Lithium ion battery pack self-discharge diagnosis algorithm oriented to multi-stage dynamic charging scene and based on residual chargeable capacity prediction

Abstract

The invention discloses an energy storage battery pack self-discharge abnormity diagnosis method based on short-time charging data, and belongs to the technical field of battery safety monitoring. In order to overcome the defects that an existing method depends on a complete charging and discharging period and is difficult to adapt to a multi-stage variable working condition scene, multi-stage charging data features are extracted, self-adaptive extraction of multi-working condition features is achieved in combination with an auto-encoder, a data driving model fusing GCN-BiLSTM (Graph Convolution-Bidirectional Long Short-Term Memory) is constructed, and a multi-stage variable working condition scene is obtained. According to the method, the RCC (Remaining Charge Capability) of the single battery is accurately estimated, and the RCC of the single battery is accurately estimated. And finally, on the basis of the size, the distribution and the change condition of each single RCC in the battery pack, the overall inconsistency of the battery pack, the SOC inconsistency of the single batteries and the self-discharge fault can be diagnosed. And a guarantee is provided for safe operation of the energy storage system.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent management of energy storage batteries, and specifically studies a diagnostic method for abnormal self-discharge of individual battery cells in a battery pack under multi-stage charging conditions, where the battery pack contains more than 400 large-capacity battery cells. Background Art

[0002] Lithium-ion batteries have been widely used in the energy storage field due to their superior performance such as high energy density, high power density, long cycle life, and low self-discharge rate. However, since the voltage and power of a single lithium-ion battery are difficult to meet the requirements of an energy storage system, it is usually necessary to construct a battery pack by connecting hundreds of single batteries in series or in parallel to meet the requirements of large-scale energy storage systems for energy storage and output. To achieve efficient energy storage and conversion, the key lies in the consistency between single lithium-ion batteries, which is also an important basis for constructing a high-performance energy storage battery pack. During the process of evaluating the electrical performance differences of single lithium-ion batteries, self-discharge is a key detection index. Generally speaking, batteries with a higher self-discharge rate will show faster capacity decay, worse cycle characteristics, and shorter service life, thus exacerbating the inconsistency between single battery cells. This inconsistency will not only significantly reduce the overall performance of the energy storage battery pack but also increase the risk of system failures during operation. Currently, the commonly used self-discharge detection methods for lithium-ion batteries mainly include the open-circuit voltage method, the capacity decay method, the model-based method, and the data-driven method. The open-circuit voltage method determines the self-discharge rate by monitoring the voltage change of the battery in the open-circuit state, but the test time is long and it is difficult to detect tiny self-discharge; the capacity decay method evaluates self-discharge by recording the capacity loss of the battery over a certain period of time, but its operation is complex and the test cycle is long; the model-based method uses equivalent circuit model parameters to fit the self-discharge behavior. Although the test efficiency is high, it has high requirements for model accuracy and parameter identification, and there may be errors in actual applications. These methods are difficult to meet the requirements of high efficiency, accuracy, and strong adaptability at the same time. In recent years, artificial intelligence technology has developed rapidly and has been effectively applied in the field of battery safety warning and fault diagnosis. Various data-driven methods are established based on the historical operation data of the battery without exploring its complex fault mechanism. The basic steps of data-driven modeling include extracting data related to battery self-discharge in the domain and carefully selecting algorithms suitable for specific applications. For data-driven methods, the key is to develop advanced algorithms and feature extraction methods for specific applications. Based on the data-driven method, the present invention proposes a method for locating and diagnosing abnormal self-discharge of individual in-service energy storage battery packs based on short-term charging data. This method can quantify the evolution trend of self-discharge and can distinguish the inconsistency of SOC (State of Charge) from the inconsistency of self-discharge. Summary of the Invention

[0003] With the significant growth in the current demand for energy storage systems, due to the excellent performance of lithium batteries, new lithium battery energy storage systems have been greatly developed. The performance of the battery pack is crucial for the overall operating efficiency and reliability of the energy storage system. Due to reasons such as production and use, the lithium battery pack will show inconsistencies. If not diagnosed and processed in time, there will be serious consequences. In particular, the self-discharge of lithium batteries is often caused by internal short circuits due to metal impurities, burrs, etc. piercing the separator, posing a risk of thermal runaway. Therefore, accurately and timely locating and diagnosing abnormal battery cells has become a key indicator for measuring the economic value and operating life of energy storage systems, with important practical significance. However, since self-discharge is an inherent characteristic of electrochemical systems, batteries will experience a certain degree of self-discharge during normal use. Therefore, it becomes difficult to accurately obtain the changing trend of the self-discharge degree of individual batteries. In addition, existing evaluation methods usually rely on long charging and discharging data or fixed feature extraction methods, which face significant limitations in actual multi-stage variable operating conditions of energy storage scenarios. The present invention proposes a method for diagnosing self-discharge anomalies in energy storage battery packs based on short-term charging data. By extracting the changing data of multi-stage charging and combining with an autoencoder to adaptively extract multi-condition features, a data-driven model of GCN-BiLSTM (Graph Convolutional Network - Bidirectional Long Short-Term Memory) is constructed to accurately estimate the RCC (Remaining charging Capacity) of individual batteries, effectively solving the problems that traditional methods rely on complete charging and discharging cycles and are difficult to adapt to complex charging and discharging strategies. This technology uses the changing trend of the RCC value to monitor the self-discharge degree of the battery in real time, combines sliding window statistics to distinguish self-discharge anomalies caused by capacity attenuation and internal short circuits, and realizes early risk warning (such as triggering an alarm when the weekly decrease of RCC exceeds 5%). At the same time, it is compatible with multi-stage charging scenarios, providing a high-precision and low-latency safety diagnosis solution for energy storage systems, and having significant application value in fields such as cascade utilization screening, dynamic balancing optimization, and thermal runaway prevention.

