Data-driven lithium-ion battery internal short circuit fault diagnosis method

Abstract

This application discloses a data-driven method for diagnosing internal short-circuit faults in lithium-ion batteries, comprising: constructing a simulation model of internal short-circuit faults in lithium-ion batteries; generating external characteristic data of the battery containing internal short-circuit faults through simulation, and constructing a training dataset based on the external characteristic data and actual battery operating data; acquiring voltage time-series data of the battery during the relaxation phase; converting the voltage time-series data into two-dimensional image data using the Gram angle field algorithm, and extracting the fault features corresponding to the two-dimensional image data; projecting the fault features into multiple different representation subspaces through a multi-head attention mechanism, and outputting the estimated internal short-circuit resistance. This data-driven method for diagnosing internal short-circuit faults elevates the fault discrimination mechanism of the battery management system from over-limit alarms based on fixed thresholds to a method architecture capable of actively perceiving early, subtle fault characteristics.

Description

Technical Field

[0001] This application relates to the field of battery short-circuit diagnosis, specifically to a data-driven method for diagnosing internal short-circuit faults in lithium-ion batteries. Background Technology

[0002] Lithium-ion batteries boast advantages such as environmental friendliness, high specific energy output, and high efficiency, making them widely used in energy storage power stations. However, as an energy storage device involving complex electrochemical reactions, lithium-ion batteries pose significant safety risks. On one hand, overcharging, over-discharging, and overheating during actual operation can lead to abnormal performance degradation and, in severe cases, internal short circuits (ISCs) and other safety hazards. On the other hand, to meet the requirements of different application scenarios, multiple battery cells are often connected in series or parallel to form battery packs, resulting in a large number of connected components, increasing system complexity and the probability of various battery system failures. Lithium-ion batteries have a low tolerance for safety hazards; therefore, early detection and identification of any potential faults are crucial. Ensuring safe operation is the primary concern for lithium-ion batteries in energy storage systems.

[0003] Although existing battery management systems (BMS) have the ability to alarm for exceeding limits and handle various significant faults, they still suffer from overly simplistic diagnostic criteria, incomplete range of measurable fault types, and poor robustness. For example, the aging mechanism of individual cells within a lithium-ion battery pack is complex, exhibiting high nonlinearity and uncertainty during charge and discharge operation, influenced by numerous factors. Furthermore, the measurement system in the BMS has inherent errors, making even minor faults within the battery system highly concealed. Early diagnosis of processes such as internal short circuits evolving into battery thermal runaway remains a challenge.

[0004] Traditional short-circuit fault diagnosis methods mainly include model-based methods and data analysis methods. Under normal fault-free conditions, the residual should be close to zero or fluctuate within a very small noise range. When an internal short circuit occurs, the actual behavior of the battery deviates from the battery model, causing a significant increase in the residual between the model output and the actual measured parameters or state, or the appearance of specific change patterns. Fault alarms can be triggered by setting thresholds or analyzing the statistical characteristics of the residuals. However, lithium-ion battery systems often have tens of thousands of individual cells, and applying a model to each cell would result in a massive computational burden. Signal analysis-based diagnostic methods directly analyze external signals measurable by the BMS (mainly voltage, current, and temperature), detecting ISC by observing abnormal changes in these signals in the time domain, frequency domain, or statistical characteristics, without needing to establish complex internal physical or chemical models of the battery. However, signal analysis methods are susceptible to interference from various factors, have a high false alarm rate, and their threshold settings are difficult to adapt to the dynamic changes in the state of lithium-ion batteries. Summary of the Invention

[0005] To address the aforementioned issues, this application proposes a data-driven method for diagnosing internal short-circuit faults in lithium-ion batteries, including: A simulation model of internal short-circuit fault in lithium-ion battery was constructed, and the corresponding model parameters were determined based on battery temperature and state of charge. The lithium-ion battery internal short-circuit fault simulation model generates external characteristic data of the battery containing internal short-circuit faults, and a training dataset is formed based on the external characteristic data of the battery and the actual battery operation data; wherein, the training dataset is used to train a deep learning model, the deep learning model including a two-dimensional convolutional neural network and a multi-head attention mechanism. Obtain the voltage time series data of the battery during the relaxation phase, and extract the initial features related to the internal short circuit fault based on the voltage time series data; The voltage time series data is converted into two-dimensional image data using the Gram angle field algorithm, and the two-dimensional image data is processed by the two-dimensional convolutional neural network. Multiple learnable convolutional kernels are used to slide and scan on the two-dimensional plane of the two-dimensional image data to extract the fault features corresponding to the two-dimensional image data. The fault features are projected into multiple different representation subspaces using the multi-head attention mechanism, and the estimated internal short-circuit resistance is output.

