Battery inconsistency detection and grading method combining image and time sequence characteristics

Abstract

The invention discloses a battery inconsistency detecting and grading method combining image and time sequence characteristics. The method comprises the following steps of: firstly, converting battery time sequence data into a two-dimensional image by adopting a Gramer angle field method, extracting potential inconsistency information from the two-dimensional image, and quantifying behavior fluctuation of a battery in different time periods by combining time sequence characteristics; in addition, the invention provides a cumulative inconsistency measurement method, the inconsistency change trend is calculated through a sliding window, and performance degradation or potential problems of the battery are captured. More importantly, the invention provides a method for grading the inconsistency of the single batteries and the modules of the batteries, and the method can be used for carrying out accurate grading according to the degree of the inconsistency of the batteries. According to the method, the single batteries with inconsistency can be accurately identified, the inconsistency degree of the single batteries can be quantified, a unified evaluation standard is provided for the inconsistency performance of the batteries of different vehicles, and effective and accurate technical support is provided for the inconsistency detection of the battery pack of the electric vehicle.

Description

Technical Field

[0001] The present invention relates to a method for detecting and grading battery inconsistency by combining image and temporal features, and belongs to the field of state evaluation of lithium-ion power batteries for new energy vehicles. Background Art

[0002] The problem of battery inconsistency in lithium-ion power batteries has always been an important challenge in the design, manufacture, and application of battery systems. Lithium-ion batteries are widely used in high-power demand devices such as electric vehicles and energy storage systems, and are usually composed of multiple battery cells connected in series or in parallel. In actual use, due to factors such as manufacturing process differences, material properties, aging processes, and usage environments, the performance of each battery cell often varies to varying degrees, mainly manifested as differences in capacity, internal resistance, charge and discharge efficiency, temperature response, and attenuation rate. With the increasing global environmental pollution and energy crisis, new energy vehicles (NEVs) have gradually become a key part of the sustainable transportation system. As the core component of new energy vehicles, the performance of the battery system directly affects the range, safety, and reliability of electric vehicles. In recent years, lithium-ion batteries (LiBs) have become the mainstream batteries for electric vehicles due to their high energy density, long service life, and low self-discharge rate. However, despite the significant advantages of lithium-ion batteries, their high energy density also brings certain safety hazards. Especially when there are inconsistencies or potential failures in the battery, the safety risks may be amplified, leading to serious consequences.

[0003] The inconsistency of battery voltage has been proven to be one of the important early warning signals for battery failures. The voltage inconsistency of batteries usually manifests as the voltage difference between different single cells, reflecting the performance fluctuations and potential failures during the use of the battery pack. Although traditional battery monitoring techniques can provide a preliminary assessment of the battery state, these methods often have certain limitations in detecting early failures and subtle performance changes. Therefore, more accurate fault diagnosis for battery voltage inconsistency has become an important topic in the research of Battery Management System (BMS). Currently, the research on battery inconsistency detection mainly falls into three categories: statistical analysis-based methods, data-driven methods, and model-based methods. Traditional model-based methods predict the operating state of the battery by establishing a mathematical model of the battery system, but as the complexity of the battery system structure increases, it is often difficult to guarantee the model accuracy. In contrast, data-driven methods, especially machine learning and deep learning techniques, have received extensive attention in the application of battery health management in recent years. By analyzing the historical data of the battery, data-driven methods can directly extract features from the battery data to achieve more efficient fault diagnosis.

[0004] Statistical analysis methods help researchers understand the relationships between variables through data collection, organization, and analysis. In the assessment of battery inconsistency, this method helps evaluate the performance differences and potential failures of battery systems under different operating conditions, thereby optimizing management strategies to ensure the safe operation of battery systems. However, there are also some deficiencies in statistical analysis methods. For example, statistical analysis usually relies on a large amount of historical data and may not work well when the data is insufficient. In addition, relatively simple statistical analysis methods may be difficult to identify inconsistencies that occur in the early and mid-stages. With the development of data science, data-driven methods have been widely used in the assessment of lithium-ion battery inconsistency. This method analyzes the historical and real-time data of batteries, automatically extracts features, and identifies changes in battery performance and potential failures. Different from traditional methods, data-driven methods do not need to rely on complex physical models, but adaptively learn the behavior of battery systems through real-time data, providing more accurate and flexible assessments. However, although data-driven methods have advantages in the assessment of battery inconsistency, they also face some challenges. The quality and integrity of data directly affect the prediction accuracy. When the data is missing or insufficient, the performance of the model will decline. The need for large-scale training data makes the computational cost relatively high. When facing complex non-linear faults and high-dimensional data, data-driven methods may not be able to achieve the accuracy of traditional physical models. Model-based methods play an important role in the assessment of lithium-ion battery inconsistency. These methods establish physical, chemical, and thermodynamic models of batteries to simulate the behavior of batteries under different operating conditions, helping to understand and predict the impact of battery inconsistency on performance, lifespan, and safety. Model-based methods can calculate the influence of different parameters and provide valuable references for battery design and management. Although model-based methods provide effective tools in the assessment of battery inconsistency, there are still some deficiencies. For example, models usually rely on a large amount of experimental data for verification, and the insufficiency or inaccuracy of the data will directly affect the accuracy of the model. The complexity of the model is also a challenge. Especially when multi-physical fields are coupled, the computational cost increases significantly, and overfitting is likely to occur. Summary of the Invention

[0005] Aiming at the difficulties in early battery inconsistency detection and the limitations of analyzing inconsistency with a single feature, the present invention proposes an inconsistency detection method combining image and time-series features. First, the present invention performs interpolation processing on the voltage data of real vehicles to avoid the impact of missing data on the assessment results. Then, by visualizing the time-series voltage data through GAF to extract features, it combines with the time-series data to deeply analyze the inconsistency of single-cell and module batteries. The proposed method verifies the effectiveness and reliability through the voltage data of real vehicles with different driving mileages. At the same time, the grading of the inconsistency degree of single-cell and module batteries is discussed, providing decision-making support for subsequent battery maintenance and replacement.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for detecting and grading battery inconsistency combining image and temporal features, including two parts: inconsistency detection and inconsistency grading; the steps of the inconsistency detection part are as follows:

[0008] Step 1: Data loading and preprocessing,

[0009] Step 1.1: Receive the real-time single-cell voltage data V transmitted by the lithium battery management system celli ;

