Lithium battery soc voltage evaluation method and system

Abstract

The application relates to the technical field of lithium battery evaluation, and discloses a lithium battery SOC voltage evaluation method and system. The method comprises the following steps: acquiring multiple groups of historical charging and discharging data of the voltage, current and temperature of a lithium battery, performing curve alignment through dynamic time warping to generate a standardized charging and discharging feature set; inputting the feature set into a pre-trained deep neural network model to extract features and obtain a high-order charging and discharging feature matrix, on the basis of which a dynamic voltage evaluation baseline model is constructed and parameters are adjusted in real time; acquiring real-time charging and discharging data, comparing the real-time charging and discharging data with the dynamic baseline model to generate a voltage deviation degree score set, and detecting an abnormal voltage fluctuation period according to the voltage deviation degree score set; performing multi-scale feature decomposition on the data of the abnormal period to obtain local and global abnormal feature vectors, tracing back to an abnormal root cause to determine key influencing factors, and finally generating an SOC voltage calibration strategy. The method can comprehensively capture the dynamic changes of the battery, accurately identify abnormalities and trace back to the root cause, adapt to complex working conditions, and guarantee the operation performance of the lithium battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery evaluation, in particular to a lithium battery SOC voltage evaluation method and system. BACKGROUND

[0002] In the process of rapid development of new energy industry, as the core component of energy storage and power battery, the accurate evaluation of state-of-charge (SOC) of lithium battery is directly related to the operation efficiency and safety performance of the equipment. The SOC evaluation methods commonly used in the current industry mostly rely on single-dimensional voltage data or fixed mathematical models. However, the electrochemical characteristics of lithium battery are easily affected by various factors such as temperature change, current fluctuation and increase of cycle number during actual charging and discharging process, which makes it difficult for traditional methods to accurately capture the dynamic changes of battery state.

[0003] For example, some methods achieve evaluation by pre-setting fixed voltage-SOC corresponding curves. However, in low temperature environment, the polarization effect of lithium battery is enhanced, and the corresponding relationship between actual voltage and SOC deviates from the pre-set curve, thereby generating a large evaluation error. Another method introduces a temperature compensation mechanism, but can only correct for a specific temperature interval and cannot adapt to complex and variable actual working conditions. At the same time, traditional methods often ignore the differences in curve shape under different charging and discharging cycles during data processing, and directly use original data for model training, which leads to insufficient adaptability of the model to long-term changes such as battery aging and performance degradation.

[0004] With the expansion of lithium battery application scenarios, from portable electronic devices to electric vehicles and large-scale energy storage systems, higher requirements are put forward for the real-time and accuracy of SOC evaluation. When the evaluation error exceeds the allowed range, it may cause overcharging, overdischarging and other problems, not only shortening the service life of the battery, but also possibly leading to safety hazards such as thermal runaway. In addition, existing methods are difficult to quickly locate the root cause of voltage abnormal fluctuation, and can only achieve abnormal alarm, which cannot provide effective guidance for subsequent voltage calibration and fault troubleshooting, further limiting their application in complex systems. SUMMARY

[0005] The present application aims to provide a lithium battery SOC voltage evaluation method and system to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a lithium battery SOC voltage evaluation method, which comprises:

[0007] obtaining a plurality of groups of historical charging and discharging data of the lithium battery, the historical charging and discharging data comprising voltage curves, current curves and temperature curves;

[0008] performing dynamic time warping-based curve alignment processing on the historical charge-discharge data to generate a standardized charge-discharge feature set;

[0009] inputting the standardized charge-discharge feature set into a pre-trained deep neural network model for feature extraction to generate a high-order charge-discharge feature matrix;

[0010] constructing a dynamic voltage evaluation baseline model based on the high-order charge-discharge feature matrix and adjusting parameters of the dynamic voltage evaluation baseline model in real time;

[0011] obtaining real-time charge-discharge data and comparing the real-time charge-discharge data with the dynamic voltage evaluation baseline model to generate a voltage deviation degree score set;

[0012] performing abnormal interval detection on the real-time charge-discharge data according to the voltage deviation degree score set to identify an abnormal voltage fluctuation period;

[0013] performing multi-scale feature decomposition on charge-discharge data corresponding to the abnormal voltage fluctuation period to generate a local abnormal feature vector and a global abnormal feature vector;

[0014] performing abnormal root cause tracing based on the local abnormal feature vector and the global abnormal feature vector to determine key influencing factors causing voltage abnormalities;

[0015] generating an SOC voltage calibration strategy according to the key influencing factors.

[0016] Preferably, the dynamic time warping-based curve alignment processing on the historical charge-discharge data to generate a standardized charge-discharge feature set specifically includes:

[0017] performing time series segmentation on the voltage curve, the current curve, and the temperature curve respectively to generate a plurality of local time series segments;

[0018] calculating a dynamic time warping distance between each local time series segment and a preset reference curve;

[0019] performing time axis stretching adjustment on the local time series segments according to the dynamic time warping distance to generate an aligned standardized time series;

[0020] performing feature fusion on the aligned standardized time series to generate the standardized charge-discharge feature set.

[0021] Preferably, the inputting the standardized charge-discharge feature set into a pre-trained deep neural network model for feature extraction to generate a high-order charge-discharge feature matrix specifically includes:

[0022] performing local feature extraction on the standardized charging and discharging feature set through a convolutional layer of the deep neural network model to generate a primary feature map;

[0023] performing feature weight distribution on the primary feature map through an attention mechanism layer of the deep neural network model to generate a weighted feature map;

[0024] performing feature dimension reduction on the weighted feature map through a fully connected layer of the deep neural network model to generate the high-order charging and discharging feature matrix.

[0025] Preferably, a dynamic voltage evaluation baseline model is constructed based on the high-order charging and discharging feature matrix, and parameters of the dynamic voltage evaluation baseline model are adjusted in real time, specifically including:

[0026] performing cluster analysis on the high-order charging and discharging feature matrix to determine a plurality of typical charging and discharging modes;

[0027] establishing an initial voltage evaluation baseline model according to the typical charging and discharging modes;

[0028] iteratively updating parameters of the initial voltage evaluation baseline model through an online learning algorithm combined with real-time charging and discharging data.

