A lithium ion battery voltage data compression and reconstruction method based on low-dimensional latent representation learning and equivalent circuit mechanism constraint

Abstract

The application discloses a lithium ion battery voltage data compression and reconstruction method based on low-dimensional latent representation learning and equivalent circuit mechanism constraint, and belongs to the technical field of lithium ion battery data processing and state modeling. The method filters, normalizes and processes voltage, current, temperature and time data in the operation process of the lithium ion battery, and constructs a low-dimensional latent representation by using multi-dimensional time sequence input, so as to realize compression representation of original voltage data. The method estimates the state of charge, open circuit voltage and equivalent circuit parameters in combination with current, time and other conditional information. In the decoding and reconstruction process, the state of charge evolution law, open circuit voltage mapping relationship and equivalent circuit mechanism constraint are introduced, so as to realize reconstruction of the lithium ion battery terminal voltage data. The method can reduce data storage and transmission pressure, better retain dynamic characteristics and state information in the operation process of the battery, improve the accuracy and interpretability of the reconstruction result, and can be applied to lithium ion battery state monitoring, performance evaluation and abnormal diagnosis.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery operation data processing and modeling technology, specifically to a method for compressing and reconstructing lithium-ion battery voltage data based on low-dimensional latent representation learning and equivalent circuit mechanism constraints, applicable to lithium-ion battery operation data processing in new energy vehicles and energy storage systems. Background Technology

[0002] With the rapid development of the new energy vehicle industry and the new energy storage industry, lithium-ion batteries have become a core component supporting vehicle operation, safety management, and performance assurance. To effectively monitor, manage, and optimize the operating status of lithium-ion batteries, continuous collection, recording, and analysis of battery-related operational information are typically required during vehicle use, energy storage operation, and maintenance. As application scale expands and operating cycles extend, the amount of related data is growing rapidly, placing higher demands on the efficiency of data storage, transmission, retrieval, and application. Simultaneously, the operation of lithium-ion batteries exhibits strong dynamic changes. How to reduce the data volume while retaining as much important information as possible reflecting the battery's state evolution, and meeting the needs of subsequent applications such as state monitoring, performance analysis, and voltage reconstruction, has become a pressing technical problem to be solved in the field of lithium-ion battery operational data processing.

[0003] In practical applications, lithium-ion battery-related data not only permeates multiple stages such as vehicle operation, energy storage charging and discharging control, maintenance diagnosis, and remote monitoring, but also typically needs to be stored, transmitted, and retrieved between different application stages. With increasing monitoring frequency and application scenarios, the continuous accumulation of data can easily lead to problems such as increased storage resource consumption, heavier transmission burdens, and decreased data retrieval response efficiency, thus affecting the efficient management and application of battery operation information. On the other hand, the operating state of lithium-ion batteries is influenced by various factors such as changes in operating conditions, the usage environment, and internal state evolution; related data typically exhibits dynamism, coupling, and state characterization. Simply using data reduction methods for processing can easily result in insufficient expression of key characteristic information, hindering subsequent state identification, performance evaluation, voltage response analysis, and mechanism modeling. Especially in applications requiring long-term monitoring, remote analysis, and system operation analysis, it is necessary to reduce resource consumption during data processing while ensuring high effectiveness and usability in expressing the patterns of battery state changes. Therefore, there is an urgent need to propose an efficient processing method for lithium-ion battery operation data, so as to achieve effective expression, convenient storage, reliable transmission and subsequent efficient utilization of relevant data while taking into account data scale control, key information retention, state characterization capability and engineering application reliability.

[0004] Existing methods for processing lithium-ion battery operation data mainly include direct storage of raw data, data compression methods based on signal processing or statistical dimensionality reduction, and data representation methods based on machine learning. While direct storage of raw data can preserve information about the battery's operation relatively completely, in scenarios involving long-term operation, continuous monitoring, and multi-terminal collaborative management of lithium-ion batteries, the continuous growth in data volume can easily lead to problems such as large storage resource consumption, heavy transmission burden, and decreased data retrieval and utilization efficiency. Methods based on signal processing, feature transformation, or statistical dimensionality reduction can reduce data volume to some extent, but these methods typically focus more on the compression effect of data format and do not adequately consider the dynamic response characteristics, state evolution laws, and intrinsic mechanism correlations exhibited during lithium-ion battery operation. Therefore, they have certain limitations in terms of the expressive power of compressed data, the ability to retain key features, and the fidelity of reconstruction. Machine learning-based data representation methods can learn latent features from historical data and improve data compression and representation capabilities. However, most methods still focus on data-driven mapping modeling and lack effective integration of the internal state change mechanism and equivalent circuit constraints of lithium-ion batteries. This can easily lead to insufficient interpretability of compression results, weak consistency of reconstructed data mechanisms, and limited applicability in subsequent state analysis and engineering applications.

[0005] Especially for lithium-ion battery operating data, it not only carries information on changes in external operating conditions but also reflects complex characteristics such as the evolution of the battery's internal state of charge, polarization behavior, open-circuit voltage changes, and terminal voltage response. Therefore, existing general data compression methods or purely data-driven modeling methods are insufficient to simultaneously meet the demands of data compression, mechanism consistency reconstruction, state representation, and preservation of key information. Based on this, it is necessary to propose a lithium-ion battery voltage data compression and reconstruction method that integrates low-dimensional latent representation learning with equivalent circuit mechanism constraints. This method can effectively reduce the data storage and transmission pressure of lithium-ion battery operating data in new energy vehicles and energy storage systems while achieving efficient compression and high-fidelity reconstruction of operating data, providing a reliable data foundation and technical support for lithium-ion battery state monitoring, performance analysis, mechanism modeling, and subsequent state analysis.

[0006] To address the problems of continuously growing raw data size, high storage resource consumption, high transmission bandwidth pressure, and limited data retrieval efficiency in the long-term monitoring and analysis of lithium-ion battery voltage data, and considering the significant time-series dynamic characteristics, state evolution characteristics, and multivariate coupling characteristics of lithium-ion battery voltage-related data, existing processing methods still suffer from insufficient retention of key information, low reconstruction accuracy after compression, and weak consistency of mechanisms. To resolve these issues, this invention proposes a lithium-ion battery voltage data compression and reconstruction method based on low-dimensional latent representation learning and equivalent circuit mechanism constraints.

[0007] This invention first filters, standardizes, and constructs samples from raw operating data collected during the operation of lithium-ion batteries, forming multivariate time-series samples that meet the model input requirements. Then, based on a low-dimensional latent representation learning method, the multivariate time-series samples are encoded with temporal features and mapped using low-dimensional compression to obtain a low-dimensional latent representation that characterizes the operating state and dynamic evolution of the lithium-ion battery, thus achieving a compact expression of the original high-dimensional operating data. Subsequently, based on the low-dimensional latent representation and combined with current, time, and other conditional information during the lithium-ion battery's operation, the internal state and mechanism parameters of the lithium-ion battery are jointly estimated to obtain a state-of-charge sequence, an open-circuit voltage sequence, and a set of equivalent circuit parameters. Furthermore, during the decoding and reconstruction process, the evolution law of the state of charge, the open-circuit voltage mapping relationship, and the equivalent circuit mechanism constraints are introduced to achieve a consistent reconstruction of the lithium-ion battery terminal voltage mechanism. A joint optimization training mechanism is constructed by combining reconstruction error constraints, dynamic difference constraints, trend constraints, smoothing constraints, peak constraints, and latent spatial distribution constraints to improve the ability of the reconstruction results to express the changing trends, dynamic characteristics, and intrinsic mechanisms of the original data.