[0010] Step 1.2: Use one of Newton interpolation method, Lagrange interpolation method, and linear interpolation method to perform data preprocessing on the missing data;

[0011] Step 2: Initialize the sliding window W,

[0012] Step 3: Extraction of single-cell battery inconsistency features,

[0013] Step 3.1: Calculate the mean, standard deviation, and fluctuation range of each single-cell battery within the current sliding window, and perform time series feature extraction;

[0014] Step 3.2: Imageize the voltage data through Gramian Angular Field (GAF),

[0015] Step 4: Calculate the inconsistency metric through the anomaly score obtained from the voltage data and the reconstruction error obtained from the image data,

[0016] Step 4.1: For each single-cell battery, calculate its anomaly score S on the temporal features;

[0017] Step 4.2: Adopt a Deep Convolutional Autoencoder (DCAE) as the image feature extraction model;

[0018] Step 4.3: Calculate the inconsistency metric;

[0019] Step 4.4: Determine the single-cell batteries with inconsistency through the results of the inconsistency metric. The single-cell battery with a larger inconsistency metric compared to the other batteries is the one with inconsistency;

[0020] The steps of the inconsistency grading part are as follows:

[0021] Step B1: Calculate the cumulative inconsistency metric,

[0022] Step B2: Perform an overall inconsistency rating by calculating the weighted average inconsistency score.

[0023] A method for detecting and grading battery inconsistency by combining image and time series features, the inconsistency detection is as follows:

[0024] Step 1.2: Interpolation was performed on the data with missing values using the Newton interpolation formula (1) or the like.

[0025]

[0026] where P n (x) is an n-degree interpolation polynomial constructed from the given data points, and f[x0,...,x k is the k-th order divided difference, is the continuous product basis function with respect to x.

[0027] A method for detecting and grading battery inconsistency by combining image and time series features, Step 2: Initialize the sliding window, set the window size to W, and the step size for each slide to S. Then the position of the sliding window can be expressed as:

[0028] W = {x i , x i+1 ,..., x i+W-1} Equation (2);

[0029] where x i represents the data points in the time series, W is the window size, and represents the number of data points within each window.

[0030] A method for detecting and grading battery inconsistency by combining image and time series features,

[0031] Step 3.1: Calculate the mean, standard deviation, and fluctuation range of each single battery within the current sliding window to extract time series features:

[0032] The mean mentioned above represents the overall level of the battery voltage and reflects the working state of the battery over a period of time;

[0033]

[0034] where x i represents the i-th value of the battery voltage data, and N is the total number of data samples;

[0035] The standard deviation mentioned above describes the degree of fluctuation of the battery voltage and reflects the stability of the battery; abnormal and drastic fluctuations in the battery voltage can be captured through the standard deviation;

[0036]

[0037] where σ is the standard deviation of the single cell at the current moment, μ is the mean value of the single cell at the current moment, and x i

[0038] represents the i-th value of the battery voltage data, and N is the total number of data samples; the fluctuation range refers to the difference between the maximum value and the minimum value of the battery voltage over a period of time;

[0039] Range = max(x i ) - min(x i ) Equation (5);

[0040] Step 3.2: Image the time series voltage data through Gramian Angular Field (GAF):

[0041] 3.2.1 Data normalization: Normalize the time series x = [x1, x2,..., x n , linearly map the original data x to the interval [0, 1], and obtain the normalized time series x′. The formula is:

[0042]

[0043] where min(x) and max(x) are the minimum and maximum values in the time series respectively, and x′ is the normalized data, with a value range between [0, 1];

[0044] 3.2.2 Calculate the cosine similarity between time points: By calculating the cosine similarity between any two time points x′ i and x′ j in the normalized time series, construct an n*n symmetric matrix G; each element G ij of the matrix G is defined as:

[0045]

[0046] where cos(θ ij ) is the angle between i and time point j, reflecting the correlation between each time point in the time series; 3.2.3 Image generation: By selecting an appropriate color mapping, visually display the values in the symmetric matrix G as colors in the image; the color mapping function is expressed as:

[0047] f: [0, 1] → RGB Equation (8);

[0048] where RGB represents a coordinate in the red, green, and blue three-channel color space;

[0049] 3.2.4 Image Channel Adjustment: To use the generated GAF as the input of the DCAE, the single-channel G matrix is converted into a three-channel image; this process is achieved by replicating the single-channel matrix three times, thus generating an image with three identical channels; this process is expressed as:

[0050] G RGB = Repeat(G, axis = 10, time = 3);

[0051] where G RGB represents the expanded three-channel image with a shape of 3 * n * n; the Repeat function repeats the single-channel matrix G three times to obtain a three-channel RGB image.

[0052] In the battery inconsistency detection and grading method combining image and time series features, step 4.1: For each single battery cell, calculate its anomaly score S in the time series features, which is measured by calculating the standardized difference between the time series features of each battery and the mean of the time series features of all batteries;

[0053] Step 4.2: The DCAE is used as the image feature extraction model. The DCAE consists of an encoder and a decoder, and altogether contains five convolutional layers and five transposed convolutional layers, and performs low-dimensional representation of data through a latent space;

[0054] The encoder part is composed of five convolutional layers. Each layer uses a 3×3 convolutional kernel, a stride of 2, and a padding of 1 to achieve layer-by-layer feature extraction and gradual reduction of the spatial dimension;

[0055] The latent space is located between the encoder and the decoder and is the low-dimensional representation of data; the dimension of the latent space is 1024 channels, that is, a total of 4096 features;

[0056] The decoder part is composed of 5 transposed convolutional layers. Its structure is symmetric to that of the encoder, gradually restoring the spatial dimension of the feature map layer by layer, and at the same time gradually reducing the number of channels from 1024 back to 3 to reconstruct an image similar to the original input;

[0057] By comparing the output reconstructed image with the original GAF image, calculate the difference between the input and the output. This difference is called the reconstruction error (RE); the reconstruction error is quantified by the mean squared error (MSE);

[0058] Step 4.3: Calculate the inconsistency metric; the inconsistency metric is used to evaluate the consistency and abnormality of the battery by combining the reconstruction error and the deviation of the time series features; it is obtained by calculating the weighted sum of squares of the reconstruction error and the deviation of the time series features.

[0059] Step 4.4: Identify the inconsistent single cells based on the result of the inconsistency metric; the single cell with a larger inconsistency metric compared to the other batteries is the inconsistent one.