[0029] Preferably, real-time charging and discharging data are acquired, and the real-time charging and discharging data are compared with the dynamic voltage evaluation baseline model to generate a voltage deviation degree score set, specifically including:

[0030] inputting the real-time charging and discharging data into the dynamic voltage evaluation baseline model to calculate a matching degree with each typical charging and discharging mode;

[0031] determining a deviation degree between the real-time charging and discharging data and the dynamic voltage evaluation baseline model according to the matching degree;

[0032] generating the voltage deviation degree score set based on the deviation degree.

[0033] Preferably, the real-time charging and discharging data are detected for an abnormal interval according to the voltage deviation degree score set to identify an abnormal voltage fluctuation period, specifically including:

[0034] performing sliding window statistical analysis on the voltage deviation degree score set to detect an abnormal score interval exceeding a preset threshold;

[0035] locating a corresponding real-time charging and discharging data time range according to the abnormal score interval to generate the abnormal voltage fluctuation period.

[0036] Preferably, the abnormal voltage fluctuation period corresponding to the charge and discharge data is subjected to multi-scale feature decomposition to generate a local abnormal feature vector and a global abnormal feature vector, specifically including:

[0037] The voltage data in the abnormal voltage fluctuation period is subjected to wavelet transform decomposition to extract fluctuation features of different time scales;

[0038] The high-frequency fluctuation features are aggregated to generate the local abnormal feature vector;

[0039] The low-frequency fluctuation features are aggregated to generate the global abnormal feature vector.

[0040] Preferably, the abnormal root cause is traced based on the local abnormal feature vector and the global abnormal feature vector to determine the key influencing factor causing the voltage abnormality, specifically including:

[0041] The local abnormal feature vector is matched with a preset battery aging feature library to identify potential aging factors;

[0042] The global abnormal feature vector is matched with a preset environmental influence feature library to identify potential environmental factors;

[0043] According to the combined weight of the aging factors and the environmental factors, the key influencing factor is determined.

[0044] Preferably, the SOC voltage calibration strategy is generated according to the key influencing factor, specifically including:

[0045] According to the type of the aging factor, a corresponding compensation algorithm is selected, and according to the intensity of the environmental factor, a compensation coefficient is determined;

[0046] The SOC voltage calibration strategy is generated in combination with the compensation algorithm and the compensation coefficient.

[0047] Preferably, the present application further includes a lithium battery SOC voltage evaluation system, the system includes a memory, a processor and a computer program stored in the memory and running on the processor, when the processor executes the computer program, the steps of the above lithium battery SOC voltage evaluation method are realized.

[0048] Compared with the prior art, the present application has the following advantages:

[0049] In the data acquisition phase, the voltage curve, current curve and temperature curve are collected synchronously. Compared with single voltage data acquisition method, it can more comprehensively reflect the state change of lithium battery during charging and discharging process, and provide more abundant information support for subsequent evaluation. Through the curve alignment processing based on dynamic time warping, the differences of curves in time dimension under different charging and discharging cycles can be eliminated, so that the standardized charging and discharging feature set is more consistent, and the feature extraction deviation caused by inconsistent curve form is avoided.

[0050] The standardized feature set is input into the pre-trained deep neural network model for high-order feature extraction, which can excavate the hidden nonlinear correlation features in the data, such as the coupling relationship between voltage, current and temperature, and the performance degradation trend of the battery in the long-term cycle process. Compared with traditional linear models, it can better adapt to the complex electrochemical characteristics of lithium batteries. The constructed dynamic voltage evaluation baseline model is not fixed, but can be adjusted in real time according to real-time data, which can track the dynamic changes of battery state in real time. Whether it is short-term temperature fluctuation, current mutation, or long-term battery aging, the accuracy of evaluation can be maintained through the dynamic adjustment of the baseline model, avoiding the accumulation of evaluation errors caused by model solidification.

[0051] In the abnormality detection link, the voltage deviation score set is generated by comparing the real-time charging and discharging data with the dynamic baseline model, which can accurately identify the abnormal voltage fluctuation period. Compared with the traditional threshold detection method, it can more sensitively capture subtle voltage abnormal changes and reduce missed detection and false detection. The abnormal period data is decomposed into multi-scale features to generate local abnormal feature vectors and global abnormal feature vectors. Local features can reflect instantaneous abnormalities (such as voltage fluctuations caused by sudden current impact), and global features can reflect systematic abnormalities (such as overall voltage deviation caused by internal structure aging of the battery). Through the combination of the two, the key influencing factors of different types of abnormalities can be distinguished, rather than just staying at the identification level of abnormal phenomena.

[0052] The SOC voltage calibration strategy based on key influencing factors can specifically solve the voltage deviation problem caused by different root causes, such as adjusting the temperature compensation parameters for temperature factors, and updating the voltage-SOC corresponding relationship model for battery aging. It realizes the closed-loop management from abnormality identification to precise calibration. In addition, the entire evaluation process does not need to rely on complex hardware device upgrade, but only through data processing algorithm and model optimization can the evaluation performance be improved, which has strong practicality and economy, and can adapt to different types and different application scenarios of lithium batteries, providing protection for the safe and efficient operation of lithium batteries in various devices. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1The working principle diagram of the lithium battery SOC voltage evaluation method described in the present application;

[0054] Figure 2 The flow chart of the curve alignment processing based on dynamic time warping;

[0055] Figure 3 The flow chart of the deep neural network feature extraction. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] Please refer to Figure 1 The present application provides a lithium battery SOC voltage evaluation method, which comprises: obtaining a plurality of groups of historical charge and discharge data of a lithium battery, including voltage curves, current curves and temperature curves, and performing curve alignment processing based on dynamic time warping on the historical charge and discharge data, thereby generating a standardized charge and discharge feature set. The standardized charge and discharge feature set is then input into a pre-trained deep neural network model for deep-level feature extraction, to generate a high-order charge and discharge feature matrix capable of representing the complex behavior of the battery. Based on the feature matrix, a dynamic voltage evaluation baseline model is constructed, the parameters of which can be adjusted in real time according to the actual running situation to adapt to the evolution of the battery performance. In the online evaluation stage, the system obtains real-time charge and discharge data and compares it with the dynamic voltage evaluation baseline model, to calculate and generate a voltage deviation degree score set. According to the score set, abnormal interval detection is performed on the real-time charge and discharge data, to accurately identify the time period during which abnormal voltage fluctuation occurs. For the identified abnormal period, further multi-scale feature decomposition is performed on the corresponding charge and discharge data, to respectively generate local abnormal feature vectors reflecting instantaneous details and global abnormal feature vectors reflecting long-term trends. Based on the two types of feature vectors, root cause analysis is performed to determine the key influencing factors leading to voltage abnormality, and finally, targeted SOC voltage calibration strategies are generated according to the key influencing factors, to realize accurate compensation and correction of the SOC estimation value.

