recovery points that are continuous, have reasonable amplitude, and are related to operating conditions. This allows the subsequent model to decouple the long-term degradation trend component and the short-term capacity recovery component. Therefore, it avoids mistakenly removing real capacity recovery as noise and prevents local capacity increases from interfering with the prediction of long-term capacity degradation trends.

[0045] Step 1.2: Clean the data from the loop.

[0046] Data cleaning is performed on the collected source domain cyclic running data and target domain cyclic running data to reduce the impact of missing records, duplicate records, abnormal time order, and obvious erroneous sampling points on subsequent model training.

[0047] Specifically, the cyclic operation data is sorted according to the number of cycles and the start time of the cycle. Missing cycle numbers, duplicate records, and abnormal time sequences are checked and processed. Obvious abnormal values ​​in capacity, voltage, current, temperature, and operating condition parameters are identified, and it is determined whether they are due to sampling errors, recording errors, or equipment malfunctions based on adjacent cycle data and operating conditions. Data points confirmed to be erroneous records can be removed, interpolated, or marked. Continuous missing data segments that cannot be reliably recovered are not included in the construction of subsequent sliding window samples. In this embodiment, a cycle refers to the vehicle power battery completing one effective charge-discharge process according to preset test conditions or actual operating conditions, or forming a data recording unit corresponding to an effective capacity record that can be used for capacity degradation analysis.

[0048] It should be noted that this step does not directly discard all local capacity increase points as outliers. Vehicle power batteries may experience short-term capacity rebounds under conditions of resting, polarization mitigation, reversible capacity release, or operating condition switching; these rebounds have practical implications. Capacity rebound points that are continuous, of reasonable magnitude, and related to operating conditions are retained in this step for subsequent decoupling modeling of long-term degradation trend components and short-term capacity rebound components in the capacity rebound sensing dual-branch variational autoencoder decoupling and reconstruction subnetwork.

[0049] Step 1.3: Perform regularization and calculate the degree of normalization degradation.

[0050] The cleaned cyclic operation data is regularized to unify the data scale for different individual batteries, batches, and operating conditions.

[0051] For capacity, voltage, current, temperature, resting time, and other auxiliary operating condition parameters, maximum-minimum normalization or standardization methods can be used to obtain the capacity sequence and the auxiliary operating condition characteristic sequence. The auxiliary operating condition characteristic sequence includes one or more of the following: voltage sequence, current sequence, temperature sequence, resting time sequence, and other auxiliary operating condition parameter sequences.

[0052] As a specific implementation method, any data sequence to be regularized can be processed using the minimum-max normalization method, the expression of which is:

[0053]

[0054] in, This represents the normalized data value; This represents the original data value after cleaning; This represents the minimum value in the corresponding data sequence. This represents the maximum value in the corresponding data sequence, which can be a capacity sequence, voltage sequence, current sequence, temperature sequence, resting time sequence, or other auxiliary operating condition parameter sequence.

[0055] After regularization, the normalization degradation degree is further calculated based on the capacity value.

[0056] As one specific implementation method, the first The normalized degradation degree corresponding to each cycle position It can be represented as:

[0057]

[0058] in, Indicates the initial capacity; Indicates the first The capacity corresponding to each cycle position; This indicates the end-of-life capacity threshold, which can be set according to battery type, application scenario, battery management system strategy, or test standards. Indicates the first The normalized degradation degree corresponding to each cyclic position. In this embodiment, the cyclic position refers to the index position of the cyclic running data in the capacity sequence or sliding window sample, which is used to characterize the relative position of the cyclic running data record in the time series.

[0059] The normalized degradation degree is used to characterize the actual degradation stage of the current cycle position and serves as the input to the subsequent degradation degree conditional position encoding module. This enables the model to perceive not only the cycle time position when making time-series predictions, but also different degradation stages such as early slow decay, stable degradation, or rapid degradation of the vehicle power battery.

[0060] After regularization and normalization degradation calculation, the capacity sequence, normalized degradation sequence, and auxiliary working condition characteristic sequence are finally obtained.

[0061] Step 1.4: Generate sliding window samples and degradation inflection point labels.

[0062] Based on the capacity sequence, normalized degradation degree sequence, and auxiliary operating condition feature sequence, a sliding window sample is constructed according to the preset input window length, output window length, and sliding step size. The input window refers to the continuous historical cyclic data interval that serves as the model input, and the output window refers to the continuous future cyclic data interval that follows the input window and serves as the supervision target for capacity prediction.

[0063] Specifically, assuming the preset input window length is... The output window length is The sliding step size is Starting from the beginning of the capacity sequence and auxiliary operating condition characteristic sequence, extract continuous... The data of each cycle is used as the input sequence, and then the subsequent continuous... The capacity value of each loop is used as the predicted output sequence; then, according to the sliding step size... Move the window backward and repeat the above process until it is no longer possible to form complete input and output samples.

[0064] Each sliding window sample includes an input sequence, a predicted output sequence, a normalized degradation degree sequence within the window, an auxiliary operating condition feature sequence, and a degradation inflection point label. The input sequence represents historical capacity degradation information and operating condition information; the predicted output sequence supervises the model's learning of future capacity degradation trends; the normalized degradation degree sequence within the window is obtained by arranging the normalized degradation degrees corresponding to each cyclic position within the input window in temporal order, and is used to generate degradation stage perception features; the auxiliary operating condition feature sequence helps the model distinguish between true capacity recovery and random noise fluctuations.

[0065] To train the subsequent inflection point detection auxiliary subnetwork, this step also generates a degradation inflection point label for each sliding window sample. If a degradation inflection point appears within a future preset loop window, the degradation inflection point label is recorded as 1; if no degradation inflection point appears within the future preset loop window, the degradation inflection point label is recorded as 0. The future preset loop window refers to the continuous future loop data interval located after the end position of the input window, used to determine whether a degradation inflection point has occurred. Its length can be the same as the output window length, or it can be set separately according to the early warning lead time requirement. The degradation inflection point can be determined based on the first-order difference, second-order difference, local slope change rate, piecewise fitting error, or the degree of abrupt change in the capacity decay rate of the capacity degradation curve. The above methods for determining the degradation inflection point can be selected or combined according to the battery type, sampling period, data noise level, and early warning lead time requirement.

[0066] Step 1.5: Divide the sample set.

[0067] The preprocessed samples obtained in step 1.4, i.e. the sliding window samples with degradation inflection point labels, are divided into source domain samples and target domain samples respectively to obtain the source domain training sample set, the source domain validation sample set, the target domain fine-tuning sample set, and the target domain test sample set.

[0068] Among them, the source domain training sample set is used to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model with capacity recovery perception, so that the model learns the general capacity degradation law, capacity recovery decoupling feature, degradation stage perception feature and degradation inflection point recognition feature.

[0069] The source domain validation sample set is used to validate the prediction performance, training stability, and early stopping conditions of the source domain pre-trained model.

[0070] The target domain fine-tuning sample set is used to perform transfer fine-tuning on the source domain pre-trained model, so that the model can adapt to the differences in capacity decay characteristics, capacity recovery rate, operating condition distribution and degradation stage of the target domain vehicle power battery.

[0071] The target domain test sample set is used to evaluate the prediction accuracy, remaining service life prediction effect, degradation inflection point identification ability, and prediction confidence interval reliability of the target domain vehicle power battery capacity prediction model.

[0072] Through the above processing, the original cyclic operation data is transformed into time-series samples with uniform structure and consistent scale, containing information on degradation stages and inflection point supervision information. This provides a data foundation for the subsequent construction of vehicle power battery degradation prediction models with capacity recovery perception, multi-loss joint pre-training, domain adaptive transfer fine-tuning, and uncertainty-aware online prediction.

[0073] Step 2: Construct a vehicle power battery degradation prediction model with capacity recovery sensing. This vehicle power battery degradation prediction model includes:

[0074] A capacity recovery sensing dual-branch variational autoencoder decouples and reconstructs a subnetwork, which is used to decouple the long-term degradation trend component and the short-term capacity recovery component, and outputs the reconstructed capacity sequence.

[0075] The degradation degree conditional position encoding module is used to fuse the cyclic time position encoding and the degradation degree position encoding, and output the conditional position encoding corresponding to each cyclic position;

[0076] Transformer predictive backbone subnetwork is used to extract long-range temporal features;

[0077] An inflection point detection auxiliary subnetwork is used to identify and output the probability of a degradation inflection point occurring within a future preset loop window;

[0078] The prediction output module is used to output future capacity prediction sequences and degradation risk auxiliary information.

[0079] Step 2 specifically includes the following processes:

[0080] Step 2.1: Construct a capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork.

[0081] like Figure 2 As shown, the capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork constructed in this step is used to decouple the long-term degradation trend component and the short-term capacity recovery component of the input sequence in the sliding window sample generated in step 1, so as to avoid deleting the real capacity recovery as random noise or passing the short-term capacity recovery as a long-term capacity upward trend to the subsequent prediction network.