[0060] The described method for detecting and grading battery inconsistency by combining image and time series features

[0061] In Step 4.4, the inconsistency grading is as follows:

[0062] Step 4.4.1: Calculate the cumulative inconsistency metric. Normalize the inconsistency metric based on the current window step count so that the metric value of each battery can adapt to the window step count, thereby realizing the grading of the inconsistency of single cells within the same window. The formula for the cumulative inconsistency metric is as follows:

[0063]

[0064] where C i,j represents the cumulative inconsistency metric of the j-th battery in the i-th window, with a value range greater than or equal to 0, t is the current time step, and M t,j is the inconsistency metric of the j-th battery at time step t, with a value range between [0, 1].

[0065] Step 4.4.2: Conduct an overall inconsistency rating by calculating the weighted average inconsistency score, enabling a comprehensive evaluation of the performance of the batteries in each vehicle. The weighted average inconsistency score is calculated using the following formula:

[0066]

[0067] where W avg represents the weighted average inconsistency score, with a value range between [0, 3.5], I i represents the inconsistency level of the i-th battery in the current window, with a value range between [0, 1], and ω i represents the inconsistency score weight of the i-th battery; the calculation method is as follows:

[0068]

[0069] where x i is the weighting coefficient set according to the inconsistency level of the single cell, with a value range between [0, 1].

[0070] The beneficial effects of the present invention are as follows:

[0071] (1) The combination of Gram Angle Field and Deep Convolutional Autoencoder is applied to inconsistency detection: The innovative combination of GAF and DCAE is applied to battery inconsistency detection, providing a new method for this field.

[0072] (2) The combination of image and time series features: The combination based on image and time series features enhances the extraction and analysis of battery data, improving the accuracy and robustness of inconsistency detection.

[0073] (3) Inconsistency grading strategy: According to the quantified reconstruction error and cumulative inconsistency index, detailed inconsistency grading is performed on individual batteries and battery packs. This provides an accurate assessment of the inconsistency level, providing data-driven support and decision-making guidance for optimizing battery maintenance and management strategies.

[0074] (4) Using real vehicle data for experiments and verification: The proposed method is verified in an actual environment using the actual operation data of various electric vehicle models to ensure the practicality and reliability of the method. Description of the Drawings

[0075] Figure 1 It is a schematic diagram of the overall process of the method of the present invention.

[0076] Figure 2 It is a schematic diagram of the driving mileage of 300 cars included in the dataset used.

[0077] Figure 3 is a schematic diagram of the voltage data of single batteries of six new energy vehicles.

[0078] Figure 4 It is a flowchart of data preprocessing.

[0079] Figure 5 It is a schematic diagram of a sliding window.

[0080] Figure 6 is a graph of the cumulative inconsistency results of a single battery.

[0081] Figure 7 It is a graph of the results of the inconsistency grading of a single battery.

[0082] Figure 8 It is a box plot of the cumulative inconsistency distribution of a single battery.

[0083] Figure 9 It is a comparison graph of the results of the inconsistency grading of vehicle battery packs under different weight coefficients.

[0084] Figure 10 It is a practical application diagram of the method proposed by the invention. Detailed implementation manners

[0085] The following combines the accompanying drawings and embodiments to describe the detailed implementation manners of the present invention in detail. The data set used in the present invention comes from the large-scale publicly available lithium-ion battery data set of new energy vehicles provided by the National Key Laboratory of Intelligent Vehicle Safety Technology (IVST). This data set comes from a total of 300 operating vehicles that have been discontinued, with an operating cycle range of 0.5 - 4 years, covering all driving conditions information such as vehicle charging, discharging, and standing still, Figure 2 is the driving mileage of all vehicles.

[0086] Table 1 Selected vehicle data information

[0087]

[0088] In order to verify the accuracy of the inconsistency detection and grading method proposed in the present invention, this paper randomly selects the charge and discharge voltage data of 6 vehicles from three different mileage segments of low, medium, and high for verification. The detailed information of the selected vehicles is shown in Table 1. Compared with the battery charge and discharge data in the laboratory, it is very difficult to find a complete charge and discharge segment during actual operation. Therefore, this paper selects relatively complete charge and discharge segments from the actual operation data of 6 vehicles. Taking vehicle 2 in Table 1 as an example, the selected charge and discharge segment is 66410.9 - 66661.9 (km). The initial state of charge (SOC) represents the remaining SOC when starting to charge, and the end SOC represents the SOC after the vehicle is charged from the initial SOC to 100% and then starts to discharge until the remaining SOC before the next charge. As shown in Figure 3, (a)-(f) are vehicles numbered 2, 145, 192, 210, 217, and 288 respectively, and it can be seen from the figure that except for vehicle 192 used for comparison, there are problems of under-voltage or over-voltage inconsistencies in the individual cells of the other vehicles.

[0089] Due to potential sensor failures or connection problems, there will be a small number of missing values in the data. However, data with missing values is not conducive to the inconsistency detection of individual cells. Therefore, it is essential to preprocess the data. As Figure 4 shown, in this paper, interpolation is performed on the data with missing values using Newton interpolation method, etc.

[0090] As Figure 5As shown, a fixed step size is selected to slide the window in the present invention. The step size represents the distance by which the window slides each time. The choice of the step size determines the degree of data overlap. A smaller step size will increase the data overlap and may capture the minute changes in the data more meticulously, while a larger step size reduces the data overlap and may lead to information loss. The selection of the sliding window and the step size needs to be balanced between performance and computational efficiency. Considering that the objective in the present invention is online detection, real-time performance becomes a key factor in the selection of the sliding window and the step size. In an online detection scenario, the model must be able to process data in real time and output results promptly. Therefore, the selection of the window size and the step size not only needs to ensure sufficient feature capture accuracy but also needs to ensure the efficiency of the computational process and avoid delays caused by excessive computational burden. Through reasonable settings of the window size and the step size, the requirements for real-time processing can be met on the premise of ensuring accuracy.

[0091] The overall process of the method for detecting and grading the inconsistency of lithium-ion power batteries by fusing image and time-series features is as Figure 1 shown. The specific steps for detecting the inconsistency of this method are as follows:

[0092] Step 1: Data loading and preprocessing,

[0093] Step 1.1: Accept the real-time single-cell voltage data V transmitted by the lithium battery management system celli ;

[0094] Step 1.2: Use Newton interpolation method, Lagrange interpolation method, and linear interpolation method to preprocess the missing data. In the present invention, for the data with missing values, interpolation is performed using Equation (1) Newton interpolation formula, etc.

