Data enhancement method for improving reliability of battery system
By combining adaptive filtering and generative AI models, the problems of traditional filtering methods failing to effectively remove noise and data-driven models lacking physical constraints are solved. The generated battery data reduces noise and conforms to electrochemical laws, thereby improving the reliability and lifespan of the battery system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-04-14
AI Technical Summary
In existing battery system data processing technologies, traditional filtering methods cannot effectively remove noise without damaging key features. Data-driven models lack physical constraints, resulting in generated data that does not conform to the internal laws of the battery, affecting the reliability and accuracy of the battery management system.
This paper adopts a method that combines adaptive filtering with generative AI model. The adaptive filtering algorithm suppresses noise and retains key features, and the data is augmented by a hybrid model of Transformer architecture and physical information neural network. Electrochemical physical laws are embedded as constraints to generate data that is both low in noise and electrochemically reasonable.
The generated data reduces noise and conforms to the internal laws of the battery, improving the operational reliability and lifespan of the battery system, and providing data support with high signal-to-noise ratio and electrochemical consistency.
Smart Images
Figure CN121856798A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy vehicle battery data analysis technology, specifically to a data augmentation method for improving the reliability of battery systems. Background Technology
[0002] In the full lifecycle operation and management of battery systems, continuously collected time-series data such as voltage, current, and temperature are crucial for battery status assessment, health diagnosis, and safety management. In-depth analysis of this operational data allows for accurate estimation of battery health and remaining lifespan, enabling optimization of charge / discharge strategies to extend lifespan and providing early warnings of serious safety hazards such as thermal runaway, ultimately improving the reliability and economy of the entire battery system. However, in practical engineering applications, the data acquisition process is rarely ideal. The interplay of sensor quantization errors, power frequency interference from the grid, and random noise from complex electromagnetic environments results in widespread high-frequency oscillations and transient spikes in the raw signal. These noise components highly overlap with the actual characteristics of battery electrochemical state changes (such as voltage drops caused by internal micro-short circuits or abnormal charging curve plateaus due to lithium plating) in both the time and frequency domains, forming a fundamental obstacle to subsequent advanced data analysis.
[0003] To address the aforementioned data quality issues, traditional methods typically rely on classical signal processing techniques, particularly various forms of low-pass filters. While these methods can suppress high-frequency noise to some extent, their inherent design principles have significant flaws. The selection of the filter's cutoff frequency and order often presents a dilemma: if set too leniently, noise removal is incomplete, still interfering with model judgments; if set too strictly, while filtering noise, it inevitably smooths out or even completely erases key transient features characterizing early battery failures, internal phase transitions, or charge-discharge phase transitions. This damage to true physical characteristics renders subsequent data-driven evaluation models built on distorted information, fundamentally limiting their accuracy and reliability.
[0004] To overcome the limitations of traditional filtering, researchers have recently begun exploring data-driven artificial intelligence methods, particularly deep learning models, to directly learn complex patterns from massive datasets and perform data augmentation or reconstruction. These methods, such as those based on generative adversarial networks or variational autoencoders, exhibit powerful nonlinear fitting capabilities. However, they reveal new problems in the physically constrained scenario of batteries: these purely data-driven models lack an understanding of the fundamental electrochemical principles within batteries. The data they generate may appear statistically reasonable but severely violate core physical laws such as mass conservation, charge conservation, and diffusion kinetics. This "physical inconsistency" means that while the generated data may have low noise, the battery behavior it reflects is impossible in the real physical world. State assessment models trained with this data will inevitably produce biases or even serious misjudgments in practical applications, failing to meet the stringent requirements of high-reliability battery management systems for data authenticity and physical feasibility. Summary of the Invention
[0005] The present invention aims to provide a data augmentation method to improve the reliability of battery systems, which can ensure that the augmented data has both low noise characteristics and electrochemical rationality, thereby helping to comprehensively improve the operational reliability and service life of battery systems.