[0082] Specifically, the capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork constructed in this step includes a shared encoder, a learnable gated fusion unit, and two parallel branches: a trend branch and a recovery branch. The trend branch includes trend branch latent variables and a trend decoder, while the recovery branch includes recovery branch latent variables and a recovery decoder. The shared encoder extracts shared intermediate features from the input sequence in the sliding window samples. The input sequence specifically includes the capacity sequence and auxiliary operating condition feature sequence within the input window, and the shared intermediate features are input to the two branches respectively. The trend branch latent variables characterize the long-term capacity degradation trend based on the shared intermediate features, and the trend decoder reconstructs the long-term degradation trend component. The recovery branch latent variables characterize the short-term capacity recovery component based on the shared intermediate features, and the recovery decoder reconstructs the short-term capacity recovery component. The learnable gated fusion unit adjusts the retention degree of the short-term capacity recovery component in the reconstructed capacity sequence and fuses the long-term degradation trend component and the short-term capacity recovery component to obtain the reconstructed capacity sequence.

[0083] As a specific implementation method, the learnable gating fusion unit generates gating coefficients based on shared intermediate features, the expression of which is:

[0084]

[0085] in, Indicates the gating coefficient. This represents the shared intermediate features output by the shared encoder. Indicates the gating weight parameter. Indicates the gating bias parameter. This represents the Sigmoid activation function.

[0086] The learnable gating fusion unit fuses the long-term degradation trend component with the short-term capacity recovery component after gating adjustment to obtain the reconstructed capacity sequence, the expression of which is:

[0087]

[0088] in, Indicates the reconstructed capacity sequence. This indicates the component representing a long-term degradation trend. This indicates the extent of the short-term capacity recovery. This indicates element-wise multiplication.

[0089] The capacity recovery-aware dual-branch variational autoencoder decoupled reconstruction subnetwork can highlight the long-term degradation trend while retaining reasonable capacity recovery information, outputting the reconstructed capacity sequence, long-term degradation trend component, and short-term capacity recovery component. These outputs are then combined with auxiliary operating condition features in subsequent steps to form decoupled capacity degradation features, which serve as the feature basis for the Transformer prediction backbone subnetwork to perform temporal modeling.

[0090] Step 2.2: Construct a degradation degree conditional location encoding module.

[0091] The degradation degree conditional location encoding module is used to introduce the normalized degradation degree sequence within the window obtained in step 1 into the time series prediction process, so that the Transformer prediction backbone subnetwork can simultaneously perceive the cycle time position and the actual degradation stage of the battery in the subsequent time series modeling process.

[0092] In actual operation of automotive power batteries, even batteries with the same number of cycles may be at different degradation stages due to differences in batch, operating conditions, temperature, rate capability, and consistency. Therefore, using only the cycle number for location coding is insufficient to accurately represent the true degradation state of the battery.

[0093] As a specific implementation method, this embodiment is based on the first... The normalized degradation degree corresponding to each cycle position Generate the corresponding degradation level position code and fuse it with the cyclic time position code to obtain the first... The conditional position encoding corresponding to each loop position is expressed as follows:

[0094]

[0095] in, Indicates the first The conditional position code corresponding to each loop position; Indicates according to the first The normalized degradation degree corresponding to each cycle position The generated degradation level location code; Indicates the first The loop time position code corresponding to each loop position; This represents the fusion weight coefficient, which can be a preset constant or a learnable parameter.

[0096] Through the degradation-conditional location encoding module, the model can distinguish the capacity change patterns under the early slow decay stage, the stable degradation stage, and the rapid degradation stage. The generated conditional location encoding is used to add decoupled capacity degradation features in subsequent steps to form the input of the Transformer prediction backbone sub-network, thereby improving the prediction adaptability in cross-batch and multi-condition scenarios.

[0097] Step 2.3: Construct the Transformer prediction backbone subnetwork.

[0098] The Transformer predictive backbone subnetwork is used to perform long-range temporal modeling of decoupled capacity degradation features and output temporal features for future capacity prediction and degradation inflection point identification.

[0099] Specifically, the reconstructed capacity sequence, long-term degradation trend component, short-term capacity recovery component, and auxiliary operating condition features obtained in step 2.1 are combined to form decoupled capacity degradation features; then the conditional position encoding obtained in step 2.2 is added to the decoupled capacity degradation features to form the input of the Transformer prediction backbone subnetwork.

[0100] The Transformer prediction backbone subnetwork comprises several layers of temporal feature encoding structures to capture long-term dependencies between different cyclic positions and extract long-range temporal features related to future capacity changes. These long-range temporal features are input into the prediction output module to generate the future capacity prediction sequence within the output window; and into the inflection point detection auxiliary subnetwork to determine whether a degradation inflection point is likely to occur within a preset future cyclic window.

[0101] With the above settings, the capacity prediction task and the degradation inflection point identification task share the same long-term time series feature base, enabling the model to focus on local change features before the arrival of the rapid degradation stage while predicting the capacity decay trend.

[0102] Step 2.4: Construct an inflection point detection auxiliary subnetwork.

[0103] The inflection point detection auxiliary subnetwork is used to identify and output the probability of a degradation inflection point occurring within a preset cyclic window, based on the long-range temporal features output by the Transformer prediction backbone subnetwork.

[0104] The inflection point detection auxiliary subnetwork takes the long-range temporal features output by the Transformer prediction backbone subnetwork as input, and generates the probability of degradation inflection points through a feature aggregation layer, a nonlinear mapping layer, and a binary classification output layer. The feature aggregation layer can employ any of the following methods: average pooling, max pooling, attention pooling, or hidden state extraction at a specific time step; the binary classification output layer can use the Sigmoid function to output probability values.

[0105] By setting up an inflection point detection auxiliary subnetwork, the prediction model can output rapid degradation risk information simultaneously with the output capacity prediction sequence. The degradation inflection point detection auxiliary task can also guide the Transformer prediction backbone subnetwork to focus on changes in capacity decay rate, local fluctuation patterns, and degradation stage transition characteristics during the training phase, thereby improving the advance warning capability of subsequent graded capacity degradation.

[0106] Step 2.5: Construct the prediction output module.

[0107] The prediction output module is used to generate a future capacity prediction sequence within the output window based on the long-range time-series characteristics of the output of the Transformer prediction backbone subnetwork. It also organizes the future capacity prediction results and the degradation inflection point occurrence probability output of the inflection point detection auxiliary subnetwork to form the output result of the vehicle power battery degradation prediction model. The output result includes the future capacity prediction sequence within the output window and degradation risk auxiliary information associated with the degradation inflection point occurrence probability, and is used for subsequent uncertainty perception online prediction and graded capacity degradation early warning.

[0108] Specifically, the prediction output module includes a capacity prediction output head and a result processing unit. The capacity prediction output head takes the long-range temporal characteristics of the Transformer prediction backbone subnetwork output as input, and outputs a length of [length missing] through a fully connected layer, a nonlinear mapping layer, or a temporal decoding layer. The future capacity prediction sequence is then processed. The result processing unit associates the future capacity prediction sequence with the degradation inflection point occurrence probability output by the inflection point detection auxiliary subnetwork in step 2.4 to form degradation risk auxiliary information. This degradation risk auxiliary information includes at least the probability of a degradation inflection point occurring within a preset future loop window, and may also include risk status indicators for subsequent tiered capacity degradation early warning, prediction window position information, or auxiliary judgment information corresponding to the capacity prediction sequence. It should be noted that the degradation inflection point occurrence probability is identified and output by the inflection point detection auxiliary subnetwork; the prediction output module does not repeatedly perform degradation inflection point identification, but only outputs the capacity prediction result and the degradation inflection point occurrence probability in a unified manner.

[0109] When the output window length When the capacity prediction output head outputs the capacity prediction value for the next cycle position or the next prediction time, the output window length... At that time, the capacity prediction output head outputs the capacity prediction sequence for multiple consecutive future cycle positions, i.e., the future capacity prediction sequence.

[0110] Step 2.6: Develop a vehicle power battery degradation prediction model that senses capacity recovery.

[0111] The above-constructed capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork, degradation degree conditional position encoding module, Transformer prediction backbone subnetwork, inflection point detection auxiliary subnetwork, and prediction output module are connected according to the data flow direction to finally form a capacity recovery sensing vehicle power battery degradation prediction model.

[0112] Through the above model structure, the present invention forms a continuous processing link of "capacity recovery decoupling - degradation stage perception - long-term time series prediction - inflection point assisted identification - capacity prediction output". In this link, each module works together around the capacity degradation prediction task and cooperates with each other to form an organic whole.

[0113] Step 3: Using the source domain training sample set as training data, perform multi-loss joint pre-training on the vehicle power battery degradation prediction model to obtain the source domain pre-trained model.

[0114] Step 3 specifically includes the following processes:

[0115] Step 3.1: Determine the training input for multi-loss joint pre-training.

[0116] This step is used to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model with capacity recovery perception constructed in step 2, using the source domain training sample set obtained in step 1 as training data, so that the model can obtain a general capacity degradation representation capability before entering the target domain for fine-tuning.

[0117] Specifically, the source domain training sample set is input into the vehicle power battery degradation prediction model for capacity recovery sensing, and the source domain validation sample set is used for performance verification and early shutdown judgment during the training process. Each training sample in the source domain training sample set includes an input sequence, a predicted output sequence, a normalized degradation degree sequence within the window, an auxiliary operating condition feature sequence, and a degradation inflection point label.

