A power battery degradation prediction method and system, a terminal device, and a medium

Abstract

This invention discloses a method, system, terminal device, and medium for predicting power battery degradation, relating to the field of power battery degradation prediction technology. The method includes: acquiring multi-source battery cycle data and extracting capacity information for each cycle according to type; constructing training labels and input / output windows based on the capacity information; constructing and pre-training an encoder, then fixing it, and mapping the window data to a historical latent state sequence via an input adapter; based on the historical latent state sequence, obtaining a student sequence through a historical representation branch and a teacher sequence through a future target anchor point branch, and constructing a representation alignment loss; starting from the historical latent state, given future actions and exogenous conditions, inferring future latent states and outputting prediction results; and training a degradation prediction model based on the prediction results and training labels through a joint loss function. This invention achieves multi-task prediction of battery degradation through student and teacher representation alignment and action condition inference, improving accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of power battery degradation prediction technology, and in particular to a power battery degradation prediction method, system, terminal equipment and medium. Background Technology

[0002] Power batteries gradually degrade during charge and discharge cycles. Accurately predicting their health degradation trajectory, inflection point, and remaining lifespan is crucial for battery safety management and operation and maintenance decisions.

[0003] In existing technologies, battery degradation prediction methods mainly rely on historical data to train a single prediction model, directly outputting a health status value or remaining life estimate. However, these methods have the following shortcomings: First, the data sources are diverse, and battery cycle data tables generated under different experimental or operating conditions vary significantly in recording granularity and field semantics. Existing methods lack the ability to identify and adapt to different data types, and directly mixing multi-source data can easily lead to labeling errors and information leakage. Second, existing methods can only provide a single numerical prediction based on historical data, and cannot answer the counterfactual question of how battery degradation results will change under different charge and discharge strategies, lacking the ability to extrapolate the battery aging process. Third, existing methods lack a verification mechanism during training to verify whether the potential state representation learned by the model effectively retains future degradation information, making it difficult to prove the reliability of the prediction results.

[0004] Therefore, there is an urgent need for a battery degradation prediction method that can be compatible with multi-source data, has the ability to extrapolate action conditions, and ensures prediction reliability through a characterization alignment mechanism, in order to fill the gap in existing technologies. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, in the field of power battery degradation prediction technology, existing methods cannot effectively process cycling data from multi-source heterogeneous batteries, lack the ability to extrapolate the battery degradation process under given future charge-discharge strategies, and struggle to verify whether the latent state representation learned by the model truly retains effective information reflecting future degradation. Therefore, an effective solution is urgently needed to address these technical problems.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for predicting the degradation of a power battery, the method comprising: Acquire multi-source battery cycle data, identify the data type of the multi-source battery cycle data, and extract capacity information of each cycle from the multi-source battery cycle data based on the extraction rules corresponding to the data type. Based on the capacity information, training labels are constructed, and historical cyclic input windows and future prediction target windows are built; wherein, the training labels include health status labels and lifespan-related labels; An encoder is constructed and pre-trained to obtain a solidified encoder; an input adapter is constructed at the input end of the solidified encoder, and the battery cycle data in the historical cycle input window is mapped to a historical latent state sequence through the input adapter and the solidified encoder; Based on the historical latent state sequence, a student latent state sequence is obtained through a historical representation branch; based on the real future data in the future prediction target window, a teacher anchor sequence is obtained through a future target anchor sequence; based on the difference between the student latent state sequence and the teacher anchor sequence, a representation alignment loss is constructed. Starting with the historical potential state sequence, and given the future action sequence and exogenous condition sequence, the future potential state sequence is deduced, and the prediction result is output based on the future potential state sequence. Based on the prediction results and the training labels, the model is trained using a joint loss function to obtain a trained battery degradation prediction model; wherein, the joint loss function is at least composed of a weighted sum of a loss term corresponding to the representation alignment loss and a task loss term determined based on the difference between the prediction results and the training labels.

[0007] In one implementation, acquiring multi-source battery cycle data, identifying the data type of the multi-source battery cycle data, and extracting capacity information for each cycle from the multi-source battery cycle data based on extraction rules corresponding to the data type includes: Obtain battery cycle data tables from different data sources, identify the data types of each data table, and the data types characterize the record granularity and field semantics of the data tables; Based on the identified data type, the capacity extraction rules corresponding to that data type are used to extract the capacity information corresponding to each battery cycle from each data table. Based on preset capacity thresholds or degradation index thresholds, the extracted capacity information is filtered for validity to obtain valid capacity information for subsequent steps.

[0008] In one implementation, the step of constructing training labels based on the capacity information and building a historical cyclic input window and a future prediction target window includes: Based on the capacity information, the battery cycle sequence is checked for cycle continuity and missing values ​​are counted, and cycles with abnormalities or missing values ​​are marked. Based on the capacity information, health status labels and lifespan-related labels are constructed; wherein, the health status labels include the health status values ​​of each cycle, and the lifespan-related labels include at least one of lifespan end, inflection point location, and degradation stage category; The availability of labels is marked to generate a label availability mask, which is used to mask invalid or unavailable labels during training; Based on the preset historical window length and predicted field of view length, the historical cycle input window and the future prediction target window are constructed by sliding from the battery cycle sequence.

[0009] In one implementation, the process of constructing and pre-training an encoder to obtain a fixed encoder, constructing an input adapter at the input end of the fixed encoder, and mapping the battery cycle data in the historical cycle input window to a historical latent state sequence through the input adapter and the fixed encoder, includes: Construct an encoder that encodes multidimensional time-series data from a single battery cycle into a fixed-dimensional latent state vector. Based on self-supervised learning, the encoder is pre-trained using at least one of the mask feature recovery task and the adjacent cyclic dynamic prediction task to obtain pre-trained encoder parameters. The pre-trained encoder parameters are fixed so that the encoder parameters are no longer updated; An input adapter is constructed at the input end of the solidified encoder. The input adapter is used to map raw battery cycling data of different dimensions to an input dimension acceptable to the solidified encoder. The parameters of the input adapter can be learned during the training process of downstream tasks.

[0010] In one implementation, the step of obtaining a student latent state sequence through a historical representation branch based on the historical latent state sequence, obtaining a teacher anchor sequence through a future target anchor branch based on the real future data in the future prediction target window, and constructing a representation alignment loss based on the difference between the student latent state sequence and the teacher anchor sequence includes: Construct a historical representation branch, which includes a cascaded student mapping module and a student projection module, for mapping the historical latent state sequence to a sequence of student latent states; A future target anchor branch is constructed, which includes a target projection module for generating a future descriptor from the real future data in the future prediction target window and generating a teacher anchor vector based on the future descriptor; wherein, the teacher anchor vector stops gradient backpropagation during training, and the future target anchor branch is only used during the model training phase; Calculate the vector distance between the vector of the sequence of the student's latent states projected by the student projection module and the teacher's anchor vector, and construct the representation alignment loss based on the vector distance weighted by the labeled availability mask.

[0011] In one implementation, the step of using the historical potential state sequence as a starting point, and under the given conditions of a future action sequence and an exogenous condition sequence, to deduce a future potential state sequence, and output a prediction result based on the future potential state sequence, includes: A shared decoding backbone is constructed, and multiple prediction task output heads are mounted on the shared decoding backbone; wherein, the multiple prediction task output heads include at least two of the following: capacity trajectory output head, health status output head, degradation stage output head, inflection point detection output head, and remaining lifetime distribution output head; An action-condition latent transition module is constructed, which includes an action coding branch and a gated residual fusion unit; wherein, the action coding branch is used to encode the future action sequence and the exogenous condition sequence, and the gated residual fusion unit is used to fuse the action coding result with the current latent state to generate the latent state of the next time step; Using the potential state of the last time step of the historical potential state sequence as the initial state, and given the future action sequence and the exogenous condition sequence, the future potential state sequence is obtained by recursively deducing step by step using the action condition potential transition module. The sequence of future potential states is input into the shared decoding backbone, and the corresponding prediction results are output in parallel through the output heads of each prediction task.

[0012] In one implementation, the method further includes: Obtain historical cycle data of the battery to be predicted within the most recent preset number of cycles, as well as preset future action sequences and exogenous condition sequences; The historical cycle data, the future action sequence, and the exogenous condition sequence are input into the trained battery degradation prediction model. The historical latent state sequence is obtained through the input adapter, the solidified encoder, and the historical representation branch. Starting from the historical latent state sequence, the future latent state sequence is recursively deduced step by step under the conditions of the future action sequence and the exogenous condition sequence. Based on the future latent state sequence, the prediction result of the battery to be predicted under multiple future cycles is output. The prediction results include at least one of the following: capacity degradation trajectory, health status change curve, degradation stage sequence, inflection point probability distribution, and remaining lifetime distribution.

[0013] Secondly, embodiments of the present invention also provide a power battery degradation prediction system, the system comprising: The data acquisition module is used to acquire multi-source battery cycle data, identify the data type of the multi-source battery cycle data, and extract the capacity information of each cycle from the multi-source battery cycle data based on the extraction rules corresponding to the data type. The training sample construction module is used to construct training labels based on the capacity information, and to build a historical cyclic input window and a future prediction target window; wherein, the training labels include health status labels and lifespan-related labels; A latent state encoding module is used to construct an encoder and perform pre-training to obtain a solidified encoder; an input adapter is constructed at the input end of the solidified encoder, and the battery cycle data in the historical cycle input window is mapped to a historical latent state sequence through the input adapter and the solidified encoder; The representation alignment loss construction module is used to obtain the student latent state sequence through the historical representation branch based on the historical latent state sequence; to obtain the teacher anchor sequence through the future target anchor branch based on the real future data in the future prediction target window; and to construct the representation alignment loss based on the difference between the student latent state sequence and the teacher anchor sequence. The latent state deduction and prediction module is used to deduce the future latent state sequence starting from the historical latent state sequence, given the future action sequence and the exogenous condition sequence, and output the prediction result based on the future latent state sequence. A joint training module is used to train the model using a joint loss function based on the prediction results and the training labels to obtain a trained battery degradation prediction model; wherein the joint loss function is at least composed of a weighted loss term corresponding to the representation alignment loss and a task loss term determined based on the difference between the prediction results and the training labels.