[0006] The basic solution provided by this invention is: a data augmentation method for improving the reliability of a battery system, comprising the following steps: S1. Signal preprocessing step: Normalize the collected raw battery operation data. S2. Adaptive filtering step: Based on the normalized data, an adaptive filtering algorithm is used to dynamically suppress noise and retain key transient features during battery charging and discharging, thereby obtaining basic data. S3. Physical constraint data enhancement step: The basic data is input into the generative AI model that integrates physical constraints; The generative AI model reconstructs and enhances the input data by embedding the electrochemical physical laws describing the internal processes of the battery as constraints into the model training and inference process, generating enhanced battery data that has both low noise characteristics and electrochemical rationality.
[0007] Furthermore, the generative AI model that integrates physical constraints is a hybrid model based on the Transformer architecture and the Physical Information Neural Networks (PINNs); wherein, the Transformer architecture is used to extract long-term temporal dependency features of battery data, and the PINNs are used to use the electrochemical partial differential equations (PDEs) describing the internal processes of the battery as the loss function of the soft-constraint embedded model.
[0008] Furthermore, the use of electrochemical PDE as the loss function of the soft-constraint embedding model specifically includes: The electrochemical PDE is discretized and converted into a differentiable operator; Based on the model predictions and the differentiable operator, the physical loss is calculated, which is used to quantify the degree to which the model predictions violate the electrochemical PDE and boundary conditions. The physical loss is added as a penalty term to the model's total loss function.
[0009] Furthermore, the total loss function is expressed as: ; in, For data reconstruction loss, ; The original input data for the model. The reconstructed data output by the model; For physical loss, ; It is a time constant; The number of discrete points in space; For hyperparameters, ; This represents the squared residuals of the PDE governing equations. The residual squared term for the boundary conditions; For regularization loss, ; These are the learnable parameters of the model. The square of the L2 norm; , , These are dynamically adjusted weighting coefficients; S3 also includes a loss balancing step, which dynamically adjusts the loss balance based on the training phase. and The ratio is used to strike a balance between data fit fidelity and adherence to physical laws.
[0010] Furthermore, the adaptive filtering step employs the Normalized Least Mean Square (NLMS) algorithm, and the update formula for its filter weights is as follows: ; in, This refers to the k-th weight coefficient at time n. Step size factor; This is an instantaneous error; For input signals; The input vector; The square of the L2 norm; It is a regularization constant less than 1.
[0011] Furthermore, after S1 and before S2, the method further includes: based on a noise characteristic model, decomposing the noise in the original battery operating data into power frequency interference, transient pulses, and environmental Gaussian noise components.
[0012] Furthermore, after S2 and before S3, a data slicing and standardization step is also included: the continuous filtered data is sliced according to a preset window length and step size, and Z-score standardization is performed on each data window to generate standardized time-series data segments.
[0013] Furthermore, after S2 and before S3, the method further includes the step of: performing sine / cosine position encoding on the timestamp information in the time-series data segment, and adding the encoding result to the data segment to enable the generative AI model to perceive absolute temporal relationships.
[0014] Furthermore, after S3, an enhanced data evaluation step is included: the signal-to-noise ratio is calculated and the electrochemical consistency is verified on the generated enhanced battery data.
[0015] Furthermore, the raw battery operating data includes voltage, current, and temperature data.
[0016] The working principle and advantages of this invention are as follows: This invention provides a data augmentation method for improving the reliability of battery systems. The augmented data ensures both low noise characteristics and electrochemical rationality, contributing to a comprehensive improvement in the operational reliability and lifespan of the battery system. The key points are: First, this solution constructs a closed-loop collaborative enhancement architecture, which connects front-end signal processing and back-end generative AI into an organic whole, achieving complementary advantages and synergistic effects between the two technical paradigms.
[0017] The signal processing layers (S1-S2) efficiently suppress noise while preserving key features, providing high-quality input with significantly reduced noise levels and maximum preservation of key features for subsequent generative AI models. This greatly alleviates the learning burden and difficulty of the AI models. The generative AI layer (S3) performs deep reconstruction, imputation, and enhancement of data based on physical constraints. This ensures that the data reconstruction and enhancement process always operates within scientific boundaries, guaranteeing that the output data not only statistically approximates reality but is also physically reliable. Furthermore, the output data can be fed back to optimize the front-end processing or used directly as the final result. The synergy between these two layers ensures that the final generated battery data possesses both a high signal-to-noise ratio and rigorous electrochemical rationality, overcoming the limitations of traditional single-method approaches.