[0058] Embodiment 1: Please refer to Figure 2, the voltage curve, the current curve and the temperature curve obtained are subjected to detailed time sequence segmentation operation, the segmentation can be based on fixed time interval length, for example, taking every ten minutes as a time window, or dynamically divided according to key event points in the charging and discharging process, such as the inflection point of constant current charging to constant voltage charging, the discharge cut-off voltage point, etc., so as to cut each continuous curve into multiple representative local time sequence segments, each segment captures the behavior characteristics of the battery in a certain short time, and the segmented data facilitates subsequent more detailed comparative analysis, avoiding the information loss that may be caused by long sequence overall alignment. After completing the segmentation, the dynamic time warping distance between each local time sequence segment and a preset reference curve needs to be calculated, the preset reference curve is usually selected from the stable performance battery charging and discharging curve obtained under the standard laboratory conditions, which represents the ideal working mode, the core of dynamic time warping distance calculation is to find the optimal nonlinear alignment path between two time sequences with different lengths, which makes the two sequences most similar in shape, the calculation process gradually fills the local distance between each point pair by constructing a cumulative distance matrix, and finds the path with the minimum total cost by backtracking, and the total length of the path is the dynamic time warping distance, which quantifies the difference between the local segment and the reference curve in shape, rather than simply the corresponding error of time points. According to the dynamic time warping distance of each local time sequence segment calculated, the time axis is stretched and adjusted, the adjustment process bends or stretches the time stamp of the segment according to the optimal alignment path, so that the key feature points in the segment, such as the voltage platform starting point and the current peak point, can be accurately aligned with the corresponding feature points on the reference curve in time, thereby effectively eliminating the time axis distortion caused by actual charging and discharging rate fluctuation, non-uniform sampling interval of data acquisition system or battery individual difference, generating the aligned standardized time sequence, which has consistency in time dimension. The aligned standardized time sequence needs to be fused to generate the final standardized charging and discharging feature set, the fusion operation can splice the voltage, current and temperature three-dimensional aligned data at the same time stamp to form a multi-dimensional feature vector, each vector contains voltage value, current value and temperature value at the same time, Alternatively, more comprehensive feature representation can be generated by calculating the interaction features between dimensions, such as instantaneous power, energy efficiency or temperature derivative of voltage change, etc., to form a standardized charging and discharging feature set, which provides a time-aligned, multi-dimensional correlated high-quality data set for subsequent deep feature extraction.

[0059] The specific operation of time series segmentation needs to consider the actual characteristics of the data and the application requirements. The segmentation strategy can adopt the equal-length window method, that is, a fixed-length sliding window is used to intercept the sequence along the time axis. The selection of the window size needs to balance the information content and the computational complexity of the segment. If the window size is too small, the segment may not contain a complete charging and discharging process. If the window size is too large, it may mask local details. Alternatively, an adaptive segmentation method based on change points can be used. The segmentation boundary is determined by detecting the mutation points of the first or second derivative of the voltage or current curve. For example, in the constant current charging stage, the current remains constant and the voltage rises slowly. When switching to the constant voltage stage, the current begins to decay, which is a natural segmentation point. This adaptive method can more accurately capture the physical stage changes in the charging and discharging process. When calculating the dynamic time warping distance, the selection of the preset reference curve is crucial. This curve should be able to represent the typical behavior of the battery under healthy state. It is usually obtained by averaging a large amount of normal historical data or selecting a most representative curve. The local distance metric used in distance calculation usually uses Euclidean distance or Manhattan distance to quantify the difference between two data points. The dynamic time warping algorithm recursively calculates the cumulative distance matrix through dynamic programming to find a path from the start to the end of the sequence, so that the sum of the local distances of all points on the path is minimized. This minimum sum is the dynamic time warping distance. It is robust to stretching and shifting of the time axis and can effectively handle the problem of inconsistent curve length under different charging and discharging rates. The stretching adjustment of the time axis based on dynamic time warping distance is a nonlinear transformation process. The adjustment algorithm maps the time index of the original local segment through the optimal alignment path, and performs resampling or interpolation. For example, if the alignment path shows that a certain time interval in the segment needs to be stretched to match the longer interval of the reference curve, interpolation is performed to increase the data points, and vice versa. The goal of adjustment is to synchronize the segment and the reference curve in key pattern features. The length of the adjusted sequence may be different from the original sequence, but the pattern similarity is the highest. In the feature fusion stage, the multi-source nature of the data needs to be considered. Simple vector splicing preserves the original information of each dimension, but the dimension may be high. Therefore, feature dimension reduction techniques such as principal component analysis can be introduced. After feature extraction of the aligned sequences in each dimension, the main components are retained before fusion. Alternatively, more advanced fusion strategies such as attention mechanism-based weighted fusion can be used. According to the importance of each dimension feature to the final evaluation task, the weight is dynamically allocated to generate a standardized charging and discharging feature set with higher information density.

[0060] In the segmentation process, the processing of voltage curve, current curve and temperature curve needs to keep the consistency in time, to ensure that the data segments from the three dimensions of the same charge-discharge event are corresponding in the time range, otherwise the fused features will lose physical meaning, so the segmentation operation is usually based on a main dimension, for example, the event points of the current curve are taken as the basis for segmentation, and then the corresponding voltage and temperature data are cut synchronously to ensure the time correlation between multi-dimensional data. The calculation efficiency of dynamic time warping distance is a factor that needs to be considered when processing a large amount of historical data. Some acceleration techniques can be used, such as using a sliding window to limit the range of path search, or using a lower bound function to terminate the calculation of paths that are unlikely to produce better paths, to balance the accuracy and computational overhead. The standardized time series generated by the time axis adjustment link may no longer be equally spaced, depending on the degree of curvature of the alignment path, so a subsequent resampling step may be needed to unify it to a standard time grid to facilitate subsequent model processing. The resampling method can be linear interpolation or spline interpolation to smoothly generate equally spaced data points. Feature fusion is not limited to simple data merging. Domain knowledge can be further introduced, such as calculating the internal state estimation value of the battery at each time point, such as polarization voltage, ohmic resistance, etc., and incorporating these derived features into the feature set, so that the standardized charge-discharge feature set not only contains original observation values, but also contains implicit features reflecting the internal physical and chemical processes of the battery.