[0008] Therefore, while ensuring the effective preservation of key information in lithium-ion battery voltage data, this invention can achieve low-dimensional compressed representation and mechanism-consistent reconstruction of the original data, improving the ability of the compressed representation to characterize battery dynamic characteristics and state evolution information, and enhancing the fidelity, physical interpretability, and engineering applicability of the reconstruction results. The low-dimensional latent representation, state-of-charge sequence, open-circuit voltage sequence, equivalent circuit parameters, and reconstructed voltage results can provide a data foundation for subsequent lithium-ion battery state analysis, performance evaluation, and anomaly identification, and are applicable to the monitoring and analysis of lithium-ion battery operating status in new energy vehicles and energy storage systems.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for compressing and reconstructing lithium-ion battery voltage data based on low-dimensional latent representation learning and equivalent circuit mechanism constraints includes the following steps.

[0010] S1: Screening, standardization, and sample construction of lithium-ion battery operating data.

[0011] S11: Collect and store raw multivariate time-series data during the operation of the lithium-ion battery; the raw time-series data includes at least voltage sequence, current sequence, temperature sequence and time sequence.

[0012] S12: Perform sample validity screening and boundary constraint verification on the original time series data, remove invalid samples that are missing, abnormal, out of bounds, or whose length does not meet the requirements, and retain valid time series samples that meet the preset sequence length and physical constraint conditions; the preset physical value range includes at least the voltage value range, the temperature value range, and the optional current value range.

[0013] S13: Perform normalization mapping on the selected voltage, current, temperature and time series to unify different physical quantities into a preset numerical space and obtain standardized time series data.

[0014] S14: Based on standardized time-series data, a fixed-length truncation is performed according to a preset sequence length to construct a fixed-length time-series input sample, and auxiliary features characterizing the dynamic change process are extracted; the auxiliary features preferably include the first-order changes of the current sequence and voltage sequence between adjacent sampling times, and the auxiliary features are respectively expressed as: (1) (2) Where, Δ I k Δ represents the current differential characteristic corresponding to the k-th sampling time. V k This represents the voltage differential characteristic corresponding to the k-th sampling time. I k and V k They represent the first k The normalized current value and normalized voltage value corresponding to each sampling time.

[0015] S15: Standardize time-series data and corresponding dynamic change characteristics are uniformly organized to form a standardized input sample set, and output to the subsequent low-dimensional latent representation construction, internal state and mechanism parameter estimation and voltage reconstruction process.

[0016] S2: Constructing low-dimensional latent representations based on multi-dimensional time-series samples.

[0017] S21: Standardized voltage sequence obtained from step S1 V Current sequence I Temperature sequence T and time series t The samples are organized according to a uniform time sequence and combined along the feature dimension to form multivariate time-series input samples. X Its expression is: (3) in, L For the preset sequence length, V i , I i , T i and t i They represent the first iThe voltage, current, temperature, and time values ​​corresponding to each sampling time point, after scale mapping processing. Using this method, the runtime timing data of a lithium-ion battery can be organized into multivariate time-series input samples.

[0018] S22: Input the multidimensional time series sample X The input encoding network is used to extract local dynamic features and multidimensional coupling relationships between different time steps using a temporal feature extraction module, thus obtaining deep temporal features. F c The formula for its calculation is: (4) in, This represents the temporal feature encoding mapping function. Then, the deep temporal features are transformed and mapped to obtain a compact feature vector. F s : (5) in, Represents a compact feature mapping function. F s This is used to characterize the comprehensive operational status features of the current sample. In one embodiment, the transformation and mapping process includes flattening the deep temporal features before performing a fully connected mapping.

[0019] S23: Based on the compact feature vector F s Calculate the mean vector of each input sample in the latent space. μ And logarithmic variance vector log σ ², its expression is: (6) (7) in, W μ and W σ These are the weight matrices for the mean branch and the variance branch, respectively. b μ and b σ These are the bias terms; μ and log σ ² These parameters collectively characterize the probability distribution of the input sample in the latent space, and are used to describe the low-dimensional latent distribution features corresponding to the input sample.

[0020] S24: Based on mean vector μ And logarithmic variance vector log σ ², a reparameterized sampling mechanism is used to generate low-dimensional latent representations. zThe calculation process is as follows: (8) (9) Here, ⊙ represents element-wise multiplication. ε This is random noise that follows a standard normal distribution. Through the above steps, the original multivariate high-dimensional time series data can be compressed and mapped into a low-dimensional latent representation. z This allows for a compact representation of the original operating data while retaining key information about the lithium-ion battery's operating status, and provides a low-dimensional characterization basis for subsequent internal state estimation, mechanism parameter estimation, and voltage reconstruction.

[0021] S3: Estimation of internal state and mechanism parameters of lithium-ion batteries based on the fusion of low-dimensional latent representation and operating condition information.

[0022] S31: The low-dimensional latent representation obtained in step S2 z Joint modeling is performed with conditional information during the operation of the lithium-ion battery. This conditional information includes the current sequence. I and time series t And organize the conditional input sequence according to a unified time order. C : (10) in, L For the preset sequence length, I i and t i They represent the first i The current and time values ​​corresponding to each sampling time point are obtained after scale mapping. Then, a conditional temporal feature extraction network is used to process the conditional input sequence. C Encode to obtain conditional time series features H c : (11) in, This represents the conditional temporal feature extraction mapping function. H c This represents the extracted conditional temporal features. In one embodiment, the conditional temporal feature extraction network includes a bidirectional long short-term memory network and a mapping layer.

[0023] S32: Based on low-dimensional latent representation z Generate characteristic modulation parameters and conditional temporal features. H c Adaptive modulation is performed to fuse latent state information with external operating condition information. The calculation formula is as follows: (12) (13) (14) in, and β These represent the scaling and offset parameters generated from the low-dimensional latent representation, respectively. , W β , and b β These are the corresponding weight matrices and bias terms, respectively. H f The denoting symbol represents the fused temporal features, and ⊙ represents element-wise multiplication. Through this method, the conditional temporal features can be dynamically adjusted using low-dimensional latent representations, thereby enhancing the model's ability to jointly represent different battery states and operating conditions.

[0024] S33: Perform feature aggregation on the fused temporal features Hf to obtain a global fused feature vector representing the overall operating state of the current sample. F g Based on the globally fused feature vector, the state of charge sequence of the lithium-ion battery is predicted. SOC First, predict the initial state of charge. SOCi nit and the magnitude of the change in state of charge Δ SOC : (15) (16) Then, predict the normalized incremental weights for each time step. w : (17) Furthermore, a complete sequence of charged states is constructed: (18) in, f 1 (·) f 2 (·)and f 3 (·) represent the corresponding mapping functions. SOC init Indicates the initial state of charge, Δ SOC Indicates the magnitude of the change in state of charge. wThis represents the normalized incremental weight corresponding to each time step, and cumsum(·) represents the cumulative summation operation. Using this method, a lithium-ion battery state-of-charge sequence with a constrained evolutionary form can be obtained, reducing the irregular fluctuations caused by independent time-by-time predictions and improving the temporal consistency of the state-of-charge estimation results.