[0006] In one example, a simulation model of an internal short-circuit fault in a lithium-ion battery is constructed, and the corresponding model parameters are determined based on the battery temperature and state of charge, specifically including: Based on the principles of battery electrochemistry and equivalent circuit theory, a coupled model describing the normal state and internal short-circuit state of a battery is constructed as a simulation model for internal short-circuit faults in lithium-ion batteries. The simulation model for internal short-circuit faults in lithium-ion batteries includes a parallel equivalent circuit model and an ISC branch. The model parameters in the simulation model of internal short circuit fault of the lithium-ion battery are set based on the function related to battery temperature and state of charge.

[0007] In one example, the lithium-ion battery internal short-circuit fault simulation model is used to simulate and generate battery external characteristic data that includes internal short-circuit faults, specifically including: In the simulation environment, different degrees of internal short-circuit faults are simulated by dynamically adjusting the corresponding short-circuit resistance parameters. Based on the internal short-circuit fault, the simulation model of the internal short-circuit fault of the lithium-ion battery is run to output the external characteristic data of the battery under various standard operating conditions, different ambient temperatures, and initial states of charge; the external characteristic data of the battery includes the time series of voltage, current, and battery temperature.

[0008] In one example, voltage time-series data of the battery during the relaxation phase is obtained, and initial features related to internal short-circuit faults are extracted based on the voltage time-series data, specifically including: Based on the battery management system, if the load current drops to a preset threshold and continues for more than a preset duration, it is determined that the system has entered the relaxation phase, and the corresponding voltage time series data is collected. The voltage time series data is subjected to moving average filtering, and initial features related to internal short-circuit faults are extracted; the initial features include initial voltage drop, relaxation time constant, energy at a specific time period, and inflection point curvature of the voltage curve.

[0009] In one example, the voltage time series data is converted into two-dimensional image data using the Gram angle field algorithm, specifically including: For the voltage time series data, perform minimum-maximum normalization; Each point in the normalized voltage time series data is treated as a single angle, and the point is mapped to polar coordinate space using the inverse cosine function; The cosine value of the sum of the angles between two points is determined using the Gram angle field algorithm, and an image matrix is ​​generated.

[0010] In one example, the two-dimensional image data is processed by the two-dimensional convolutional neural network. Multiple learnable convolutional kernels slide across the two-dimensional plane of the image data to extract fault features corresponding to the two-dimensional image data. Specifically, this includes: The two-dimensional convolutional neural network is defined to include four convolutional blocks; wherein, a single convolutional block includes two consecutive convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer; the number of convolutional kernels corresponding to each convolutional block increases sequentially; The high-dimensional feature vector corresponding to the two-dimensional image data is extracted by sliding scan on the two-dimensional plane of the two-dimensional image data through the convolution kernel in the two-dimensional convolutional neural network, and is used as the fault feature.

[0011] In one example, the fault features are projected into multiple different representation subspaces using the multi-head attention mechanism to output an estimated internal short-circuit resistance, specifically including: The fault characteristics are reshaped into corresponding feature sequences; The feature sequence is processed through the multi-head attention mechanism, which linearly projects the feature sequence into the corresponding query matrix, key matrix, and value matrix, and performs parallel scaling dot product attention on multiple corresponding heads, and transforms it into the corresponding probability distribution. The probability distributions output by each head are concatenated and linearly projected to obtain the final feature enhanced with global context information. The final feature is input into the regressor to obtain the estimated internal short-circuit resistance.