[0004] To achieve the above object, the present invention provides the following technical solutions: A fault diagnosis algorithm that uses deep learning algorithms based on historical operating data to evaluate the RCC of batteries and then detect abnormal self-discharging monomers in the battery pack. Specifically, it includes the following six parts: S1: During the charging and operation of the energy storage system, the battery data is collected in real time through the BMS (Battery Management System) and data preprocessing is performed. The specific process includes:

[0005] S11: Compared with the discharge conditions of the battery pack in the energy storage system, the charging process is easier to control. Use the BMS of the energy storage system to collect the charging data of the battery pack, including time, total voltage, current, maximum cell voltage, minimum cell voltage, maximum cell temperature, minimum cell temperature, battery cell voltage, etc., and record and store the above data.

[0006] S12: Before fault diagnosis, the battery cluster data needs to be cleaned to improve data quality and increase the accuracy and reliability of the algorithm. Preprocess the above data, specifically including: removing redundant duplicate values, sorting out-of-order values, eliminating outliers, and filling in missing values. Then store and record the selected valid battery data for subsequent use.

[0007] S2: Construct samples for subsequent model training. The specific process includes: S21: There is rich inconsistent information at the current switching point. Therefore, the voltage and current data at the current switching point can be selected as the input. Screen the cells in the battery pack that reach the upper cut-off voltage, define the current switching event detection condition, and trigger data interception when the charging current change rate exceeds the preset threshold ΔI_threshold; screen the single cells in the battery pack that reach the upper cut-off voltage, and symmetrically intercept the three-dimensional voltage-current-temperature sequence of n sampling points centered on the switching point, mainly including the single cell voltage sequence V = V 1 , V 2 ,…, V n , current sequence I = I 1 , I 2 ,…, I n and temperature sequence T = T 1 , T 2 ,…, T n . The final construction of the sample can be expressed as:

[0008] S22: During the charging process of the battery pack, consider the single cells that reach the upper cut-off voltage as being close to full charge. At this time, its remaining charge capacity can be calculated by Ampere integration of the charging data after the sample end point. The specific steps are as follows: First, determine the time range. For each single cell that reaches the upper cut-off voltage, define the time range T 0, T1]; Starting time T 0: The starting time of the remaining charging segment of the single cell, determined by the end point of the selected sample in S21; Cut-off time T 1: The time point when the actual charging of the single cell ends, that is, the time when the charging current drops to the termination threshold or charging stops after constant current changes to constant voltage; Then extract the charging data, within the time range T 0, T 1], extract the current data sequence of this single cell { I ( t 0), I ( t 1), …, I ( t b )}, where b represents the number of sampling points in the remaining charging segment; Subsequently, calculate the remaining charging capacity. According to the Ampere integral formula, calculate the remaining charging capacity of the single cell Q :

[0009] Finally, calculate the remaining charging capacity of each single cell that reaches the upper cut-off voltage to form the remaining charging capacity matrix of the single cell: Q = Q 1 ，Q 2 ， … ，Q N , where N is the number of battery cells.

[0010] S23: Normalize the obtained samples. Use the maximum-minimum normalization for voltage, and use the method of dividing by the maximum value for current and remaining charging capacity to facilitate subsequent model training.

[0011] S3: Construct an encoder-decoder network. The input is the three-channel time series of voltage U, current I, and temperature T of the charging data sample; The encoder extracts features through 3 one-dimensional convolutional layers and compresses them into the latent space, and outputs an intermediate feature vector with a dimension of p * q ( p is the preset feature dimension, q is the preset feature length); The decoder reconstructs the input data through the transposed convolutional layer, and the loss function is defined as the sum of the mean square errors of the three channels of the input and the reconstructed output: Among them, x k refers to the input sample, x krecon Refers to the reconstructed sample; the intermediate feature is mapped to a standardized dimension by a fully connected layer p*q of the feature vector, which serves as the input feature for the following capacity prediction model.