[0095]

[0096] Among them, P n (x) is the nth interpolation polynomial constructed through the given data points, and f[x0,...,x k is the kth-order divided difference quotient, is the continued product basis function with respect to x.

[0097] Step 2: Initialize the sliding window. Set the window size to W and the step size for each slide to S. Then the position of the sliding window can be expressed as:

[0098] W = {x i , x i+1 ,..., x i+W-1} Equation (2);

[0099] Among them, x i represents the data points in the time series, and W is the window size, indicating the number of data points within each window.

[0100] Step 3: Extraction of the inconsistency characteristics of single cells

[0101] Step 3.1: Calculate the mean, standard deviation, and fluctuation range of each single cell within the current sliding window, and perform time series feature extraction. The specific principle is as follows:

[0102] (1) The mean represents the overall level of the battery voltage and can reflect the working state of the battery over a period of time. If the battery voltage deviates from the normal mean, it may indicate an unstable state or a fault in the battery.

[0103]

[0104] where x i represents the i-th value of the battery voltage data, and N is the total number of data samples;

[0105] (2) The standard deviation describes the degree of fluctuation of the battery voltage and reflects the stability of the battery. Through the standard deviation, the drastic fluctuations of the battery voltage can be captured, and these fluctuations may be early signals of abnormalities;

[0106]

[0107] where σ is the standard deviation of the single cell at the current moment, μ is the mean of the single cell at the current moment, x i represents the i-th value of the battery voltage data, and N is the total number of data samples.

[0108] (3) The fluctuation range refers to the difference between the maximum and minimum values of the battery voltage over a period of time. It can reveal the amplitude change of the battery voltage during this period and reflect whether the battery has experienced large voltage fluctuations during operation. Large-scale voltage fluctuations may be closely related to the health status of the battery, especially when the battery is affected by external environmental changes or aging.

[0109] Range = max(x i ) - min(x i ) Equation (5).

[0110] Step 3.2: Image the time series voltage data through the Gramian Angular Field (GAF). The principle is as follows:

[0111] (1) Data normalization: Normalize the time series x = [x1, x2,..., x n , linearly map the original data x to the interval [0, 1], and obtain the normalized time series x′. The specific formula is:

[0112]

[0113] Among them, min(x) and max(x) are the minimum and maximum values in the time series respectively. x' is the normalized data, and its value range is between [0, 1];

[0114] (2) Calculate the cosine similarity between time points: By calculating the cosine similarity between any two time points x' i and x' j in the normalized time series, a symmetric matrix G of n*n is constructed. Each element G ij of the matrix G is defined as:

[0115]

[0116] Among them, cos(θ ij ) is the angle between i and time point j, reflecting the correlation between each time point in the time series;

[0117] (3) Image generation: By selecting an appropriate color mapping, the values in the symmetric matrix G can be intuitively displayed as colors in the image. The color mapping function can be expressed as:

[0118] f: [0, 1] → RGB Equation (8);

[0119] Among them, RGB represents a coordinate in the color space of the red, green, and blue channels;

[0120] (4) Image channel adjustment: In order to use the generated GAF as the input of the DCAE, the single-channel G matrix needs to be converted into a three-channel image. This process can be achieved by replicating the single-channel matrix three times, thus generating an image with three identical channels. This process can be expressed as:

[0121] G RGB = Repeat(G, axis = 10, time = 3)

[0122] Among them, G RGB represents the expanded three-channel image, with a shape of 3*n*n; The Repeat function repeats the single-channel matrix G three times to obtain a three-channel RGB image.

[0123] Step 4: Calculate the inconsistency measure based on the anomaly score obtained from the time series data and the reconstruction error obtained from the image data;

[0124] Step 4.1: For each single-cell battery, calculate its anomaly score S on the time series characteristics. This score is measured by calculating the normalized difference between the time series characteristics of each battery and the mean of the time series characteristics of all batteries. The specific calculation formula is as follows:

[0125]

[0126] Among them, F k,i represents the k-th time series feature of the i-th battery, and u' k and σ' k are the mean and standard deviation of the k-th feature respectively.

[0127] Step 4.2: In the present invention, DCAE is adopted as the image feature extraction model. DCAE consists of an encoder and a decoder, and altogether includes five convolutional layers and five deconvolutional layers, and performs low-dimensional representation of data through a latent space.

[0128] The encoder part is composed of five convolutional layers. Each layer uses a 3×3 convolutional kernel, a stride of 2, and a padding of 1 to achieve layer-by-layer feature extraction and gradual reduction of the spatial dimension. Specifically:

[0129] The first convolutional layer converts the input 3-channel GAF image into 64 channels. The size of the feature map is reduced from 50×50 to 25×25, and nonlinearity is introduced through the ReLU activation function to preliminarily extract the basic features of the input data;

[0130] The second convolutional layer converts 64 channels into 128 channels, and the size of the feature map is further reduced to 13×13, and the ReLU activation function is applied;

[0131] The third convolutional layer converts 128 channels into 256 channels, and the size of the feature map is reduced to 7×7, and continues to be processed through the ReLU activation function;

[0132] The fourth convolutional layer converts 256 channels into 512 channels, and the size of the feature map is reduced to 4×4, and the activation function is still ReLU;

[0133] The fifth convolutional layer converts 512 channels into 1024 channels, and the size of the feature map is reduced to 2×2, and the ReLU activation function is applied to complete the feature extraction process of the encoder.

[0134] The convolution operation is expressed as:

[0135] Z (l) = ReLU(W (l) * Z (l-1) + b (l) ) Equation (10);

[0136] Among them, Z (l-1) is the input GAF of the (l-1)-th layer, W (l) and b (l) are the weight and bias of the l-th convolutional layer respectively, * represents a two-dimensional convolution operation, and ReLU is the activation function.

[0137] The Latent Space is located between the encoder and the decoder and is a low-dimensional representation of the data. In this paper, the dimension of the Latent Space is 1024 channels, that is, a total of 4096 features. This setting aims to ensure that the Latent Space can fully express the complex features of the input data while maintaining an appropriate compression ratio to balance the feature expression ability and computational efficiency.