[0018] Second, this solution employs adaptive filtering technology, which can accurately preserve key transient features.
[0019] The adaptive filtering algorithm used in this solution can track the changes in the statistical characteristics of signals and noise in real time and dynamically adjust their filtering weights. It can effectively filter out high-frequency noise while accurately distinguishing and retaining the real transient features that are crucial for battery status assessment. This avoids the shortcomings of traditional fixed low-pass filters that easily smooth out these key information, and provides a more informative data foundation for subsequent AI models.
[0020] Third, this plan has strong physical rationality.
[0021] This approach introduces a hybrid architecture combining a physical information neural network (PEN) and a Transformer model. It utilizes the Transformer's multi-head attention mechanism to capture long-term temporal dependencies in battery data, simulating their cross-timescale interactions. Simultaneously, the PSN discretizes the electrochemical partial differential equations describing internal battery processes (such as lithium-ion concentration distribution and potential changes) into differentiable operators, which are then embedded as a soft constraint loss function within the neural network. By calculating this "physical loss," the degree of violation of physical laws is quantified, and this value guides the model's optimization. The enhanced data generated by this architecture not only fits the observed patterns but also adheres to the battery's intrinsic electrochemical principles, ensuring the data's physical feasibility and authenticity, and significantly improving its credibility and reliability.
[0022] This solution possesses absolute temporal relationship awareness capabilities.
[0023] The charging and discharging process of a battery is a highly time-dependent dynamic process. Its state changes exhibit inherent periodicity, relaxation effects, and memory characteristics, and a simple sequence of data points cannot fully describe its causal relationships over time. This solution creatively embeds sine and cosine positional encodings into the timestamp information before inputting the time-series data into the Transformer model. This step infuses the model with a sense of the absolute temporal order and relative time intervals of the data points, enabling the model to understand whether a data point is in the early or late stages of charging, and the length of the interval between two cycles. This allows for a more accurate modeling of the evolution of battery behavior over time, such as the relaxation recovery process of open-circuit voltage and the trend of polarization voltage changes at different charging rates. Through this awareness of absolute temporal relationships, the time-series data fragments generated by this solution are not only numerically reasonable but also dynamically consistent with the actual operating characteristics of the battery, providing higher-quality data support for building accurate lifetime prediction models and adaptive charging and discharging strategies. Attached Figure Description
[0024] Figure 1 This is a schematic flowchart of an embodiment of a data augmentation method for improving the reliability of a battery system according to the present invention. Figure 2This is a schematic diagram of the generative AI model operation process of an embodiment of a data augmentation method for improving the reliability of a battery system according to the present invention. Detailed Implementation
[0025] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A data augmentation method for improving the reliability of a battery system includes the following steps: S1. Signal preprocessing step: Normalize the collected raw battery operating data.
[0026] The raw battery operating data includes voltage, current, and temperature data.
[0027] Specifically, in this embodiment, the normalization process includes the following operations: Missing value imputation: To address potential signal loss during data acquisition, time series prediction methods are used for imputation. For example, an ARIMA(1,1,1) model (autoregressive order p=1, differencing order d=1, moving average order q=1) is used. The model parameters are chosen based on the analysis of autocorrelation and partial autocorrelation plots of a large number of battery steady-state operation data sequences, selecting a generally applicable order. For each signal channel with a missing point (e.g., voltage U), the 50 valid data points before and after the missing point are used as the training set to build an ARIMA model and predict the value of the missing point, thus completing the imputation.
[0028] Sliding normalization: To eliminate the influence of different physical dimensions such as voltage, current, and temperature, and to suppress slowly changing baseline drift, sliding window normalization is used. For example: setting a window with a length of... A sliding window, long enough to cover several typical charge / discharge pulses. For data points within the window... Its normalized value is ,in, and These are the mean and standard deviation of the data within the window, respectively. This operation sets the mean to 0 and the standard deviation to 1 within each local window, preserving the relative variation characteristics.
[0029] Downsampling: Considering that battery state changes are typically a slow process, excessively high sampling rates introduce unnecessary computational complexity and high-frequency details (mostly noise). After completing the above steps, the data is downsampled by averaging adjacent points, reducing the sampling rate from 1Hz to 0.2Hz. This frequency is set based on the typical requirements of the battery management system for state update frequency (e.g., once every 1-5 seconds).