[0118] Specifically, the input sequence is used to decouple the long-term degradation trend component and the short-term capacity recovery component in the capacity recovery-aware dual-branch variational autoencoder decoupling and reconstruction sub-network; the predicted output sequence is used to supervise the Transformer prediction backbone sub-network to learn the future capacity degradation trend within the output window; the normalized degradation degree sequence within the window is used to generate conditional position encoding; the auxiliary operating condition feature sequence is used to characterize the differences in operating conditions; and the degradation inflection point label is used to supervise the inflection point detection auxiliary sub-network to learn the capacity decay rate change, local slope change, and degradation stage transition features before the arrival of the rapid degradation stage.

[0119] Step 3.2: Set the multi-loss joint pre-training objective.

[0120] This step is used to construct the joint training objective for the model source domain pre-training stage, so that the model not only learns the capacity prediction results, but also learns capacity sequence reconstruction, latent variable distribution constraints, sparse representation of capacity recovery, decoupling of trend and recovery, physical constraints of long-term trend, and identification of degradation inflection points.

[0121] The objectives of the multi-loss joint pre-training include capacity prediction loss, variational autoencoder (VAE) reconstruction loss, KL divergence loss, recovery component sparsity loss, latent variable orthogonal decoupling loss, monotonic soft constraint loss, and inflection point detection cross-entropy loss, enabling the model to acquire the capabilities of capacity recovery decoupling, degradation stage perception, degradation trend prediction, degradation inflection point identification, and physical constraint prediction.

[0122] Specifically, the capacity prediction loss is used to constrain the difference between the future capacity prediction sequence output by the model and the actual capacity sequence; the VAE reconstruction loss is used to constrain the ability of the capacity recovery-aware dual-branch variational autoencoder decoupled reconstruction subnetwork to reconstruct the input window content sequence; the KL divergence loss is used to constrain the posterior distribution of the latent variables of the trend branch and the recovery branch to be close to the preset prior distribution; the recovery component sparsity loss is used to limit the response intensity of the recovery branch at non-recovery positions; the latent variable orthogonal decoupling loss is used to reduce the information coupling between the latent variables of the trend branch and the recovery branch; the monotonicity soft constraint loss is used to suppress unreasonable capacity increases that exceed the tolerance band; and the inflection point detection cross-entropy loss is used to train the inflection point detection auxiliary subnetwork to output the probability of a degenerate inflection point occurring within the preset future cyclic window.

[0123] It should be noted that the monotonicity soft constraint does not mean forcing the capacity sequence to be strictly monotonically decreasing, nor does it contradict the capacity recovery component retention mechanism. The recovery branch in the decoupled reconstruction subnetwork of the capacity recovery sensing dual-branch variational autoencoder is used to retain short-term capacity recovery information with continuity and reasonable amplitude; the monotonicity soft constraint is mainly used to constrain the long-term prediction trend, suppressing abnormal capacity increases or continuous unreasonable increases that exceed the preset tolerance band, thereby ensuring that the prediction results conform to the long-term degradation law of vehicle power batteries on a macro scale.

[0124] Step 3.3: Construct the joint pre-trained total loss function of multiple losses.

[0125] After determining each loss term, the capacity prediction loss, VAE reconstruction loss, KL divergence loss, recovery component sparsity loss, latent variable orthogonal decoupling loss, monotonic soft constraint loss, and inflection point detection cross-entropy loss are weighted and summed to form the total loss function for joint pre-training of the model with multiple losses.

[0126] As a specific implementation method, the total loss function can be expressed as:

[0127]

[0128] in, Represents the total loss function. Indicates capacity prediction loss. This represents the VAE reconstruction loss. This represents the KL divergence loss. This indicates a sparse loss of the rebound component. This represents the orthogonal decoupling loss of latent variables. This represents the monotonic soft-constraint loss. This represents the cross-entropy loss for inflection point detection; , , , , , and These represent the weight coefficients corresponding to each loss term. The weight coefficients can be set or dynamically adjusted according to the training stage, sample size, capacity recovery rate, proportion of samples at the degradation inflection point, and model convergence.

[0129] By adopting the above total loss function, the vehicle power battery degradation prediction model can simultaneously learn capacity sequence reconstruction, capacity recovery decoupling, capacity degradation trend prediction, latent variable distribution constraints, dual-branch feature decoupling, long-term trend physical constraints, and degradation inflection point auxiliary identification during the source domain pre-training stage, thus avoiding structural fragmentation caused by isolated training of each network module.

[0130] Step 3.4: Perform multi-loss joint pre-training.

[0131] Based on the source domain training sample set, the degradation prediction model for automotive power batteries is iteratively trained to enable the model to acquire the capabilities of capacity recovery decoupling, degradation trend prediction, and degradation inflection point identification. During each training iteration, the source domain training samples sequentially pass through the capacity recovery sensing dual-branch variational autoencoder decoupling and reconstruction sub-network, the degradation degree conditional location encoding module, the Transformer prediction backbone sub-network, the inflection point detection auxiliary sub-network, and the prediction output module to obtain the reconstructed capacity sequence, the future capacity prediction sequence, and the probability of degradation inflection point occurrence.

[0132] Subsequently, the training loss is calculated based on the total loss function constructed in step 3.3, and the parameters of the shared encoder, trend branch, rise branch, learnable gated fusion unit, degradation degree conditional position encoding module, Transformer prediction backbone subnetwork, inflection point detection auxiliary subnetwork and prediction output module are updated through backpropagation.

[0133] During training, the model's capacity prediction error, reconstruction performance, degradation inflection point identification performance, and training stability are validated using a source domain validation sample set. Pre-training stops when the validation loss stabilizes, a preset stopping condition is met, or a preset upper limit is reached in the number of training epochs.

[0134] Step 3.5: Save the source domain pre-trained model.

[0135] After completing the multi-loss joint pre-training, the model structure and weight parameters are saved to obtain the source domain pre-trained model. The source domain pre-trained model includes the parameters of the capacity recovery-aware dual-branch variational autoencoder decoupled reconstruction sub-network, the parameters of the degradation degree conditional location encoding module, the parameters of the Transformer prediction backbone sub-network, the parameters of the inflection point detection auxiliary sub-network, and the parameters of the prediction output module.

[0136] The source domain pre-trained model obtained through this step already possesses the capabilities of capacity recovery decoupling, degradation stage perception, capacity degradation trend prediction, long-term trend physical constraints, and degradation inflection point assisted identification. It can serve as the initialization model basis for the target domain transfer fine-tuning in step 4.

[0137] Step 4: Use the target domain fine-tuning sample set to perform transfer fine-tuning on the source domain pre-trained model, and introduce hierarchical differential learning rate configuration and maximum mean difference domain alignment loss to obtain the target domain vehicle power battery capacity prediction model; wherein, hierarchical differential learning rate configuration is used to set different learning rate ratios according to the function of network layers to retain the general capacity degradation knowledge of the source domain and adapt to the degradation characteristics of the target domain; maximum mean difference domain alignment loss is used to reduce the distribution offset between the source domain and the target domain in the latent feature space.

[0138] Step 4 specifically includes the following processes:

[0139] Step 4.1: Load the source domain pre-trained model.

[0140] This step is used to call the source domain pre-trained model obtained in step 3 and use it as the initialization basis for the target domain vehicle power battery capacity prediction model.

[0141] Specifically, the structure and weight parameters of the source domain pre-trained model are read, including the parameters of the capacity recovery-aware dual-branch variational autoencoder decoupling and reconstruction sub-network, the parameters of the degradation degree conditional position encoding module, the parameters of the Transformer prediction backbone sub-network, the parameters of the inflection point detection auxiliary sub-network, and the parameters of the prediction output module.

[0142] By loading a pre-trained model from the source domain, the target domain model can inherit the general capacity degradation rules, capacity recovery decoupling ability, degradation stage perception ability, and degradation inflection point assisted identification ability learned from the source domain data. This reduces the amount of data required for de novo training under small sample conditions in the target domain and improves the convergence stability of the transfer fine-tuning process.

[0143] Step 4.2: Configure the tiered differential learning rate.

[0144] This step is used to set different learning rate ratios based on the differences in the roles of different functional modules in the source domain pre-trained model, so as to improve the adaptability of the target domain while retaining the general knowledge of the source domain.

[0145] Specifically, a first learning rate is set for the shared encoder to maintain the basic capacity degradation characteristics obtained during the source domain pre-training stage; a second learning rate is set for the trend decoder, the recovery decoder, and the learnable gated fusion unit to adapt to the target domain capacity decay rate, short-term capacity recovery magnitude, and local temporal fluctuation characteristics; a third learning rate is set for the bottom encoding layer of the Transformer prediction backbone subnetwork to adjust the temporal feature representation of the target domain; and a fourth learning rate is set for the top encoding layer, prediction output module, and inflection point detection auxiliary subnetwork of the Transformer prediction backbone subnetwork to enhance the target domain capacity prediction and degradation inflection point identification capabilities. The first, second, third, and fourth learning rates increase sequentially, and their specific values ​​can be adjusted based on the target domain sample size, the degree of difference between the source and target domains, model convergence, and validation results.

[0146] By configuring differentiated learning rates across layers, we can avoid the source domain knowledge forgetting or target domain underfitting caused by using the same learning rate for all network layers, thus enabling the source domain pre-trained model to maintain good transfer stability under small sample target domain conditions.