[0025] S34: Based on the fusion feature representation F g Predicting equivalent circuit model parameters P In one embodiment, the parameter set includes R 0. Resistance of the first polarization branch R 1. Resistance of the second polarization branch R 2. Capacitor of the first polarization branch C 1 and second polarization branch capacitors C 2, its expression is: (19) (20) in, f p (·) represents the equivalent circuit parameter prediction function. To ensure the rationality of the parameter values, the prediction results can be constrained to a preset physical range through bounded mapping.

[0026] S35: Based on the predicted state of charge sequence SOC In addition, it integrates feature representations to construct a mapping relationship between the state of charge and the open-circuit voltage, thereby obtaining the open-circuit voltage sequence. OCV In this invention, a parameterized nonlinear mapping function is preferably used. SOC - OCV The relationship is modeled, and in one implementation, it is represented by a double Sigmoid function, the expression of which is: (twenty one) in, a , b , c , d , e , f and g These are the mapping parameters obtained through adaptive prediction based on the fused feature representation. Through the above steps, the lithium-ion battery state-of-charge sequence, open-circuit voltage sequence, and equivalent circuit parameters can be obtained simultaneously based on the fusion results of the low-dimensional latent representation and conditional information, providing an internal state characterization and mechanistic parameter basis for subsequent battery terminal voltage reconstruction.

[0027] S4: Voltage reconstruction and joint optimization training based on equivalent circuit mechanism constraints.

[0028] S41: Open-circuit voltage sequence obtained from S3 OCV Current sequence I Time series t and equivalent circuit model parameters R 0、 R 1. R 2. C 1 and C 2. Construct a lithium-ion battery terminal voltage reconstruction model; wherein, the equivalent circuit model includes an ohmic internal resistance branch and two polarizations. RC The polarization branches are used to characterize the ohmic voltage drop and polarization response of the lithium-ion battery under dynamic operating conditions. Then, based on the current input, time interval, and equivalent circuit parameters, the dynamic response of each polarization branch is discretely recursively solved to obtain the voltage of the first polarization branch. V 1 and second polarization branch voltage V 2. A lithium-ion battery terminal voltage reconstruction sequence is constructed by combining open-circuit voltage and ohmic voltage drop. V rec : (twenty two) in, V rec This represents the reconstructed terminal voltage sequence. Through the above method, the mechanistic constraints of the original voltage curve of a lithium-ion battery can be reconstructed. In one embodiment, the current and time sequences involved in the terminal voltage calculation are first restored to their corresponding physical quantity spaces before the terminal voltage is calculated.

[0029] S42: Calculate the time interval Δ between adjacent sampling times based on the time series t. t k The dynamic responses of the two polarization branches are calculated using the parameters of the second-order equivalent circuit model; the time constants of the first and second polarization branches are as follows: (twenty three) (twenty four) The corresponding polarization voltage recursive processes are expressed as follows: (25) (26) in, V 1,k and V 2,k They represent the first k The voltages of the first polarization branch and the second polarization branch corresponding to each sampling time, Δ t k = t k -tk−1 The above method enables discrete-time modeling of the dynamic polarization process of lithium-ion batteries.

[0030] S43: Terminal voltage reconstruction sequence calculated based on steps S41 and S42 V rec By mapping the output space of the model, the reconstructed voltage sequence of the model output is obtained, and its expression is: (27) in, V min and V max These represent the preset lower voltage limit and upper voltage limit, respectively. Through the above processing, the reconstructed output and the original input samples are located in the same numerical space, facilitating error calculation and model training. In one embodiment, the output space mapping is consistent with the scaling method used for the voltage sequence in step S1.

[0031] S44: Construct a joint loss function to train and optimize the model; wherein, the joint loss function includes voltage reconstruction loss. L rec Differential loss L diff Smoothing loss L smooth Trend loss L trend Peak loss L peak and potential spatial regularization loss L KL The total loss function is expressed as: (28) (29) (30) (31) (32) (33) (34) in, λ 1 to λ 6 represents the weighting coefficient for each loss term. V Represents the actual voltage sequence. This represents the reconstructed voltage sequence. In one embodiment, the joint loss function can be extended according to the reconstruction accuracy requirements, dynamic consistency requirements, and application scenario requirements to introduce additional constraint terms.

[0032] Based on the total loss function, the lithium-ion battery data compression and reconstruction model jointly constructed by low-dimensional latent representation learning and equivalent circuit mechanism constraints is iteratively trained and its parameters are updated. After training, the encoding end is used to realize low-dimensional compression of the original multivariate high-dimensional time series data of lithium-ion batteries, and the decoding end is used to realize high-fidelity reconstruction of the compressed data in combination with mechanism constraints. This reduces the data storage and transmission burden of lithium-ion batteries while providing basic support for subsequent lithium-ion battery state analysis and performance evaluation.

[0033] S5: Lithium-ion battery health status estimation, degradation trajectory prediction and fault diagnosis based on low-dimensional time-series feature representation and equivalent circuit reconstruction information.

[0034] S51: Construct historical samples for health status assessment and degradation trend analysis, and perform time-alignment and capacity-alignment processing on the historical operating sequences of lithium-ion batteries. Obtain historical operating characteristic data from multiple lithium-ion battery samples. X ={ X i} and capacity label data Y ={ Y i}, where the feature data of each lithium-ion battery sample X i Represented as a multidimensional temporal feature tensor formed after dividing according to a preset time window, wherein the preset time window is a continuous time segment divided according to a uniform rule, and the capacity is labeled with data. Y i This is represented as the capacity observation sequence corresponding to each time window.

[0035] For feature data X i Using the sample time window length as the time scale, time alignment was performed on different lithium-ion battery samples, and the maximum window size was selected. T max As a uniform length, and to construct a standard time grid G t For samples with insufficient window length, linear interpolation is used to extend them to a uniform length; for capacity label data... Y i Based on the standard time grid Gt, alignment is performed using at least one of the following methods: local regression interpolation, spline interpolation, piecewise cubic Hermite interpolation, multinomial regression interpolation, or weighted average interpolation, to obtain the capacity time series. In this implementation, the capacity alignment result is represented as: (35)

[0036] in, Γ(·) denotes the capacity interpolation function based on the standard time grid. Further, taking the first... T in The multidimensional feature sequence and capacity sequence at each time step are used as model input, and subsequent steps are taken. T out Using the capacity sequence at each time step as the prediction target, a sample for historical health status analysis and decline trajectory prediction is constructed: (36) (37) (38) in, T in Indicates the length of the input time step. T out This indicates the output prediction time step length. Through the above method, a unified alignment of historical capacity evolution sequences of different lithium-ion batteries is achieved, laying the foundation for subsequent... SOH The estimation and recession trajectory prediction provide a unified historical sample basis and time scale basis.

[0037] S52: Lithium-ion batteries based on low-dimensional timing feature characterization and equivalent circuit reconstruction information SOH Estimation methods.