[0030] S1.1 Noise characteristic modeling step: Based on the noise characteristic model, the noise in the original battery operation data is decomposed into power frequency interference, transient pulses and environmental Gaussian noise components.
[0031] Specifically, to perform targeted filtering, the noise components are mathematically described. Based on the analysis of interference sources in the battery testing environment, the total noise is... The model is a superposition of three parts: ; in, For power frequency interference; t is the time variable, and A is the amplitude of the power frequency interference. The frequency is the power frequency, typically the standard frequency of the power grid. In this embodiment, .
[0032] For transient pulses; K is the total number of transient pulses. It is the Dirac function; The time when the k-th transient pulse occurs. The intensity (amplitude) of the k-th transient pulse.
[0033] The environment is Gaussian noise; the model is zero-mean Gaussian white noise with variance. It is estimated by calculating the variance of the signal during the quiescent period without charging or discharging activity (which can be determined by the current signal).
[0034] S2. Adaptive filtering step: Based on the normalized data, an adaptive filtering algorithm is used to dynamically suppress noise and retain key transient features during battery charging and discharging, thereby obtaining basic data.
[0035] The adaptive filtering step employs the Normalized Least Mean Square (NLMS) algorithm, and the update formula for its filter weights is as follows: ; The filtered signal output by the filter is: ; in, This refers to the k-th weight coefficient at time n. Step size factor; This is an instantaneous error; For input signals; The input vector; The square of the L2 norm; It is a regularization constant less than 1. This represents the length of the filter.
[0036] In this embodiment, This can balance convergence speed and stability; This is used to prevent numerical instability caused by excessively low energy in the input vector.
[0037] The NLMS algorithm adjusts the weights through error e(n) feedback, which enables it to effectively suppress the noise modeled in S2. At the same time, due to its adaptive characteristics, its response to real transients such as charging and discharging inflection points is far superior to that of a low-pass filter with fixed coefficients.
[0038] S2.1 Data Slicing and Standardization Steps: Slice the continuous filtered data into slices according to the preset window length and step size, and perform Z-score standardization on each data window to generate standardized time-series data segments.
[0039] Specifically, the data is divided into fixed-length windows. The window length is set. Step length This window length is sufficient to capture a basic charging and discharging process, and the step size is smaller than the window length, ensuring data continuity and generating a sufficient number of training samples.
[0040] Data within each slice window Standardization can be carried out independently: .in, and These are the mean and standard deviation of the window. This operation ensures that each input sample has a uniform distribution before being fed into the AI model, accelerating model training convergence.
[0041] S2.2. Encode the timestamp information in the time series data segment using sine / cosine positions, and add the encoding result to the data segment to give the generative AI model (specifically the Transformer architecture) the ability to perceive absolute temporal relationships.
[0042] Specifically, for a length of The sequence has a position index pos of [ .
[0043] For each position pos and dimension index i, compute the position encoding vector P: ; ; Among them, model feature dimensions It can achieve a good balance between expressive power and computational cost.
[0044] The location code P is compared with the standardized data slice in S2.1. Add point by point: At this point, each data point contains its numerical information and its absolute / relative position in the sequence.
[0045] S3. Physical constraint data enhancement step: The basic data is input into the generative AI model that integrates physical constraints; The generative AI model reconstructs and enhances the input data by embedding the electrochemical physical laws describing the internal processes of the battery as constraints into the model training and inference process, generating enhanced battery data that has both low noise characteristics and electrochemical rationality.
[0046] like Figure 2 As shown, the generative AI model that integrates physical constraints is a hybrid model based on the Transformer architecture and the Physical Information Neural Network (PINNs). The Transformer architecture is used to extract long-term temporal dependency features of battery data, and the PINNs are used to use the electrochemical partial differential equations (PDEs) describing the internal processes of the battery as the loss function of the soft-constraint embedded model.
[0047] Specifically, we first use a multi-head attention mechanism to extract long-term dependency features.
[0048] Will Input a Transformer encoder layer to capture long-term timing dependencies between voltage, current, and temperature.
[0049] For input Perform three linear transformations to generate the query matrix Q, the key matrix K, and the value matrix V.