[0061] When implementing the curve alignment process based on dynamic time warping, it is necessary to process the noise and abnormal points that may exist in the historical data, because these will interfere with the calculation of the dynamic time warping distance and the alignment accuracy, so before segmentation and alignment, the original voltage, current and temperature curves usually need to be preprocessed, such as smoothing the data using a moving average filter or wavelet denoising technology to remove high-frequency noise while preserving key morphological features, and the preprocessed data can improve the robustness of the dynamic time warping algorithm. The construction of the preset reference curve can also be a collection of multiple curves rather than a single curve to cover the normal behavior patterns of the battery under different health states or different environmental temperatures, in which case the minimum dynamic time warping distance of the local segment to multiple reference curves is calculated, and the reference curve with the minimum distance is selected as the alignment reference, so that the alignment process can adapt to the normal drift of the battery state. The standardized time series generated after the time axis stretching adjustment may change the distribution of the data points, and the statistical properties of the adjusted sequence need to be checked, such as whether the mean and variance have changed significantly, and if necessary, standardized processing is performed to make all sequences have zero mean and unit variance to eliminate the dimension effect and facilitate the processing of the subsequent feature extraction model. The standardized charging and discharging feature set generated by feature fusion should have a data structure that is convenient for the subsequent deep neural network model to read, and is usually organized in the form of a tensor, such as a three-dimensional tensor, where the first dimension represents the sample index, the second dimension represents the time step, and the third dimension represents the feature dimension. This tensor can be directly input into a convolutional neural network for processing.

[0062] The entire dynamic time warping-based curve alignment processing flow needs to iteratively optimize parameters such as segmentation window size, dynamic time warping path constraint window width, and fused feature selection. These parameters can be tuned on the historical data set through cross-validation to find a general setting that performs well on different types of lithium battery data, thereby ensuring that the generated standardized charging and discharging feature set has high representativeness and consistency. During processing, metadata generated during the alignment process, such as the reference curve identifier used by each segment and the adjustment scaling factor, may be recorded, which may be helpful for subsequent anomaly diagnosis or model interpretation. The generation of the standardized charging and discharging feature set marks the completion of the data preprocessing phase, providing clean, aligned, and multi-dimensional fusion input data for deep neural network models.

[0063] Embodiment 2: see Figure 3The standardized charging and discharging feature set is input into a pre-trained deep neural network model for feature extraction, and a high-order charging and discharging feature matrix is generated. The process starts with inputting the aligned multi-dimensional time series data into the convolutional layer of the network, which is usually composed of a set of one-dimensional convolutional kernels that slide and convolve on the input sequence along the time dimension. The core function of each convolutional kernel is local perception, and each kernel is responsible for detecting a certain local pattern or feature in the input data. For example, one kernel may be sensitive to a short drop in the voltage curve, and another kernel may focus on identifying a slowly varying platform in the current. The convolution operation generates a series of feature maps by weighting and summing the data points within the local time window through the filter and adding a bias term, and then transforming through a nonlinear activation function such as ReLU. These feature maps constitute a preliminary abstract representation of the original data, capturing various transient phenomena and short-term correlations in the charging and discharging process. After the convolutional layer, the feature maps are sent to the attention mechanism layer for processing. The core idea of the attention mechanism is to simulate the selective attention of human cognitive processes. In this layer, the system calculates a weight coefficient for each time step or each feature channel in the primary feature map, which represents the importance of this information for the final judgment of the battery state. The calculation of the weight is usually achieved through a small learning neural network or scaled dot-product attention, which considers the content of the feature itself and its context relationship. For key time points or key feature channels that can significantly indicate battery abnormalities or specific working states, the attention mechanism assigns higher weights, making them more dominant in subsequent calculations. Conversely, for parts that contain redundant information or more noise, lower weights are assigned. The information density and discriminability of the feature maps generated after this weighted processing are enhanced.

[0064] The weighted feature maps are passed to the fully connected layer for feature dimension reduction and high-level integration. The fully connected layer is characterized by the fact that each neuron in the layer is connected to all the outputs of the previous layer. This global connection allows the layer to consider all local features after convolution and attention weighting, and model the complex nonlinear relationships between them. The fully connected layer performs linear transformation on the high-dimensional feature vector input through a large parameter matrix, and applies a nonlinear activation function again. This process is essentially a spatial transformation and compression of distributed feature representation, mapping the information-rich features to a lower-dimensional but more discriminative vector space. Each dimension of the low-dimensional vector represents a high-order abstract feature extracted from the original data, which may be related to the internal chemical state, aging degree or working mode of the battery, etc.

[0065] The pre-training phase of the deep neural network model is usually performed on a large amount of labeled or unlabeled historical charging and discharging data. For convolutional layers and fully connected layers, the training objective can be to learn feature representations by reconstructing the input data, or to perform supervised training associated with a state estimation task. The parameters of the attention mechanism layer are usually optimized together with other layers during the overall network training process to minimize the prediction error. The trained network has the ability to automatically extract discriminative features from multi-dimensional charging and discharging curves. In actual deployment, a standardized set of charging and discharging features is fed into the pre-trained model in the form of a sliding window. The forward propagation process of the model sequentially passes through the convolution, attention, and fully connected layers, and finally obtains a high-order charging and discharging feature vector corresponding to each input time window at the output end of the fully connected layer. Organizing the feature vectors of all time windows in chronological order forms a high-order charging and discharging feature matrix, which is a highly refined representation of the original time series data. The design parameters of the convolutional layer include the number, length, and step of the convolution kernel. The number of kernels determines the richness of the feature types extracted from the data, the length of the kernel determines the size of the receptive field, i.e., the range of the time series observed each time, and the step controls the interval of the sliding convolution. These parameters need to be adjusted according to the time characteristics of the charging and discharging data. For example, to capture the subtle fluctuations in the charging and discharging plateau, multiple convolution kernels of different scales may be needed. Attention mechanism calculations can occur in different dimensions. Temporal attention focuses on which time points are important, and channel attention focuses on which feature channels are important. In practical applications, multiple attention mechanisms can be used sequentially or in parallel to improve feature quality from different angles. The output dimension of the fully connected layer, i.e., the length of the high-order feature vector, is a hyperparameter that needs to be balanced. Too high a dimension may lead to increased computational burden and overfitting risk, while too low a dimension may not be able to retain sufficient information. A suitable compression ratio usually needs to be determined through experimental exploration.