[0012] S4: Construct a data-driven model of GCN-BiLSTM to estimate the remaining charge capacity of a single battery cell. Specifically, it includes: S41: During the data construction process, first divide the intermediate feature into w equal time segments according to a fixed length. Each segment obtains a node feature vector by calculating the mean value. These nodes together form the vertex set of the graph. Subsequently, establish the connection relationship of the edges by calculating the cosine similarity of the features between nodes. When the similarity between two nodes exceeds the preset threshold, a directed edge is established between these two nodes. All node pairs that meet the conditions will establish edge connections, finally forming a directed graph structure based on the similarity of time series data, where the nodes retain the local statistical features of the original time series data, and the edges reflect the correlation strength between different time segments.

[0013] S42: This model adopts a hybrid architecture design of a graph convolutional neural network (GCN) and a bidirectional long short-term memory network (BiLSTM). During the construction process, first use two layers of graph convolutional network (GCNConv) to perform feature extraction and message passing on the input graph data. After each layer of GCN, a ReLU activation function is connected to increase the non-linear expression ability. Subsequently, perform graph-level feature pooling through global_mean_pool to aggregate the features of all nodes into a unified graph representation. Then input the pooled features into the BiLSTM layer for time series modeling to capture the long-term dependencies in the sequence. Finally, map the hidden state of the BiLSTM to the prediction target space through a fully connected layer, thereby completing the end-to-end mapping process from the graph structure features to the final prediction value. The training of the entire network model is to optimize the loss function. Select MAE (Mean Absolute Error) as the loss function to quantify the difference between the predicted value and the true value. The formula is as follows:

[0014] where, y i is the true value, y i p is the model predicted value, m is the number of samples.

[0015] S43: During the training implementation of the model, the entire training process is built based on the PyTorch framework. The model structure adopts a hybrid architecture of GCN and BiLSTM. The GCN part contains two layers of graph convolutional networks, each followed by a ReLU activation function. The BiLSTM layer is used for sequence modeling. During training, the Adam optimizer is used for parameter optimization, the initial learning rate is set to 0.005, and the total number of training epochs is 1000. The model is trained in a batch processing manner, with batch_size set to 512, and data is loaded in batches and randomly shuffled through DataLoader. In each training epoch, the model first performs forward propagation to calculate the predicted values, uses L1Loss to calculate the loss, and then updates the model parameters through backpropagation.

[0016] S5: Based on the trained model, calculate the remaining charge capacity (RCC) of each cell for each cycle:

[0017] S51: According to the method of S21, extract the input samples from the charging data of each charging cycle and then input the samples into the trained model to estimate the remaining charging capacity of each cell when it reaches the upper cut-off voltage in the j th cycle Q j = Q 1j ，Q 2j ，…，Q ij ，Q Nj , where N represents that the battery pack has a total of N battery cells, Q ij represents the i th cell's remaining charging capacity after reaching the upper cut-off voltage in the j th cycle.

[0018] S52: Due to different battery charging strategies, the selection at the current switching point is not exactly the same. Thus, the starting points for calculating the remaining charge capacity are not consistent. To ensure that the calculated RCC values are relatively stable each time. Therefore, we need to perform normalization on Q j so that the RCC values of the cells that cannot reach the upper cut-off voltage can be obtained, eliminating systematic biases and improving the stability of the calculation results. The specific calculation formula is as follows: where, RCC ij : The remaining charge capacity of the i th cell in the jth cycle; min( Q j ) is the jThe minimum remaining charge capacity of all monomers in a cycle. Obtain the estimated value of the RCC of all monomers in the j-th cycle RCC j = RCC 1j ， RCC 2j , …, RCC Nj :

[0019] S5: The RCC of a lithium battery can effectively reflect the inconsistency between battery monomers. Generally speaking, for a monomer with a relatively small RCC, its SOC is relatively high and its capacity is relatively small, while for a monomer with a relatively large RCC, its SOC is relatively low and its capacity is relatively large. Based on the magnitude, distribution, and variation of the RCC of each monomer in the battery pack, we can diagnose the overall inconsistency of the battery pack, the SOC inconsistency of battery monomers, and self-discharge faults. The specific process is as follows:

[0020] S51: First, measure the overall inconsistency of the battery pack. To quantify the overall inconsistency of the battery pack, the range of the RCC value is used as a measurement index. The range is defined as:

[0021] Where: max( RCC j ):The maximum value of the monomer RCC in the j -th cycle; min( RCC j ):The minimum value of the monomer RCC in the j -th cycle. The range Range RCC directly reflects the range of differences in monomer RCC within the battery pack and can be used to characterize the consistency of the battery pack. The larger the range, the more serious the inconsistency of the monomers within the battery pack.

[0022] S52: Then there is the SOC inconsistency of the monomer. The difference between the RCC value of the battery monomer and the median of the RCC can be calculated. The selection of the median has noise resistance and can accurately reflect the central tendency of the data. The formula is as follows:

[0023] Where: D ij :The difference between the RCC of the j -th cycle and the median for the i -th monomer; Median( RCC j ):The median of the RCC for the j ​Median of the RCC set in a cycle. Monomer difference D ij Characterizes the deviation of its RCC from the concentration level of the battery pack, thereby reflecting the SOC inconsistency of the monomer.