[0012] The data-driven method for diagnosing internal short-circuit faults in lithium-ion batteries proposed in this application can bring the following benefits: 1. A data-driven internal short-circuit fault diagnosis method is proposed, which elevates the fault discrimination mechanism of the battery management system from over-limit alarms based on fixed thresholds to a methodological architecture capable of proactively sensing early, subtle fault characteristics. This not only expands the fault diagnosis functional boundaries of the battery management system but also provides key state input parameters for redefining system-level safety boundaries, predicting lifespan degradation trajectories, and optimizing energy management strategies. It serves as an important technical support for building the next-generation intelligent battery management system.

[0013] 2. A simulation model of internal short-circuit fault in lithium-ion batteries is constructed. By introducing battery temperature and state of charge as control variables for model parameters, the mapping relationship between key parameters and operating conditions is established. By adjusting the parameter values ​​of the short-circuit resistance branch in the model, the external response characteristics of the battery under different fault levels are simulated, realizing data augmentation under fault conditions and providing a large-scale labeled sample set covering multiple operating conditions and multiple fault levels for subsequent data-driven model training and generalization.

[0014] 3. The terminal voltage response curve during the battery relaxation stage is selected as the analysis object. Leveraging the fact that this stage's signal is unaffected by charging protocols and load disturbances, and is easily obtained from the battery management system, manual statistical features with physical correlation to internal short-circuit faults are extracted. Simultaneously, the Gram angle field algorithm is used to encode the relaxation voltage time series into a two-dimensional image. While fully preserving temporal dependencies, image domain features with high-dimensional, multi-level expressive capabilities are constructed, providing structured input for the deep learning model.

[0015] 4. The constructed deep learning model is composed of a two-dimensional convolutional neural network and a multi-head attention mechanism. The two-dimensional convolutional neural network is responsible for extracting local spatial hierarchical features in the image, while the multi-head attention mechanism is responsible for modeling the long-range dependencies of feature sequences and selectively enhancing key information. The sequential fusion of the two achieves complementary functional coupling between feature extraction and global context modeling, thereby improving the network's quantitative regression accuracy for fault resistance and enabling the model to differentiate important historical information in the feature sequence. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the data-driven internal short-circuit fault diagnosis method for lithium-ion batteries in this application embodiment. Figure 2 This is a schematic diagram of the circuit structure of a simulation model of an internal short-circuit fault in a lithium-ion battery under one scenario in an embodiment of this application. Figure 3 This is a flowchart illustrating a two-dimensional convolutional neural network in one scenario of an embodiment of this application. Figure 4 This is a schematic diagram of a multi-head attention mechanism in one scenario of an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] Due to the long-term service life and harsh operating environment of battery packs, lithium-ion batteries may experience internal short-circuit faults in individual cells, leading to a decrease in their lifespan and even serious safety accidents. Therefore, compared to fault diagnosis of other components, internal short-circuit diagnosis of lithium-ion battery packs is of greater significance for improving battery pack safety.

[0020] The purpose of this application is to provide a data-driven method for diagnosing internal short circuit faults in lithium-ion batteries, overcoming the shortcomings and experimental limitations of traditional methods, enabling high-quality and accurate diagnosis of soft and hard ISC in lithium-ion batteries, improving the detection sensitivity of early minor faults, achieving rapid estimation of fault resistance, and enhancing the safety of lithium-ion battery systems.

[0021] like Figure 1 As shown in the figure, this application provides a data-driven method for diagnosing internal short-circuit faults in lithium-ion batteries, including: S101: Construct a simulation model of internal short-circuit fault in a lithium-ion battery, and determine the corresponding model parameters based on battery temperature and state of charge.

[0022] This application addresses battery diagnostic faults, primarily including internal short-circuit faults. Specifically, an internal short-circuit fault in a lithium-ion battery refers to a fault state caused by factors such as separator failure, metal dendrite puncture, or impurity contact, resulting in an abnormal direct or indirect electrical connection between the positive and negative electrodes within the battery, thereby triggering a continuous, localized high-current discharge path.