[0066] The preprocessed standardized data is sometimes further formatted before inputting into the convolutional layer, for example, organized into single-channel or multi-channel one-dimensional sequences similar to images, where the number of channels corresponds to the feature dimension, and the model can learn joint representations from voltage, current, and temperature signals simultaneously. Pooling layers, especially max-pooling or average-pooling, are usually added after the convolutional layer to downsample the feature maps, thereby reducing the computational load and introducing certain translation invariance, but the pooling operation may lose precise time position information, so it needs to be used carefully in time series analysis. The calculation process of attention weights is interpretable itself, and by visualizing these weights, we can understand which parts of the input sequence the model focuses on when making a judgment, which provides certain insights into the model's decision-making. A flattening operation may be included before the fully connected layer to convert the two-dimensional feature maps output by the convolution and attention layers into one-dimensional vectors for fully connected operations. For very deep networks, mechanisms such as skip connections may also be introduced to alleviate the problem of gradient vanishing and ensure the effectiveness of training. The generated high-order charge-discharge feature matrix is the final output of the entire feature extraction process, and each row represents a high-order feature vector for a time window. These feature vectors will be used as input to build a dynamic voltage evaluation baseline model. Since they have been refined by a deep network and contain rich state information, they can significantly improve the performance and robustness of the subsequent baseline model. The entire feature extraction process is end-to-end, and once the model is trained, the deployment stage only needs to forward propagate to quickly obtain high-order features, meeting the real-time evaluation requirements. The deep neural network model is not immutable, and its parameters can be fine-tuned using newly collected normal data through online learning, allowing the feature extraction capability to adapt to slow changes in battery performance.

[0067] Embodiment 3: involves constructing a dynamic voltage assessment baseline model based on high-order charge-discharge feature matrix, and adjusting its parameters in real time, and using the model to process real-time charge-discharge data to generate a set of voltage deviation scores, the starting point of the process is to perform clustering analysis on the high-order charge-discharge feature matrix extracted by the deep neural network, the clustering analysis aims to discover the internal and recurring pattern structure from a large number of historical feature vectors, the algorithm used can be K-means algorithm based on centroid, DBSCAN algorithm based on density or Gaussian mixture model based on probability distribution, the choice of algorithm depends on the data distribution characteristics of the feature space and different assumptions about the cluster shape, for example, K-means assumes that clusters are convex and of similar size, while DBSCAN can discover clusters of arbitrary shape and is robust to noise points, through clustering, a large number of historical charge-discharge behavior feature vectors are divided into a number of limited categories, each category represents a typical charge-discharge mode, for example, the high-current mode in the initial stage of fast charging, the low-current decay mode in the final stage of constant voltage charging, the steady-state discharge mode under different load conditions, etc., these modes constitute the behavior baseline of the normal operation of the battery. After identifying multiple typical charge-discharge modes, a baseline voltage behavior description needs to be established for each mode, i.e. to construct an initial voltage assessment baseline model, for each mode category determined by clustering, collect all high-order feature vectors belonging to the category and their corresponding original voltage data segments verified as normal, for the voltage behavior under this mode, multiple statistical learning methods can be used for modeling, one direct method is to calculate the statistical confidence interval of the voltage sequence under this mode, for example, taking the mean plus or minus three times the standard deviation of the voltage value at each time point as the normal fluctuation range, another more detailed method is to use a regression model to predict the normal voltage curve or key voltage feature value corresponding to the high-order feature vector, the initial model is a collection of these sub-models, covering all identified typical operating states.

[0068] After the initial model is built, it is not fixed and needs to be iteratively updated by online learning algorithm combined with the new charging and discharging data transmitted in real time. Online learning allows the model to continuously adapt to the slow drift of battery performance caused by factors such as cycle aging, environmental temperature changes, etc. The update strategy can be based on the idea of incremental learning. For example, when new real-time data is determined to belong to a certain typical mode and its voltage behavior is normal, then the data point is used to update the baseline description corresponding to the mode. For models based on confidence intervals, the mean and standard deviation can be updated online. For regression models, online optimization algorithms such as stochastic gradient descent can be used to fine-tune model parameters. This gradual updating mechanism enables the baseline model to track the normal evolution of the battery life cycle, maintaining its timeliness and accuracy of evaluation. In the model application stage, the real-time charging and discharging data obtained is first subjected to the same preprocessing and feature extraction process as the historical data, and is converted into a real-time high-order feature vector. Then the vector is input into the dynamic voltage evaluation baseline model. The model needs to calculate the matching degree between the real-time feature vector and each typical charging and discharging mode. The calculation of matching degree can be based on distance or similarity measure, for example, calculating the Euclidean distance or cosine similarity of the real-time feature vector to each mode cluster center. The smaller the distance or the higher the similarity, the higher the matching degree with the mode. A specific matching quantization method can use the following relationship:

[0069] ,

[0070] wherein the symbol represents the matching degree score of the real-time feature vector to the th typical mode, the value closer to 1 indicates the higher matching degree; the symbol represents the real-time high-order feature vector to be evaluated; the symbol represents the cluster center vector of the th typical charging and discharging mode obtained by cluster analysis; the symbol represents the distance measure between the vectors and , usually using Euclidean distance; the symbol is a scale parameter greater than zero, used to control the decay rate of the score with increasing distance.