[0147] Step 4.3: Construct joint training batches.

[0148] This step is used to simultaneously utilize source domain samples and target domain samples during the target domain migration fine-tuning process, so that the model retains general capacity degradation knowledge while adapting to the specific degradation characteristics of automotive power batteries in the target domain.

[0149] Specifically, samples are selected from the source domain training sample set and the target domain fine-tuning sample set according to a preset ratio, namely, source domain samples and target domain samples, and a joint training batch is constructed using the source domain samples and target domain samples. The joint training batch includes source domain samples and target domain samples. The source domain samples are used to maintain the model's memory of general degradation trends, capacity recovery decoupling rules, and inflection point change characteristics, including the source domain input sequence, source domain predicted output sequence, normalized degradation degree sequence within the source domain window, source domain auxiliary operating condition feature sequence, and source domain degradation inflection point label. The target domain samples are used to guide the model to adapt to batch differences, operating condition differences, capacity recovery magnitude differences, and degradation stage distribution differences of the target automotive power battery, including the target domain input sequence, target domain predicted output sequence, normalized degradation degree sequence within the target domain window, target domain auxiliary operating condition feature sequence, and target domain degradation inflection point label. The ratio of source domain samples to target domain samples can be adjusted according to the number of target domain samples and the model validation effect.

[0150] Step 4.4: Introduce the maximum mean difference domain alignment loss.

[0151] This step is used to reduce the distribution offset between source domain samples and target domain samples in the latent feature space, thereby improving the effectiveness of transferring pre-trained knowledge from the source domain to the target domain.

[0152] Specifically, during the joint fine-tuning process, latent feature representations corresponding to source domain samples and target domain samples are extracted respectively, and the distribution difference between the two is measured by the maximum mean difference domain alignment loss.

[0153] As a specific implementation method, the maximum mean difference domain alignment loss is applied to the latent variables of the trend branch to align the long-term capacity degradation trend characteristics of the source and target domains. The maximum mean difference domain alignment loss can be expressed as:

[0154]

[0155] in, This represents the maximum mean difference domain alignment loss; Indicates the number of samples in the source domain; Indicates the number of samples in the target domain; This represents the latent variables of the trend branches corresponding to the source domain samples; This represents the latent variables of the trend branches corresponding to the samples in the target domain; Represents the kernel mapping function; Represents the regenerating nucleus Hilbert space. In the regenerating nucleus Hilbert space The norm squared, calculated in the standard, is used to measure the distributional difference between the mean embedding of latent features in the source domain and the mean embedding of latent features in the target domain.

[0156] In other implementations, the maximum mean difference domain alignment loss can also be applied to the shared intermediate features of the shared encoder output, the latent variables of the rising branch, the intermediate timing features of the Transformer, or a combination of the above features.

[0157] Step 4.5: Construct the target domain migration fine-tuning total loss function.

[0158] This step combines the multi-loss joint training objective from step 3 with the maximum mean difference domain alignment loss to form the overall training objective for the target domain transfer fine-tuning stage.

[0159] As a specific implementation method, the total loss function in the target domain migration fine-tuning stage can be expressed as:

[0160]

[0161] in, This represents the total loss function during the migration fine-tuning phase. This represents the total loss function constructed in step 3 for the joint pre-training phase of multiple loss functions; This represents the maximum mean difference domain alignment loss; This represents the weighting coefficient corresponding to the maximum mean difference domain alignment loss. It can be set or dynamically adjusted based on the distribution difference between the source and target domains, the number of samples in the target domain, and the prediction performance on the validation set. This total loss function... The model can simultaneously maintain capacity recovery decoupling, capacity degradation prediction, physical constraints, inflection point assisted identification, and domain distribution alignment capabilities during fine-tuning of the target domain.

[0162] Step 4.6: Perform target domain migration fine-tuning and obtain the target domain vehicle power battery capacity prediction model.

[0163] The source domain pre-trained model is iteratively fine-tuned based on the target domain fine-tuning sample set, source domain samples, hierarchical differentiated learning rate configuration, and target domain transfer fine-tuning total loss function.

[0164] Specifically, source domain samples and target domain samples are input into the vehicle power battery degradation prediction model for capacity recovery perception to obtain the reconstructed capacity sequence, future capacity prediction sequence, probability of degradation inflection point occurrence, and potential feature representation; the training loss is calculated based on the transfer fine-tuning total loss function, and the parameters of each network module are updated according to the hierarchical differentiated learning rate.

[0165] During fine-tuning, the model's prediction performance is monitored using validation samples from the target domain or a validation subset derived from the fine-tuning sample set in the target domain. The test sample set in the target domain is used for performance evaluation after fine-tuning. When the model's prediction performance meets the preset requirements, or when the training process meets the preset stopping conditions, the transfer fine-tuning is stopped, and the fine-tuned model structure and weight parameters are saved, resulting in the target domain vehicle power battery capacity prediction model.

[0166] The target domain vehicle power battery capacity prediction model obtained through the above process retains the general capacity degradation knowledge obtained in the source domain pre-training stage, and adapts to the batch differences, operating condition differences, capacity recovery characteristics and degradation stage distribution differences of target domain vehicle power batteries, providing a model basis for uncertainty perception online prediction in step 5.

[0167] Step 5: Deploy the target domain vehicle power battery capacity prediction model in an online application environment, and use the Monte Carlo Dropout uncertainty-aware inference mechanism to perform online capacity degradation prediction for the vehicle power battery to be predicted. In the inference stage, multiple random forward propagations are performed on the same input sample to obtain multiple predicted capacity sequences and multiple degradation inflection point probabilities. The predicted capacity mean sequence, predicted confidence interval sequence, and final degradation inflection point probability are calculated. The remaining service life estimate of the vehicle power battery to be predicted is determined based on the relationship between the predicted capacity mean sequence and the preset end-of-life capacity threshold.

[0168] Step 5 specifically includes the following processes:

[0169] Step 5.1: Deploy the target domain vehicle power battery capacity prediction model.

[0170] This step deploys the target domain vehicle power battery capacity prediction model obtained in step 4 to an online application environment, enabling it to receive cyclic operation data of the vehicle power battery to be predicted and output capacity degradation prediction results and prediction uncertainty information. The target domain vehicle power battery capacity prediction model includes a capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork fine-tuned by target domain migration, a degradation degree conditional position encoding module, a Transformer prediction backbone subnetwork, an inflection point detection auxiliary subnetwork, and a prediction output module.

[0171] Specifically, the target domain vehicle power battery capacity prediction model can be deployed on any one of the following: the local controller of the battery management system, the on-board computing unit, the edge computing device, or the cloud server. The deployment method can be selected based on on-board computing resources, communication conditions, and real-time requirements; no specific limitations are imposed here.

[0172] Step 5.2: Acquire and preprocess the real-time cyclic operation data of the vehicle power battery to be predicted.

[0173] This step is used to obtain real-time cyclic operation data of the vehicle power battery to be predicted and process it into online input samples consistent with the model training phase.

[0174] Specifically, in the online prediction phase, real-time cyclic operation data of the vehicle power battery to be predicted is continuously acquired. This real-time cyclic operation data includes the number of cycles, cycle start time, discharge capacity, charging capacity, charging and discharging voltage, charging and discharging current, battery surface temperature, resting time, and operating parameters related to the operating status of the vehicle power battery.

[0175] The aforementioned real-time looping data undergoes the same online preprocessing as in step 1, including data cleaning, regularization, degradation degree calculation, and sliding window sample construction, to obtain the input samples to be predicted. Specifically, data cleaning handles obvious sampling errors, missing records, and temporal anomalies; regularization unifies the data scale between the online input data and the training samples; degradation degree calculation obtains the normalized degradation degree corresponding to the current loop; and sliding window sample construction extracts data from several recent consecutive loops to form the online inference input for the model.

[0176] It should be noted that during the online preprocessing process, local capacity recovery points with continuity, reasonable amplitude, and relevant operating conditions are still retained and are not directly deleted as outliers. This ensures that the capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork can continue to decouple the long-term degradation trend component and the short-term capacity recovery component.

[0177] When the object to be predicted is a single cell of a vehicle power battery, an online input sample can be directly constructed based on the single cell's cycle operation data; when the object to be predicted is a vehicle power battery module or battery pack, an online input sample can be constructed based on the single cell's capacity characteristics, the module's statistical characteristics, the battery pack's operating condition characteristics, or a combination thereof.

[0178] Step 5.3: Using the Monte Carlo Dropout uncertainty-aware reasoning mechanism.

[0179] This step is used to introduce a prediction uncertainty metric into the online inference stage of the target domain vehicle power battery capacity prediction model, so that the model not only outputs the capacity prediction value, but also the confidence range of the prediction result.

[0180] Specifically, during the inference phase, the preset Dropout layer in the target domain vehicle power battery capacity prediction model is kept active. The preset Dropout layer can be set at any one or more locations in the capacity recovery sensing dual-branch variational autoencoder decoupling and reconstruction subnetwork, the Transformer prediction backbone subnetwork, the inflection point detection auxiliary subnetwork, and the prediction output module, to form a random deactivation state during multiple forward propagation processes.