[0024] S53: When a self-discharge fault occurs in a monomer of the battery pack, its RCC value will show an abnormal decrease, and the slope change of its RCC value will be very different from that of a normal monomer. Based on this principle, we can use the sliding window local consistency analysis method to calculate the RCC slope of each monomer in the battery pack, and then determine the self-discharging monomer. The sliding window local consistency analysis method is an efficient anomaly detection method, especially suitable for time series data with high noise or large volatility. Through the sliding window local consistency analysis method, we can further diagnose the battery self-discharge fault.

[0025] First, define the sliding window: According to the sampling frequency and analysis requirements of the battery pack cycle data, set a reasonable window length ( W ). The window size can be set to 5 or 10 cycles, depending on the volatility of the data. And select an appropriate sliding step ( Δ ) such as 1 or 2 to ensure that the window covers all data points. The range of the sliding window is T k = j , j + W - 1], and the formula for calculating the change rate of RCC is:

[0026] △ RCC i,k is a matrix representing the change rate of the RCC of each battery monomer within the window T k .

[0027] Calculate the local mean within the time window T k , and based on this, calculate the local deviation of the monomer within this time window to measure the consistency evaluation of each monomer's change within this time window. The specific formula is as follows:

[0028] where: represents the local mean of each time window within the time window T K , and S i,k is the local deviation, representing the difference between the monomer change and the local average change.

[0029] S54: Finally, the range of the RCC obtained in each cycle can be calculated based on historical data RCC , the difference between the RCC of a single cell and the median D ij to set the threshold according to the 95th percentile thrshold1 , thrshold2 to evaluate the overall inconsistency of the battery pack and the inconsistency of the SOC of single battery cells. And calculate the local deviation degree based on historical data S i,k to set the threshold according to the 95th percentile thrshold3 and the setting of exceeding the threshold in multiple consecutive time windows thrshold4 to locate and diagnose the self-discharging single cells.

[0030] Generally speaking, compared with the prior art, the above technical solutions conceived by the present invention can achieve the following beneficial effects: The present invention proposes a method for locating and detecting self-discharging single cells in a large-scale energy storage battery cluster containing more than 400 single cells.

[0031] The present invention uses short-time charging data to generate accurate label data to support the deep learning model, breaking through the limitations of traditional methods that rely on long-time complete cycle data, and significantly reducing the data acquisition time and calculation cost. At the same time, this method greatly expands the applicable working condition range of the model, including different charge and discharge rates and complex operating environments, ensuring its robustness and universality in practical scenarios. Through the combination of rapid label generation and deep learning model, the rapid diagnosis of lithium battery self-discharge faults is realized, effectively improving the operation efficiency of the energy storage system; The present invention constructs RCC labels based on real collected data, improves the accuracy of RCC estimation and the credibility of diagnosis, and overcomes the deficiencies of traditional methods that rely on simulation data or ideal conditions. At the same time, the sliding window technique is adopted to generate multiple samples in a single cycle, extract more useful information from the local features of the time series, and enhance the data utilization efficiency and the accuracy performance of the model. The introduction of the sliding window method not only improves the ability to identify noise and anomalies, but also expands the adaptability of the model to dynamic working conditions, making it more sensitive and reliable in practical applications; The present invention constructs a multi-condition adaptive feature extraction model with an encoder-decoder architecture, maps the input data under different working conditions to the same latent space by reconstructing the input, and extracts intermediate features, which can adapt to the changes in working conditions; The present invention constructs a time-series graph based on a graph structure by segmenting battery cell time-series data and using cosine similarity. Combining a graph convolutional network (GCN) to extract graph features and a bidirectional long short-term memory network (BiLSTM) to capture time dependencies, an end-to-end prediction framework is formed, avoiding cumbersome manual feature design. At the same time, the accuracy and efficiency of the model are improved through global feature pooling and an optimized training strategy, demonstrating a powerful ability to model complex time-series relationships and graph data. The present invention constructs a hierarchical diagnosis framework through the RCC index, systematically quantifying and diagnosing the battery state from the overall consistency of the battery pack to the inconsistency of the individual SOC, and then to the self-discharge fault. It includes introducing the RCC range to measure the overall consistency, using the deviation from the median to evaluate the inconsistency of the individual SOC, and detecting the self-discharge fault through sliding window local consistency analysis, enhancing the fineness and robustness of the diagnosis. At the same time, multi-layer thresholds are dynamically set based on historical data, effectively improving the adaptability and noise resistance of the method, and it is applicable to the health management and predictive maintenance of complex battery packs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required to be used in the embodiments of the present application. Obviously, the following described drawings are only some embodiments of the present application, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts: Figure 1 It is a flowchart for detecting self-discharge abnormal monomers in a battery cluster. Figure 2 It is an algorithm flowchart for estimating the remaining charge capacity. Figure 3 It is a framework diagram for hierarchical diagnosis using RCC. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] To achieve the above object, the present invention provides the following technical solutions: A fault diagnosis algorithm based on historical operation data, using deep learning algorithms to evaluate the RCC of a battery and detect self-discharge abnormal monomers in a battery pack. As Figure 1 shown, it specifically includes the following 6 steps.