[0071] According to the calculated matching degree, the deviation between the real-time charge and discharge data and the dynamic voltage evaluation baseline model can be determined. If the real-time data has a high matching degree with a certain typical mode and its actual voltage value is within the expected normal range of the mode, the deviation is low. Conversely, if the real-time data has a low matching degree with all modes, or although it has a certain matching degree with a certain mode, its actual voltage value significantly exceeds the expected range corresponding to the mode, the deviation is high. The deviation degree can be quantified by combining the matching degree score and the voltage residual. Based on the deviation degree, the system generates a voltage deviation score set, which can be designed as a continuous numerical value, such as between 0 and 1, where 0 represents complete normality and 1 represents severe deviation; or as discrete levels, such as "normal", "attention", "warning", and "serious". The score set forms a new time series over time, which intuitively reflects the abnormality of the battery voltage behavior relative to the historical normal baseline.

[0072] The determination of the number of clusters in clustering analysis is a key step. The elbow rule, silhouette coefficient and other indicators can be used to assist in selecting the optimal number of clusters to ensure that the diversity of battery behavior is fully captured and excessive meaningless fine patterns are avoided. When updating the model parameters online, a forgetting mechanism or sliding window needs to be set. Too distant historical data may not represent the current state of the battery, and its influence should be gradually weakened. For example, exponential weighted moving average can be used to update the statistics, making the model more focused on recent data. When calculating the matching degree, in addition to considering the distance from the cluster center, the goodness of fit of the real-time feature vector to the entire cluster distribution can also be considered, such as calculating its posterior probability of belonging to the Gaussian mixture model component of the cluster. The calculation of the deviation degree requires the definition of a comprehensive index, which may integrate multiple factors, such as the distance to the nearest mode, the root mean square error of the voltage observation value and the mode expected value, and the duration of the anomaly. The generated voltage deviation score needs to be smoothed, such as using a moving average filter, to eliminate false positives caused by transient jitter, and thus more robustly reflect the true abnormal trend.

[0073] The entire dynamic baseline modeling and score generation process constitutes a closed-loop system, with the baseline model providing a reference for real-time evaluation, and the data confirmed as normal in the results of real-time evaluation being used to update the baseline in turn, making the system adaptive and able to cope with long-term slow changes in battery performance without misjudging normal performance degradation as a fault, while also being able to sensitively detect sudden and abnormal voltage fluctuations to provide reliable input for subsequent fine diagnosis. The effectiveness of this method depends on the quality and representativeness of the historical data, as well as the reasonable setting of the clustering and online learning algorithm parameters, which need to be fully validated and optimized in the offline stage.

[0074] Example 4: Abnormal interval detection according to voltage deviation score set, identify abnormal voltage fluctuation period, and multi-scale feature decomposition of voltage data in this period, generate local and global abnormal feature vectors, the process begins with sliding window statistical analysis of continuously generated voltage deviation score time series, assuming that the system produces a deviation score between 0 (normal) to 1 (severe abnormality) at a frequency of one per second, these scores constitute a discrete time series, in order to detect continuous abnormalities rather than transient spikes, a sliding time window of appropriate length needs to be set, for example a window of 60 seconds, the window moves back one unit per second, for each window position, calculate the arithmetic mean of all 60 scores in the window, by observing the trajectory of this moving average, random jitter in the score sequence can be smoothed out, making it easier to find abnormal intervals with consistently high scores, set a threshold, for example when the moving average exceeds 0.7 for more than 10 seconds, then determine that this time interval is an abnormal score interval.

[0075] According to the detected abnormal score interval, the corresponding timestamp range can be traced back to the original real-time charge and discharge data stream, so as to accurately frame the abnormal voltage fluctuation period, for example, assuming that the system detects that the sliding average deviation score exceeds the standard continuously between time stamps t=1000s to t=1100s, then mark this 100s time period as the abnormal voltage fluctuation period that needs to be analyzed in-depth, this period is the direct data source for subsequent multi-scale feature decomposition. Referring to Table 1, a simplified abnormal score interval detection example is shown, where "window start time" represents the start time of the sliding window, "score set in window" is the 60 score values covered by the window (here for illustration, only 5 representative values are listed), "window average score" is the calculated key indicator, and "determination result" is given based on the threshold (e.g. 0.7).

[0076] Table 1: sliding window abnormal score detection:

[0077]

[0078] After successfully identifying the abnormal voltage fluctuation period, the original voltage data recorded in this period needs to be decomposed into multi-scale features to reveal different manifestations of the abnormality. The wavelet transform is used in the decomposition process. The wavelet transform can decompose a signal into different frequency subbands, thus providing localized information in both time and frequency. A suitable mother wavelet function, such as the Daubechies wavelet or the Symlets wavelet, is selected, and the voltage signal in the abnormal period is decomposed into multiple levels. Each level of decomposition will decompose the signal into an approximation coefficient sequence and a detail coefficient sequence. The approximation coefficients capture the low-frequency profile of the signal, i.e., the overall trend, while the detail coefficients capture the high-frequency details of the signal, i.e., the local fluctuations.

[0079] The high-frequency detail coefficients obtained by wavelet decomposition are aggregated to generate a local abnormal feature vector. The purpose of aggregation is to convert a series of coefficients into a fixed-length vector that can represent the high-frequency characteristics of this period. The aggregation method can be to calculate the statistical features of the detail coefficients at each level, such as calculating the standard deviation of the first-level detail coefficients to measure the intensity of the finest-scale fluctuations, calculating the energy to reflect the overall power of the high-frequency components, and calculating the Shannon entropy to measure the irregularity or randomness of the fluctuations. For deeper decomposition levels, the statistical quantities of the detail coefficients are also calculated. All these statistical features extracted from high-frequency components at different scales together form the local abnormal feature vector, which mainly describes the instantaneous and rapid change characteristics of the voltage signal in the abnormal period, such as whether there are high-frequency oscillations, glitches, or steep jumps. The low-frequency approximation coefficients obtained by wavelet decomposition are aggregated to generate a global abnormal feature vector. Generally, after multi-level decomposition, the approximation coefficients of the last level represent the core low-frequency trend of the signal after removing high-frequency details. The same statistical feature extraction is performed on these low-frequency approximation coefficient sequences, such as calculating the mean of the trend sequence to understand the overall shift of the voltage level, calculating the slope to determine whether the voltage is slowly decreasing or increasing, and calculating the curvature to observe whether the trend has a bend. These features extracted from the low-frequency trend together form the global abnormal feature vector, which describes the long-term and systematic change patterns of the voltage behavior in the abnormal period, such as whether there is a slow voltage decay platform or a persistent voltage drift.