[0014] Thirdly, embodiments of the present invention also provide a terminal device, the terminal device including a memory, a processor, and a power battery degradation prediction program stored in the memory and executable on the processor, wherein when the processor executes the power battery degradation prediction program, it implements the steps of the power battery degradation prediction method described in any of the above schemes.

[0015] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a power battery degradation prediction program. When the power battery degradation prediction program is executed by a processor, it implements the steps of the power battery degradation prediction method described in any of the above schemes.

[0016] Beneficial Effects: This invention discloses a method, system, terminal device, and medium for predicting power battery degradation, relating to the field of power battery degradation prediction technology. The method first acquires multi-source battery cycling data, identifies the data type of the multi-source battery cycling data, and extracts capacity information for each cycle from the multi-source battery cycling data based on extraction rules corresponding to the data type. Then, based on the capacity information, training labels are constructed, and a historical cycle input window and a future prediction target window are built, wherein the training labels include health status labels and lifespan-related labels. Subsequently, an encoder is constructed and pre-trained to obtain a fixed encoder. An input adapter is constructed at the input end of the fixed encoder, and the battery cycling data in the historical cycle input window is mapped to a historical latent state sequence through the input adapter and the fixed encoder. Then, based on the historical latent state sequence, a student latent state sequence is obtained through a historical representation branch, and a teacher anchor sequence is obtained through a future target anchor branch based on the real future data in the future prediction target window. Based on the difference between the student latent state sequence and the teacher anchor sequence, a representation alignment loss is constructed. Next, starting with the historical latent state sequence, and given the future action sequence and exogenous condition sequence, the future latent state sequence is deduced, and a prediction result is output based on the future latent state sequence. Finally, based on the prediction result and the training labels, the model is trained using a joint loss function to obtain a trained battery degradation prediction model; wherein the joint loss function is at least a weighted sum of the loss term corresponding to the representation alignment loss and the task loss term determined based on the difference between the prediction result and the training labels.

[0017] This invention employs a student-teacher representation alignment mechanism, utilizing a future target anchor branch that provides supervision signals only during the training phase. This allows the historical representation branch to learn a latent state representation that retains future degradation information, thus avoiding future information leakage. Furthermore, through action-conditional latent deduction, the model can recursively deduce the battery degradation process given different future charge / discharge strategies and exogenous conditions, achieving multi-task prediction and supporting counterfactual comparisons. In addition, by integrating representation alignment loss and multi-task loss through a joint loss function, the various modules of the model are synergistically optimized, improving the accuracy and reliability of battery degradation prediction. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a specific implementation of the power battery degradation prediction method provided in this invention.

[0019] Figure 2 This is a schematic diagram of the overall process for predicting the degradation of power batteries using a world model, as provided in an embodiment of the present invention.

[0020] Figure 3This diagram illustrates the encoder source, student and target anchor points, and multi-task transfer structure.

[0021] Figure 4 This diagram illustrates the relationship between battery characterization levels and the predicted object.

[0022] Figure 5 This is a comparison chart of the model and baseline for the mean absolute error of the health status in the twenty-cycle capacity and health status during historical engineering verification.

[0023] Figure 6 This is a schematic diagram illustrating the improvement of the mean absolute error of the residual calibration model relative to the baseline for different predicted spans during historical engineering verification.

[0024] Figure 7 This is a schematic diagram of the power battery degradation prediction system provided in an embodiment of the present invention.

[0025] Figure 8 This is a block diagram illustrating the internal structure of the terminal device provided in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0028] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, the first control information and the second control information are only used to distinguish different control information and do not limit their order.

[0030] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0031] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0032] Lithium-ion batteries are widely used in new energy vehicles, energy storage systems, aerospace power, and industrial equipment. With increasing charge-discharge cycles, batteries exhibit phenomena such as capacity decay, increased internal resistance, enhanced polarization, abnormal temperature rise, and degradation mode migration. Accurately characterizing the current degradation state of the battery and predicting the degradation consequences under different future operating conditions or control actions is fundamental to developing maintenance strategies, reducing safety risks, and improving overall lifecycle economics.

[0033] Existing technologies typically regress health status, capacity, or remaining lifetime directly from historical observables, with outputs often being single point-in-time or single scalar quantities. These methods struggle to answer the counterfactual question of "what differences will arise when different future actions are taken in the same current state," and also find it difficult to verify whether hidden representations retain useful information about future voltage, current, temperature, and curve morphology. World models offer a unified technical approach: forming latent states from historical observations, driving the evolution of latent states through actions, and then decoding future observations from latent states.

[0034] The raw battery data also suffers from various issues, including the coexistence of multiple table structures, mixing of cumulative and cycle data, duplicate cycle numbers across segments, missing complete cycle segments, unobserved end-of-life, and short-term capacity rebound caused by measurement noise. If capacity is uniformly generated based on the intra-phase range, or if the window spans missing segments, dataset boundaries, or discontinuous cycles, incorrect labels or future information leakage will occur. Therefore, before training the world model, protocol-aware data reconstruction, cycle continuity auditing, censored label masking, and training boundary isolation must be performed.

[0035] Furthermore, the input dimensions, sampling methods, and semantic spaces of external underlying time series models are often inconsistent with battery cycle data. Simply freezing an undefined encoder and connecting it to a regression head fails to clarify the encoder source, input adaptation method, gradient path, timing of teacher branch usage, and parameter sharing relationships among prediction branches. Therefore, this invention provides reproducible structural constraints on encoder pre-training and solidification, input adaptation, training-period target anchors, shared decoding backbone, independent task heads, and potential action condition transitions.

[0036] Therefore, this embodiment aims to provide a trainable, verifiable, and deployable world model-based method for predicting the degradation of power batteries. This method enables the model to form continuous potential battery states from historical observables, gradually deducing the degradation process under controllable actions and exogenous conditions, and using future observable states as direct supervision and verification of the sufficiency of the representation. Simultaneously, it outputs capacity or health state trajectories, potential degradation stages, kneepoint events, and distributions of end-of-life (EOL) or remaining useful life (RUL) through a unified potential trajectory. The overall technical approach is as follows: Figure 2 As shown, from top to bottom, it is divided into three stages: data and sample construction, representation learning and latent dynamics, and multi-task output and verification.

[0037] This embodiment provides a method for predicting the degradation of power batteries, such as... Figure 1 As shown, the specific steps include the following: Step S100: Obtain multi-source battery cycle data, identify the data type of the multi-source battery cycle data, and extract the capacity information of each cycle from the multi-source battery cycle data based on the extraction rules corresponding to the data type.

[0038] In this embodiment, the multi-source battery cycle data refers to battery charge-discharge cycle records from different experimental platforms, operating conditions, or data acquisition systems. The data format can be a structured data table containing fields such as channel number, global cycle number, internal cycle number, step number, step status, sampling time, voltage, current, power, temperature, capacity, energy, state of charge, and protocol settings. Due to the diverse data sources, these data tables differ significantly in recording granularity and field semantics. Indiscriminate mixing and use of these tables will lead to semantic errors in capacity labels and future information leakage.

[0039] The data type represents the record granularity and field semantics of the data table. Specifically, it can be divided into three types based on the table header structure, record granularity, and field monotonicity: detailed process table, step summary table, and processing result table.

[0040] The detailed process table records the changes in physical quantities such as voltage, current, and temperature within each charge-discharge cycle in time series form, sampling point by sampling point; the step summary table records the cumulative charge, energy, and time of each charge-discharge step in steps as the basic unit; and the processing result table directly provides post-processed conclusive data such as the final capacity and state of health (SOH) of each cycle.

[0041] The extraction rules refer to the capacity value acquisition logic formulated for each data type, consistent with its recording semantics. For detailed process tables, within the same channel, the same global cycle, and the same charge / discharge phase, phase capacity is generated based on the effective range difference or final value of the accumulated amount, and resampled curves can be retained for subsequent curve feature extraction; for step summary tables, cycle-level capacity is generated using the maximum value or final value of the charging or discharging steps, and curve reconstruction is not performed for non-existent process points; for processing result tables, matching is performed according to segment identifier, channel number, and global cycle number, and these are given priority as sources of capacity tags to improve tag quality.

[0042] The capacity information refers to the effective capacity value corresponding to each battery cycle extracted from various data tables, with discharge capacity usually being the primary reference.

[0043] Through protocol-aware data reconstruction processing, battery cycling data from different sources and in different formats are unified into a capacity sequence with consistent semantics. This effectively avoids semantic errors in capacity labels caused by mixed table structures and provides a reliable data foundation for the accurate construction of training labels.

[0044] In one implementation, the steps of acquiring multi-source battery cycle data, identifying the data type of the multi-source battery cycle data, and extracting capacity information for each cycle from the multi-source battery cycle data based on extraction rules corresponding to the data type specifically include the following steps: Step S110: Obtain battery cycle data tables from different data sources, identify the data types of each data table, and the data types characterize the record granularity and field semantics of the data tables; Step S120: Based on the identified data type, use the capacity extraction rule corresponding to the data type to extract the capacity information corresponding to each battery cycle from each data table; Step S130: Based on a preset capacity threshold or degradation index threshold, the extracted capacity information is filtered for validity to obtain valid capacity information for subsequent steps.