[0038] For the historical operating feature sequence and capacity tag sequence constructed in step S51, an input feature vector for characterizing the health state of the lithium-ion battery is constructed by further combining the low-dimensional time-series feature representation results obtained in step S3 and the equivalent circuit reconstruction information obtained in step S4. The input feature vector includes at least one or more of the following: low-dimensional time-series features, reconstructed voltage-related features, internal state estimates, equivalent circuit parameters, and capacity evolution statistical features.

[0039] Based on this, establish a lithium-ion battery health status assessment system for the current time window. SOH The estimation model maps the input feature vector to an estimated health status value within the corresponding time window. SOH The health status estimation model is a supervised learning model. In one embodiment, it can be implemented using a Gaussian process regression model, a random forest model, a support vector regression model, a linear regression model, a gradient boosting regression model, or other machine learning regression models. Let the i-th lithium-ion battery sample be in the i-th position. t The health status feature vector constructed under each time window is: f i,t Then its corresponding SOH The estimation result can be expressed as: (39) in, G(·)express SOH Estimation model. Further, the stated... SOH The label can be obtained based on the ratio of capacity to rated capacity, expressed as: (40) in, y i,t Indicates the first i The capacity observation value of a lithium-ion battery sample at the t-th time window. y rated This represents the rated capacity. By minimizing the error between the estimated and actual values, model training and parameter optimization are achieved, thereby obtaining the lithium-ion battery's performance under various time windows. SOH The estimation results are used to achieve a quantitative assessment of the current health status of the battery.

[0040] S53: A method for predicting the degradation trajectory of lithium-ion batteries based on time-series feature encoding and sequence learning models.

[0041] Based on the historical operating feature sequence and capacity tag sequence constructed in step S51, and further combined with the low-dimensional time-series feature characterization results obtained in step S3 or the consistent time-series feature encoding results, the future capacity degradation trajectory of lithium-ion batteries is predicted. Unlike step S52, which estimates the health status for the current time window, this step predicts the capacity degradation process over multiple future time steps. Let the... i A lithium-ion battery sample at the input time window length T in The following feature sequence is The corresponding historical capacity sequence is The feature matrix at each time step. ,in N This indicates the number of individual units or the number of local observation points. d This represents the feature dimension at a single time step. First, for the feature matrix at each time step... x i,t A convolutional feature encoder is used for feature extraction to obtain the corresponding time-step feature representations: (41) in, Enc cnn (·) represents the convolutional feature encoding function. Further, the historical capacity sequence... y i,Tin Corresponding time step feature representation Concatenate the features along the feature dimension to construct the sequence model input: (42) will sequence The input sequence is used to learn the model for time series modeling, thereby obtaining a global time series representation that characterizes the battery degradation and evolution. h i : (43) in, F seq (·) represents a sequence learning model. In one embodiment, the sequence learning model is a bidirectional long short-term memory network. Based on the global temporal representation... h i Generate the future through decoding mapping T out Capacity prediction sequence at each time step: (44) in, Dec (·) represents the capacity trajectory decoding function consisting of a repeating vector layer and a fully connected layer. In one implementation, a mean squared error loss function is used to constrain the predicted capacity sequence and the true capacity sequence: (45) The above methods enable the prediction of the future capacity degradation trajectory of lithium-ion batteries, thereby characterizing their degradation trend and subsequent lifespan evolution.

[0042] S54: An anomaly identification and fault early warning method based on low-dimensional temporal feature representation and reconstructed response information.

[0043] In addition to their applications in health status estimation and degradation trajectory prediction, the aforementioned low-dimensional time-series feature representation, internal state estimation results, and voltage reconstruction response information can be further used for anomaly identification and fault early warning. Specifically, lithium-ion battery charging segment data is acquired and resampled at equal time intervals, with key points extracted to form fixed-length charging segment sequence samples. The voltage, current, temperature, and time sequences from each charging segment are input into a pre-trained low-dimensional representation and voltage reconstruction model. Low-dimensional time-series feature representations corresponding to each sample are extracted, and the corresponding voltage reconstruction results, internal state estimation results, and reconstruction deviation information are obtained. (46) in, Enc vae (·) denotes the encoding function of the low-dimensional representation model. V , I , T , τLet represent voltage, current, temperature, and time series, respectively. Let z represent the low-dimensional temporal feature representation corresponding to the sample. This low-dimensional temporal feature representation can serve as the unified feature basis for health status estimation in step S52, decay trajectory prediction in step S53, and anomaly identification in this step. The low-dimensional temporal feature representations corresponding to the same individual at different time steps are organized into a feature matrix in chronological order, and a three-dimensional temporal feature tensor is further constructed: (47) Where T represents the time step, C represents the number of units, and D represents the feature dimension. Further, the feature tensors are cumulatively summed over the time dimension to obtain the cumulative evolution features: (48) in, This represents the cumulative characterization of the latent features of a single cell over time. Through this method, the low-dimensional representation extracted from the voltage reconstruction model is transformed into an evolutionary feature input suitable for multi-cell time-series fault diagnosis.

[0044] Based on this, for each time step and each individual entity, comprehensive diagnostic features are extracted by combining its low-dimensional temporal characteristics, reconstruction error features, dynamic response bias, internal state estimate, and relative differences with other entities at the same time step. These comprehensive diagnostic features include at least the current time-time statistical features, historical time window statistical features, temporal trend and rate of change features, relative bias features, and historical window change features. Let ft,c be the comprehensive diagnostic feature corresponding to the t-th time step and the c-th entity, then: (49) in, f stat Indicates statistical characteristics, f temp Representing temporal characteristics, f rel Indicates relative characteristics, f chg The change characteristics are represented. Further, after standardization and dimensionality reduction of the comprehensive diagnostic features, anomaly scoring is performed using the Isolation Forest algorithm, the single-class Support Vector Machine algorithm, and the Mahalanobis distance method, respectively, to obtain the corresponding anomaly scores: (50) (51) (52) in, g IF (·) g SVM (·) g Mah(·) represent the anomaly detection functions: Isolation Forest, Single-Class Support Vector Machine, and Mahalanobis distance, respectively. The three anomaly scores are normalized and then integrated to obtain the final anomaly score. sens : (53) in, , , These represent the normalized anomaly scores, respectively. Furthermore, a health score is constructed based on the anomaly scores. H t,c : (54) in, H t,c Indicates the first t The time step, the first c The health score corresponds to each individual. H t,c This is a diagnostic health assessment index constructed based on anomaly scoring, used to rank the degree of individual anomalies and assist in early warning determination; it represents the abnormal risk status of an individual in a diagnostic sense and can be compared with the data obtained in step S52. SOH The estimation results complement each other and together constitute the health status assessment results of lithium-ion batteries.

[0045] Based on this, a dynamic threshold sequence θt is constructed for each individual anomaly score sequence. In one embodiment, the dynamic threshold is obtained based on the maximum value of the anomaly score of non-anomaly individuals at the same time step or an interpolated and smoothed threshold sequence: (55) in, When the cumulative sum statistic satisfies: At that time, the judgment of the first c One individual unit triggers a fault warning, among which... λ This indicates the preset alarm threshold. Furthermore, the fault warning result is output, including at least one of the following: alarm unit number, alarm start time, alarm duration, and alarm count. By employing the above method, the impact of random fluctuations at a single moment on the anomaly detection results can be suppressed, improving the stability and reliability of lithium-ion battery fault warnings. Beneficial effects of the present invention

[0046] The beneficial effects of this invention are as follows.