[0034] Step 1: During the charging and operation of the energy storage system, the battery data is collected in real time through the BMS (Battery Management System) and data preprocessing is performed.

[0035] Compared with the discharge conditions of the battery pack in the energy storage system, the charging process is easier to control. The BMS of the energy storage system is used to collect the charging data of the battery pack, including time, total voltage, current, maximum cell voltage, minimum cell voltage, maximum cell temperature, minimum cell temperature, battery cell voltage, etc., and record and store the above data. Then, before fault diagnosis, the battery cluster data needs to be cleaned to improve the data quality and increase the accuracy and reliability of the algorithm. The above data is preprocessed, specifically including: removing redundant duplicate values, sorting out-of-order values, removing outliers, and filling in missing values. Then, the selected valid battery data is stored and recorded for subsequent use.

[0036] Step 2: Construct samples for subsequent model training. The specific process includes: First: There is rich inconsistent information at the current switching point. Therefore, the voltage and current data at the current switching point can be selected as the input. Screen the cells in the battery pack that reach the upper cut-off voltage, define the current switching event detection condition, and trigger data interception when the charging current change rate exceeds the preset threshold ΔI_threshold; screen the single cells in the battery pack that reach the upper cut-off voltage, and symmetrically intercept the three-dimensional voltage-current-temperature sequence of n sampling points centered on the switching point, mainly including the single cell voltage sequence V = V 1 , V 2 ,…, V n , current sequence I = I 1 , I 2 ,…, I n and temperature sequence T = T 1 , T 2 ,…, T n . The final construction of the sample can be expressed as:

[0037] Then construct the output sample. During the charging process of the battery pack, the single cells that reach the upper cut-off voltage are regarded as being close to full charge. At this time, its remaining charge capacity can be calculated by Ampere integration of the charging data after the sample end point. The specific steps are as follows: First, determine the time range. For each single cell that reaches the upper cut-off voltage, define the time range T 0, T 1]; start time T0: The starting time of the remaining charging segment of the single cell, determined by the end point of the selected sample in S21; cut-off time T 1: The time point when the actual charging of the single cell ends, that is, the time when the charging current drops to the termination threshold or stops charging after constant current changes to constant voltage; then extract the charging data, within the time range T 0, T 1], extract the current data sequence of this single cell { I ( t 0), I ( t 1), …, I ( t b )}, where b represents the number of sampling points in the remaining charging segment; then calculate the remaining charging capacity According to the Ampere integration formula, calculate the remaining charging capacity of the single cell Q :

[0038] Finally, calculate the remaining charging capacity of each single cell that reaches the upper cut-off voltage to form the remaining charging capacity: Q = Q 1 ，Q 2 ， … ，Q N , where N is the number of battery cells. Subsequently, the obtained samples are normalized. Use the maximum-minimum normalization for voltage, and use the method of dividing by the maximum value for current and remaining charging capacity for normalization to facilitate subsequent model training.

[0039] Step three: As Figure 2 shown, construct an encoder-decoder network, and the input is the three-channel time series of voltage U, current I, and temperature T of the charging data sample; the encoder extracts features through 3 one-dimensional convolutional layers and compresses them into the latent space, and the output dimension is p * q intermediate feature vector ( p is the preset feature dimension, q is the preset feature length); the decoder reconstructs the input data through the transposed convolutional layer, and the loss function is defined as the sum of the mean square errors of the three channels of the input and the reconstructed output: Among them, x k refers to the input sample, x k reconRefers to the reconstructed sample; the intermediate feature "feature" is mapped to a standardized p*q feature vector through a fully connected layer, which serves as the input feature for the following capacity prediction model.

[0040] Step 4: As Figure 2 shown, construct a data-driven model of GCN-BiLSTM to estimate the remaining charging capacity of a single battery cell. Specifically, it includes: During the data construction process, first divide the intermediate feature "feature" into w equal time segments according to a fixed length. Each segment obtains a node feature vector by calculating the mean value, and these nodes together form the vertex set of the graph; subsequently, establish the connection relationship of the edges by calculating the cosine similarity of the features between nodes. When the similarity between two nodes exceeds a preset threshold, a directed edge is established between these two nodes, and all node pairs that meet the conditions will establish edge connections, ultimately forming a directed graph structure based on the similarity of time series data, where the nodes retain the local statistical features of the original time series data, and the edges reflect the correlation strength between different time segments. This model adopts a hybrid architecture design of a graph convolutional neural network (GCN) and a bidirectional long short-term memory network (BiLSTM). During the construction process, first use two layers of graph convolutional network (GCNConv) to perform feature extraction and message passing on the input graph data. After each layer of GCN, a ReLU activation function is connected to increase the non-linear expression ability; subsequently, global_mean_pool is used for graph-level feature pooling to aggregate the features of all nodes into a unified graph representation; then the pooled features are input into the BiLSTM layer for time series modeling to capture the long-term dependencies in the sequence; finally, a fully connected layer maps the hidden state of the BiLSTM to the prediction target space, thus completing the end-to-end mapping process from the graph structure features to the final prediction value. The training of the entire network model is to optimize the loss function. Select MAE (Mean Absolute Error) as the loss function to quantify the difference between the predicted value and the true value. The formula is as follows:

[0041] Among them, y i is the true value, is the model predicted value, mis the number of samples. Then, during the training implementation of the model, the entire training process is built based on the PyTorch framework. The model structure adopts a hybrid architecture of GCN and BiLSTM. The GCN part contains two layers of graph convolutional networks, each followed by a ReLU activation function. The BiLSTM layer is used for sequence modeling. During training, the Adam optimizer is used for parameter optimization, the initial learning rate is set to 0.005, and the total number of training epochs is 1000. The training of the model adopts a batch processing method, with batch_size set to 512, and data is loaded in batches and randomly shuffled through DataLoader. In each epoch of training, the model first performs forward propagation to calculate the predicted values, uses L1Loss to calculate the loss, and then updates the model parameters through backpropagation.

[0042] Step Five: Based on the trained model, calculate the remaining charge capacity (RCC) of each cell for each cycle: First, according to the method of S21, extract the input samples from the charging data of each charging cycle, and then input the samples into the trained model to estimate the remaining charging capacity of each cell when it reaches the upper cut-off voltage in the j th cycle Q j = Q 1j ，Q 2j ，…，Q ij ，Q Nj , where N represents that the battery pack has N battery cells, Q ij represents the i th cell, and j represents the remaining charging capacity of the

[0043] Then, due to different battery charging strategies, the selection at the current switching point is not exactly the same. Thus, the starting points for calculating the remaining charge capacity are not consistent. To ensure that the calculated RCC values are relatively stable each time. Therefore, we need to perform normalization processing on Q j so as to obtain the RCC values of the cells that cannot reach the upper cut-off voltage, eliminate systematic biases, and improve the stability of the calculation results. The specific calculation formula is as follows: where, RCC ij : the remaining charging capacity of the i th cell in the jth cycle; min( Q j ) is the jThe minimum of the remaining charge capacities of all monomers in a cycle. Obtain the estimated values of the RCC of all monomers in the j-th cycle RCC j = RCC 1j ， RCC 2j , …, RCC Nj :

[0044] Step 6: The RCC of the lithium battery can effectively reflect the inconsistency between battery monomers. Generally, for a monomer with a relatively small RCC, its SOC is relatively high and its capacity is relatively small, while for a monomer with a relatively large RCC, its SOC is relatively low and its capacity is relatively large. Based on the magnitude, distribution, and variation of the RCC of each monomer in the battery pack, we can diagnose the overall inconsistency of the battery pack, the SOC inconsistency of battery monomers, and self-discharge faults, as Figure 3 shown below:

[0045] First, measure the overall inconsistency of the battery pack. To quantify the overall inconsistency of the battery pack, the range of the RCC values is used as a measurement index. The range is defined as:

[0046] where: max( RCC j ):The maximum value of the monomer RCC in the j -th cycle; min( RCC j ):The minimum value of the monomer RCC in the j -th cycle. The range Range RCC directly reflects the range of differences in monomer RCC within the battery pack and can be used to characterize the consistency of the battery pack. The larger the range, the more serious the inconsistency of the monomers within the battery pack.

[0047] Next is the monomer SOC inconsistency. The difference between the monomer RCC value and the median of the RCC can be calculated. The selection of the median has noise resistance and can accurately reflect the central tendency of the data. The formula is as follows:

[0048] where: D ij :In the j -th cycle, the difference between the RCC of the i -th monomer and the median; Median( RCC j ):In the j ​Median of the RCC set in a cycle. Monomer difference D ij Characterizes the deviation of its RCC from the concentration level of the battery pack, and thus reflects the SOC inconsistency of the monomer.

[0049] Secondly, when a self-discharge fault occurs in a monomer of the battery pack, its RCC value will show an abnormal decrease, and the slope change of its RCC value will be very different from that of a normal monomer. Based on this principle, we can use the sliding window local consistency analysis method to calculate the RCC slope of each monomer in the battery pack, and then determine the self-discharging monomer. The sliding window local consistency analysis method is an efficient anomaly detection method, especially suitable for time series data with high noise or large volatility. Through the sliding window local consistency analysis method, we can further diagnose the battery self-discharge fault.