[0080] The size and moving step of the sliding window are parameters that need to be adjusted according to the application scenario. A too long window may cause the positioning of the abnormal start and end to become sluggish, and a too short window may not effectively smooth out transient interference and is prone to false positives. The moving step is usually smaller than the window length to achieve dense coverage detection of the time series. The number of wavelet decomposition layers determines the range of analysis scales. The more the number of layers, the richer the analysis scale, but the larger the amount of calculation, and the lower the time resolution corresponding to the deepest approximation coefficient. The appropriate decomposition layer number needs to be selected according to the expected duration of the anomaly and the focus frequency band, for example, for an anomaly lasting for several minutes, 4 to 5 layers of decomposition may be more appropriate. The specific composition of the local and global feature vectors can be customized according to the diagnostic needs. In addition to the above statistical quantities, other features such as zero-crossing rate, peak count, etc. can be introduced to more comprehensively describe the morphology of the abnormal signal. The generated local and global abnormal feature vectors serve as quantitative descriptions of the abnormal voltage fluctuation pattern, providing accurate and calculable data inputs for subsequent root cause tracing, enabling the system to distinguish between different types of abnormalities, such as transient voltage drop caused by internal micro-short circuit or trend voltage reduction caused by overall internal resistance increase. The entire abnormal interval detection and feature decomposition process realizes the conversion from continuous deviation score to discrete abnormal event, and then to quantitative abnormal pattern features. It converts an abnormal signal in time into a point in a feature space, which contains key information of the anomaly in time and frequency dimensions, laying a foundation for knowledge-based or data-based diagnostic algorithms. The value of this method lies not only in indicating the time of abnormal occurrence, but more importantly, in deeply and multi-angle describing the inherent pattern of the anomaly.

[0081] Example 5: Abnormal root cause tracing based on local abnormal feature vector and global abnormal feature vector, and generating SOC voltage calibration strategy accordingly, the process starts from matching the decomposed local abnormal feature vector with a pre-set battery aging feature library, which is a knowledge base constructed based on a large number of experiments, historical failure data and electrochemical mechanism knowledge, in which various known battery aging modes corresponding to typical local voltage feature signatures are systematically stored, for example, micro-short circuit caused by lithium dendrite growth usually causes a short but sharp voltage drop on the voltage curve, its local feature vector may show very high high-frequency energy and pulse pattern; active material loss or lithium inventory decay may show subtle shrinkage of charge-discharge platform capacity, and its local feature may reflect the change of platform voltage fluctuation characteristics; the increase of internal resistance caused by the thickening of solid electrolyte interface film, its local feature may show the change of voltage step pattern at the moment of current loading, the matching process is realized by calculating the similarity between the real-time abnormal corresponding local abnormal feature vector and each aging mode template feature vector in the feature library, cosine similarity or inverse Euclidean distance can be used as a measure method. Match the global abnormal feature vector with a pre-set environmental influence feature library, which is also a carefully constructed knowledge base, which summarizes the typical trend changes of battery voltage under different external environmental conditions, for example, when the battery is in a low temperature environment, the conductivity of the electrolyte decreases, the ion migration rate slows down, resulting in increased overall polarization of the battery, its voltage curve will show a significant decrease in platform voltage, early charging and discharging cutoff and other trend characteristics, these characteristics will be reflected in the global abnormal feature vector, such as voltage average value moving down, trend slope being negative, etc.; on the contrary, in high temperature environment, the acceleration of side reactions may cause the slow drift of voltage platform or the continuous change of coulomb efficiency, its global feature vector may show different patterns, the matching process is similar to the aging feature matching, through similarity calculation to determine which typical mode of environmental factors is closest to the current abnormality.

[0082] After obtaining the matching results from the aging factor and environmental factor dimensions respectively, the influence degree of these potential factors on the current anomaly needs to be comprehensively evaluated, that is, to determine their combined weights. The allocation of weights is not a simple average, but a decision based on multi-source information. For example, the matching similarity of a certain aging pattern feature is as high as 0.95, while the matching similarity of all environmental patterns is lower than 0.6, so the weight of this aging factor will be given a higher value. Another case is that if the system monitors that the environmental temperature indeed drops sharply and is highly consistent with the time of the anomaly, then even if the matching degree of a certain aging pattern is not low, the weight of the environmental factor may be adjusted to be high. The weight allocation can be based on a rule base, or a pre-trained small classifier can be used, which takes the matching similarity, environmental sensor readings, anomaly duration, etc. as input, and outputs the confidence weight of each factor. According to the determined key influencing factors and their weights, the SOC voltage calibration strategy is generated. The generation of the strategy is targeted. If the key influencing factor is determined to be "micro-short circuit caused by lithium dendrite", since it is manifested as instantaneous voltage drop, the calibration strategy may select a dynamic internal resistance compensation algorithm, which monitors the voltage transient response at the moment of current change in real time, and corrects the ohmic voltage drop involved in the SOC estimation accordingly. At the same time, according to the severity of this anomaly (which may be represented by the deviation score or the feature vector amplitude), the size of the compensation coefficient is determined. If the key influencing factor is "internal resistance increase caused by low temperature", the temperature-internal resistance relationship model is selected for compensation, and the compensation coefficient is determined according to the measured temperature value. The compensation coefficient may be a function value negatively correlated with temperature. The final generated calibration strategy instruction contains the identifier of the selected compensation algorithm and the calculated compensation coefficient. The instruction is sent to the SOC estimation module of the battery management system.