[0045] In this embodiment, data type identification can be automatically performed based on the column name characteristics and field contents of each data table. The detailed process table typically includes a sampling time column and voltage and current sequences recorded point by point, with its timestamps exhibiting monotonically increasing characteristics; the typical characteristic of the step summary table is that each step has only one record, containing cumulative quantity fields for the start and end times of the step; the processing result table typically uses the channel number and cycle number as primary keys, directly providing the final value of the capacity or discharge capacity for each cycle.

[0046] In one specific implementation, the capacity extraction rule obtains the discharge capacity in the following manner:

[0047] in, Indicates the battery or channel number. Indicates the global cyclic number. This represents the set of discharge records confirmed by step state and current direction in this cycle. For the discharge capacity field in this record, This is the final effective cycle discharge capacity. When the field is a cumulative amount within a cycle, the final or maximum value is taken; when the field is a cumulative amount within a detailed process and a reset occurs, the effective range difference is calculated within the same phase, i.e., the phase termination value minus the starting value. To ensure that the results of different rules do not interfere with each other, capacity information obtained from different rules must not be mixed without auditing.

[0048] In the validity screening, the preset capacity threshold can be set according to the battery's rated capacity; for example, capacity values ​​exceeding 120% of the rated capacity or falling below 20% are considered abnormal. The degradation index threshold can be determined by combining the quantiles of the capacity decay rate in the training set. The filtered valid capacity information serves as input for subsequent label construction, while metadata such as source file, channel, start and end loops, and target loops are retained for each sample for verification. Through the above processing, heterogeneous data from different sources are unified into capacity sequences with consistent semantics, eliminating data quality issues at the source and ensuring a reliable data foundation for the labels used in subsequent model training.

[0049] In one specific implementation, after fully reading all original tables, a three-tiered index—segmented, channel-based, and global loop-based—is established, prioritizing the matching of capacity labels in the processing results table. The detailed process table is used to generate resampling curves to extract loop-level curve features, while the step summary table is only used to extract loop-level statistical indicators.

[0050] Step S200: Based on the capacity information, construct training labels and build a historical cyclic input window and a future prediction target window; wherein, the training labels include health status labels and lifespan-related labels.

[0051] In this embodiment, the training labels refer to the set of real target values ​​used to supervise model learning. The health status label characterizes the degree of battery degradation in the current cycle and is defined as the ratio of each cycle's capacity value to the rated capacity. Specifically, it is stored in decimal form; for example, a health status of 92% is recorded as 0.92. The rated capacity can be given by the battery model specifications or determined by a specified baseline cycle capacity, but must remain consistent within the same experimental protocol.

[0052] The lifespan-related labels include various annotation information closely related to the battery's entire life cycle degradation process, including the end of life, the location of the inflection point, and the type of degradation stage.

[0053] For lifetime endpoints, a judgment rule based on continuously low thresholds and verifiable cycle intervals is used. Samples for which lifetime endpoints have not been observed are considered right-censored samples, and their remaining lifetimes must not be forcibly set to zero. Right censoring is a concept in survival analysis, referring to a situation where, at the end of the observation period, the target event has not yet occurred for the studied object; therefore, it is only known that the event occurred after the current observation period, but the specific timing is unknown. For example, a battery may have run 400 cycles at the end of the dataset, with a health status of 92%, but has not yet reached its lifetime endpoint (e.g., its health status drops below 80%). At this point, it is only known that its lifetime is greater than 400 cycles, but its remaining lifespan is unknown; this is a right-censored sample.

[0054] Inflection point labels are obtained through segmented slope change detection. When the minimum number of samples is satisfied on both sides of a candidate inflection point and the absolute value of the decay slope after the inflection point increases significantly, the inflection point is marked as identifiable. If the fit improvement is insufficient or the candidate point is at the boundary, it is marked as unidentifiable.

[0055] The degradation stage labels are constructed by combining the local decay rate, temperature rise change, coulombic efficiency, and current stress after the training set capacity curve is robustly smoothed. For example, they are divided into categories such as stable degradation stage, accelerated degradation stage, high-speed degradation stage, and short-term recovery stage.

[0056] The historical cycle input window refers to the multi-dimensional feature sequence extracted from a preset number of consecutive battery cycles, including the cycle statistical features and curve-level features of each cycle. The input window only contains the feature information of the current and historical cycles. Future capacity, future health status, future curves, future stage labels and life events must not be included in the historical input. This constraint ensures that there is no leakage of future information during the training and inference process.

[0057] Specifically, let the single-cycle feature vector be... , Indicates the first There are global loops, and the feature dimension is . Each cycle stage's characteristics can encompass statistical indicators such as normalized discharge capacity, coulombic efficiency, average temperature, average voltage, and internal resistance estimation, and may also include resampled voltage-capacity curve characteristics or capacity increment curve characteristics. The historical cycle input window consists of continuous... The eigenvectors of each cycle are arranged in an ordered manner, and are represented as follows:

[0058] in, For the length of the history window, The dimension of the unicyclic feature.

[0059] The future prediction target window refers to the set of prediction objects consisting of a preset number of cycles after the current cycle. Both are constructed in a sliding manner on the complete battery cycle sequence.

[0060] Specifically, let the predicted span be... The unified prediction object set can be represented as:

[0061] in, This is the predicted value for the capacity trajectory. For health status trajectory prediction values, For the probability distribution of inflection point events, This represents the probability distribution of lifespan end events. For future observable state sequences, This represents the probability distribution for future degradation stages. This represents the difference in degradation consequences between the two action sequences. In practice, one or more of these tasks can be enabled as prediction tasks based on label availability. Output headers for tasks that are not enabled should not be considered as having completed evaluation. The correspondence between the above multi-level prediction objects is as follows: Figure 4 As shown, it covers the complete hierarchy from bottom to top, from raw observations, trajectory representation, degradation state to lifetime events and decision outputs.

[0062] Through the above label construction and window building process, each training sample has a clear historical input boundary and future prediction target boundary, and the availability of various labels is explicitly marked.

[0063] In one implementation, the step of constructing training labels based on the capacity information and building a historical cyclic input window and a future prediction target window specifically includes the following steps: Step S210: Based on the capacity information, perform cycle continuity checks and missing value statistics on the battery cycle sequence, and mark cycles with abnormalities or missing values. Step S220: Based on the capacity information, construct a health status label and a lifespan-related label; wherein, the health status label includes the health status value of each cycle, and the lifespan-related label includes at least one of the lifespan end point, inflection point location, and degradation stage category; Step S230: Mark the availability of the labels and generate a label availability mask. The label availability mask is used to mask invalid or unavailable labels during training. Step S240: Based on the preset historical window length and predicted field of view length, slide to construct the historical cycle input window and the future prediction target window from the battery cycle sequence.

[0064] In this embodiment, consistency audits are performed on physical quantities such as voltage, current, temperature, capacity, energy, and time. The range check requires each physical quantity to fall within the physically permissible range of the battery system; for example, the cell voltage should be within the chemical window, and the temperature should be within a reasonable operating range. The direction check verifies whether the recorded current direction is consistent with the charge / discharge phase marked in the step status; inconsistencies are flagged as abnormal. The monotonicity check examines the overall monotonically decreasing characteristic of the discharge capacity over the number of cycles, excluding short-term recovery effects. The sampling order check ensures that the timestamps within the same cycle are strictly monotonically increasing, without disorder or duplication. The power integral consistency check compares the accumulated energy obtained by integrating the product of voltage and current over time with the recorded energy field; if the deviation exceeds a preset threshold, the cycle is flagged as physically unreliable. Abnormal cycles flagged by the physical consistency audit, along with anomalies found in subsequent cycle continuity checks, are entered into the sample filtering process.

[0065] Subsequently, data quality is ensured through a cycle continuity check. Specifically, after establishing a global cycle index for each channel, the beginning and end of segments, cycle gaps, cross-segment repetitions, missing phases, and conflicts when multiple sources of data exist within the same cycle are explicitly recorded. For short gaps, interpolation is only allowed to fill them within the same cycle and the same step; for gaps in complete cycle segments, any form of filling is prohibited, and no training window or label window can cross the boundary of the gap. After the above audit, metadata such as source file identifier, channel number, window start and end global cycle number, target cycle number, adjacent cycle difference sequence, continuity flag, and label availability mask are saved for each valid sample. This allows for backtracking of the data source and audit results of samples as needed during training, and also enables source tracing of abnormal prediction results after model deployment.

[0066] Each input window stores the window's start and end global cycle numbers, adjacent cycle differences, missing boundary markers, and target index. If the difference between any adjacent global cycle numbers is greater than 1, or if the window crosses the dataset partition boundary, the sample is not included in training or evaluation. For the end of lifetime, remaining lifetime, and inflection point, the label value, whether the event occurred, the censoring type, and the supervision mask are stored respectively, so as to precisely control the participation of each loss term according to the label availability mask during training.

[0067] The training, validation, and test sets are divided according to battery, channel, operating condition, or time boundary, without random shuffling. Standardized parameters, zero-variance deletion, feature selection, and other data-driven thresholds are obtained only from the training set and then used as fixed values ​​for the validation and test sets to avoid cross-set information leakage.

[0068] Regarding tag construction, the first Health status labels in a global loop Defined as the effective capacity of each cycle With rated capacity The ratio is expressed as:

[0069] The rated capacity can be given by the battery model or determined by the specified baseline cycle capacity, but it must be consistent within the same experimental protocol. When stored in decimal system, 0.01 corresponds to one percentage point.