[0181] For the same input sample to be predicted, multiple random forward propagations are performed. Each random forward propagation sequentially passes through the capacity recovery sensing dual-branch variational autoencoder decoupling and reconstruction subnetwork, the degradation degree conditional location encoding module, the Transformer prediction backbone subnetwork, the inflection point detection auxiliary subnetwork, and the prediction output module, outputting a set of predicted capacity sequences within an output window and the probability of degradation inflection points occurring within a future preset cyclic window. The number of random forward propagations can be set according to online computing resources, prediction real-time requirements, and confidence interval stability requirements; no limit is imposed here.

[0182] Because the random inactivation state of the pre-set Dropout layer is different in each random forward propagation, the same input sample to be predicted can yield multiple prediction results, including multiple prediction capacity sequences and multiple probabilities of occurrence of degradation inflection points. These multiple prediction results are used to characterize the degree of prediction dispersion of the model under the current input sample, thereby achieving uncertainty-aware online inference.

[0183] Step 5.4: Calculate the predicted capacity mean sequence.

[0184] This step is used to calculate the average predicted capacity at each future cycle position within the output window based on the capacity prediction results obtained from multiple random forward propagations, and use it as the main prediction result for the future capacity degradation trend of the vehicle power battery to be predicted.

[0185] Suppose that the same input sample to be predicted is executed... The second random forward propagation, the... The nth random forward propagation obtained The prediction capacity for each future cycle position is Then the first The mean predicted capacity of each future cycle position can be expressed as:

[0186]

[0187] in, Indicates the first The mean of the predicted capacity for each future cycle position. Indicates the number of random forward propagations. Indicates the first The nth random forward propagation obtained The predictive capacity of the future cycle positions.

[0188] When the output window length is 1, the predicted capacity mean represents the capacity prediction result for the next cycle position; when the output window length is greater than 1, a sequence of predicted capacity mean values ​​for multiple consecutive cycle positions can be obtained.

[0189] Step 5.5: Calculate the predicted confidence interval sequence.

[0190] This step is used to calculate the prediction standard deviation and prediction confidence interval based on multiple predicted capacity sequences obtained from multiple random forward propagations, in order to quantify the uncertainty of the model's prediction of future capacity.

[0191] No. The predicted standard deviation of each future cycle position can be expressed as:

[0192]

[0193] in, Indicates the first The standard deviation of the prediction for each future cyclic position.

[0194] As one specific implementation method, the first The prediction confidence interval for each future cycle position can be expressed as:

[0195]

[0196] in, Indicates the first The confidence interval for predicting a future cycle position. This represents the confidence coefficient. Optionally, when using an approximate 95% confidence level, A value of 1.96 is acceptable; in other embodiments, It can also be configured according to the risk strategy and application requirements of the battery management system.

[0197] The prediction confidence interval is used to characterize the range of confidence of the capacity prediction results. The narrower the confidence interval, the more stable the model prediction results; the wider the confidence interval, the higher the uncertainty of the current prediction.

[0198] Step 5.6: Calculate the probability of the final degradation inflection point occurring.

[0199] This step is used to calculate the probability of the final degradation inflection point occurring within a future preset cyclic window, based on the probability of occurrence of multiple degradation inflection points obtained from multiple random forward propagations.

[0200] Let the first The probability of the degradation inflection point obtained from the random forward propagation is: The probability of the final degradation inflection point occurring can be expressed as:

[0201]

[0202] in, This indicates the probability of the final degradation inflection point occurring. This represents the number of random forward propagations.

[0203] The probability of the final degradation inflection point is used to represent the risk level of the vehicle power battery to be predicted entering a rapid degradation stage within a preset cycle window. It should be noted that this step only calculates and outputs the probability of the final degradation inflection point, and does not perform an inflection point imminent warning judgment.

[0204] Step 5.7: Output the online capacity degradation prediction results and determine the estimated remaining useful life.

[0205] This step is used to summarize the predicted mean capacity sequence, the predicted confidence interval sequence, and the probability of the final degradation inflection point to form the online capacity degradation prediction result of the vehicle power battery to be predicted.

[0206] Specifically, the online capacity degradation prediction results include the predicted capacity mean sequence within the output window, the predicted confidence interval sequence corresponding to the predicted capacity mean sequence, and the probability of the final degradation inflection point occurring within a future preset cyclic window.

[0207] The online capacity degradation prediction results can be output to the battery management system, vehicle display terminal, edge computing platform, or cloud health management platform for subsequent remaining capacity estimation, remaining service life estimation, and graded capacity degradation early warning in step 6. The remaining service life of the vehicle power battery to be predicted can be estimated based on the relationship between the predicted average capacity sequence and the preset end-of-life capacity threshold.

[0208] As an optional implementation, the future cycle position where the predicted average capacity first falls below the preset end-of-life capacity threshold after the current cycle position is taken as the predicted end-of-life position, and the difference in cycle number between this predicted end-of-life position and the current cycle position is used as the estimated remaining lifespan. When the predicted average capacity within the current prediction window is not lower than the preset end-of-life capacity threshold, a rolling prediction method can be used to continue extrapolation. This involves updating the latest prediction result or subsequently collected actual operating data to a new input window, repeatedly performing online capacity degradation prediction until the predicted average capacity falls below the preset end-of-life capacity threshold, and determining the estimated remaining lifespan accordingly. The preset end-of-life capacity threshold can be set according to battery type, vehicle application scenario, battery management system strategy, maintenance safety requirements, or testing standards, and is not limited here.

[0209] Through step 5 above, the target domain vehicle power battery capacity prediction model can simultaneously output the capacity degradation trend, prediction confidence range, remaining service life estimation results, and degradation inflection point risk probability during the online operation phase, achieving online capacity degradation prediction with uncertainty awareness.

[0210] Step 6: Establish graded capacity degradation early warning judgment conditions. Based on the comparison results of the predicted capacity mean sequence, predicted confidence interval sequence, probability of the final degradation inflection point, estimated remaining service life, and graded capacity degradation early warning judgment conditions, graded capacity degradation early warning is performed on the vehicle power battery to be predicted. The graded capacity degradation early warning includes at least one or more of the following: uncertainty warning, inflection point approaching warning, and end of life warning. The graded capacity degradation early warning results are generated.

[0211] Step 6 specifically includes the following processes:

[0212] Step 6.1: Set the judgment conditions for graded capacity degradation early warning.

[0213] This step is used to establish the judgment conditions for the graded capacity degradation early warning of the vehicle power battery to be predicted, based on the predicted average capacity, predicted confidence interval, probability of the final degradation inflection point and estimated remaining service life output in step 5.

[0214] Specifically, the predicted average capacity sequence, the predicted confidence interval sequence, the probability of the final degradation inflection point occurring within the preset future cyclic window, and the estimated remaining useful life are used as inputs for early warning judgment. Among them, the predicted average capacity is used to characterize the future capacity degradation trend, the predicted confidence interval is used to characterize the uncertainty of the prediction result, and the probability of the final degradation inflection point occurring is used to characterize the risk level of entering the rapid degradation stage within the preset future cyclic window.

[0215] In this embodiment, three types of judgment conditions are set: uncertainty warning, inflection point imminent warning, and end-of-life warning. The uncertainty warning is used to determine whether the credibility of the current prediction result meets the requirements of online application; the inflection point imminent warning is used to determine whether the vehicle power battery to be predicted is at risk of entering a rapid degradation stage; the end-of-life warning is used to determine whether the vehicle power battery to be predicted is likely to approach or reach the end of its life state based on the predicted capacity average sequence, the preset end-of-life capacity threshold, and the estimated remaining lifespan.

[0216] The tiered capacity degradation early warning criteria include: triggering an uncertainty warning when the half-width of the predicted confidence interval at any future cycle position within a preset future cycle window is greater than a preset uncertainty threshold, or when the half-width of the confidence intervals at multiple consecutive future cycle positions is greater than the preset uncertainty threshold; triggering an inflection point imminent warning when the probability of the final degradation inflection point is greater than a preset inflection point threshold; and triggering a lifespan termination warning when the predicted average capacity at any future cycle position is lower than a preset lifespan end capacity threshold, or when the predicted average capacity sequence is lower than the preset lifespan end capacity threshold at several consecutive future cycle positions, or when the estimated remaining lifespan is less than a preset remaining lifespan warning threshold.

[0217] Among them, the preset uncertainty threshold, preset inflection point threshold, preset end-of-life capacity threshold, and preset remaining life warning threshold can be set according to the type of vehicle power battery, vehicle usage scenario, battery management system strategy, maintenance safety requirements, or historical operating data statistics, and can also be dynamically adjusted according to actual operating conditions.

[0218] Step 6.2: Trigger uncertainty warning.

[0219] This step is used to determine whether there is high uncertainty in the current capacity degradation prediction results based on the prediction confidence interval.

[0220] Specifically, the half-width of the confidence interval is calculated based on the predicted confidence interval obtained in step 5. An uncertainty warning is triggered when the half-width of the confidence interval at any future cycle position within the preset future cycle window is greater than a preset uncertainty threshold, or when the half-width of the confidence interval at multiple consecutive future cycle positions is greater than the preset uncertainty threshold.