[0047] This invention proposes a low-dimensional compressed characterization method for lithium-ion battery voltage data. This method constructs a unified input sample based on multi-variable time-series information such as voltage, current, temperature, and time, and achieves compact encoding of high-dimensional operational data through low-dimensional latent representation learning. This reduces data storage overhead and transmission burden while preserving key information about battery operating status and dynamic evolution.

[0048] This invention proposes a method for estimating the internal state and mechanistic parameters of a battery based on the fusion of operating conditions. This method jointly models a low-dimensional latent representation with operating condition information such as current and time, and further estimates the state-of-charge sequence, open-circuit voltage sequence, and equivalent circuit parameters, thereby achieving a synergistic characterization of the internal state evolution and external operating condition response characteristics of a lithium-ion battery.

[0049] This invention proposes a voltage reconstruction and joint optimization method that integrates equivalent circuit mechanism constraints. This method introduces equivalent circuit mechanism constraints during the decoding and reconstruction process to achieve lithium-ion battery terminal voltage reconstruction, and improves reconstruction accuracy, dynamic consistency, and physical interpretability through joint loss training.

[0050] This invention proposes an application method for estimating the health status of lithium-ion batteries, predicting degradation trajectories, and diagnosing faults. Based on lithium-ion battery voltage data compression and reconstruction, this method further combines historical capacity evolution prediction and potential characteristic anomaly analysis to achieve health status estimation, degradation trajectory prediction, and single-cell fault diagnosis and early warning. This enables its application in lithium-ion battery life assessment, safety monitoring, and operation and maintenance management. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the overall process of the lithium-ion battery voltage data compression and reconstruction method described in this invention. Figure 2 This is a flowchart illustrating the process of screening, standardizing, and constructing standardized samples for lithium-ion battery operating data as described in this invention. Figure 3 This is a roadmap of the core technologies for constructing low-dimensional latent representations and fusing conditional information based on multi-dimensional temporal input as described in this invention. Figure 4 This is a roadmap for the voltage reconstruction technology based on internal state estimation and equivalent circuit mechanism constraints as described in this invention. Figure 5 The graph shows the reconstructed voltage data and dynamic operating condition response results of the lithium-ion battery described in this invention. Figure 6 This is a graph showing the results of the lithium-ion battery health status estimation and degradation trajectory prediction described in this invention. Figure 7 This is a diagram showing the fault diagnosis results of the lithium-ion battery described in this invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. For example... Figure 1As shown in the figure, this embodiment provides a method for compressing and reconstructing lithium-ion battery voltage data based on low-dimensional latent representation learning and equivalent circuit mechanism constraints. The method includes lithium-ion battery operation data screening, normalization processing and sample construction, low-dimensional latent feature extraction, internal state and mechanism parameter estimation, voltage reconstruction and joint optimization training based on second-order equivalent circuit model, as well as health status estimation, degradation trajectory prediction and fault diagnosis.

[0053] Step 1: Screening, standardization, and sample construction of lithium-ion battery operating data.

[0054] like Figure 2 As shown, raw time-series data of a lithium-ion battery during actual operation is acquired. This raw time-series data includes at least voltage, current, temperature, and time sequences, and the corresponding time segments are recorded and stored. The raw time-series data undergoes validity screening, removing invalid samples that are missing, out of bounds, exhibit abnormal fluctuations, or whose sequence length does not meet preset requirements, to ensure the integrity, reliability, and consistency of the input data. The screened voltage, current, temperature, and time sequences are standardized to reduce the impact of differences in the dimensions and value ranges of different physical quantities on subsequent model training and parameter learning. The screened time-series data is truncated and organized according to a preset sequence length to form fixed-length time-series samples that meet the model input requirements. Furthermore, the change characteristics between adjacent sampling times are extracted based on the current and voltage sequences as auxiliary characterization information reflecting dynamic response characteristics, thereby improving the ability to describe the battery's time-series changes. Based on the above processing results, standardized input samples containing voltage, current, temperature, and time series, as well as corresponding change representation information, are formed, which are used for subsequent low-dimensional latent representation construction, internal state and mechanism parameter estimation, and voltage reconstruction.

[0055] Step 2: Construct a low-dimensional latent representation based on multi-dimensional time series samples.

[0056] like Figure 3 As shown, the standardized samples obtained in step one are first organized according to a uniform time order, and the voltage sequence V, current sequence I, temperature sequence T, and time sequence t are combined along the feature dimension to form a multivariable time-series input sample for lithium-ion batteries, which is expressed as follows: (1) in, L For the preset sequence length, V i , I i , T i and t i They represent the firsti The voltage, current, temperature, and time values ​​corresponding to each sampling moment are processed by scale mapping. In this way, the lithium-ion battery runtime sequence data is organized into unified multi-dimensional time-series samples to facilitate subsequent time-series correlation feature encoding and low-dimensional compressed representation construction.

[0057] Then, the multidimensional time-series input samples are input into the encoding network to extract features of local change patterns and multivariate coupling relationships in the time dimension, thereby obtaining a deep time-series representation characterizing the evolution of the battery's operating state. The encoding process can be represented as follows: (2) in, X This represents the input multidimensional time series sample. Represents the temporal feature encoding mapping function. F This represents the deep temporal features obtained after encoding.

[0058] Subsequently, the deep time-series features are transformed and mapped to obtain a compact feature representation characterizing the overall operating state of the lithium-ion battery, and the mean vector μ and logarithmic variance vector log of the latent distribution are further output. σ ², The calculation process is as follows: (3) (4) in, W μ and W σ These are the weight matrices for the mean branch and the variance branch, respectively. b μ and b σ These are the bias terms; μ and log σ The parameters that jointly characterize the probability distribution of the input sample in the latent space are used to describe the low-dimensional latent distribution features corresponding to the input sample. Further, based on a reparameterization mechanism, the latent space is randomly sampled to obtain a low-dimensional latent representation. z The calculation formula is as follows: (5) in, σ =exp(0.5log σ ²), ε For random noise that follows a standard normal distribution, ⊙ denotes element-wise multiplication.

[0059] Through the aforementioned reparameterization process, a probabilistic compressed representation of the original multivariate high-dimensional time-series data is achieved while ensuring model trainability. Finally, the low-dimensional latent representation z is used as a compressed characterization of the lithium-ion battery's operating state and dynamic evolution characteristics for subsequent conditional fusion reconstruction, internal state and mechanism parameter estimation, and voltage reconstruction modeling. Step two effectively maps the original high-dimensional, multi-dimensional coupled lithium-ion battery operating data into low-dimensional latent features, reducing data storage and transmission costs and providing a unified low-dimensional representation basis for subsequent mechanism consistency reconstruction processing that combines internal state evolution laws, open-circuit voltage mapping relationships, and equivalent circuit mechanism constraints.

[0060] Step 3: Estimation of battery internal state and mechanism parameters based on the fusion of low-dimensional latent representation and operating condition information.