[0050] Set attention count The Q, K, and V components are divided into four dimensions: Submatrices, parallel computation of scaled dot product attention: For each head i: ; ; in, This is a scaling factor used to prevent the gradient from vanishing in the Softmax function. Let be the query weight of the i-th head.
[0051] The outputs of the four attention heads are concatenated and fused through a linear layer. Then, they are passed through a layer normalization (LayerNorm) and a two-layer feedforward neural network (FFN, with 512 hidden layers), using the ReLU activation function. The final output is denoted as... This represents a time-series feature rich in long-term dependencies: .
[0052] Furthermore, the use of electrochemical PDE as the loss function of the soft-constraint embedding model specifically includes: The electrochemical PDE (partial differential equation) is discretized and transformed into a differentiable operator; Based on the model predictions and the differentiable operator, the physical loss is calculated, which is used to quantify the degree to which the model predictions violate the electrochemical PDE and boundary conditions. The physical loss is added as a penalty term to the model's total loss function.
[0053] The following explanation uses a one-dimensional lithium-ion solid-phase diffusion equation as an example. This equation describes the lithium-ion concentration in the battery cathode material. The patterns of change over time and space are one of the core PDEs in battery electrochemical models.
[0054] Electrochemical PDE equation: ; Parameter meaning: Lithium ion concentration is a spatial location and time The function; The solid-phase diffusion coefficient of lithium ions in the electrode material is assumed to be constant. Spatial coordinates, with a range of values. ,in Electrode thickness; : Time variable.
[0055] Boundary condition 1 (electrode surface, ): ; in To determine based on the external current of the battery The calculated surface concentration satisfies: ; here This represents the effective surface area of the electrode.
[0056] Boundary condition 2 (electrode center, ): ; This indicates that the concentration gradient is zero at a position symmetrical to the center of the electrode.
[0057] Discretization process: spatial interval Evenly divided into The time interval is divided into points. Let there be a step size. ; ; in , For time step.
[0058] Discretize the diffusion term using the central difference scheme: ; Using the forward difference discretization time term: ; Substituting into the original PDE, we obtain the discretized equation: ; After organizing, it can be written as: ; in The residuals are the control equations.
[0059] Discretization of boundary conditions: for (surface): ; for (center): ; Specifically, the total loss function is expressed as: ; in, For data reconstruction loss, ; The original input data for the model. The reconstructed data output by the model; For physical loss, ; It is a time constant; The number of discrete points in space; For hyperparameters, ; This represents the squared residuals of the PDE governing equations. The residual squared term for the boundary conditions; For regularization loss, ; These are the learnable parameters of the model. The square of the L2 norm; , , These are dynamically adjusted weighting coefficients.
[0060] Furthermore, by fusing temporal features and physical constraints through a fully connected decoder, enhanced data with low noise and physical decoupling is output.
[0061] Specifically, time series characteristics Decoding is performed through a fully connected layer, reconstructing the input... Augmented data with the same dimensions During this process, the physical loss calculated by S8 will affect the parameter updates of the decoder through backpropagation, causing... It not only resembles clean data in form, but also adheres to the underlying physical laws of electrochemistry.
[0062] 3.1 Loss balancing steps are dynamically adjusted according to the training phase. and The ratio is used to strike a balance between data fit fidelity and adherence to physical laws. For example, a smaller ratio can be set in the initial stage. This allows the model to learn the basic patterns of the data first; in the later stages of training, the model can be dynamically increased through strategies. To strengthen physical constraints.
[0063] S4. Enhanced data evaluation step: Calculate the signal-to-noise ratio and verify the electrochemical consistency of the generated enhanced battery data.
[0064] The signal-to-noise ratio calculation includes: Treat it as a signal, This is considered residual noise. Calculate the ratio of signal power to noise power and convert it to decibels (dB). The signal-to-noise ratio should be significantly improved compared to the original data (e.g., more than 10 dB).
[0065] The electrochemical consistency verification includes: The current and voltage data are substituted into a high-precision battery electrical model (such as an equivalent circuit model, ECM) to estimate the state of charge (SOC) trajectory. The SOC trajectory is then checked to ensure it is smooth, monotonic, and without abrupt changes. Simultaneously, during the constant current phase, the voltage curve is examined to confirm that it conforms to the typical polarization characteristics at that current. Any phenomenon that violates these fundamental electrochemical principles indicates that the physical constraints are not fully effective.