[0138] The decoder part consists of 5 Transposed Convolutional Layers, whose structure is symmetric to that of the encoder. It gradually restores the spatial dimension of the feature map layer by layer, while gradually reducing the number of channels from 1024 back to 3 to reconstruct an image similar to the original input. Specifically:

[0139] The first transposed convolutional layer restores the 1024-channel latent representation to 512 channels, expands the feature map size from 2×2 to 4×4, and passes through the ReLU activation function;

[0140] The second transposed convolutional layer reduces the 512 channels to 256 channels, expands the feature map size to 7×7, and continues to apply the ReLU activation function;

[0141] The third transposed convolutional layer reduces the 256 channels to 128 channels, restores the feature map size to 13×13, and processes it through the ReLU activation function;

[0142] The fourth transposed convolutional layer reduces the 128 channels to 64 channels, expands the feature map size to 25×25, and applies the ReLU activation function;

[0143] The fifth transposed convolutional layer restores the number of channels to 3 through a 3×3 convolutional kernel, a stride of 2, a padding of 1, and no additional output padding. The feature map size reaches 50×50, and the output is limited to the range [0,1] through the Sigmoid activation function to ensure that the pixel values of the reconstructed image are consistent with those of the input image.

[0144] The transposed convolution operation is expressed as:

[0145] Z (l) =ReLU(W (l) *Z (l+1) +b (l) ) Equation (11);

[0146] where Z (l+1) is the input GAF of the (l + 1)-th layer.

[0147] For the last transposed convolutional layer, the settings are as follows:

[0148] Y=Sigmoid(W (5) *Z (5) +b(5) ) Equation (12);

[0149] where Y is the finally reconstructed output image, W (5) and b (5) are the convolutional kernel weights and bias parameters of the fifth transposed convolutional layer respectively, * represents the transposed convolution operation, and Sigmoid is the activation function;

[0150] By comparing the output reconstructed image with the original GAF image, the difference between the input and the output can be calculated, and this difference is called the reconstruction error (RE). The reconstruction error is quantified by the mean squared error (MSE), and the specific formula is as follows:

[0151]

[0152] where N is the total number of pixels of the input image, x j and represent the values of the input image and the reconstructed image at the jth pixel respectively.

[0153] Step 4.3: Calculate the inconsistency metric. The inconsistency metric is used to evaluate the consistency and abnormality of the battery by combining the reconstruction error and the deviation degree of the temporal characteristics. It is obtained by calculating the weighted sum of squares of the reconstruction error and the deviation degree of the temporal characteristics. The specific calculation formula is as follows:

[0154]

[0155] where M i is the inconsistency metric, with a value range between [0, 1], RE i is the reconstruction error, S i is the anomaly score. When the inconsistency metric of a certain battery is relatively large compared with the other batteries, it indicates that the performance of this battery in data reconstruction and temporal characteristics is quite different from that of other single cells, so it is considered that this battery has inconsistency.

[0156] Step 4.4: Identify the single cells with inconsistency through the results of the inconsistency metric. The single cells with an inconsistency metric larger than the other batteries are the single cells with inconsistency.

[0157] The specific steps for grading the inconsistency of this method are as follows:

[0158] Step 1: Calculate the cumulative inconsistency metric, normalize the inconsistency metric based on the current window step, so that the metric value of each battery can adapt to the window step, thereby realizing the grading of the inconsistency of single cells within the same window. The formula for the cumulative inconsistency metric is as follows:

[0159]

[0160] Among them, C i,j represents the cumulative inconsistency metric of the j-th battery in the i-th window, with a value range greater than or equal to 0. t is the current time step, and M t,j is the inconsistency metric corresponding to the j-th battery at time step t, with a value range between [0, 1].

[0161] Step 2: Conduct an overall inconsistency rating by calculating the Weighted Average Inconsistency, enabling a comprehensive evaluation of the performance of each vehicle's battery. The weighted average inconsistency score is calculated using the following formula:

[0162]

[0163] Among them, W avg represents the weighted average inconsistency score, with a value range between [0, 3.5]. I i represents the inconsistency level of the i-th battery in the current window, with a value range between [0, 1]. ω i represents the inconsistency score weight of the i-th battery. The calculation method is as follows:

[0164]

[0165] Among them, x i is the weighting coefficient set according to the inconsistency level of the single battery, with a value range between [0, 1].

[0166] The following combines an example using the single-cell voltage data of 6 pure electric passenger vehicles to verify the accuracy and reliability of the lithium-ion power battery inconsistency detection and grading method that fuses image and time-series features. By measuring the inconsistency of single cells, the present invention can quantify the inconsistency degree of each battery in the battery pack. Considering that during the long-term use of the battery, its inconsistency may gradually deepen over time, we calculated the cumulative inconsistency metric of single battery i and showed the trend of these metrics changing with the number of sliding times in Figure 6. The cumulative inconsistency metric is obtained by accumulating and normalizing the changes in the inconsistency metric in different windows. (a)-(f) in Figure 6 show the cumulative inconsistency scores of different batteries in different time windows. It can be seen from the figure that the cumulative inconsistency metric of the battery changes with the increase in the number of sliding times, and the differences between different batteries are significant. These differences reveal the degree of voltage fluctuation of the battery under the same test conditions.

[0167] By comparing Figure 6 and Figure 2 ,Figure 2 Batteries that exhibit significant fluctuations in the original voltage data also show significantly different cumulative inconsistency metrics from the rest of the batteries in Figure 6. In particular, in Figure 6(a), the cumulative inconsistency metric of cell 83 is high, indicating that the performance of this battery fluctuates significantly over multiple time periods and exhibits relatively obvious anomalies. In Figures 6(b) and (e), the driving mileage of vehicles 145 and 217 is in the range of 1276.6 km to 1396.6 km and 9544.5 km to 9775.2 km respectively, indicating that the vehicles are in the early operating state. However, it can still be clearly observed from the figure that cells 25, 63, and 82 also show relatively high cumulative inconsistency metrics, further verifying the effectiveness of this method for detecting early inconsistent batteries. It should be noted that in Figure 6(c), since the inconsistency levels of the individual cells of this vehicle are all relatively high, no significantly inconsistent outlier individual cells are found. This phenomenon indicates that the inconsistency metric can effectively capture the fluctuations in battery performance over different time periods and can make precise evaluations based on these changing trends. In addition, cell 82 in Figure 6(e) and cell 33 in Figure 6(f) show a decrease in the cumulative inconsistency metric. This phenomenon indicates that although some batteries show an increase in the inconsistency metric in the current window, the overall inconsistency metric of these batteries has decreased. This change highlights that the inconsistency metric considers the relative performance of all batteries rather than the absolute fluctuations of individual batteries. Specifically, although some batteries show significant fluctuations, the overall performance of the battery pack is relatively stable, and the cumulative inconsistency metric of individual cells will show a downward trend.