[0061] like Figure 3 As shown, the low-dimensional latent representation obtained in step two is first... z This is fused with conditional information from the lithium-ion battery's operation, including current sequence I and time sequence. t The current sequence and time sequence are combined step-by-step to form a conditional input sequence, which is expressed as follows: (6) in, L For the preset sequence length, I i and t i They represent the first i The current and time values ​​corresponding to each sampling time point are processed by scale mapping. Based on the conditional input sequence, a conditional temporal feature extraction network is used to model the temporal dependencies and conditional change patterns during the operation of the lithium-ion battery, obtaining conditional temporal features related to subsequent state estimation and voltage reconstruction. The calculation process can be expressed as follows: (7) in, This represents the conditional temporal feature extraction mapping function. H c This represents the extracted conditional temporal features.

[0062] Then, based on low-dimensional latent representation z Feature modulation parameters are generated, and the conditional time-series features are adaptively modulated to achieve effective fusion of latent state information and external operating condition information. The modulation process employs a feature modulation mechanism, the calculation formula of which is as follows: (8) (9) (10) in, and β These represent the scaling and offset parameters generated from the low-dimensional latent representation, respectively. , W β , and b β These are the corresponding weight matrices and bias terms, respectively. H f The denoting symbol represents the fused temporal features, and ⊙ represents element-wise multiplication. Through this method, the conditional temporal features can be dynamically adjusted using low-dimensional latent representations, thereby enhancing the model's ability to jointly represent different battery states and operating conditions.

[0063] Subsequently, the fused time-series features are globally aggregated and mapped to obtain a global fused feature vector representing the overall operating state of the current sample. Based on the global fused feature vector, the lithium-ion battery state-of-charge sequence is predicted. SOC In this invention, the initial state of charge (SOC) and the magnitude of SOC change are first predicted. Then, a complete SOC evolution sequence is constructed by combining a normalized weight distribution. The calculation process is as follows: (11) (12) (13) (14) in, f 1 (·) f 2 (·)and f 3 (·) represent the corresponding mapping functions. SOC init Indicates the initial state of charge, Δ SOC Indicates the magnitude of the change in state of charge. w This represents the normalized incremental weights corresponding to each time step, and cumsum(·) represents the cumulative summation operation. Using this method, a lithium-ion battery state-of-charge sequence with a constrained evolutionary form can be obtained, thereby reducing the irregular fluctuations caused by independent time-by-time predictions.

[0064] Furthermore, based on the fused feature representation, the parameters of the second-order equivalent circuit model are predicted, including the ohmic internal resistance. R 0. Resistance of the first polarization branch R 1. Resistance of the second polarization branch R 2. Capacitor of the first polarization branchC 1 and second polarization branch capacitors C 2. The calculation process is expressed as follows: (15) (16) in, f p (·) represents the equivalent circuit parameter prediction function. P This represents the set of equivalent circuit parameters obtained from the prediction. To ensure the rationality of the parameter values, the prediction results can be constrained to a preset physical range through bounded mapping.

[0065] Finally, based on the predicted sequence of charged states SOC In addition, it integrates feature representations to construct a mapping relationship between the state of charge and the open-circuit voltage, thereby obtaining the open-circuit voltage sequence. OCV In this invention, a parameterized nonlinear mapping function is preferably used. SOC - OCV The relationship is modeled, and in one implementation, it is represented by a double Sigmoid function, the expression of which is: (17) in, a , b , c , d , e , f and g These are the function parameters obtained through adaptive prediction based on fused feature representation and constraint mapping. Through the above steps, based on the fusion results of low-dimensional latent representation and conditional information, the state-of-charge sequence, open-circuit voltage sequence, and equivalent circuit parameters of the lithium-ion battery can be obtained simultaneously, thus providing an internal state characterization basis and a mechanism parameter basis for subsequent battery terminal voltage reconstruction.

[0066] Step 4: Voltage reconstruction and joint optimization training based on equivalent circuit mechanism constraints.

[0067] like Figure 4 As shown, firstly, based on the open-circuit voltage sequence obtained in step three... OCV Current sequence I Time series t Based on the equivalent circuit model parameters, a lithium-ion battery terminal voltage reconstruction model is constructed; wherein, the equivalent circuit model includes an ohmic internal resistance branch and two polarizations. RCThe polarization branches are used to characterize the ohmic voltage drop and polarization response of lithium-ion batteries under dynamic operating conditions. Then, the time interval between adjacent sampling moments is determined based on the time series, and the dynamic response of each polarization branch is discretely recursively solved based on the current input and circuit parameters to obtain the voltage of the first polarization branch. V 1 and second polarization branch voltage V 2. A lithium-ion battery terminal voltage reconstruction sequence is constructed by combining open-circuit voltage and ohmic voltage drop. V rec : (18) in, V rec This represents the reconstructed terminal voltage sequence. R 0 represents the instantaneous voltage drop caused by the ohmic internal resistance. V 1 and V 2 represents the dynamic polarization voltage corresponding to the two polarization branches. Through the above method, the mechanistic constraints of the original terminal voltage curve of the lithium-ion battery can be reconstructed. In one embodiment, the current sequence and time sequence are first restored to the corresponding physical quantity space before participating in the terminal voltage solution. The obtained terminal voltage is then mapped to the model output space to maintain consistency between mechanistic calculation and data-driven modeling.

[0068] Subsequently, to ensure that the reconstructed voltage sequence numerically approximates the original voltage curve and maintains the dynamic variation characteristics of the original data, this invention constructs a joint loss function to train and optimize the model. The joint loss function includes at least voltage reconstruction loss, difference loss, smoothing loss, trend loss, peak loss, and latent space regularization loss. Specifically, the voltage reconstruction loss constrains the overall error between the reconstructed voltage and the true voltage; the difference loss constrains the variation amplitude of the reconstructed voltage between adjacent time steps to maintain the dynamic variation trend of the voltage curve; the smoothing loss reduces local jitter and abnormal fluctuations in the reconstructed curve; the trend loss maintains the overall direction of change of the reconstructed curve; the peak loss enhances the ability to preserve the maximum and minimum characteristics of the original voltage curve; and the latent space regularization loss constrains the distribution characteristics of the low-dimensional latent representation to improve the continuity and generalization of the latent space. Based on the above losses, a total loss function is constructed, the expression of which is as follows: (19) in, L Represents the total loss function. L rec Indicates voltage reconstruction loss, L diff Indicates the difference loss. L smooth Indicates smoothing loss. L trendIndicates trend loss. L peak Indicates peak loss. L KL Denotes the Kullback-Leibler divergence regularization term of the latent space. λ 1 to λ 6 represents the weighting coefficients corresponding to each loss term. In one embodiment, the total loss function can be extended according to the reconstruction accuracy requirements, dynamic consistency requirements, and application scenario needs to introduce additional constraint terms.

[0069] Furthermore, the total loss function is used to iteratively train the lithium-ion battery data compression and reconstruction model constructed based on low-dimensional latent representation learning and equivalent circuit mechanism constraints. During training, the model parameters are optimized and updated to ensure consistency between low-dimensional compressed representation, internal state estimation, and voltage reconstruction. After model training, the original multivariate lithium-ion battery operating data can be compressed into a low-dimensional latent representation using the encoding side, and the compressed data can be reconstructed using the decoding side in conjunction with mechanism constraints. This reduces the burden of data storage and transmission for lithium-ion battery operating data while achieving high-fidelity reconstruction of the original voltage curve with consistent mechanism.

[0070] Step 5: Lithium-ion battery health status estimation, degradation trajectory prediction, and fault diagnosis based on low-dimensional time-series feature representation and equivalent circuit reconstruction information.