[0023] Based on the principles of battery electrochemistry and equivalent circuit theory, a coupled model describing the normal state and internal short-circuit state of a battery is constructed as a simulation model for internal short-circuit faults in lithium-ion batteries. The simulation model for internal short-circuit faults in lithium-ion batteries includes a parallel equivalent circuit model and an ISC branch. Based on the function related to battery temperature and state of charge, the model parameters in the simulation model for internal short-circuit faults in lithium-ion batteries are set.

[0024] This application constructs a simulation model of internal short-circuit faults to generate battery external characteristic data with different short-circuit fault levels, achieving data augmentation. The simulation model of internal short-circuit faults in lithium-ion batteries can consist of a first-order equivalent circuit model and self-discharge internal resistance. For example... Figure 2 As shown, an extended Thevenin equivalent circuit model can be used as a basis here, integrating a variable ISC branch. The Thevenin equivalent circuit model includes series voltage sources. Ohmic internal resistance RC network (including parallel polarization resistors) , ), where SOC and T correspond to the State of Charge (SOC) and battery temperature, respectively. The negative pole is the negative terminal, and the positive pole is through... The RC network is connected to the positive terminal. The ISC branch is connected in parallel with the Thevenin equivalent circuit model, including a variable resistor that is related to the state of charge and temperature. It is used to simulate internal short-circuit faults of different severity and evolution stages.

[0025] At this point, the model parameters include the ohmic internal resistance. Polarization resistance Both the time constant and the time constant are experimentally calibrated as continuous functions of battery temperature T (range: -10℃~50℃) and instantaneous SOC. For example, ,in, , , , The corresponding coefficients were obtained by fitting the test data of Hybrid Pulse Power Characterization (HPPC) under different temperature and charge states through multiple linear regression.

[0026] S102: Using the lithium-ion battery internal short-circuit fault simulation model, simulate and generate battery external characteristic data containing internal short-circuit faults, and construct a training dataset based on the battery external characteristic data and actual battery operation data; wherein, the training dataset is used to train a deep learning model, the deep learning model including a two-dimensional convolutional neural network and a multi-head attention mechanism.

[0027] In the simulation environment, by dynamically adjusting the corresponding short-circuit resistance parameters, different degrees of internal short-circuit faults can be simulated, including simulations of different fault degrees and evolution processes from minor internal short circuits to severe internal short circuits. For example, this parameterized model (simulation model of internal short-circuit fault in lithium-ion batteries) can be implemented in the MATLAB / Simulink environment.

[0028] Based on internal short-circuit faults, a simulation model of internal short-circuit faults in lithium-ion batteries is run, and the external characteristic data of the battery under various standard operating conditions, different ambient temperatures, and initial states of charge are output. The external characteristic data of the battery includes time series of voltage, current, and battery temperature.

[0029] Running this model generates batches of battery external characteristic data, including time series data such as voltage, current, and surface temperature, under various standard operating conditions (such as the Urban Dynamometer Driving Schedule (UDDS) and the Worldwide Harmonized Light Vehicles Test Cycle (WLTC)) and at different ambient temperatures and initial SOCs. For example, the internal short-circuit resistance can be controlled programmatically. Batch simulations are performed with a wide range of variations (e.g., from 100Ω to simulate a minor leakage to 1Ω to simulate a severe short circuit), combined with standard driving cycles (such as UDDS and WLTC) and different ambient temperatures and initial SOC conditions. Each simulation synchronously outputs the battery terminal voltage V, current I, and battery temperature (i.e., the battery surface temperature). The time series, and record the real time series. The value serves as the data label.

[0030] The massive amount of data generated by the above simulation is precisely The fault data tagged (i.e., battery external characteristic data, such as 100,000 time series) are integrated, cleaned and standardized with normal battery operation data (such as 20,000 sets) collected from actual battery management systems (including data collected from actual vehicles or laboratory benches) to jointly construct a large-scale, multi-condition training dataset that covers normal and different degrees of faults, providing a foundation for data-driven diagnostic models.

[0031] When a normal battery experiences an early ISC (Internal Short Circuit) failure, the short circuit between the positive and negative terminals can be modeled using the branch current ISC with the larger ISC resistance. This can be expressed mathematically as follows: ; Where t is time, Let be the load current at time t. For ohmic resistance, For polarization internal resistance, For the equivalent internal short-circuit fault short-circuit internal resistance, This refers to the open circuit voltage (OCV) of the battery in its current state. The polarization voltage of the RC circuit. This is the terminal voltage.