[0221] As one specific implementation method, the first The half-width of the confidence interval for each future cycle position can be expressed as:

[0222]

[0223] in, Indicates the first The half-width of the confidence interval for predicting a future cycle position. Indicates the first The upper bound of the prediction confidence interval for each future cyclic position. Indicates the first The lower bound of the prediction confidence interval for each future cyclic position.

[0224] When satisfied At that time, an uncertainty warning is triggered. Among them, This indicates a preset uncertainty threshold.

[0225] Uncertainty warnings are used to indicate that the current prediction results are not reliable enough. They can suggest increasing observation data, increasing the data collection frequency, reducing the weight of the current prediction results in maintenance decisions, or marking the vehicle power battery to be predicted as a key tracking target.

[0226] Step 6.3: Trigger an early warning for an approaching inflection point.

[0227] This step is used to determine whether the vehicle power battery to be predicted is likely to enter a rapid degradation phase within a preset cycle window in the future, based on the probability of the final degradation inflection point.

[0228] Specifically, the probability of the final degradation inflection point obtained in step 5 is compared with a preset inflection point threshold. When the probability of the final degradation inflection point is greater than the preset inflection point threshold, an inflection point imminent warning is triggered. The judgment condition can be expressed as:

[0229]

[0230] in, This indicates the probability of the final degradation inflection point occurring. This indicates the preset inflection point threshold.

[0231] The inflection point warning is used to indicate that the vehicle power battery may transition from a stable degradation phase to a rapid degradation phase within a preset cycle window. Compared to judging solely based on whether the capacity is below the end-of-life threshold, the inflection point warning can provide early warning of the risk of a significantly increased capacity decay rate before the capacity reaches the end-of-life standard.

[0232] Step 6.4: Trigger end-of-life warning.

[0233] This step is used to determine whether the vehicle power battery to be predicted is likely to approach or reach the end of its life state based on the predicted average capacity sequence, the preset end-of-life capacity threshold, the estimated remaining lifespan, and the preset remaining lifespan warning threshold.

[0234] Specifically, the average predicted capacity at each future cycle position within the output window of step 5 is compared with a preset end-of-life capacity threshold. A life-end warning is triggered when the average predicted capacity at any future cycle position is lower than the preset end-of-life capacity threshold, or when the average predicted capacity sequence is lower than the preset end-of-life capacity threshold for several consecutive future cycle positions. A life-end warning is also triggered when the estimated remaining useful life determined in step 5 is less than a preset remaining useful life warning threshold. The life-end warning judgment condition based on the average predicted capacity can be expressed as:

[0235]

[0236] in, Indicates the first The mean of the predicted capacity for each future cycle position. This indicates the preset end-of-life capacity threshold.

[0237] As another specific implementation, the life-end warning judgment condition based on the remaining useful life estimate can be expressed as follows:

[0238]

[0239] in, This represents the estimated remaining useful life. This indicates the preset remaining lifespan warning threshold.

[0240] The preset end-of-life capacity threshold and preset remaining life warning threshold can be set according to the end-of-life standards for automotive power batteries, vehicle range requirements, power performance requirements, safety management strategies, or cascade utilization standards.

[0241] It should be noted that uncertainty warnings, inflection point approaching warnings, and end-of-life warnings can be triggered independently or simultaneously. When multiple warnings are triggered simultaneously, the final warning status can be output according to the warning level with the highest risk, or multiple warning labels can be output at the same time.

[0242] Step 6.5: Generate graded capacity degradation early warning results.

[0243] This step is used to generate a pre-warning result for the graded capacity degradation of the vehicle power battery to be predicted, based on the judgment results of steps 6.2 to 6.4.

[0244] Specifically, when no uncertainty warning, inflection point approaching warning, or end-of-life warning is triggered, the output is a normal state; when an uncertainty warning is triggered, the output is a state of insufficient prediction credibility; when an inflection point approaching warning is triggered, the output is a state of rapid degradation risk; and when an end-of-life warning is triggered, the output is a state of end-of-life risk.

[0245] As an optional implementation method, the graded capacity degradation early warning results can be divided into four different levels: normal, attention, risk, and severe risk. The normal level corresponds to a state where no warning has been triggered; the attention level corresponds to an uncertainty warning; the risk level corresponds to an inflection point approaching warning; and the severe risk level corresponds to a lifespan end warning. When multiple warning conditions are met simultaneously, a comprehensive early warning result can be output based on preset priorities or risk levels.

[0246] Through the aforementioned graded capacity degradation early warning, this embodiment transforms the predicted mean capacity sequence, predicted confidence interval sequence, probability of the final degradation inflection point, and estimated remaining service life into health management information that can be used by the battery management system, providing a basis for capacity degradation monitoring, maintenance plan formulation, and safety risk control.

[0247] Step 7: Evaluate the performance of the target domain vehicle power battery capacity prediction model based on the target domain test sample set, online running data, or subsequent collected capacity data, and trigger model parameter updates, online incremental fine-tuning, data resampling, or warning threshold adjustment based on the evaluation results.

[0248] Step 7 specifically includes the following processes:

[0249] Step 7.1: Evaluate the performance of the target domain vehicle power battery capacity prediction model.

[0250] This step is used to evaluate the performance of the target domain vehicle power battery capacity prediction model based on the target domain test sample set, online running data, or subsequent collected actual capacity data.

[0251] Specifically, in the offline testing phase, the predictive performance of the target domain vehicle power battery capacity prediction model is evaluated using the target domain test sample set obtained in step 1; in the online application phase, the actual capacity data collected subsequently is compared and evaluated with the previously output prediction results.

[0252] Model performance evaluation includes capacity prediction accuracy evaluation, remaining useful life prediction accuracy evaluation, degradation inflection point identification performance evaluation, and prediction confidence interval reliability evaluation. Capacity prediction accuracy evaluation measures the error between the predicted mean capacity and the actual capacity value. Remaining useful life prediction accuracy evaluation measures the difference between the estimated remaining useful life, the predicted end-of-life position, and the actual end-of-life position. The predicted end-of-life position is determined based on the location where the predicted mean capacity sequence obtained in step 5 first falls below a preset end-of-life capacity threshold. The actual end-of-life position is determined based on the location where the actual capacity in the target domain test sample set, online operating data, or subsequently collected capacity data first falls below the preset end-of-life capacity threshold. Degradation inflection point identification performance evaluation measures the ability of the inflection point detection auxiliary subnetwork to identify rapidly degrading near-terminal states. Prediction confidence interval reliability evaluation measures the coverage of the actual capacity value falling within the prediction confidence interval and the reasonableness of the prediction confidence interval width.

[0253] As an optional implementation, capacity prediction accuracy can be evaluated using indicators such as root mean square error, mean absolute error, mean absolute percentage error, or coefficient of determination; degradation inflection point identification performance can be evaluated using indicators such as accuracy, recall, precision, F1 score, or area under the receiver operating characteristic curve; remaining useful life prediction accuracy can be evaluated using indicators such as absolute error of remaining cycles, relative error of remaining cycles, or deviation of lifespan end location; and prediction confidence interval reliability can be evaluated using one or more of the following indicators: prediction interval coverage, average prediction interval width, standardized average prediction interval width, or comprehensive coverage width evaluation. Specifically, prediction interval coverage characterizes the proportion of actual capacity values ​​falling within the prediction confidence interval, average prediction interval width characterizes the average width of the prediction confidence interval, and the comprehensive coverage width evaluation indicator comprehensively measures the prediction interval coverage capability and the reasonableness of the interval width.

[0254] Step 7.2: Trigger model parameter updates, online incremental fine-tuning, data resampling, or early warning threshold adjustments based on the evaluation results.

[0255] This step is used to update the target domain vehicle power battery capacity prediction model and early warning judgment conditions based on the model performance evaluation results.

[0256] Specifically, the model update mechanism is triggered when any of the following conditions are met: capacity prediction error is greater than a preset error threshold, remaining useful life prediction error is greater than a preset useful life error threshold, degradation inflection point identification effect is lower than a preset identification requirement, or prediction confidence interval reliability does not meet a preset coverage requirement.

[0257] The model update mechanism includes model parameter updates, online incremental fine-tuning, target domain sample supplementation, and data resampling or re-transfer fine-tuning. As a specific implementation, subsequently collected actual capacity data can be added to the target domain fine-tuning sample set. While retaining the hierarchical differentiated learning rate configuration and the maximum mean difference domain alignment mechanism from step 4, incremental fine-tuning is then performed on the target domain vehicle power battery capacity prediction model.

[0258] When the overall predictive capability of the model meets the requirements, but the frequency of early warning triggers does not match the actual maintenance needs, the early warning threshold adjustment mechanism is triggered. Specifically, based on actual operating data, maintenance records, capacity failure standards, and safety policies, the preset uncertainty threshold, preset inflection point threshold, preset end-of-life capacity threshold, and preset remaining lifespan early warning threshold are adjusted.

[0259] Through the above-mentioned evaluation and feedback mechanism, the present invention can adapt to changes in capacity decay characteristics, capacity recovery rate, degradation stage distribution and operating conditions during the long-term operation of vehicle power batteries, and continuously maintain the ability to predict and classify capacity degradation.

[0260] Step 7.3: Output capacity degradation prediction, graded early warning and evaluation feedback results.