[0071] The historical operating characteristic sequence and capacity observation sequence of lithium-ion batteries are time-aligned and interpolated to construct a health status assessment and degradation trajectory prediction sample under a unified time scale. The multidimensional operating characteristic sequence and capacity sequence of the first Tin time steps are taken as input, and the capacity sequence of the subsequent Tout time steps is taken as the prediction target, as expressed respectively: (20) (twenty one) (twenty two) in, Indicates the input feature sequence. Represents the input capacity sequence. This indicates the target sequence for output capacity prediction. T in Indicates the length of the input time step. T out This indicates the output time step length. Through the above method, a unified alignment of historical operating data and capacity degradation sequences from different lithium-ion battery samples is achieved, providing a unified sample basis and time scale for subsequent health status estimation and degradation trajectory prediction.

[0072] Furthermore, combining the low-dimensional time-series feature representation results obtained in step 3 and the equivalent circuit reconstruction information obtained in step 4, an input feature vector for characterizing the health state of the lithium-ion battery is constructed. The input feature vector includes at least one or more of the following: low-dimensional time-series features, reconstructed voltage-related features, internal state estimates, equivalent circuit parameters, and capacity evolution statistical features. Based on the input feature vector, a lithium-ion battery health state estimation model oriented towards the current time or the current time window is established to obtain the health state estimate at the corresponding time or time window. (twenty three) in, y i,t Indicates the first i The capacity observation value of a lithium-ion battery sample at the t-th time window. y rated This indicates the rated capacity. The above method allows for a quantitative assessment of the current health status of lithium-ion batteries.

[0073] For the historical operating feature sequence and capacity label sequence constructed above, the future capacity degradation trajectory of lithium-ion batteries is predicted by further combining the low-dimensional temporal feature representation results obtained in step 3 or the temporal feature encoding results consistent with them. Unlike the aforementioned health status estimation, which assesses the current moment or the current time window, this step predicts the capacity degradation process over multiple future time steps. First, the feature matrix xi,t at each time step is represented and extracted using a convolutional feature encoder to obtain the corresponding time step feature representation. Then, the encoded feature representation is concatenated with the capacity input value at the corresponding time step to construct a joint temporal input sequence: (twenty four) The joint temporal input sequence is used to perform temporal modeling on the input sequence learning model to obtain a global temporal representation characterizing the battery degradation evolution law: (25) in, F seq (·) represents a sequence learning model. In one embodiment, the sequence learning model is a bidirectional long short-term memory network. Based on the global temporal representation... h i Generate the future through decoding mapping T out Capacity prediction sequence at each time step: (26) in, Dec(·) represents the capacity trajectory decoding function consisting of a repeating vector layer and a fully connected layer. In one implementation, a mean squared error loss function is used to constrain the predicted capacity sequence and the true capacity sequence: (27) The above methods enable the prediction of the future capacity degradation trajectory of lithium-ion batteries, thereby characterizing their degradation trend and lifespan evolution characteristics.

[0074] In addition to their applications in health status estimation and degradation trajectory prediction, the aforementioned low-dimensional time-series feature representations, internal state estimation results, and mechanism reconstruction response information can be further used for anomaly identification and fault early warning. Specifically, lithium-ion battery charging segment data is acquired and resampled at equal time intervals, with key points extracted to form fixed-length charging segment sequence samples. The voltage, current, temperature, and time sequences from each charging segment are input into a pre-trained low-dimensional representation and voltage reconstruction model. The low-dimensional time-series feature representation corresponding to each sample is extracted, and the corresponding voltage reconstruction results, internal state estimation results, and reconstruction deviation information are obtained. (28) in, Enc vae (·) denotes the encoding function of the low-dimensional representation model. V , I , T , τ Let represent voltage, current, temperature, and time series, respectively. Let z represent the low-dimensional temporal feature representation corresponding to the sample. This low-dimensional temporal feature representation can serve as the unified feature basis for health status estimation in step S52, decay trajectory prediction in step S53, and anomaly identification in this step. The low-dimensional temporal feature representations corresponding to the same individual at different time steps are organized into a feature matrix in chronological order, and a three-dimensional temporal feature tensor is further constructed: (29) in, T Indicates the number of time steps. C Indicates the number of individual units. D This represents the feature dimension. Further, the feature tensor is summed cumulatively over the time dimension to obtain the cumulative evolutionary features: (30) in, This represents the cumulative characterization of the latent features of a single cell over time. Through this method, the low-dimensional representation extracted from the voltage reconstruction model is transformed into an evolutionary feature input suitable for multi-cell time-series fault diagnosis.

[0075] Based on the aforementioned low-dimensional temporal feature representation, cumulative evolution features, and reconstruction response information, statistical features, temporal trend features, relative deviation features, and rate of change features are extracted to construct a comprehensive anomaly diagnostic feature. The comprehensive anomaly diagnostic feature is then scored using Isolation Forest, Single-Class Support Vector Machine, and Mahalanobis distance methods, respectively. Finally, the anomaly scores are normalized and integrated to obtain the final anomaly score. (31) in, , , These represent the normalized anomaly scores, respectively. Furthermore, a health score is constructed based on the anomaly scores. H t,c : (32) in, H t,c Indicates the first t The time step, the first c The health score corresponds to each individual. H t,c This is a diagnostic health assessment index constructed based on anomaly scoring, used to rank the degree of individual anomalies and assist in early warning determination; it mainly characterizes the abnormal risk status of an individual in a diagnostic sense, and can be combined with the aforementioned... SOH The estimation results complement each other and together constitute the health status assessment results of lithium-ion batteries.

[0076] Based on this, a dynamic threshold sequence θt is constructed for each individual anomaly score sequence. In one embodiment, the dynamic threshold is obtained based on the maximum value of the anomaly score of non-anomaly individuals at the same time step or an interpolated and smoothed threshold sequence: (33) in, When the cumulative sum statistic satisfies: At that time, the judgment of the first c One individual unit triggers a fault warning, among which... λ This indicates the preset alarm threshold. Further, it outputs fault warning results, including at least one of the following: alarm cell number, alarm start time, alarm duration, and alarm count. The results of lithium-ion battery health status assessment, degradation trajectory prediction, and fault warning are as follows: Figure 6 and Figure 7 As shown.