[0032] When the charging and discharging current is canceled, the terminal voltages of the normal battery and the ISC battery when the battery enters the resting process (zero input state) from the charging and discharging state can be expressed as: ; ;in, For the battery Voltage and time constant of the RC circuit when entering zero-input state .

[0033] As the resting time increases, the polarization process of the battery gradually disappears. The terminal voltage of a normal battery... Terminal voltage can be expressed as: ; .

[0034] As shown in the above formula, when a battery experiences an ISC (Independent Circuit Capacitance) fault, its terminal voltage will be lower than that of a normal battery. Based on this, this application utilizes the relaxed voltage feature and constructs a quantitative diagnostic method for early ISC faults using a deep learning algorithm. The obtained training dataset is used to train the deep learning model, which is mainly used to extract fault features corresponding to lithium-ion batteries. This model includes a two-dimensional convolutional neural network and a multi-head attention mechanism, which will be described in detail below.

[0035] S103: Obtain the voltage time series data of the battery during the relaxation phase, and extract the initial features related to the internal short circuit fault based on the voltage time series data.

[0036] By mapping the impact of ISC faults on the internal characteristics of lithium-ion batteries using voltage time series data (also known as voltage curves) during the relaxation phase, initial features highly correlated with ISC faults can be extracted to support ISC fault diagnosis.

[0037] The charging phase curve is affected by the initial state of charge (SOC) and charging strategy, while the relaxation phase voltage curve is relatively unaffected by the charging process and can be easily obtained from the battery management system. During relaxation, the battery's dynamic response gradually shifts from a charge transfer region with a small time constant to a diffusion region with a larger time constant.

[0038] The voltage change during battery relaxation time consists of the instantaneous voltage change caused by ohmic internal resistance and dynamic changes caused by kinetic effects and ion transfer. Due to the slow internal chemical and physical reactions, it takes a long time for the battery to reach its initial voltage range (OCV). The battery's relaxation voltage takes several hours to reach equilibrium, and the voltage curve exhibits strong nonlinear characteristics.

[0039] During dynamic voltage changes, the decrease in battery terminal voltage is mainly caused by internal dynamic changes. The magnitude of the dynamic voltage can be obtained by subtracting OCV from the terminal voltage. During the relaxation phase, the dynamic voltage at time t... It can be represented as: .

[0040] Based on this, the battery management system detects that the load current has dropped to a preset threshold and remains there for more than a preset duration, thus determining that the system has entered the relaxation phase and collecting the corresponding voltage time series data. A trigger mechanism is set in the battery management system (BMS) so that whenever the load current is detected to drop below a preset threshold (e.g., |I|<0.05C) and remains there for a preset duration, the system determines that the system has entered the relaxation phase and initiates high-sampling-rate (e.g., 1Hz) voltage acquisition.

[0041] A moving average filter is applied to the voltage time series data, and initial features related to internal short-circuit faults are extracted. These initial features include initial voltage drop, relaxation time constant, energy at a specific time interval, and the inflection point curvature of the voltage curve. A fixed-duration segment of voltage time series data (e.g., 30 minutes) is extracted after the start of the relaxation phase (referred to as the voltage series). The voltage sequence is filtered using a moving average to suppress measurement noise. A set of initial characteristics characterizing the relaxation dynamics and sensitive to ISC is calculated, including the voltage drop within the first 30 seconds after relaxation begins, i.e., the initial voltage drop (…). The voltage curve is fitted to a single exponential decay function. Obtain the relaxation time constant (τ), where, Let be the battery terminal voltage at time t. Let A be the final stable value of the voltage, and A be the amplitude. Calculate the integral value of the voltage curve over a specific time window (e.g., from minute 5 to 10) to obtain the energy (E) for that specific time period. Calculate the curvature (K) at the inflection points of the voltage curve to reflect changes in the relaxation rate. These features constitute the initial eigenvector F=[ ,τ,E,K].