[0261] This step is used to summarize and output the online capacity degradation prediction results obtained in step 5, the hierarchical capacity degradation early warning results generated in step 6, and the model performance evaluation results.

[0262] Specifically, the output includes the mean sequence of predicted capacity, the sequence of predicted confidence intervals, the probability of the final degradation inflection point occurring within the future preset cyclic window, the estimated remaining useful life, the uncertainty warning status, the inflection point approaching warning status, the end of life warning status, the comprehensive warning level, and the model performance evaluation results.

[0263] The aforementioned outputs can be sent to battery management systems, in-vehicle display terminals, edge computing platforms, or cloud-based health management platforms to display the capacity degradation trend, prediction reliability, degradation inflection point risk, end-of-life risk, and current model performance status of automotive power batteries. By comprehensively evaluating the accuracy of capacity prediction, remaining lifespan prediction, degradation inflection point identification, and prediction confidence interval reliability, and triggering model updates or threshold adjustments based on the evaluation results, the method can adapt to changes in degradation characteristics during the long-term operation of automotive power batteries.

[0264] Through steps 1-7 above, the present invention forms a closed-loop processing procedure of "online prediction - hierarchical early warning - performance evaluation - feedback update", which enables the target domain vehicle power battery capacity prediction model to continuously adapt to changes in data distribution during long-term operation and provides a stable decision-making basis for vehicle power battery health management and safety maintenance.

[0265] The proposed method for predicting the degradation of automotive power batteries based on capacity recovery sensing in this embodiment processes the cyclic operation data of automotive power batteries in a unified manner, and combines capacity recovery sensing decoupled modeling, degradation degree conditional location encoding, degradation inflection point assisted identification, domain adaptive migration fine-tuning, and uncertainty sensing reasoning to achieve comprehensive prediction of capacity degradation trend, remaining service life estimation, and degradation risk status. This improves the online health status assessment and hierarchical early warning capabilities under multiple operating conditions, small sample, and cross-batch scenarios.

[0266] like Figure 3 As shown, another embodiment of the present invention also provides a vehicle power battery degradation prediction system with capacity recovery sensing. This system includes, in sequence: a data acquisition and preprocessing unit, a model building unit, a multi-loss joint pre-training unit, a domain adaptive transfer fine-tuning unit, an uncertainty-aware online prediction unit, a graded capacity degradation early warning unit, and an evaluation feedback unit. This system is used to implement the above-described vehicle power battery degradation prediction method with capacity recovery sensing, and each unit is used to implement the corresponding steps in the above method embodiments.

[0267] Specifically, the data acquisition and preprocessing unit is used to collect cyclic operation data of the source domain and target domain vehicle power batteries, and perform preprocessing, including data cleaning, regularization, degradation degree calculation, sliding window sample generation, and degradation inflection point label generation, to obtain preprocessed samples. These preprocessed samples are then divided into source domain training sample sets, source domain validation sample sets, target domain fine-tuning sample sets, and target domain test sample sets. The cyclic operation data includes the number of cycles, cycle start time, discharge capacity, charging capacity, charge / discharge voltage, charge / discharge current, battery surface temperature, resting time, and operating parameters related to the vehicle power battery's operating state.

[0268] The model building unit is used to construct a vehicle power battery degradation prediction model with capacity recovery sensing capability. This vehicle power battery degradation prediction model includes:

[0269] A capacity recovery sensing dual-branch variational autoencoder decouples and reconstructs a subnetwork, which is used to decouple the long-term degradation trend component and the short-term capacity recovery component, and outputs the reconstructed capacity sequence.

[0270] The degradation degree conditional position encoding module is used to fuse the loop time position encoding and the degradation degree position encoding, and output the conditional position encoding corresponding to each loop;

[0271] Transformer predictive backbone subnetwork is used to extract long-range temporal features;

[0272] An inflection point detection auxiliary subnetwork is used to identify and output the probability of a degradation inflection point occurring within a future preset loop window;

[0273] The prediction output module is used to output future capacity prediction sequences and degradation risk auxiliary information.

[0274] The multi-loss joint pre-training unit is used to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model with capacity recovery perception using the source domain training sample set as training data, resulting in a source domain pre-trained model. The multi-loss joint pre-training unit sets the total loss function of multi-loss joint pre-training as a weighted sum of capacity prediction loss, VAE reconstruction loss, KL divergence loss, recovery component sparsity loss, latent variable orthogonal decoupling loss, monotonic soft constraint loss, and inflection point detection cross-entropy loss, enabling the model to acquire the capabilities of capacity recovery decoupling, degradation stage perception, capacity degradation trend prediction, long-term trend physical constraints, and degradation inflection point auxiliary identification.

[0275] The domain adaptive transfer fine-tuning unit is used to transfer and fine-tune the source domain pre-trained model using the target domain fine-tuning sample set. It introduces hierarchical differential learning rate configuration and maximum mean difference domain alignment loss to obtain the target domain vehicle power battery capacity prediction model. The hierarchical differential learning rate configuration is used to set different learning rate ratios according to the functions of the network layers to retain the general capacity degradation knowledge of the source domain and adapt to the degradation characteristics of the target domain. The maximum mean difference domain alignment loss is used to reduce the distribution offset between the source domain and the target domain in the latent feature space.

[0276] The uncertainty-aware online prediction unit is used to deploy the target domain vehicle power battery capacity prediction model in an online application environment. It uses a Monte Carlo Dropout uncertainty-aware inference mechanism to perform online capacity degradation prediction for the vehicle power battery under test. Specifically, during the inference phase, the uncertainty-aware online prediction unit performs multiple random forward propagations on the same input sample to be predicted, obtaining multiple predicted capacity sequences and multiple probabilities of degradation inflection points. It then calculates the predicted mean capacity sequence, the predicted confidence interval sequence, and the probability of the final degradation inflection point. Based on the relationship between the predicted mean capacity sequence and a preset end-of-life capacity threshold, it determines the estimated remaining lifespan of the vehicle power battery to be predicted.

[0277] The graded capacity degradation early warning unit is used to establish graded capacity degradation early warning judgment conditions. Based on the comparison results of the predicted average capacity sequence, the predicted confidence interval sequence, the probability of the final degradation inflection point, the estimated remaining service life, and the graded capacity degradation early warning judgment conditions, it performs graded capacity degradation early warning for the vehicle power battery to be predicted. The graded capacity degradation early warning includes at least one or more of the following: uncertainty warning, inflection point imminent warning, and end-of-life warning, generating graded capacity degradation early warning results. Specifically, when the half-width of the confidence interval at any future cycle position within a preset future cycle window is greater than a preset uncertainty threshold, or when the half-width of the confidence interval at multiple consecutive future cycle positions is greater than the preset uncertainty threshold, an uncertainty warning is triggered; when the probability of the final degradation inflection point is greater than a preset inflection point threshold, an inflection point imminent warning is triggered; when the predicted average capacity at any future cycle position is lower than a preset end-of-life capacity threshold, or when the predicted average capacity sequence is lower than the preset end-of-life capacity threshold at several consecutive future cycle positions, or when the estimated remaining service life is less than a preset remaining service life warning threshold, an end-of-life warning is triggered. The graded capacity degradation early warning unit can output the normal state, the state with insufficient prediction reliability, the state of rapid degradation risk, the state of end-of-life risk, or a comprehensive early warning level.

[0278] Furthermore, the system also includes an evaluation feedback unit, which evaluates the performance of the target domain vehicle power battery capacity prediction model based on the target domain test sample set, online operating data, or subsequently collected actual capacity data. Based on the evaluation results, it triggers model parameter updates, online incremental fine-tuning, data resampling, or warning threshold adjustments. The performance evaluation by the evaluation feedback unit includes capacity prediction accuracy evaluation, remaining service life prediction accuracy evaluation, degradation inflection point identification effect evaluation, and prediction confidence interval reliability evaluation. This embodiment uses the evaluation feedback unit to comprehensively evaluate capacity prediction accuracy, remaining service life prediction accuracy, degradation inflection point identification effect, and prediction confidence interval reliability, and triggers model updates or threshold adjustments based on the evaluation results, enabling the system to adapt to changes in degradation characteristics during the long-term operation of vehicle power batteries.

[0279] The aforementioned units can be implemented through software programs, hardware circuits, or a combination of both, and can be deployed in battery management systems, vehicle computing platforms, edge computing devices, or cloud servers. The data acquisition devices, storage devices, computing devices, communication devices, and display terminals involved in the system can all be implemented using existing equipment. Specific models, interface types, and deployment methods can be selected based on the actual application scenario and are not limited here.

[0280] Through the above system structure, this invention forms a continuous processing link for data acquisition and preprocessing, model building, multi-loss joint pre-training, domain adaptive transfer fine-tuning, uncertainty-aware online prediction, graded capacity degradation early warning, and evaluation feedback. This ensures that each unit of the system corresponds one-to-one with the method steps, enabling the prediction of capacity degradation, estimation of remaining capacity, estimation of remaining service life, identification of degradation inflection point risk, and graded early warning for automotive power batteries under multiple operating conditions, small sample sizes, and cross-batch scenarios.