[0168] By comparing the results of the original voltage data with the cumulative inconsistency metric, we can effectively identify the batteries that exhibit significant fluctuations and anomalies during the test. The cumulative inconsistency metric not only provides a quantitative assessment of the battery voltage fluctuations but also provides a strong basis for subsequent battery performance optimization. This method enables the quantification of the inconsistency of each battery in the battery pack, thus providing more accurate and reliable decision-making support for the BMS.

[0169] The present invention compares the calculation times of different sliding step sizes and calculation window sizes in the inconsistency detection of six vehicles, and focuses on analyzing the calculation time within each calculation window to select the best combination of window and step size. According to the data in Table 2, when the sliding step size is 15 and the calculation window size is 50, the lowest calculation time is exhibited among all vehicles, and this configuration shows consistent efficiency for different vehicles. Specifically, for vehicles 145, 192, 210, 217, and 288, the calculation times are from 0.0428 seconds to 0.0435 seconds, while for vehicle 288, the calculation time is 0.0458 seconds. This setting is significantly superior to the combinations of the remaining step sizes and calculation windows.

[0170] Table 2 Comparison results of different sliding step sizes and calculation windows

[0171]

[0172] This result indicates that the combination of a sliding step size of 15 and a calculation window of 50 can achieve the shortest calculation time within a single calculation window, which is particularly important for applications with high real-time requirements in online detection systems. On the contrary, when the calculation window size increases, the calculation time increases significantly, which may affect the response speed of real-time detection. Therefore, based on the experimental results of the present invention, in practical applications, selecting a configuration with a sliding step size of 15 and a calculation window of 50 has the shortest calculation time within a single calculation window and can effectively meet the real-time requirements in online detection.

[0173] To facilitate the verification of the practicality of the method proposed in the present invention, vehicle 2 is selected as an example, and the cumulative inconsistency metric under the current window is calculated every 50 sliding windows, and the inconsistency rating of each single battery at each window position is performed to reveal the voltage fluctuations of the battery during long-term operation. Figure 7 Shows the cumulative inconsistency metric of single batteries in vehicle 2 at different numbers of sliding windows (N sw ). As the number of window slides increases, the inconsistency performance of each battery gradually emerges. Especially for battery 83, its inconsistency metric continuously increases during the entire evaluation process, indicating that the performance of this battery fluctuates greatly and shows obvious inconsistencies in multiple time periods.

[0174] In Figure 7 (a), when the number of sliding windows N sw = 50, the cumulative inconsistency metric of battery 83 has reached 0.98, showing a significant fluctuation trend. As the number of slides increases, in Figure 7 (b) and (c), the cumulative inconsistency metric of this battery rises to 0.99, indicating that the degree of its inconsistency further intensifies. In Figure 7Within the window range from (d) to (e), the cumulative inconsistency metric of battery 83 remains at a relatively high level and reaches 1.00 at N sw = 300, indicating that the inconsistency within the entire time window has reached the highest level. Further observation of Figure 7 (g) to (j) reveals that the cumulative inconsistency metric of battery 83 always remains at 1.00, showing continuous and stable high inconsistency. This trend indicates that over time, the inconsistency phenomenon of this battery gradually intensifies and maintains high volatility over a long period, showing significant anomalies compared to other batteries. In contrast, the inconsistency metrics of other batteries are relatively low at different window numbers and do not show significant fluctuation trends, indicating that the operating states of the remaining single cells are relatively stable without obvious inconsistency changes. Additionally, from the overall distribution, the inconsistency ratings of each battery gradually tend to be stable, but the inconsistency scores of individual batteries may still fluctuate over time. This phenomenon indicates that the inconsistency metric can not only reveal the performance fluctuations of batteries at different time windows but also quantify the long-term voltage inconsistency trend of individual batteries. The cumulative inconsistency metrics of the remaining vehicles with high-inconsistency single cells within each window are shown in Table 3, and the blanks in the table are due to different initial data lengths.

[0175] Table 3 Cumulative Inconsistency Metrics of Different Vehicles within Different Windows

[0176]

[0177] To better monitor the inconsistency degree of single cells, the present invention adopts a grading method based on the cumulative inconsistency metric, aiming to timely detect potential problems of the battery and take corresponding measures to ensure the safety and performance stability of the battery pack. Therefore, we set three intervals to divide the health status of the battery: when the cumulative inconsistency metric is less than 0.6, the battery is considered to be in a normal state; when the metric value is between 0.6 and 0.8, it indicates mild inconsistency; and when the metric value is between 0.8 and 1.0, it indicates severe inconsistency. The setting of these thresholds is based on an in-depth analysis of the cumulative inconsistency metrics of the last window of six vehicles and is verified using Figure 8 box plots.

[0178] Figure 8A box plot showing the cumulative inconsistency metrics of six vehicles, with the set thresholds marked. The box plot can help us intuitively understand the health status of the batteries of different vehicles and clearly distinguish the levels of battery inconsistency according to the set thresholds. Each box in the figure represents the interquartile range of the data for that vehicle, showing the distribution of the middle 50% of the data. The line in the box is the median. The whiskers represent the range of the data, usually extending to the maximum and minimum data points outside 1.5 times the interquartile range. The green dashed line represents the mild inconsistency threshold, while the red dashed line represents the severe inconsistency threshold. As can be seen from the figure, the cumulative inconsistency metrics of most single cells are below 0.6, indicating that the performance of these batteries is still within the normal range and no obvious problems have occurred. However, the metrics of some single cells exceed 0.6 and enter the mild inconsistency range (0.6 to 0.8). In addition, in vehicles 2, 210, 217, and 288, the cumulative inconsistency metric of one battery significantly exceeds 0.8, reaching severe inconsistency, indicating that these batteries may have suffered serious failures, affecting the normal operation of the vehicle, and immediate inspection and repair measures must be taken. In addition, although there are no single cells in vehicle 192 that enter the severe inconsistency range, the range whiskers of its battery pack have exceeded the mild threshold line, indicating that the overall performance of the battery pack of this vehicle is poor and the cumulative inconsistency of some single cells is high, which can also be verified in Figure 8 as well.