[0070] The lifespan end is determined using a method that employs continuously low thresholds and verifiable cycle intervals. If continuous... The number of cycles is below the threshold And the difference between adjacent global loop numbers is no greater than 1, which can be represented as:

[0071]

[0072] in, Indicates the end-of-life cycle. This is the lifespan threshold, which can be set to 0.8, meaning that when the health status drops to 80%, the lifespan end determination is triggered. To confirm the continuous length, for example, take To eliminate occasional capacity fluctuations, The difference between adjacent global cycle numbers. This is an indicator function. Remaining useful life. Take the end-of-life loop and the current loop The difference is not less than zero.

[0073] When no end-of-life condition meeting the above criteria is observed in the battery cycling sequence, a supervision mask is used. Marking the sample as zero indicates that the lifetime end label for that sample is unavailable; this is the supervision mask for the inflection point. Also used to mark whether an inflection point is identifiable, represented as:

[0074] in For the set of observable lifetime ends, This is the set of identifiable inflection points.

[0075] Inflection points can be detected through continuous line regression or equivalent piecewise slope changes, denoted as:

[0076] in, The candidate inflection points are numbered cyclically. The capacity decay slope before the inflection point. and The sum of these is the decay slope after the inflection point. This is the random error term. The minimum sample size requirement must be met on both sides of the candidate inflection point, and the absolute value of the slope after the inflection point must be significantly greater than that before the inflection point. The label is marked as unidentifiable when there is a cross-cycle gap, insufficient improvement in fit, or when the candidate point is at the sequence boundary.

[0077] The number of stage categories is denoted as The 70th and 90th quantiles of the decay rate are estimated only on the training set and used as a reference for the stage division threshold. Let... To end the loop Given the complete historical observation sequence, after the model outputs the normalized posterior probabilities for each stage, the final stage label is determined by the following formula:

[0078] in, The total number of stage categories, To end the loop The probability distribution of each stage under historical observation. Labels for potential degradation stages. In addition to the above-mentioned method of directly constructing labels based on statistics such as local decay rate, weak labels can also be obtained by using hidden Markov models, switching state space models, or clustering algorithms. However, regardless of the label generation method used, the label generation algorithm and its parameter thresholds must be fixed on the training set before the validation and test sets are processed according to certain rules. The label generation logic must not be dynamically adjusted during the validation or testing phase.

[0079] Each sample is associated with a corresponding binary mask for each label type. A value of 1 indicates that the label is available for supervised training, while a value of 0 indicates that the corresponding loss term is not included in normalization. This label availability mask ensures that the model does not use missing lifetime endpoints, unidentifiable inflection points, or unobserved degradation events as erroneous supervisory signals. During window construction, samples are generated cyclically across the entire battery cycle sequence in a sliding manner based on a preset historical window length and prediction field of view length (also known as prediction span or prediction step size), while ensuring that the window of any sample does not cross the dataset partition boundaries or discontinuous cycle segments.

[0080] Step S300: Construct an encoder and perform pre-training to obtain a solidified encoder; construct an input adapter at the input end of the solidified encoder, and map the battery cycle data in the historical cycle input window into a historical potential state sequence through the input adapter and the solidified encoder.

[0081] In this embodiment, the encoder is the component in the overall model architecture responsible for compressing the multidimensional time-series data of a single battery cycle into a low-dimensional latent representation. Its role is to automatically extract key features closely related to the battery degradation state from the original charge-discharge curves. The original battery cycle data typically contains long-term sequences of multiple channels such as voltage, current, and temperature, sampled at frequencies of one second or even higher. A single cycle may contain hundreds to thousands of sampling points. This high-dimensional raw data simultaneously contains degradation signals, measurement noise, and operating condition fluctuation information. The encoder maps this high-dimensional complex data into a compact, fixed-dimensional vector, which serves as the latent state of the cycle, condensing its degradation characteristics.

[0082] The pre-training stage is the first phase of the overall model training process, employing self-supervised learning to independently initialize and optimize the encoder's parameters. During pre-training, the inherent patterns of battery cycling data are used to construct the training objective, ensuring the encoder has a basic understanding of the battery data's distribution and temporal evolution patterns before tackling specific downstream tasks. The pre-trained encoder parameters, obtained after pre-training, provide a high-quality parameter initialization starting point for downstream multi-task training, significantly accelerating the convergence speed of downstream training and improving the final prediction accuracy.

[0083] Solidified encoders preserve and freeze the encoder parameters acquired during the pre-training phase, preventing them from being updated via gradient backpropagation during downstream multi-task training. Solidification protects the encoder's ability to extract general battery degradation features learned during pre-training, preventing specific loss signals from downstream tasks from compromising its fundamental representational capabilities and avoiding catastrophic forgetting. Furthermore, a learnable input adapter is introduced on top of the solidified encoder to handle data format bridging, allowing the model to flexibly adapt to battery cycling data from different sources and dimensions without altering its core encoding capabilities. The downstream loss function values ​​flow through the solidified encoder's forward propagation and then propagate normally back to the upper-layer input adapter.

[0084] The input adapter is a learnable preprocessing module built before the input of the solidified encoder. It maps raw battery cycling data of different dimensions to a fixed input dimension acceptable to the solidified encoder. This raw battery cycling data may come from different experimental platforms and have different numbers of channels and sampling frequencies. Since the pre-training stage is completed on a specific data format, the input dimension of the solidified encoder is fixed, while the data dimension in actual applications may be inconsistent with this. The input adapter acts as a bridge between the two. The parameters of the input adapter can be learned and participate in gradient updates during downstream multi-task training, thus automatically adjusting the data mapping strategy according to specific task requirements.

[0085] The historical latent state sequence refers to a series of latent state vectors arranged in chronological order, obtained after multiple consecutive battery cycle data in the historical input window are sequentially processed by the input adapter's dimension mapping and the solidified encoder's feature extraction. Each latent state vector corresponds to a historical cycle within the window, and all vectors together form an ordered sequence that fully records the battery's degradation state evolution process within the historical window period. The historical latent state sequence is used for representation learning, action condition inference, and prediction decoding of the entire model. The historical representation branch extracts student latent states from this sequence, and the action condition latent transition module recursively infers forward from its final state.

[0086] Through encoding and mapping, the original high-dimensional heterogeneous battery cycle data is transformed into a latent state representation with unified structure and continuous semantics, providing a representational basis for the simulation and deduction of the world model.

[0087] In one implementation, the process of constructing and pre-training an encoder to obtain a solidified encoder, constructing an input adapter at the input end of the solidified encoder, and mapping the battery cycle data in the historical cycle input window to a historical latent state sequence through the input adapter and the solidified encoder, specifically includes the following steps: Step S310: Construct an encoder, which is used to encode multidimensional time-series data of a single battery cycle into a latent state vector of fixed dimensions; Step S320: Based on self-supervised learning, the encoder is pre-trained using at least one of the mask feature recovery task and the adjacent cyclic dynamic prediction task to obtain pre-trained encoder parameters. Step S330: Solidify the pre-trained encoder parameters so that the encoder parameters are no longer updated; Step S340: Construct an input adapter at the input end of the solidified encoder. The input adapter is used to map raw battery cycling data of different dimensions to an input dimension acceptable to the solidified encoder. The parameters of the input adapter can be learned during the training process of downstream tasks.

[0088] In this embodiment, the encoder is responsible for compressing the raw battery cycle data into a low-dimensional latent representation. Its function is to map the multi-dimensional temporal data of a single battery cycle into a fixed-dimensional latent state vector, which condenses the key degradation features of that cycle. The encoder can be built based on a Transformer architecture with a self-attention mechanism, or alternatively, a temporal convolutional network or a long short-term memory network can be used.

[0089] The pre-training refers to the independent parameter initialization training of the encoder using self-supervised learning before downstream multi-task training. The pre-training phase does not require any capacity or lifetime labels; instead, it utilizes the inherent patterns of battery cycling data to construct the training objective. The joint objective of pre-training is:

[0090] in, The masked feature recovery loss works by randomly masking a portion of the input features and then requiring the encoder to recover the masked feature values ​​from the context, thus forcing the encoder to learn the intrinsic correlation of the battery cycle data. To dynamically predict loss between adjacent cycles, the encoder is required to predict the change in the potential representation of the next cycle from the potential representation of the current cycle, so that the encoder can capture the temporal evolution of the battery degradation process. The curve reconstruction loss requires reconstructing the original charge-discharge curve from the latent representation, ensuring that the latent representation does not lose key information. , and These are the non-negative weights of the three losses mentioned above.

[0091] The pre-trained encoder parameters are obtained after pre-training. This parameter is then fixed. Fixed parameter refers to the encoder parameters being fixed during downstream multi-task training. Updates are no longer performed via gradient backpropagation. Stabilizing the encoder parameters only means that their parameter gradients are not computed; the encoder's forward computation graph must still be fully preserved so that the downstream loss can be backpropagated correctly to the input adapter's parameters. This gradient design ensures that the adapter can still learn while the encoder is stabilized.

[0092] The input adapter maps raw battery cycle data of different dimensions to a fixed input dimension acceptable to the solidified encoder. A preferred structure is a two-layer multilayer perceptron, which vectorizes the historical cycle input window, maps it to a 512-dimensional hidden layer, and then maps it to a 256-dimensional latent input. The historical cycle window is represented as:

[0093] in Indicates the length of the history window. Indicates the first i The observed inputs correspond to each time step. To preserve the temporal state information during the historical loop, this paper first performs independent feature mapping on the input at each time step within the historical window. The input adapter is represented as:

[0094] Expanded to:

[0095] in, Indicates input adapter, Its trainable parameter set , This indicates that the historical loop input window is flattened into a vector. For layer normalization, The activation function is a Gaussian error linear unit. After adapter mapping, it is fed into the solidified encoder to obtain the latent state. The adapter parameters... It can be learned during downstream task training.