Claims

1. A method for compressing and reconstructing lithium-ion battery voltage data based on low-dimensional latent representation learning and equivalent circuit mechanism constraints, characterized in that, Includes the following steps: S1: Obtain raw multivariate time-series data during the operation of lithium-ion batteries and construct standardized input samples; Specifically, it includes: S11: Collect and store raw multivariate time-series data during the operation of the lithium-ion battery, wherein the raw multivariate time-series data includes at least voltage sequence, current sequence, temperature sequence and time sequence; S12: Perform sample validity screening and boundary constraint verification on the original multivariate time series data, remove invalid samples that are missing, abnormal, out of bounds, or whose length does not meet the requirements, and retain valid time series samples that meet the preset sequence length and physical constraint conditions. S13: Perform normalization mapping on the screened voltage, current, temperature, and time series to obtain standardized time series data; S14: Based on the standardized time series data, a fixed-length truncation is performed according to the preset sequence length to construct a fixed-length time series input sample, and auxiliary features characterizing the dynamic change process are extracted; S15: The standardized time series data and its corresponding dynamic change auxiliary features are uniformly organized to form a standardized input sample set, and output to the subsequent low-dimensional latent representation construction, internal state and mechanism parameter estimation and voltage reconstruction process; S2: Based on low-dimensional latent representation learning, the multivariate time-series input samples are compressed and encoded to obtain a low-dimensional latent representation characterizing the dynamic state evolution features of the lithium-ion battery; specifically including: S21: Based on the standardized voltage sequence, current sequence, temperature sequence and time sequence obtained in step S1, organize them in a unified time order and combine them along the feature dimension to form multivariable time series input samples; S22: Input the multivariate temporal input samples into the encoding network, extract local dynamic features and multidimensional coupling features, and obtain deep temporal features and compact feature vectors; S23: Calculate the mean vector and log-variance vector of the input sample in the latent space based on the compact feature vector to characterize the low-dimensional latent distribution features corresponding to the input sample; S24: Based on the mean vector and log-variance vector, a reparameterized sampling mechanism is used to generate a low-dimensional latent representation, thereby achieving compressed mapping of the original multivariate high-dimensional time series data and providing a low-dimensional representation basis for subsequent internal state estimation, mechanism parameter estimation, and voltage reconstruction. Its expression is: (1) in, Z For low-dimensional potential representation, μ Let log be the vector of the mean of the latent distribution. σ 2 Let the log-variance vector be the latent distribution. ε For a random disturbance term that follows a standard normal distribution, ⊙ denotes element-wise multiplication; S3: Based on the physical information fusion mechanism, constrained decoding is performed on the low-dimensional latent representation to estimate the internal state and mechanism parameters of the lithium-ion battery, and the terminal voltage is reconstructed; specifically including: S31: Based on the mean vector and log-variance vector, a reparameterized sampling mechanism is used to generate a low-dimensional latent representation to achieve compressed mapping of the original multivariate high-dimensional time series data, and to provide a low-dimensional representation basis for subsequent internal state estimation, mechanism parameter estimation and voltage reconstruction. S32: Generate feature modulation parameters based on the low-dimensional latent representation, and adaptively modulate the conditional time series features to achieve the fusion of latent state information and external operating condition information, and obtain the fused time series features; S33: Perform feature aggregation on the fused time-series features to obtain a global fused feature vector, and predict the state of charge sequence of the lithium-ion battery based on the global fused feature vector; S34: Predict equivalent circuit model parameters based on the global fusion feature vector, wherein the equivalent circuit model parameters include at least ohmic internal resistance, polarization branch resistance, and polarization branch capacitance. S35: Based on the charge state sequence and the fused feature representation, construct the mapping relationship between the charge state and the open-circuit voltage to obtain the open-circuit voltage sequence, the expression of which is: (2) in, OCV Open circuit voltage, SOC In a charged state, a , b , c , d , e , f and g These are the mapping parameters obtained from adaptive prediction based on fused feature representations; S4: Voltage reconstruction and joint optimization training based on equivalent circuit mechanism constraints; S41: Based on the open-circuit voltage sequence, current sequence, time sequence and equivalent circuit model parameters obtained in step S3, construct a lithium-ion battery terminal voltage reconstruction model. S42: Based on the time series calculation, the time interval between adjacent sampling moments is calculated, and the dynamic response of the polarization branch is discretely and recursively solved using the equivalent circuit model parameters to obtain the voltage of each polarization branch. The expression for this voltage is: (3) (4) in, V 1,k and V 2,k They represent the first k The voltages of the first polarization branch and the second polarization branch corresponding to each sampling time, Δ t k = t k -t k−1 The parameters of the second-order equivalent circuit model include at least the ohmic internal resistance. R 0. Resistance of the first polarization branch R 1. Resistance of the second polarization branch R 2. Capacitor of the first polarization branch C 1 and second polarization branch capacitors C 2. Using the above method, discrete-time modeling of the dynamic polarization process of lithium-ion batteries can be achieved; S43: Based on the open-circuit voltage, ohmic voltage drop and voltage of each polarization branch, construct the terminal voltage reconstruction sequence and perform output space mapping to obtain the reconstructed voltage sequence; (5) in, V rec This represents the reconstructed terminal voltage sequence. Through this method, the mechanistic constraints of the original voltage curve of a lithium-ion battery can be reconstructed. S44: Construct a joint loss function to jointly train and optimize the low-dimensional latent representation learning module, the internal state and mechanism parameter estimation module, and the voltage reconstruction module, so as to achieve low-dimensional compressed expression and mechanism constraint reconstruction of lithium-ion battery voltage data. S5: Lithium-ion battery health status estimation, degradation trajectory prediction and fault diagnosis based on low-dimensional time-series feature representation and equivalent circuit reconstruction information; S51: Construct historical samples for health status assessment and degradation trend analysis, and perform time alignment and capacity alignment processing on the historical operating sequences of lithium-ion batteries; S52: Based on the low-dimensional time-series feature representation results, internal state estimation results, and equivalent circuit reconstruction information, input features for lithium-ion battery health state estimation are constructed, and health state estimation results are obtained. S53: Based on historical operating characteristic sequences, capacity tag sequences, and low-dimensional time series characteristic characterization results, predict the future capacity degradation trajectory of lithium-ion batteries; S54: Based on low-dimensional time-series feature representation, reconstruction error information, dynamic response deviation and internal state estimation results, extract comprehensive diagnostic features to perform anomaly identification, health scoring and fault early warning judgment.

2. The lithium-ion battery voltage data compression and reconstruction method according to claim 1, characterized in that, The low-dimensional latent representation in step S2 is a compact representation obtained by jointly encoding the multivariate coupled time-series features composed of voltage, current, temperature and time series. It is used to retain the state evolution information of lithium-ion batteries under dynamic operating conditions and to realize the compressed expression of the original multivariate high-dimensional time-series data.

3. The lithium-ion battery voltage data compression and reconstruction method according to claim 1, characterized in that, In step S3, the low-dimensional latent representation is fused with the operating condition information, and the charge state sequence, open circuit voltage sequence and second-order equivalent circuit model parameters are estimated based on the fused time-series features to achieve physical information fusion representation for lithium-ion battery terminal voltage reconstruction.

4. The lithium-ion battery voltage data compression and reconstruction method according to claim 1, characterized in that, In step S4, a mechanism-constrained voltage reconstruction model is constructed based on the mapping relationship between the state of charge and the open-circuit voltage and the dynamic response of the second-order equivalent circuit. The distribution stability of the low-dimensional latent representation, the terminal voltage reconstruction error, and the consistency of dynamic changes are simultaneously constrained by the joint loss function, so as to achieve compressed storage and transmission of lithium-ion battery operating data and high-fidelity reconstruction of the terminal voltage.

5. The lithium-ion battery voltage data compression and reconstruction method according to claim 1, characterized in that, In step S5, based on the low-dimensional time-series feature characterization results obtained in step S2 and the internal state estimation results, equivalent circuit parameters, and voltage reconstruction information obtained in steps S3 and S4, input features are constructed for lithium-ion battery health state estimation, capacity degradation trajectory prediction, anomaly identification, and fault warning. The input features are used to realize lithium-ion battery health state estimation, future capacity degradation process prediction, and anomaly scoring, health scoring, and fault warning determination.

Images