[0179] Table 4 presents the distribution of single cells in each vehicle under different inconsistency ranges. Through the statistical analysis of these data, the present invention can further understand the status of the batteries of each vehicle and provide a basis for subsequent battery management and maintenance. These data not only help us verify the previously set inconsistency levels, but also contribute to the monitoring of battery inconsistency and the formulation of preventive measures.

[0180] Table 4 Inconsistency Level Single Cell Statistics

[0181]

[0182] The present invention conducts an overall inconsistency rating on the batteries of different vehicles, aiming to evaluate the performance consistency of batteries of the same model on different vehicles. Although the batteries adopt the same specifications and designs, environmental factors, driving behaviors, charge-discharge modes, etc. in actual applications can lead to differences in the performance of the batteries of each vehicle. By uniformly rating the inconsistency of batteries of the same model on different vehicles, these differences can be effectively revealed, and a basis for cross-vehicle battery performance comparison can be provided. To ensure the accuracy of the evaluation results, the present invention adopts a weighted average inconsistency scoring method. This method not only considers the performance fluctuations of each battery cell but also combines its weight to reflect the importance of different batteries in the overall evaluation. By allocating weights according to the battery performance fluctuations and abnormal fluctuations, it is ensured that in the evaluation process, the impact of inconsistent batteries on the score is fully reflected.

[0183] As Figure 9 shown, the present invention uses 4 sets of different weighting coefficients to compare the rationality of the overall inconsistency scores on different vehicles. During the weighting process, batteries with severe inconsistency are given higher weights, while batteries with mild inconsistency and normal batteries are given lower weights. The impact of different coefficient sets on the score is significantly different. Setting smaller coefficients (0.5, 1, 1.5 and 1, 1.5, 2) fails to effectively highlight the performance differences between batteries and may not effectively amplify the impact of inconsistent batteries, reducing the evaluation sensitivity. Setting moderate coefficients (1, 2, 3) balances the battery performance differences and the weighting effect and can steadily reflect the differences in vehicle batteries. While larger coefficients (1.5, 2.5, 3.5) can highlight battery differences, especially for batteries with large performance fluctuations, but may cause the minor fluctuations of batteries with consistent performance to be over-amplified, affecting the stability of the evaluation.

[0184] Therefore, based on Figure 9 the analysis of the inconsistency degree of single vehicle batteries, the present invention selects a moderate coefficient set (1, 2, 3) to balance the battery volatility and inconsistency, ensuring the scientific nature and stability of the evaluation results. This weighted average inconsistency scoring method provides a solid basis for battery performance optimization and comparison of batteries of the same model on different vehicles, and effectively avoids the deviation impact of the performance of a single battery on the results.

[0185] The comparison between the inconsistency detection and grading results of the true single cell voltage data of 6 vehicles and the actual battery voltage verification results shows the wide applicability of the present invention under different vehicles and conditions, providing valuable insights for further analysis and actions. This confirms the accuracy and reliability of the present invention and highlights the practical advantages of real-time evaluation. In addition, the grading strategy for inconsistent single cell batteries and battery packs helps vehicle owners timely understand the status of vehicle batteries and replace them immediately when severe inconsistencies occur in the batteries, ensuring driving safety.

[0186] In view of the difficulties in early battery inconsistency detection and the limitations of analyzing inconsistency with a single feature, this study proposes an inconsistency detection method that combines image and temporal features. First, the present invention interpolates the voltage data of real vehicles to avoid the influence of data missing on the evaluation results. Then, by visualizing the temporal voltage data through GAF to extract features, the inconsistency of single-cell and module batteries is deeply analyzed in combination with the temporal data. The proposed method is verified for effectiveness and reliability through the voltage data of real vehicles with different driving mileages. Meanwhile, the classification of the inconsistency degree of single-cell and module batteries is discussed, providing decision support for subsequent battery maintenance and replacement. The summary is as follows: 1) The battery data of vehicles in the early, middle, and late stages are respectively selected for comparison and analysis to verify the effectiveness and feasibility of the proposed method. 2) Through experimental comparison, it is determined that the best calculation speed is achieved when the window size is 50 and the sliding step is 15, ensuring the real-time performance of online detection. 3) The study classifies the inconsistent batteries, providing a reference for the operation safety of actual vehicle batteries.

[0187] Figure 10 The actual application scenario of the method proposed by the present invention is shown. In this system, the vehicle exchanges data with other vehicles, traffic lights, and base stations through wireless communication. The BMS on each vehicle is responsible for monitoring the battery health status and calculating the consistency of battery performance. The GAF conversion is only carried out within the current time window, and the image is deleted after the calculation is completed, thus reducing the required computing resources and enabling real-time analysis in an in-vehicle system under the Python environment. The battery data is transmitted between vehicles through a wireless network to ensure that each vehicle can synchronously update the battery status of each other. When battery inconsistency is detected, this information is transmitted to the cloud through the base station for further analysis and processing, thereby diagnosing the health status of the battery system and optimizing the battery management strategy. In addition, the system can dynamically adjust the sliding step and calculation window according to the real-time data uploaded by different vehicle models, thereby improving the real-time performance and adaptability of the evaluation algorithm and ensuring the safety of new energy vehicles during driving.

[0188] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, which all fall within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalent technical solutions thereof.

Claims

1. A method for detecting and grading battery inconsistency by combining image and temporal features, characterized in that, It includes two parts: inconsistency detection and inconsistency grading; Among them, the steps of the inconsistency detection part are as follows: Step 1: Data loading and preprocessing, Step 1.1: Receive the real-time single-cell voltage data transmitted by the lithium battery management system Step 1.2: Use one of Newton interpolation method, Lagrange interpolation method, and linear interpolation method to perform data preprocessing on the missing data; Step 2: Initialize the sliding window W, Step 3: Extract the inconsistency features of single cells, Step 3.1: Calculate the mean, standard deviation, and fluctuation range of each single cell within the current sliding window to extract time series features; Step 3.2: Imageize the voltage data through Gramian Angular Field (GAF), Step 4: Calculate the inconsistency metric through the anomaly score obtained from the voltage data and the reconstruction error obtained from the image data, Step 4.1: For each single cell, calculate its anomaly score S on the time series features; Step 4.2: Use a Deep Convolutional Autoencoder (DCAE) as the image feature extraction model; Step 4.3: Calculate the inconsistency metric; Step 4.4: Identify the single cells with inconsistency through the results of the inconsistency metric. The single cells with a larger inconsistency metric compared to the other cells are the ones with inconsistency; Among them, the steps of the inconsistency grading part are as follows: Step B1: Calculate the cumulative inconsistency metric, Step B2: Conduct an overall inconsistency rating by calculating the weighted average inconsistency score.