[0096] The data mapped by the input adapter is sent to the firmware encoder to obtain the potential states of each cycle:

[0097] in, This represents the forward mapping function of the firmware encoder. The encoder parameters are obtained and fixed during the pre-training phase. This means that the encoder parameters remain numerically unchanged from the pre-training results. This indicates that the gradients of the encoder parameters are always zero during downstream task training. The stop gradient operation, which covers the entire encoder forward propagation, is not used to replace the above parameter freezing method because it would cut off the gradient propagation path of the loss function to the input adapter parameters, preventing the adapter parameters from being updated by downstream tasks. The correct implementation is to freeze the encoder parameter set in the optimizer, or to apply zero gradient updates only to the encoder parameters after backpropagation, while maintaining the integrity of the forward computation graph.

[0098] The encoder focuses on learning a general representation of battery degradation features during the pre-training phase, while the adapter takes on the task of bridging specific data formats to the encoder's semantic space.

[0099] Besides the aforementioned internal self-supervised pre-training method, another approach to obtaining encoder parameters is to import a structurally compatible external pre-trained temporal encoder, and then use a battery input adapter to unify the battery cycling data to the input dimension of this external encoder. The external pre-trained model is not a necessary condition for implementing this invention, and a locally randomly initialized surrogate model that has not been pre-trained should not be presented as inheriting the capabilities of an external pre-trained model. Both approaches are handled consistently during downstream task training; that is, the encoder parameters are not updated after being fixed, and gradients only flow to the input adapter.

[0100] In one implementation, the history window length is 10 cycles. The input adapter fuses the cyclic statistical features, resampling curve features, and missing mask, then maps them uniformly to 256 dimensions. After solidifying the encoder, the gradient switch of its parameters is turned off, but the entire encoder forward pass must not be placed within a stopping gradient context; otherwise, the learning path of the adapter parameters will be cut off. As a verifiable implementation, after the first backpropagation during downstream task training, the gradient norm of the input adapter parameters is audited to see if it is greater than zero, and the gradient of the encoder parameters is empty or zero. The audit results are written to the training log. This gradient audit directly verifies whether the gradient contract between the solidified encoder and the learnable adapter is correctly implemented. The structural relationship between the encoder solidification, input adaptation, and upstream and downstream representation branches is as follows: Figure 3 As shown, the differentiated pathways between the training and inference stages are fully presented.

[0101] Step S400: Based on the historical latent state sequence, obtain the student latent state sequence through the historical representation branch; based on the real future data in the future prediction target window, obtain the teacher anchor sequence through the future target anchor branch; based on the difference between the student latent state sequence and the teacher anchor sequence, construct the representation alignment loss.

[0102] In this embodiment, the historical representation branch refers to a data processing path that exists throughout the entire model lifecycle. Its function is to extract a sequence of student latent states from historical cyclic data. The historical representation branch includes training, validation, testing, and deployment phases. This branch takes the latent features obtained from the historical window through the input adapter and the solidified encoder as input, and forms a continuous sequence of student latent states through a battery-specific mapping.

[0103] The future target anchor branch is an auxiliary supervision path enabled only during model training and is completely removed during validation, testing, and deployment. In this branch, a future descriptor not included in historical inputs is first generated from real future data in the future prediction target window. Then, a lightweight target projection module projects this descriptor into a sequence of teacher anchors. This projection operation stops gradient backpropagation during training, ensuring that the teacher anchors serve only as supervision signals and are not negatively affected by the student's learning process.

[0104] The student latent state sequence refers to the latent state representation sequence inferred solely from historical information, output from the historical representation branch; the teacher anchor sequence refers to the latent state representation sequence obtained from real future data and used as the standard answer during the training period.

[0105] The representation alignment loss is a supervisory signal constructed based on the difference between the student's latent state sequence and the teacher's anchor sequence. Its purpose is to force the student to learn a latent state that is as close as possible to the representation extracted by the teacher from real future data, given only historical information.

[0106] During training, future information provides high-quality supervisory anchors through the teacher branch, but the parameters of the teacher branch do not receive gradients from the alignment loss, and the teacher branch does not participate in any computation during the inference phase. The student branch thus learns to infer high-quality latent representations containing future degradation information from limited historical data, while completely avoiding the risk of future information entering the inference input.

[0107] In one implementation, the step of obtaining a student latent state sequence through a historical representation branch based on the historical latent state sequence, obtaining a teacher anchor sequence through a future target anchor branch based on the real future data in the future prediction target window, and constructing a representation alignment loss based on the difference between the student latent state sequence and the teacher anchor sequence, specifically includes the following steps: Step S410: Construct a historical representation branch, wherein the historical representation branch includes a cascaded student mapping module and a student projection module, used to map the historical latent state sequence into a sequence of student latent states; Step S420: Construct a future target anchor branch, wherein the future target anchor branch includes a target projection module, used to generate a future descriptor from the real future data in the future prediction target window, and generate a teacher anchor vector based on the future descriptor; wherein the teacher anchor vector stops gradient backpropagation during training, and the future target anchor branch is only used during the model training phase; Step S430: Calculate the vector distance between the vector of the sequence of the student's latent states after projection by the student projection module and the teacher's anchor vector, and construct the representation alignment loss based on the vector distance weighted by the labeled availability mask.

[0108] In this embodiment, the student mapping module in the historical representation branch can be implemented using a two-layer multilayer perceptron or a lightweight temporal decoder. Its input is the latent features output by the fixed encoder, and its output is the preliminary student latent state. The student projection module further maps the student latent state to the same semantic space as the teacher anchor point, facilitating the calculation of the alignment loss. The parameters of both the student mapping module and the student projection module can be learned during downstream training.

[0109] The target projection module in the future target anchor branch adopts a lightweight structure similar to the student projection module but with independent parameters. During training, a stopping gradient operation is performed on its output. The parameters of the target projection module can be updated using exponential moving average to maintain stability, or they can be fixed directly after initialization. This branch is only used during the model training phase and is deleted along with its weight file when the model is exported and deployed.

[0110] Specifically, the history student branch exists in the training, validation, testing, and deployment phases. It takes as input the representation obtained from the history window via the input adapter and solidified encoder, and passes through a battery-specific mapping. Forming a continuous potential battery state , is represented as:

[0111] in, Indicates The student mapping module is for parameters. For the encoder to be fixed The potential encoded output of each cycle.

[0112] The future target anchor branch is only enabled during the training phase. It first generates a future descriptor that is not included in the historical input from the real future window. Then, the supervisory anchor point for the stopping gradient is obtained using the lightweight target projector. The teacher anchor point vector is represented as:

[0113] in, Indicates the target projection module. The parameters can be updated using an exponential moving average method based on the student projection module parameters, or they can be directly fixed after initialization. This indicates that a stop gradient operation is performed on the output, so that the gradient is not backpropagated to the target projection module when calculating the alignment loss. These are descriptors extracted solely from the real future window, completely isolated in value from the data in the historical input window. Parameters of the target projection module. The stability of the teacher anchor points is ensured by using exponential moving average updates or pre-fixing them so that they do not change with training. The teacher anchor point vector is derived from the projection of the future target anchor point branch onto the real future data and processed by stopping gradients.

[0114] Then using an effective future window mask Calculate the alignment loss:

[0115] In the formula, This represents the L2 norm, i.e., the Euclidean distance. For student projectors, For the first The vector of each sample's student latent state after mapping by the student projection module, with parameters and Participate in training, parameters Exponential moving averages can be used for updates or pre-fixing. To prevent the denominator from being zero, For the first The effective future window mask for each sample is set to 1 if the future window data is complete and available, and 0 otherwise. During the validation, testing, and online deployment phases, real future descriptors and target anchors are completely removed, retaining only historical student branches; therefore, no future information enters the inference input. In this structure, "student" refers to historical branches, and "teacher" refers only to future target anchors during the training period.

[0116] pass The loss function forces the historical student branch to learn to align with the future anchor point, thereby gaining the ability to represent future degradation information without ever touching future data.

[0117] Step S500: Starting from the historical potential state sequence, under the given future action sequence and exogenous condition sequence, deduce the future potential state sequence, and output the prediction result based on the future potential state sequence.

[0118] In this embodiment, the future action sequence refers to a sequence of controllable operational variables planned to be applied to the battery within the prediction horizon, including but not limited to charge / discharge current or power commands, constant current and constant voltage protocol parameters, upper and lower limits of state of charge, cutoff voltage, resting arrangements, and cooling power. The exogenous condition sequence refers to a sequence of external environmental variables that cannot be directly controlled by the battery management system within the prediction horizon, including but not limited to ambient temperature and load requirements. The action sequence and condition sequence are stored separately in the data structure to ensure that their semantics are clear and they are not confused with each other.

[0119] The deduction refers to the process of recursively generating the potential states for each future time step from the current state of the historical potential state sequence, given the future action sequence and the exogenous condition sequence, through the action-conditional potential transition module. At each time step, the action-conditional potential transition module receives the current potential state, the action code and condition code for that step, and the step size code as input, and outputs the potential state for the next time step after gated residual fusion. Recursive deduction enables the model to possess simulator capabilities; that is, by changing the input future action sequence, the same model can deduce different future degradation trajectories from the same historical starting point.