[0050] First, define the sliding window: According to the sampling frequency and analysis requirements of the battery pack cycle data, set a reasonable window length ( W ). The window size can be set to 5 or 10 cycles, depending on the volatility of the data. And select an appropriate sliding step ( Δ ) such as 1 or 2 to ensure that the window covers all data points. The range of the sliding window is T k = j , j + W - 1], and the formula for calculating the change rate of RCC is:

[0051] △ RCC i,k is a matrix representing the change rate of the RCC of each battery monomer within the window T k .

[0052] Calculate the local mean within the time window T k , and based on this, calculate the local deviation of the monomer within this time window to measure the consistency evaluation of each monomer's change within this time window. The specific formula is as follows:

[0053] Where: represents the local mean of each time window within the time window T K , and S i,k is the local deviation, representing the difference between the monomer change and the local average change.

[0054] Finally, the range of the RCC obtained in each cycle can be calculated based on historical data RCC , the difference between the RCC of a single cell and the median D ij The threshold is set according to the 95% quantile thrshold1 , thrshold2 to evaluate the overall inconsistency of the battery pack and the inconsistency of the SOC of individual battery cells. And calculate the local deviation degree according to historical data S i,k The threshold is set according to the 95% quantile thrshold3 and the threshold setting is exceeded in multiple consecutive time windows thrshold4 to locate and diagnose faults of self-discharging single cells.

Claims

1. A method for estimating the RCC (Remaining charging Capacity) based on short-term charging data and the GCN-BiLSTM (Graph Convolutional Network - Bidirectional Long Short-Term Memory) algorithm applicable to multi-stage variable charging conditions, and then diagnosing abnormal single-cell self-discharge of a battery pack, characterized in that, Including the following steps: S1. The battery management system collects the operation data of the battery pack in real time during the charging stage and performs preprocessing. The data includes the total voltage, current, extreme values of single-cell voltage, extreme temperature values, and the single-cell voltage sequence; S2. Intercept the charging data of a fixed length at the current switching point and calculate the remaining charging capacity of the single cell that reaches the upper cut-off voltage Q As a training sample, and perform normalization processing on the sample data; S3. Construct a multi-condition adaptive feature extraction model with an encoder-decoder architecture, map the input data under different conditions to the same latent space through reconstructing the input, and extract intermediate features; S4. Construct a hybrid model based on the graph convolutional network and the bidirectional long short-term memory network. Extract the graph structure features through two layers of GCNConv layers, input them into the BiLSTM for time series modeling after global pooling, and establish the mapping relationship between the intermediate features and the remaining charging capacity; S5. Input the real-time charging data into the trained model, calculate the remaining rechargeable capacity (RCC) values of each single cell, and obtain the standardized RCC index through range normalization processing; S6. Based on the range distribution, median deviation, and sliding window local consistency analysis of the RCC values, realize the diagnosis and separation of the overall inconsistency of the battery pack, the SOC (State of Charge) inconsistency fault, and the self-discharge fault.

2. The method according to claim 1, wherein The data acquisition and preprocessing in step S1 include: recording the total voltage, total current, extreme values of single-cell voltages, extreme temperature values, and all single-cell voltage sequences of the battery cluster at a preset sampling frequency; performing multi-level cleaning on the original data; using moving window mean filtering to eliminate high-frequency noise; identifying and removing abnormal voltage jump points based on the box plot method; filling in missing data segments using linear interpolation; constructing a feature matrix X∈R^{N×T×3} with timestamp markings, where N is the number of single cells, T is the time step, and the 3D features include voltage, current, and temperature.

3. The method according to claim 1, characterized in that The sample construction method in step S2 includes: Define the detection condition for the current switching event. When the charging current change rate exceeds the preset threshold ΔI_threshold data interception is triggered; select the single cells in the battery pack that reach the upper cut-off voltage, and symmetrically intercept the voltage-current-temperature three-dimensional sequence of n sampling points centered on the switching point as the input sample: For a single cell that reaches the upper cut-off voltage, the remaining charging capacity is calculated by the Ampere integration method, where the time range T0, T1 is defined as follows: T0 is the start time of the remaining charging segment of the single cell; T1 is the actual charging end time of the single cell, that is, the time point when the charging reaches the upper cut-off voltage of charging, and then according to the Ampere integration formula, the remaining charging capacity of the single cell is calculated Q : Among them, the time range is within T0 , T1 , and extract the current data sequence of this monomer I(t) .

4. The method according to claim 1, characterized in that, The encoder-decoder architecture of step S3 includes: constructing an encoder-decoder network, with the input being the three-channel time series of voltage U, current I, and temperature T of the charging data samples; the encoder extracts features through 3 one-dimensional convolutional layers and compresses them into the latent space, outputting an intermediate feature vector of dimension p * q ( p is the preset feature dimension, q is the preset feature length); the decoder reconstructs the input data through a transposed convolutional layer, and the loss function is defined as the sum of the mean square errors of the three channels of the input and the reconstructed output: Among them, x k refers to the input sample, x k recon refers to the reconstructed sample; the intermediate feature is mapped to a standardized p*q feature vector, which serves as the input data for the following capacity prediction model.