[0042] Initial features are characteristics extracted from the relaxation voltage time series that characterize the voltage relaxation variation pattern, such as slope and inflection points. They are used to explicitly characterize the local variation characteristics of the relaxation voltage under internal short-circuit conditions. These initial features can be directly used to determine whether a corresponding fault exists, or they can be mapped and fused with deep fault features extracted through Gram angle field transformation and two-dimensional convolutional neural networks, and further input into a deep learning model.

[0043] S104: The voltage time series data is converted into two-dimensional image data using the Gram angle field algorithm, and the two-dimensional image data is processed by the two-dimensional convolutional neural network. Multiple learnable convolutional kernels are used to slide and scan on the two-dimensional plane of the two-dimensional image data to extract the fault features corresponding to the two-dimensional image data.

[0044] The voltage time series data is converted into two-dimensional image data by Gramian Angular Field (GAF), which provides multi-level, high-dimensional fault characteristics for the data-driven model while preserving the complete information of the signal and maintaining its dependence on time.

[0045] Gaussian Image Coding (GAF) is an image coding technique that uses coordinate transformations and Gram matrices to transform time series data into two-dimensional images. GAF uses polar coordinates to represent the time matrix and then uses the Gram matrix to transform the generated angles. This is mainly because the Gram matrix can preserve time dependence, but it is difficult to distinguish valuable information from Gaussian noise. Therefore, a spatial transformation is required first, the most common one being the transformation from Cartesian coordinates to polar coordinates.

[0046] The Cartesian coordinate system includes rectangular and oblique coordinate systems. The two number axes intersect at the origin, forming a planar affine coordinate system, and the units of measurement for both axes are consistent. When the two coordinate axes are perpendicular, it is called a rectangular coordinate system; otherwise, it is an oblique coordinate system. Extracted acceleration data is generally in a two-dimensional rectangular coordinate system, where the position of each point is determined by two numbers. The polar coordinate system is also a two-dimensional coordinate system. It refers to taking a point in the plane as the pole, drawing an axis as the polar axis, and determining the positive direction of unit length and angle (generally counterclockwise is taken as positive). The position of a point in polar coordinates is represented by its distance from the origin and its angle with the polar axis. Therefore, points in the Cartesian coordinate system can be mapped one-to-one with those in polar coordinates. The Gram matrix can calculate the linear dependence of a set of vectors and is commonly used in linear algebra and geometry. A Gram matrix of n vectors is a matrix composed of the dot products of pairs of vectors, as shown in the following equation: If the norms of vectors u and v are both 1, then the dot product of the two vectors can be rewritten as: .

[0047] Therefore, if we are dealing with unit vectors, their dot product is determined only by the angle, which can be expressed in radians, and the calculated value will be between -1 and 1. Assuming that all vectors are unit vectors, the Gram matrix will become: ;in, Two vectors and The angle between them.

[0048] Based on this, the acceleration response is converted into two-dimensional image data using GAF, and the process is as follows: For voltage time series data, minimum-maximum normalization is performed. Assume the sensor collects relaxation voltage data as X = { , , ,…, , The following formula is used to normalize the data and scale it to the range [-1, 1]: ;in, The voltage data before normalization. The voltage data is normalized. , These represent the minimum and maximum values ​​of the voltage data, respectively.

[0049] Each point in the normalized voltage time series data is treated as a single angle, and mapped to polar coordinate space using an inverse cosine function. The voltage time series data in the relaxation phase, scaled in the previous steps, is then converted to polar coordinates. Two key variables are considered: the voltage value and its corresponding timestamp. These variables can be represented by the angle and radius in polar coordinates, respectively. Encoding is performed in polar coordinates, encoding the scaled acceleration value using angular cosine, with values ​​between [0, π]. The corresponding timestamp is then encoded as the radius, as shown in the following formula: ;in, for The corresponding timestamp, where M is the normalization factor. , Let [0, 1] be the angle and radius corresponding to the polar coordinate space. Divide the interval [0, 1] into several parts, discard the 0s, associate the remaining points with the time series data, and ensure that each point falls within a unit circle in polar coordinates. As time increases, the voltage time series data will also be continuously distorted within this unit circle at different angles and radii. The advantages of this encoding are: the entire encoding is bijective, that is, given a time series, the result obtained in polar coordinates through this encoding method will be unique; compared with the rectangular coordinate system, polar coordinates maintain absolute time correlation, and time series at different intervals, after being rescaled and converted to polar coordinates, will also contain different angle boundaries, which also provides different information granularities for converting the one-dimensional time series of relaxed voltage into two-dimensional image data.