[0066] This embodiment provides a data augmentation method to improve the reliability of a battery system, which ensures that the augmented data has both low noise characteristics and electrochemical rationality, thus helping to comprehensively improve the operational reliability and service life of the battery system.
[0067] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.
Claims
1. A data augmentation method for improving the reliability of a battery system, characterized in that, Includes the following steps: S1. Signal preprocessing step: Normalize the collected raw battery operation data. S2. Adaptive filtering step: Based on the normalized data, an adaptive filtering algorithm is used to dynamically suppress noise and retain key transient features during battery charging and discharging, thereby obtaining basic data. S3. Physical constraint data enhancement step: The basic data is input into the generative AI model that integrates physical constraints; The generative AI model reconstructs and enhances the input data by embedding the electrochemical physical laws describing the internal processes of the battery as constraints into the model training and inference process, generating enhanced battery data that has both low noise characteristics and electrochemical rationality.
2. The data augmentation method for improving the reliability of a battery system according to claim 1, characterized in that, The generative AI model that integrates physical constraints is a hybrid model based on the Transformer architecture and Physical Information Neural Networks (PINNs). The Transformer architecture is used to extract long-term temporal dependency features of battery data, and the PINNs are used to use the electrochemical partial differential equations (PDEs) describing the internal processes of the battery as the loss function of the soft-constraint embedded model.
3. The data augmentation method for improving the reliability of a battery system according to claim 2, characterized in that, The use of electrochemical PDE as the loss function in the soft-constraint embedding model specifically includes: The electrochemical PDE is discretized and converted into a differentiable operator; Based on the model predictions and the differentiable operator, the physical loss is calculated, which is used to quantify the degree to which the model predictions violate the electrochemical PDE and boundary conditions. The physical loss is added as a penalty term to the model's total loss function.
4. The data augmentation method for improving the reliability of a battery system according to claim 3, characterized in that, The total loss function is expressed as: ； in, For data reconstruction loss, ; The original input data for the model. The reconstructed data output by the model; For physical loss, ; It is a time constant; The number of discrete points in space; For hyperparameters, ; This represents the squared residuals of the PDE governing equations. The residual squared term for the boundary conditions; For regularization loss, ; These are the learnable parameters of the model. The square of the L2 norm; , , These are dynamically adjusted weighting coefficients; S3 also includes a loss balancing step, which dynamically adjusts the loss balance based on the training phase. and The ratio is used to strike a balance between data fit fidelity and adherence to physical laws.
5. The data augmentation method for improving the reliability of a battery system according to claim 1, characterized in that, The adaptive filtering step employs the Normalized Least Mean Square (NLMS) algorithm, and the update formula for its filter weights is as follows: ； in, This refers to the k-th weight coefficient at time n. Step size factor; This is an instantaneous error; For input signals; The input vector; The square of the L2 norm; It is a regularization constant less than 1.
6. The data augmentation method for improving the reliability of a battery system according to claim 1, characterized in that, After S1 and before S2, the method further includes: based on a noise characteristic model, decomposing the noise in the original battery operating data into power frequency interference, transient pulses, and environmental Gaussian noise components.
7. The data augmentation method for improving the reliability of a battery system according to claim 1, characterized in that, After S2 and before S3, the process also includes data slicing and standardization steps: the continuous filtered data is sliced according to a preset window length and step size, and Z-score standardization is performed on each data window to generate standardized time-series data segments.
8. The data augmentation method for improving the reliability of a battery system according to claim 7, characterized in that, After S2 and before S3, the method further includes the step of: performing sine / cosine position encoding on the timestamp information in the time series data segment, and adding the encoding result to the data segment to give the generative AI model the ability to perceive absolute temporal relationships.
9. The data augmentation method for improving the reliability of a battery system according to claim 1, characterized in that, Following S3, an enhanced data evaluation step is also included: performing signal-to-noise ratio calculation and electrochemical consistency verification on the generated enhanced battery data.
10. A data augmentation method for improving the reliability of a battery system according to claim 1, characterized in that, The raw battery operating data includes voltage, current, and temperature data.