2. The battery inconsistency detection and grading method combining image and timing features according to claim 1, characterized in that The described inconsistency detection is as follows: Step 1.2: Interpolate the data with missing values using Equation (1), Newton interpolation formula, etc. Among them, P n (x) is an nth-degree interpolation polynomial constructed through given data points, f[x0,...,x k is the kth-order divided difference quotient, is the continued product basis function with respect to x.

3. A method for detecting and grading battery inconsistency combining image and time series features according to claim 2, characterized in that Step 2: Initialize the sliding window, set the window size to W, and the step size for each slide to S. Then the position of the sliding window can be expressed as: W = {x i , x i+1 ,..., x i+W-1} Equation (2); Among them, x i represents a data point in the time series, and W is the window size, representing the number of data points within each window.

4. A method for detecting and grading battery inconsistency combining image and time series features according to claim 3, characterized in that Step 3.1: Calculate the mean, standard deviation, and fluctuation range of each single cell within the current sliding window to extract time series features: The described mean represents the overall level of the battery voltage and reflects the working state of the battery over a period of time; where x i represents the i-th value of the battery voltage data, and N is the total number of data samples; The described standard deviation describes the degree of voltage fluctuation of the battery and reflects the stability of the battery; the abnormal violent fluctuation of the battery voltage is captured through the standard deviation; where σ is the standard deviation of the single cell at the current moment, μ is the mean value of the single cell at the current moment, and x i represents the i-th value of the battery voltage data, and N is the total number of data samples; the described fluctuation range refers to the difference between the maximum and minimum values of the battery voltage over a period of time; Range = max(x i ) - min(x i ) Equation (5); Step 3.2: Imageize the time series voltage data through Gramian Angular Field (GAF): 3.2.1 Data normalization: Normalize the time series \(x = [x_1, x_2,\cdots,x n \) to linearly map the original data \(x\) to the interval \([0, 1]\) to obtain the normalized time series \(x'\). The formula is as follows: Among them, min(x) and max(x) are the minimum and maximum values in the time series respectively, and x′ is the normalized data, with a value range between [0, 1]; 3.2.2 Calculate the cosine similarity between time points: By calculating the cosine similarity between any two time points x i ′ and x′ j in the normalized time series, construct an n*n symmetric matrix G; each element G ij of the matrix G is defined as: where, cos(θ ij ) is the angle between i and time point j, reflecting the correlation between each time point in the time series; 3.2.3 Image Generation: By selecting an appropriate color mapping, the values in the symmetric matrix G are visually presented as colors in an image; the color mapping function is expressed as: f: [0, 1] → RGB Equation (8); where RGB represents a coordinate in the color space of the red, green, and blue channels; 3.2.4 Image Channel Adjustment: To use the generated GAF as the input of the DCAE, the single-channel G matrix is converted into a three-channel image; this process is achieved by replicating the single-channel matrix three times, thus generating an image with three identical channels; this process is expressed as: G RGB = Repeat(G, axis = 10, time = 3) Among them, G RGB represents the expanded three-channel image with a shape of 3*n*n; the Repeat function repeats the single-channel matrix G three times to obtain a three-channel RGB image.

5. A method for detecting and grading battery inconsistency combining image and temporal features according to claim 4, characterized in that Step 4.1: For each single cell, calculate the anomaly score S of its temporal features, which is measured by calculating the normalized difference between the temporal features of each cell and the mean of the temporal features of all cells; Step 4.2: The DCAE is used as the image feature extraction model. The DCAE consists of an encoder and a decoder, and altogether contains five convolutional layers and five transposed convolutional layers, and performs low-dimensional representation of data through a latent space; The encoder part is composed of five convolutional layers. Each layer uses a 3×3 convolutional kernel, a stride of 2, and a padding of 1 to achieve gradual extraction of features and gradual reduction of the spatial dimension; The latent space is located between the encoder and the decoder and is the low-dimensional representation of data; the dimension of the latent space is 1024 channels, that is, a total of 4096 features; The decoder part is composed of 5 transposed convolutional layers. Its structure is symmetric to that of the encoder, gradually restoring the spatial dimension of the feature map layer by layer, and at the same time gradually reducing the number of channels from 1024 back to 3 to reconstruct an image similar to the original input; By comparing the output reconstructed image with the original GAF image, calculate the difference between the input and the output, and this difference is called the reconstruction error (RE); the reconstruction error is quantified by the mean squared error (MSE); Step 4.3: Calculate the inconsistency metric; the inconsistency metric is used to evaluate the consistency and anomaly of the battery by combining the reconstruction error and the deviation degree of the temporal features; it is obtained by calculating the weighted sum of squares of the reconstruction error and the deviation degree of the temporal features; Step 4.4: Determine the single cells with inconsistency through the result of the inconsistency metric. The single cells with an inconsistency metric larger than that of the other cells are the single cells with inconsistency.

6. A method for detecting and grading battery inconsistency combining image and temporal features according to claim 5, characterized in that In step 4.4, The inconsistency grading is as follows: Step 4.4.1: Calculate the cumulative inconsistency metric, and normalize the inconsistency metric based on the current window step number so that the metric value of each battery cell can adapt to the window step number, thereby realizing the grading of the inconsistency of single battery cells within the same window; the formula for the cumulative inconsistency metric is as follows: Among them, C i,j represents the cumulative inconsistency metric of the j-th battery in the i-th window, and its value range is greater than or equal to 0. t is the current time step, and M t,j is the inconsistency metric of the j-th battery corresponding to the time step t, and its value range is between [0, 1]; Step 4.4.2: Conduct an overall inconsistency rating by calculating the Weighted Average Inconsistency, so that the performance of the vehicle batteries is comprehensively evaluated; the Weighted Average Inconsistency is calculated by the following formula: Among them, W avg represents the weighted average inconsistency score, and its value range is between [0, 3.5]. I i represents the inconsistency level of the i-th battery cell in the current window, and its value range is between [0, 1]. ω i represents the inconsistency score weight of the i-th battery cell; the calculation method is as follows: Among them, x i is a weighting coefficient set according to the inconsistency level of the single cells, and its value range is between [0, 1].

Images