[0281] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0282] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for predicting the degradation of automotive power batteries based on capacity recovery sensing, characterized in that, Includes the following steps: Step 1: Collect the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery, and perform preprocessing and sample division to obtain the source domain training sample set, the source domain validation sample set, the target domain fine-tuning sample set, and the target domain test sample set. Step 2: Construct a vehicle power battery degradation prediction model with capacity recovery sensing, wherein the vehicle power battery degradation prediction model includes: A capacity recovery sensing dual-branch variational autoencoder decouples and reconstructs a subnetwork, which is used to decouple the long-term degradation trend component and the short-term capacity recovery component, and outputs the reconstructed capacity sequence. The degradation degree conditional position encoding module is used to fuse the cyclic time position encoding and the degradation degree position encoding, and output the conditional position encoding corresponding to each cyclic position; The Transformer predicts the backbone subnetwork, whose input is the decoupled capacity degradation feature with added conditional position encoding, used to perform long-range time series modeling on the decoupled capacity degradation feature and extract long-range time series features. The decoupled capacity degradation feature is obtained by combining the reconstructed capacity sequence, the long-term degradation trend component, the short-term capacity recovery component and auxiliary operating condition features. The inflection point detection auxiliary subnetwork is used to identify and output the probability of a degradation inflection point occurring within a future preset loop window; The prediction output module is used to output future capacity prediction sequences and degradation risk auxiliary information; Step 3: Use the source domain training sample set to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model to obtain the source domain pre-trained model; Step 4: Use the target domain fine-tuning sample set to perform transfer fine-tuning on the source domain pre-trained model, and introduce hierarchical differential learning rate configuration and maximum mean difference domain alignment loss to obtain the target domain vehicle power battery capacity prediction model. Step 5: Deploy the target domain vehicle power battery capacity prediction model in an online application environment, and use the Monte Carlo Dropout uncertainty perception reasoning mechanism to perform online capacity degradation prediction for the vehicle power battery to be predicted, calculate the predicted capacity mean sequence, the predicted confidence interval sequence and the probability of the final degradation inflection point, and determine the estimated remaining service life. Step 6: Establish tiered capacity degradation early warning judgment conditions. Based on the predicted mean capacity sequence, predicted confidence interval sequence, probability of the final degradation inflection point, and estimated remaining useful life, tiered capacity degradation early warning is performed. It includes at least one or more of the following: uncertainty warning, inflection point imminent warning, and end of useful life warning. The tiered capacity degradation early warning results are generated.

2. The method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to claim 1, characterized in that, The capacity recovery sensing dual-branch variational autoencoder decoupled reconstruction subnetwork includes a shared encoder, a learnable gating fusion unit, and parallel trend branches and recovery branches. The trend branch includes a trend branch latent variable and a trend decoder, and the recovery branch includes a recovery branch latent variable and a recovery decoder. The shared encoder is used to extract shared intermediate features of the input capacity sequence and auxiliary operating condition features; The trend branch latent variables are used to characterize the long-term capacity degradation trend based on the shared intermediate features, and the trend decoder is used to reconstruct the long-term degradation trend components. The latent variables of the recovery branch are used to characterize the short-term capacity recovery component based on the shared intermediate features, and the recovery decoder is used to reconstruct the short-term capacity recovery component. The learnable gating fusion unit is used to fuse the long-term degradation trend component and the short-term capacity recovery component to obtain a reconstructed capacity sequence.

3. The method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to claim 2, characterized in that, The formula for reconstructing the capacity sequence is as follows: in, Indicates the reconstructed capacity sequence; Indicates the component representing the long-term degradation trend; This indicates the extent of the short-term capacity rebound; This represents element-wise multiplication; The gating coefficient is expressed by the following formula: in, This represents the shared intermediate features output by the shared encoder; Indicates the gating weight parameter; Indicates the gating bias parameter; This represents the Sigmoid activation function.

4. A method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to claim 2 or 3, characterized in that, The hierarchical differentiated learning rate configuration includes: setting a first learning rate for the shared encoder; setting a second learning rate for the trend decoder, the rise decoder, and the learnable gated fusion unit; setting a third learning rate for the bottom encoding layer of the Transformer prediction backbone subnetwork; and setting a fourth learning rate for the top encoding layer of the Transformer prediction backbone subnetwork, the prediction output module, and the inflection point detection auxiliary subnetwork; wherein the first learning rate to the fourth learning rate increases sequentially.

5. A method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to claim 2 or 3, characterized in that, The formula for the conditional position encoding is: in, Indicates the first The conditional position code corresponding to each loop position; Indicates according to the first The normalized degradation degree corresponding to each cycle position The generated degradation level location code; Indicates the first The loop time position code corresponding to each loop position; This represents the fusion weighting coefficient.

6. A method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to any one of claims 1 to 3, characterized in that, The maximum mean difference domain alignment loss is applied to the latent variables of the trend branch to characterize the long-term capacity degradation trend of the source and target domains. The formula for the maximum mean difference domain alignment loss is: in, This represents the maximum mean difference domain alignment loss; Indicates the number of samples in the source domain; Indicates the number of samples in the target domain; This represents the latent variables of the trend branches corresponding to the source domain samples; This represents the latent variables of the trend branches corresponding to the samples in the target domain; Represents the kernel mapping function; Represents the regenerating nucleus Hilbert space. In the regenerating nucleus Hilbert space The norm squared, calculated in the standard, is used to measure the distributional difference between the mean embedding of latent features in the source domain and the mean embedding of latent features in the target domain.

7. A method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to any one of claims 1 to 3, characterized in that, The criteria for determining the graded capacity degradation early warning are as follows: An uncertainty warning is triggered when the half-width of the confidence interval at any future cycle position within the preset future cycle window is greater than the preset uncertainty threshold, or when the half-width of the confidence interval at multiple consecutive future cycle positions is greater than the preset uncertainty threshold. When the probability of the final degradation inflection point is greater than the preset inflection point threshold, an inflection point imminent warning is triggered. A life-end warning is triggered when the average predicted capacity at any future cycle position is lower than the preset life-end capacity threshold, or when the average predicted capacity sequence is lower than the preset life-end capacity threshold at several consecutive future cycle positions, or when the estimated remaining lifespan is less than the preset remaining lifespan warning threshold.

8. A method for predicting the degradation of automotive power batteries based on capacity recovery sensing according to any one of claims 1 to 3, characterized in that, Following step 6, the following is also included: Step 7: Based on the target domain test sample set, online running data, or subsequently collected capacity data, evaluate the performance of the target domain vehicle power battery capacity prediction model, and trigger model parameter updates, online incremental fine-tuning, data resampling, or warning threshold adjustment based on the evaluation results.

9. A vehicle power battery degradation prediction system with capacity recovery sensing, characterized in that, include: The data acquisition and preprocessing unit is used to acquire the cyclic operation data of the source domain vehicle power battery and the target domain vehicle power battery, and to perform preprocessing and sample partitioning to obtain the source domain training sample set, the source domain verification sample set, the target domain fine-tuning sample set, and the target domain test sample set. A model building unit is used to build a vehicle power battery degradation prediction model that senses capacity recovery. The vehicle power battery degradation prediction model includes: A capacity recovery sensing dual-branch variational autoencoder decouples and reconstructs a subnetwork, which is used to decouple the long-term degradation trend component and the short-term capacity recovery component, and outputs the reconstructed capacity sequence. The degradation degree conditional position encoding module is used to fuse the cyclic time position encoding and the degradation degree position encoding, and output the conditional position encoding corresponding to each cyclic position; The Transformer predicts the backbone subnetwork, whose input is the decoupled capacity degradation feature with added conditional position encoding, used to perform long-range time series modeling on the decoupled capacity degradation feature and extract long-range time series features. The decoupled capacity degradation feature is obtained by combining the reconstructed capacity sequence, the long-term degradation trend component, the short-term capacity recovery component and auxiliary operating condition features. The inflection point detection auxiliary subnetwork is used to identify and output the probability of a degradation inflection point occurring within a future preset loop window; The prediction output module is used to output future capacity prediction sequences and degradation risk auxiliary information; The multi-loss joint pre-training unit is used to perform multi-loss joint pre-training on the vehicle power battery degradation prediction model using the source domain training sample set to obtain the source domain pre-trained model. The domain adaptive transfer fine-tuning unit is used to transfer fine-tun the source domain pre-trained model with the target domain fine-tuning sample set, and introduces hierarchical differential learning rate configuration and maximum mean differential domain alignment loss to obtain the target domain vehicle power battery capacity prediction model. The uncertainty-aware online prediction unit is used to deploy the target domain vehicle power battery capacity prediction model in an online application environment. Through the Monte Carlo Dropout uncertainty-aware inference mechanism, it performs online capacity degradation prediction of the vehicle power battery to be predicted, calculates the predicted capacity mean sequence, the predicted confidence interval sequence and the probability of the final degradation inflection point, and determines the estimated remaining service life. The graded capacity degradation early warning unit is used to establish graded capacity degradation early warning judgment conditions. It performs graded capacity degradation early warning based on the predicted mean capacity sequence, predicted confidence interval sequence, probability of the final degradation inflection point, and estimated remaining useful life. It includes at least one or more of the following: uncertainty early warning, inflection point imminent warning, and end of useful life warning, and generates graded capacity degradation early warning results.

Images