[0120] The future potential state sequence refers to a series of potential state vectors obtained through the above recursive deduction process, with each vector corresponding to a future cycle in the prediction field. The prediction result refers to the set of results generated in parallel by the output heads of each prediction task after inputting the future potential state sequence into the shared decoding backbone. It may include one or more of the following: capacity degradation trajectory, health state change curve, degradation stage probability distribution, inflection point probability, and remaining lifetime distribution. Through the above action condition deduction mechanism, the model can not only predict the future degradation trend of the battery, but also perform counterfactual comparative analysis under different charge and discharge strategies, providing quantitative basis for the optimization of battery management strategies.

[0121] In one implementation, the step of starting with the historical potential state sequence, deducing a future potential state sequence under given future action sequences and exogenous condition sequences, and outputting a prediction result based on the future potential state sequence specifically includes the following steps: Step S510: Construct a shared decoding backbone and attach multiple prediction task output heads to the shared decoding backbone; wherein, the multiple prediction task output heads include at least two of the following: capacity trajectory output head, health status output head, degradation stage output head, inflection point detection output head, and remaining lifetime distribution output head; Step S520: Construct an action condition latent transition module, which includes an action coding branch and a gated residual fusion unit; wherein, the action coding branch is used to encode the future action sequence and the exogenous condition sequence, and the gated residual fusion unit is used to fuse the action coding result with the current latent state to generate the latent state of the next time step; Step S530: Taking the potential state of the last time step of the historical potential state sequence as the initial state, and given the future action sequence and the exogenous condition sequence, the future potential state sequence is obtained by recursively deducing step by step using the action condition potential transition module. Step S540: Input the future potential state sequence into the shared decoding backbone, and output the corresponding prediction results in parallel through the output heads of each prediction task.

[0122] In this embodiment, the shared decoding backbone receives the student's latent state sequence and performs deep information integration. An optional shared decoding backbone structure is a two-layer multilayer perceptron, configured with 256-dimensional input, 512-dimensional hidden layers, and 256-dimensional output, and equipped with GELU activation function, Dropout regularization, and layer normalization, as shown below:

[0123] in, Indicates a shared decoding backbone. , and , The weights and biases for the first and second layers are respectively. For layer normalization, The activation function is a Gaussian error linear unit. The parameters of the shared decoding backbone are shared by the output heads of all prediction tasks, which allows tasks to mutually benefit each other. For example, loops with rapid capacity decay are often also the locations where inflection points occur and degradation phases switch, and the shared backbone can naturally capture these common patterns across tasks.

[0124] Multiple prediction task output heads mounted on the shared decoding backbone each have independently trainable final layer parameters. The capacity trajectory output head predicts using a residual form relative to the current capacity, expressed as:

[0125] in, For the capacity output head corresponding to the future number The mapping function of the step, The known capacity of the current loop.

[0126] The health status trajectory is obtained by normalizing the capacity trajectory to the rated capacity, and is expressed as:

[0127] The advantage of using residuals instead of directly predicting absolute values ​​is that the model only needs to learn the change in capacity relative to the current value, which reduces the learning difficulty and improves prediction accuracy.

[0128] The future observable output head can decode the voltage, current, temperature, and capacity increment curves or their compact descriptors after the next cycle or multiple cycles as needed, represented as:

[0129] in, For the future observable output header corresponding to the future... The mapping function of the step, whose input is the shared decoding backbone at the 1st step. Output of each loop The output is the future number. Predicted values ​​of observable states for each cycle .

[0130] The output head of the degradation stage uses independent linear or two-layer multilayer perceptrons to output the probability distribution of each stage, expressed as:

[0131] in, For stage category index, Indicates the future number The degradation phase of the cycle belongs to the first... Predicted probability of class The output header for the degradation stage corresponds to the future... The mapping function of the step, whose input is the shared decoding backbone at the 1st step. Output of each loop , This indicates the category dimension of the degradation stage. Normalize the probability of each class so that the sum of the probabilities is 1.

[0132] The inflection point detection output head outputs the discrete risk rate or event probability for each cycle within the prediction span. The remaining lifetime distribution output head outputs the lifetime end risk rate and derives the survival function and event distribution from it. The independent output heads do not share the last layer parameters to avoid forced coupling between prediction objectives of different dimensions at the decision boundary; sharing occurs only in the historical representation branches after the input adapter and solidified encoder outputs, and in the shared decoding backbone.

[0133] The action-conditional latent transition module is used to implement the extrapolation capability of the world model simulator. This module consists of action encoding branches and gated residual fusion units. At each extrapolation step, the current latent state is... Embedding of controllable motion Embedding of uncontrollable conditions and step size position encoding The concatenation is performed to obtain the fused vector, which is represented as:

[0134] in, , and These are the action embedding network, conditional embedding network, and stride position embedding network, respectively. The semicolon indicates that the vectors are concatenated along the feature dimension.

[0135] Subsequently, the latent state is updated through a gated residual structure. First, a dimensionally gated vector is generated using the Sigmoid function. Then, the latent increment is calculated using a multilayer perceptron. Finally, the latent state for the next time step is obtained by combining residual connections with gating and layer normalization, as follows:

[0136]

[0137] in, and These are the learnable weight matrix and bias vector of the gated layer, respectively. This represents a dimension-wise gated vector used to control the potential increment in each dimension. The retention ratio, Indicates the potential state at the next time step. This is the Sigmoid function, whose output range is from 0 to 1. Indicates Multilayer perceptron with parameters, For the parameter set of the potential incremental network, This represents the Hadamard product, which is element-wise multiplication. The gated residual structure allows the model to selectively retain or update information across various dimensions of the latent state, maintaining stability against long-term degradation trends while allowing for flexible capture of short-term fluctuations and action responses. This transition module can be executed recursively. The process begins with the potential state of the last time step in the historical potential state sequence as the initial state, and then iterates forward step by step to generate a complete future potential state sequence. Its joint distribution can be factored into the product of the conditional probabilities of each time step, expressed as:

[0138] in, This represents a sequence of future actions. Given two different sequences of future actions starting from the same current potential state, this is relevant when two different sequences of future actions are input. and At this point, the difference in degradation consequences under the two strategies can be obtained, thus achieving counterfactual comparative analysis. The counterfactual degradation difference is expressed as:

[0139] The prediction results output in parallel by the output heads of each prediction task, together with the constructed training labels, will serve as input to the training step, driving the calculation of the joint loss function and the updating of model parameters.

[0140] In a specific example, after obtaining the current potential state from the same history window, two segments of length 1 are constructed. The future action sequence is determined. The first strategy uses a 2C charging rate in the low state of charge region, a 1C charging rate in the medium state of charge region, and a 0.5C charging rate in the high state of charge region, while enabling active cooling. The second strategy uses a uniform 1C charging rate and disables active cooling. Maintaining consistency between exogenous temperature and load scenarios, the action condition potential transition module is recursively called to obtain two potential state trajectories, a future observable sequence, and a capacity trajectory. The difference in degradation consequences between the two is then calculated. To prevent confusion between action variables and environmental variables, charging / discharging current commands, power limits, state of charge targets, cooling power, and static scheduling are stored in the action tensor, while ambient temperature, driving requirements, and uncontrollable loads are stored in the condition tensor. If historical data does not include strategy changes or if strategy selection is influenced by unobserved factors, the counterfactual output is interpreted only as a conditional prediction. Counterfactual identification needs to be supplemented through randomized experiments, known protocol switching, or offline causal correction.

[0141] To facilitate a comprehensive overview of the structural selection, parameter status, and participation stages of the above modules, Table 1 summarizes and displays the key configurations of the input adapter, fixed encoder, historical student mapping, future target anchor point, shared decoding backbone, each task output head, and action condition potential transfer module.

[0142] Table 1

[0143] Step S600: Based on the prediction results and the training labels, the model is trained using a joint loss function to obtain a trained battery degradation prediction model; wherein, the joint loss function is at least composed of a weighted sum of the loss term corresponding to the representation alignment loss and the task loss term determined based on the difference between the prediction results and the training labels.

[0144] In this embodiment, the joint loss function is a unified optimization objective driving the entire downstream multi-task training process. It is composed of at least a weighted sum of a loss term representing the alignment loss and a task loss term determined based on the difference between the prediction result and the training label. The task loss term can be further decomposed into a weighted combination of multiple sub-items such as capacity trajectory, observable state, degradation stage, inflection point event, and lifetime distribution. Each loss term is applied to the corresponding prediction task output head.

[0145] The training refers to the process of calculating the deviation between the prediction result and the training label using a joint loss function, and synchronously updating all unfixed parameters in the model using the gradient backpropagation algorithm, given that the architecture of each module and the forward propagation path are already determined. Specifically, trainable parameters include input adapter parameters, student mapping parameters and student projection parameters of the historical representation branches, shared decoding backbone parameters, independent parameters of each prediction task output head, and action condition latent transition module parameters; fixed encoder parameters and teacher projection parameters are not updated in this round of training. The battery degradation prediction model obtained after training contains the architecture of all necessary modules and the trained parameters. The model can be directly used to predict and extrapolate the degradation process of new batteries after removing the future target anchor branch.

[0146] Specifically, the complete training objective of the joint loss function consists of a weighted sum of seven losses:

[0147] in, This is the capacity trajectory loss term, used to monitor the prediction accuracy of capacity or health status trajectories. As an observable state loss term, it monitors the reconstruction quality of observable curves such as future voltage, current, and temperature. This loss term is not only used to improve prediction accuracy, but also a direct verification method to check whether the latent state representation retains enough future information. For the loss items in the degradation stage, monitor the accuracy of the degradation stage classification; For the inflection point loss item, monitor the prediction of the inflection point event time; This is a life loss item; To characterize the alignment loss term; This is the physical constraint loss term. Each loss term is first normalized according to its own effective supervision mask, and then multiplied by the corresponding task weight λ to prevent missing labels from being mistakenly treated as zero values ​​in training.