[0083] The construction of the battery aging feature library and the environmental impact feature library is a continuous process that needs to be updated and improved as the understanding of new materials and new processes deepens. Each feature template in the library needs to be supported by sufficient experimental data or field data. The similarity threshold in the matching process needs to be set carefully. A too low threshold may lead to false matching, and a too high threshold may miss some new or mixed faults. The determination method of the combination weight needs to have certain flexibility to cope with the situation where multiple factors are coupled to cause abnormalities, for example, the superposition of aging and low temperature may make the voltage abnormality more intense. The generation of the SOC voltage calibration strategy needs to consider the actual executable ability of the BMS. The compensation algorithm provided should be supported by the BMS and the calculation complexity should be acceptable. The determination of the compensation coefficient needs to have a clear mapping relationship, for example, mapping a certain function value of the deviation score to the compensation coefficient to ensure the feasibility and effectiveness of the strategy. The whole root cause tracing and strategy generation process embodies the closed loop from signal analysis to decision execution, which converts abstract voltage fluctuation abnormalities into specific impact factor judgments and finally into operable calibration commands, so that the battery management system can not only perceive abnormalities, but also understand abnormalities and take targeted compensation measures, thereby maintaining the reliability of SOC estimation in the presence of interference or aging. The effectiveness of the method depends on the completeness and accuracy of the feature library and the rationality of the weight distribution logic, which need to be fully verified and tested before system deployment.

[0084] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0085] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the SOC voltage of a lithium battery, characterized in that, The method includes: Acquire multiple sets of historical charge and discharge data for lithium batteries, including voltage curves, current curves, and temperature curves; The historical charge and discharge data are subjected to curve alignment processing based on dynamic time warping to generate a standardized set of charge and discharge features. The standardized charge and discharge feature set is input into a pre-trained deep neural network model for feature extraction to generate a high-order charge and discharge feature matrix. A dynamic voltage evaluation baseline model is constructed based on the high-order charge and discharge feature matrix, and the parameters of the dynamic voltage evaluation baseline model are adjusted in real time. Acquire real-time charge and discharge data, and compare the real-time charge and discharge data with the dynamic voltage evaluation baseline model to generate a voltage deviation score set; Based on the voltage deviation score set, abnormal interval detection is performed on the real-time charge and discharge data to identify periods of abnormal voltage fluctuations. Multi-scale feature decomposition is performed on the charging and discharging data corresponding to the abnormal voltage fluctuation period to generate local abnormal feature vectors and global abnormal feature vectors; Based on the local and global anomaly feature vectors, the root cause of the anomaly is traced to determine the key influencing factors that lead to the voltage anomaly. A SOC voltage calibration strategy is generated based on the aforementioned key influencing factors; The process of constructing a dynamic voltage evaluation baseline model based on the high-order charge-discharge feature matrix and adjusting the parameters of the dynamic voltage evaluation baseline model in real time specifically includes: Cluster analysis was performed on the high-order charge-discharge feature matrix to identify several typical charge-discharge modes; Establish an initial voltage evaluation baseline model based on the typical charge and discharge modes described above; The parameters of the initial voltage evaluation baseline model are iteratively updated by combining online learning algorithms with real-time charge and discharge data; The process of acquiring real-time charge and discharge data and comparing it with the dynamic voltage assessment baseline model to generate a voltage deviation score set specifically includes: The real-time charge and discharge data are input into the dynamic voltage evaluation baseline model to calculate the matching degree with each typical charge and discharge mode; The degree of deviation between the real-time charge / discharge data and the dynamic voltage evaluation baseline model is determined based on the matching degree. The voltage deviation score set is generated based on the degree of deviation. The step of performing multi-scale feature decomposition on the charging and discharging data corresponding to the abnormal voltage fluctuation period to generate local abnormal feature vectors and global abnormal feature vectors specifically includes: Wavelet transform decomposition is performed on the voltage data during the abnormal voltage fluctuation period to extract fluctuation features at different time scales; The high-frequency fluctuation features are aggregated to generate the local anomaly feature vector; The low-frequency fluctuation features are aggregated to generate the global anomaly feature vector.

2. The lithium battery SOC voltage evaluation method according to claim 1, characterized in that, The step of performing curve alignment processing on the historical charge-discharge data based on dynamic time warping to generate a standardized charge-discharge feature set specifically includes: The voltage curve, current curve, and temperature curve are segmented into time series to generate multiple local time series segments. Calculate the dynamic time warping distance between each local time series segment and the preset baseline curve; The local time series segments are scaled and adjusted according to the dynamic time warping distance to generate an aligned and standardized time series. The aligned and standardized time series are fused to generate the standardized charge and discharge feature set.

3. The lithium battery SOC voltage evaluation method according to claim 2, characterized in that, The step of inputting the standardized charge-discharge feature set into a pre-trained deep neural network model for feature extraction to generate a high-order charge-discharge feature matrix specifically includes: The convolutional layer of the deep neural network model is used to extract local features from the standardized charge and discharge feature set to generate a primary feature map. The attention mechanism layer of the deep neural network model is used to assign feature weights to the primary feature map to generate a weighted feature map. The weighted feature map is subjected to feature dimensionality reduction through the fully connected layer of the deep neural network model to generate the high-order charge-discharge feature matrix.

4. The lithium battery SOC voltage evaluation method according to claim 1, characterized in that, The step of detecting abnormal intervals in the real-time charge and discharge data based on the voltage deviation score set, and identifying periods of abnormal voltage fluctuations, specifically includes: A sliding window statistical analysis is performed on the voltage deviation score set to detect abnormal score intervals that exceed a preset threshold; The abnormal voltage fluctuation period is generated based on the real-time charge and discharge data time range corresponding to the abnormal scoring interval.

5. The lithium battery SOC voltage evaluation method according to claim 1, characterized in that, The step of tracing the root causes of voltage anomalies based on the local and global anomaly feature vectors to identify key influencing factors specifically includes: The local anomaly feature vectors are matched with a preset battery aging feature library to identify potential aging factors; The global anomaly feature vector is matched with a preset environmental impact feature library to identify potential environmental factors; The key influencing factors are determined based on the combined weights of the aging factors and environmental factors.

6. The lithium battery SOC voltage evaluation method according to claim 5, characterized in that, The SOC voltage calibration strategy generated based on the key influencing factors specifically includes: Select the corresponding compensation algorithm according to the type of aging factor, and determine the compensation coefficient according to the intensity of the environmental factor; The SOC voltage calibration strategy is generated by combining the compensation algorithm and the compensation coefficient.

7. A lithium battery SOC voltage evaluation system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the lithium battery SOC voltage evaluation method according to any one of claims 1 to 6.

Images