5. The method according to claim 1, characterized in that, The construction of the hybrid model in step S4 includes: a graph structure construction module: evenly dividing the input intermediate feature feature into w time segments of a fixed length. Each segment obtains a node feature vector by calculating the mean value, and these nodes together form the vertex set of the graph; subsequently, the connection relationship of the edges is established by calculating the cosine similarity of the features between the nodes. When the similarity between two nodes exceeds the preset threshold, a directed edge is established between these two nodes, and all node pairs that meet the conditions will establish edge connections, ultimately forming a directed graph structure based on the similarity of time series data, where the nodes retain the local statistical features of the original time series data, and the edges reflect the correlation strength between different time segments. A model establishment module: adopts a hybrid architecture of GCN and BiLSTM. First, graph feature extraction and message passing are performed through a two-layer GCNConv (including ReLU activation). After aggregating the graph-level features through global average pooling, the BiLSTM is input to capture the time series dependence relationship, and finally, the end-to-end target mapping prediction is realized through a fully connected layer. The training of the entire network model is to optimize the loss function: Select MAE (Mean Absolute Error) as the loss function to quantify the difference between the predicted value and the true value. The formula is as follows: Among them, y i is the true value, y i p is the model prediction value, m is the number of samples.

6. The method according to claim 1, wherein The RCC calculation in step S5 includes: dynamically correcting the original Q value; extracting input samples from the charging data of each charging cycle; symmetrically intercepting n sampling points of the input samples centered on the switching point Sample ij : representing the input sample of the i th cell in the j th cycle; then inputting the sample into the trained model to estimate the remaining charging capacity of each cell when charged to the upper cut-off voltage in the j th cycle Q j = Q 1j ，Q 2j ，…，Q ij ，Q Nj , where N represents that the battery pack has a total of N battery cells, Q ij represents the remaining charging capacity of the i th cell after reaching the upper cut-off voltage in the j th cycle; to ensure that the RCC value obtained each time is relatively stable, so we need to perform an offset adjustment on Q j so as to obtain the RCC value of the cell that cannot reach the upper cut-off voltage, eliminate the system deviation and improve the stability of the calculation result; the specific calculation formula is as follows: Among them, RCC ij : In the j-th cycle, the remaining rechargeable capacity of the i -th monomer; min( Q j ) is the minimum value of the remaining charging capacities of all monomers in the j -th cycle. The estimated value of the RCC of all monomers in the j-th cycle is obtained RCC j = RCC 1j ， RCC 2j , …, RCC Nj .​ 7. The method according to claim 1, characterized in that The diagnosis mechanism in step S6 includes: First, measure the overall inconsistency of the battery pack, and calculate the range of the RCC values as the measurement index: Among them, max( RCC j ): the maximum value of the monomer RCC in the j th cycle; min( RCC j ): the minimum value of the monomer RCC in the j th cycle, and the range Range RCC directly reflects the range of differences in the monomer RCC within the battery pack and can be used to characterize the consistency of the battery pack. The larger the range, the more serious the inconsistency of the monomers within the battery pack; then, to evaluate the inconsistency of the monomer SOC, calculate the difference between the RCC value of the battery monomer and the median of the RCC D ij : Among them, D ij : In the j th cycle, the difference between the RCC of the i th monomer and the median is Median( RCC j ): The median of the RCC set in the j th cycle; The monomer difference D ij characterizes the deviation of its RCC from the concentration level of the battery pack, and thus reflects the SOC inconsistency of the monomer; Finally, the local consistency analysis method of the sliding window is used to calculate the RCC slope of each monomer of the battery pack, and then the self-discharging monomer is determined; First, define the sliding window: According to the sampling frequency and analysis requirements of the battery pack cycle data, set a reasonable window length ( W ). The window size can be set to 5 or 10 cycles, depending on the volatility of the data. And select an appropriate sliding step size ( Δ ) such as 1 or 2 to ensure that the window covers all data points. The range of the sliding window is T k = j , j+W -1], and the formula for calculating the change rate of RCC is: Among them, △ RCC i,k is a matrix representing the change rate of each battery cell RCC within the window T k ; calculate the local mean within the time window T k and calculate the local deviation of the cell within this time window based on this, so as to measure the consistency of each cell's change within this time window. The specific formula for evaluation is as follows: Among them, ΔRCC k ave represents the local mean of each time window within the time window T K ; Among them, S i,k is the local deviation degree, representing the difference between the monomer change and the local average change; finally, the range of the RCC for each cycle is calculated based on historical data RCC , the difference between the RCC of the monomer and the median D ij The 95% quantile of thrshold1 , thrshold2 is used to evaluate the overall inconsistency of the battery pack and the inconsistency of the SOC of the battery monomers. And the local deviation degree is calculated according to historical data S i,k The threshold is set according to the 95% quantile of thrshold3 and continuous z time windows exceeding the threshold setting thrshold4 are used to locate the self-discharging monomers and diagnose faults.

Images