[0148] Among them, the life loss term adopts censorship-aware survival analysis loss, and the life endpoint branch outputs the risk rate within the discrete prediction span. This leads to the construction of the survival function and event distribution. Specifically, the survival function is expressed as:

[0149] The event distribution is represented as follows:

[0150] The survival function is the product of the complements of the risk rates at each time step, and the event distribution is the product of the survival function and the risk rate at that step. For right-censored samples, only the likelihood term of those still alive before the censoring time is used, and the time of missing events is not treated as zero value regression.

[0151] The physical constraint loss term is used to impose a soft monotonic constraint on capacity prediction to tolerate short-term capacity rebound caused by measurement noise or recovery effects.

[0152] in, The number of valid samples. and The first The sample in the future Step and the first Health status prediction value of steps This is a tolerance parameter, allowing a healthy state to occur no more than [a certain number] times between adjacent cycles. A rebound without penalty will only occur if the rebound exceeds [a certain threshold]. The squared penalty term is generated only at certain times. This soft constraint method maintains the physical rationality of the long-term degradation trend without treating measurement noise as a real degradation signal.

[0153] In terms of evaluation metrics, commonly used metrics for capacity or health status trajectories include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE).

[0154]

[0155]

[0156] in, The number of valid samples. and The first The true and predicted values ​​of each sample To avoid extremely small positive numbers that divide by zero when the true value is close to zero, the phase tasks are evaluated using macro-average F1 score, balanced accuracy, and confusion matrix; the inflection point detection task uses event-time absolute error and tolerance window hit rate; and the survival analysis task uses consistency index, time-dependent Brier score, and prediction interval coverage.

[0157] In a specific multi-task training configuration, the model is trained simultaneously on multiple tasks with prediction spans of 1, 5, 10, and 20 cycles. During training, the future target anchor branch uses true future descriptors consistent with each prediction span to construct teacher anchors, and stopping gradient processing is applied to the loss; the validation and testing code does not compute this representation alignment auxiliary loss. During the evaluation phase, the MAE, RMSE, and MAPE for capacity and health status are output for each prediction span. MAPE and error distribution can only be reliably calculated if the logs store the true values, predicted values, and sample masks for each sample. MAE for each span alone cannot directly derive trajectory-level MAPE, RMSE, or uncertainty coverage.

[0158] Through the collaborative design of the shared decoding backbone, multi-task output head, action condition potential transfer module, and joint loss function, the model achieves a complete closed loop from multi-source heterogeneous data to prediction of the overall battery degradation. It can also perform strategy comparison analysis by changing the future action sequence, providing quantifiable decision support for battery life cycle management.

[0159] The battery degradation prediction model obtained after training includes the architecture of all the necessary modules and the trained parameters. This model can be directly used to predict and extrapolate the degradation process of new batteries after removing the future target anchor branch. By unifying the optimization of representation alignment and multi-task prediction through a joint loss function, a synergistic effect is formed among the various modules of the model. The representation alignment loss ensures that the latent state retains future degradation information, while the task loss ensures that various prediction targets can be accurately decoded from the latent state. The two promote each other and avoid the suboptimal problem caused by the separation of representation learning and downstream tasks.

[0160] In a series of historical engineering validations, the residual calibration model achieved a mean absolute error of 0.060552 Ah for capacity and 0.001002 Ah for health state predictions on a 20-cycle test; the corresponding mean absolute errors were 0.064096 Ah and 0.001068 Ah for the baseline-invariant model, and 0.073372 Ah and 0.001193 Ah for the early characterization-aligned prototype. The residual calibration model improved the mean absolute errors for capacity and health state by approximately 5.53% and 6.22%, respectively, compared to the baseline-invariant model. The performance comparison between the model and baseline in the 20-cycle prediction test in historical engineering is shown below. Figure 5 As shown, a lower error value indicates higher prediction accuracy. The error improvement for different prediction spans is shown in the figure. Figure 6 As shown, this intuitively demonstrates the model's performance advantages in medium- to long-term prediction horizons. Figure 6 In the middle, the improvement rate is based on Calculations show that negative improvements can still occur even with short prediction step sizes, where E06 is the experimental data step number.

[0161] In one implementation, the method further includes the following steps: Step S710: Obtain the historical cycle data of the battery to be predicted within the most recent preset number of cycles, as well as the preset future action sequence and exogenous condition sequence; Step S720: Input the historical cycle data, the future action sequence, and the exogenous condition sequence into the trained battery degradation prediction model. Obtain the historical latent state sequence through the input adapter, the solidified encoder, and the historical representation branch. Starting from the historical latent state sequence, recursively deduce the future latent state sequence step by step under the conditions of the future action sequence and the exogenous condition sequence. Based on the future latent state sequence, output the prediction result of the battery to be predicted under multiple future cycles. The prediction results include at least one of the following: capacity degradation trajectory, health status change curve, degradation stage sequence, inflection point probability distribution, and remaining lifetime distribution.

[0162] In this embodiment, when the trained battery degradation prediction model is put into actual inference, the future target anchor branch has been completely removed, and only the student mapping module and student projection module in the historical representation branch are retained.

[0163] In the actual inference process, the historical cycle data of the battery to be predicted within the most recent preset number of cycles is first obtained. This preset number is usually consistent with the historical window length used during training. At the same time, the preset future action sequence and exogenous condition sequence are obtained. The future action sequence describes the charging and discharging strategy parameters to be used within the prediction field of view, and the exogenous condition sequence describes the expected environmental temperature and load conditions.

[0164] After the three types of inputs are fed into the model, they first undergo dimension mapping and standardization by the input adapter, then the solidified encoder extracts the latent features of each historical loop. Subsequently, the historical representation branch maps the latent feature sequence into a student latent state sequence. Using the state of the last time step of this student latent state sequence as the initial state, the action-conditional latent transition module, driven by the given future action sequence and exogenous condition sequence, recursively extrapolates forward step by step to generate a future latent state sequence covering the entire prediction horizon. Finally, the future latent state sequence passes through the shared decoding backbone and the output heads of each enabled prediction task, outputting in parallel the prediction results of the battery to be predicted in multiple future loops.

[0165] The prediction results can be flexibly configured according to actual needs. The capacity degradation trajectory depicts the decay process of battery capacity from the current value to the prediction endpoint in the form of a numerical sequence; the health status change curve shows the evolution trend of health status with the number of cycles in the form of a percentage; the degradation stage sequence identifies the degradation stage category of the battery in each cycle within the prediction field, such as stable degradation stage, accelerated degradation stage, or high-speed degradation stage; the inflection point probability distribution gives an estimate of the probability of the inflection point occurring in each cycle, providing a basis for early warning of the degradation acceleration inflection point; the remaining lifetime distribution shows the distribution of the remaining usable cycles of the battery in the form of a survival function or probability density, providing a probabilistic reference for maintenance planning and replacement decisions.

[0166] In terms of deployment, vehicle-side deployment can retain only the input adapter, fixed encoder, historical student mapping module, action-condition latent transition module, and capacity and health status output heads and risk output heads. Cloud-side deployment adds future curve output heads, degradation stage output heads, inflection point output heads, and survival distribution output heads to the vehicle-side modules. The deployment file also fixes the feature order, standardized parameters, missing mask conventions, rated capacity, lifetime threshold, action and condition field dictionaries, and model version number to prevent semantic inconsistencies between training and inference. The target anchor module during training is deleted along with its weight file when the model is exported and deployed; during inference, latent state sequences are obtained only from historical data through historical representation branches.

[0167] In summary, compared with existing technologies, this embodiment distinguishes between detailed process tables, step summary tables, and processing result tables through protocol-aware data reconstruction, avoiding semantic errors in capacity labels and leakage of cumulative quantities. Furthermore, by using historical student branches and future target anchors enabled only during the training period, future information is used for supervisory representation without entering validation, testing, and deployment inputs. By sharing the decoding backbone and independent task output layers, it maintains cross-task commonality and the independence of the dimensions and decision boundaries of each predicted object. Through gated residual transfer with separation of actions and exogenous conditions, future observables and degradation consequences can be progressively deduced, providing a calling interface for planning or model predictive control. Through continuity constraints, right censoring or interval censoring supervisory masks, and soft physical constraints, it avoids treating missing lifetimes, unidentifiable inflection points, or short-term recovery phenomena as erroneous supervision. The future observable state is both the model output and a supervisory signal verifying whether the potential battery representation retains sufficient predictable information, extending prediction from a single SOH (State of Health) or RUL (Remaining Useful Life) regression to a verifiable upstream state representation.

[0168] like Figure 7 As shown in the figure, an embodiment of the present invention provides a power battery degradation prediction system, which includes: a data acquisition module 10, a training sample construction module 20, a latent state encoding module 30, a representation alignment loss construction module 40, a latent state inference and prediction module 50, and a joint training module 60.

[0169] Specifically, the data acquisition module 10 is used to acquire multi-source battery cycle data, identify the data type of the multi-source battery cycle data, and extract the capacity information of each cycle from the multi-source battery cycle data based on the extraction rules corresponding to the data type. The training sample construction module 20 is used to construct training labels based on the capacity information, and to build a historical cyclic input window and a future prediction target window; wherein, the training labels include health status labels and lifespan-related labels; The latent state encoding module 30 is used to construct an encoder and perform pre-training to obtain a solidified encoder; an input adapter is constructed at the input end of the solidified encoder, and the battery cycle data in the historical cycle input window is mapped to a historical latent state sequence through the input adapter and the solidified encoder; The representation alignment loss construction module 40 is used to obtain a student latent state sequence based on the historical latent state sequence through a historical representation branch; to obtain a teacher anchor sequence based on the real future data in the future prediction target window through a future target anchor branch; and to construct a representation alignment loss based on the difference between the student latent state sequence and the teacher anchor sequence. The latent state deduction and prediction module 50 is used to deduce the future latent state sequence starting from the historical latent state sequence, under the condition of a given future action sequence and exogenous condition sequence, and output the prediction result based on the future latent state sequence. The joint training module 60 is used to train the model using a joint loss function based on the prediction results and the training labels to obtain a trained battery degradation prediction model; wherein the joint loss function is at least composed of a weighted loss term corresponding to the representation alignment loss and a task loss term determined based on the difference between the prediction results and the training labels.

[0170] Based on the above embodiments, the present invention also provides a terminal device, the principle block diagram of which can be as follows: Figure 8 As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for predicting the degradation of a power battery. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0171] Those skilled in the art will understand that Figure 8 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0172] In one embodiment, a terminal device is provided, including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs including instructions for performing operations as described in the embodiments of the methods above.

[0173] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0174] 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.

[0175] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting the degradation of power batteries, characterized in that, The method includes: Acquire multi-source battery cycle data, identify the data type of the multi-source battery cycle data, and extract capacity information of each cycle from the multi-source battery cycle data based on the extraction rules corresponding to the data type. Based on the capacity information, training labels are constructed, and historical cyclic input windows and future prediction target windows are built; wherein, the training labels include health status labels and lifespan-related labels; An encoder is constructed and pre-trained to obtain a solidified encoder; an input adapter is constructed at the input end of the solidified encoder, and the battery cycle data in the historical cycle input window is mapped to a historical latent state sequence through the input adapter and the solidified encoder; Based on the historical latent state sequence, a student latent state sequence is obtained through a historical representation branch; based on the real future data in the future prediction target window, a teacher anchor sequence is obtained through a future target anchor sequence; based on the difference between the student latent state sequence and the teacher anchor sequence, a representation alignment loss is constructed. Starting with the historical potential state sequence, and given the future action sequence and exogenous condition sequence, the future potential state sequence is deduced, and the prediction result is output based on the future potential state sequence. Based on the prediction results and the training labels, the model is trained using a joint loss function to obtain a trained battery degradation prediction model; wherein, the joint loss function is at least composed of a weighted sum of a loss term corresponding to the representation alignment loss and a task loss term determined based on the difference between the prediction results and the training labels.

2. The power battery degradation prediction method according to claim 1, characterized in that, The process of acquiring multi-source battery cycle data, identifying the data type of the multi-source battery cycle data, and extracting capacity information for each cycle from the multi-source battery cycle data based on extraction rules corresponding to the data type includes: Obtain battery cycle data tables from different data sources, identify the data types of each data table, and the data types characterize the record granularity and field semantics of the data tables; Based on the identified data type, the capacity extraction rules corresponding to that data type are used to extract the capacity information corresponding to each battery cycle from each data table. Based on preset capacity thresholds or degradation index thresholds, the extracted capacity information is filtered for validity to obtain valid capacity information for subsequent steps.

3. The power battery degradation prediction method according to claim 1, characterized in that, The step of constructing training labels based on the capacity information and building a historical cyclic input window and a future prediction target window includes: Based on the capacity information, the battery cycle sequence is checked for cycle continuity and missing values ​​are counted, and cycles with abnormalities or missing values ​​are marked. Based on the capacity information, health status labels and lifespan-related labels are constructed; wherein, the health status labels include the health status values ​​of each cycle, and the lifespan-related labels include at least one of lifespan end, inflection point location, and degradation stage category; The availability of labels is marked to generate a label availability mask, which is used to mask invalid or unavailable labels during training; Based on the preset historical window length and predicted field of view length, the historical cycle input window and the future prediction target window are constructed by sliding from the battery cycle sequence.

4. The power battery degradation prediction method according to claim 1, characterized in that, The process involves constructing and pre-training an encoder to obtain a fixed encoder. An input adapter is then built at the input of the fixed encoder. Through the input adapter and the fixed encoder, the battery cycle data in the historical cycle input window is mapped to a historical latent state sequence, including: Construct an encoder that encodes multidimensional time-series data from a single battery cycle into a fixed-dimensional latent state vector. Based on self-supervised learning, the encoder is pre-trained using at least one of the mask feature recovery task and the adjacent cyclic dynamic prediction task to obtain pre-trained encoder parameters. The pre-trained encoder parameters are fixed so that the encoder parameters are no longer updated; An input adapter is constructed at the input end of the solidified encoder. The input adapter is used to map raw battery cycling data of different dimensions to an input dimension acceptable to the solidified encoder. The parameters of the input adapter can be learned during the training process of downstream tasks.

5. The method for predicting the degradation of power batteries according to claim 1, characterized in that, The process involves obtaining a student latent state sequence through a historical representation branch based on the historical latent state sequence, obtaining a teacher anchor sequence through a future target anchor branch based on the real future data in the future prediction target window, and constructing a representation alignment loss based on the difference between the student latent state sequence and the teacher anchor sequence, including: Construct a historical representation branch, which includes a cascaded student mapping module and a student projection module, for mapping the historical latent state sequence to a sequence of student latent states; A future target anchor branch is constructed, which includes a target projection module for generating a future descriptor from the real future data in the future prediction target window and generating a teacher anchor vector based on the future descriptor; wherein, the teacher anchor vector stops gradient backpropagation during training, and the future target anchor branch is only used during the model training phase; Calculate the vector distance between the vector of the sequence of the student's latent states projected by the student projection module and the teacher's anchor vector, and construct the representation alignment loss based on the vector distance weighted by the labeled availability mask.

6. The method for predicting the degradation of power batteries according to claim 1, characterized in that, Starting with the historical potential state sequence, and given a future action sequence and an exogenous condition sequence, the process of deriving a future potential state sequence and outputting a prediction result based on the future potential state sequence includes: A shared decoding backbone is constructed, and multiple prediction task output heads are mounted on the shared decoding backbone; wherein, the multiple prediction task output heads include at least two of the following: capacity trajectory output head, health status output head, degradation stage output head, inflection point detection output head, and remaining lifetime distribution output head; An action-condition latent transition module is constructed, which includes an action coding branch and a gated residual fusion unit; wherein, the action coding branch is used to encode the future action sequence and the exogenous condition sequence, and the gated residual fusion unit is used to fuse the action coding result with the current latent state to generate the latent state of the next time step; Using the potential state of the last time step of the historical potential state sequence as the initial state, and given the future action sequence and the exogenous condition sequence, the future potential state sequence is obtained by recursively deducing step by step using the action condition potential transition module. The sequence of future potential states is input into the shared decoding backbone, and the corresponding prediction results are output in parallel through the output heads of each prediction task.

7. The method for predicting the degradation of power batteries according to claim 1, characterized in that, The method further includes: Obtain historical cycle data of the battery to be predicted within the most recent preset number of cycles, as well as preset future action sequences and exogenous condition sequences; The historical cycle data, the future action sequence, and the exogenous condition sequence are input into the trained battery degradation prediction model. The historical latent state sequence is obtained through the input adapter, the solidified encoder, and the historical representation branch. Starting from the historical latent state sequence, the future latent state sequence is recursively deduced step by step under the conditions of the future action sequence and the exogenous condition sequence. Based on the future latent state sequence, the prediction result of the battery to be predicted under multiple future cycles is output. The prediction results include at least one of the following: capacity degradation trajectory, health status change curve, degradation stage sequence, inflection point probability distribution, and remaining lifetime distribution.

8. A power battery degradation prediction system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source battery cycle data, identify the data type of the multi-source battery cycle data, and extract the capacity information of each cycle from the multi-source battery cycle data based on the extraction rules corresponding to the data type. The training sample construction module is used to construct training labels based on the capacity information, and to build a historical cyclic input window and a future prediction target window; wherein, the training labels include health status labels and lifespan-related labels; A latent state encoding module is used to construct an encoder and perform pre-training to obtain a solidified encoder; an input adapter is constructed at the input end of the solidified encoder, and the battery cycle data in the historical cycle input window is mapped to a historical latent state sequence through the input adapter and the solidified encoder; The representation alignment loss construction module is used to obtain the student latent state sequence through the historical representation branch based on the historical latent state sequence; to obtain the teacher anchor sequence through the future target anchor branch based on the real future data in the future prediction target window; and to construct the representation alignment loss based on the difference between the student latent state sequence and the teacher anchor sequence. The latent state deduction and prediction module is used to deduce the future latent state sequence starting from the historical latent state sequence, given the future action sequence and the exogenous condition sequence, and output the prediction result based on the future latent state sequence. A joint training module is used to train the model using a joint loss function based on the prediction results and the training labels to obtain a trained battery degradation prediction model; wherein the joint loss function is at least composed of a weighted loss term corresponding to the representation alignment loss and a task loss term determined based on the difference between the prediction results and the training labels.

9. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a power battery degradation prediction program stored in the memory and executable on the processor. When the processor executes the power battery degradation prediction program, it implements the steps of the power battery degradation prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a power battery degradation prediction program, which, when executed by a processor, implements the steps of the power battery degradation prediction method as described in any one of claims 1-7.

Images