[0050] The Gramian Angular Summation Field (GASF) algorithm is used to determine the cosine of the angle sum between two points, generating an image matrix. The voltage time series data, after undergoing polar coordinate transformation as described above, also contains temporal information; therefore, GAF can be used to reconstruct the acceleration time series. Based on different formulas, two different images can be generated: the Gramian Angular Summation Field (GASF) and the Gramian Angular Difference Field (GADF), with the following expressions: ; ;in, , These are the matrices corresponding to GASF and GADF, respectively. The generated matrix is ​​an n×n symmetric matrix, and its pixel value (i,j) encodes the joint temporal information of time i and j in the original sequence. The final output is a single-channel grayscale image.

[0051] As shown by the formulas corresponding to GASF and GADF, as time increases, the encoded position moves from the upper left corner of the matrix to the lower right corner. Therefore, the time dimension is encoded into the geometric structure of the matrix through GAF. After conversion by GAF encoding technology, the corresponding positions in the two-dimensional image are completely mapped to the acceleration time series in different feature forms such as points, lines, and colors.

[0052] The above GAF images (two-dimensional image data) are processed based on 2D CNN. Multiple learnable convolutional kernels are used to perform sliding scans on the two-dimensional plane of the input data to automatically extract the spatial hierarchical fault features of the image, thereby realizing the detection and quantitative diagnosis of ISC faults.

[0053] like Figure 3 As shown, a two-dimensional convolutional neural network (2DCNN) is defined as consisting of four convolutional blocks; each convolutional block includes two consecutive convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer; the number of convolutional kernels corresponding to each convolutional block increases sequentially.

[0054] The architecture of a 2D CNN mainly consists of an input layer, convolutional layers, activation functions, pooling layers, fully connected layers, and an output layer.

[0055] The input layer receives raw two-dimensional image data, which is usually in the form of RGB three channels, meaning that each image consists of three color channels: red, green, and blue.

[0056] Convolutional Layers: The core component of 2D CNNs. Convolutional layers perform local perception and feature extraction on the input image through a series of learnable convolutional kernels (or filters). Each convolutional kernel corresponds to a specific feature pattern and moves across the input image in a sliding window manner, calculating the dot product between the convolutional kernel and the local region of the image to obtain a feature map. This process helps to capture local features such as edges and textures in the image.

[0057] Cross-correlation: In convolutional layers, cross-correlation is often used instead of strict convolution. Cross-correlation operates by performing a dot product between the convolution kernel and a local region of the input image using a sliding window, resulting in an output feature map.

[0058] Activation functions: Activation functions are used to increase the non-linearity of a network, enabling it to fit more complex functional relationships. Common activation functions include Rectified Linear Unit (ReLU), Sigmoid, and Tanh. The ReLU function is widely used in 2D CNNs due to its simplicity, effectiveness, and relatively mild gradient vanishing problem.

[0059] Pooling layers: Pooling layers reduce the size of feature maps by downsampling, thereby lowering computational complexity and improving feature robustness. Common pooling methods include max pooling (MP) and average pooling (AP). Max pooling selects the maximum value within the pooling window as the output, while average pooling calculates the average value of the elements within the pooling window as the output.

[0060] Fully connected layers: Located at the end of the network, fully connected layers are used to perform weighted calculations and classification of the extracted features. Each neuron in a fully connected layer is connected to all neurons in the previous layer. By calculating the weighted sum and applying an activation function, the final classification result is obtained.

[0061] Output layer: The output layer typically uses the Softmax function to normalize the output of the fully connected layer, obtaining the probability distribution for each category. Finally, the category with the highest probability is selected as the prediction result.

[0062] By using convolutional kernels in a 2D convolutional neural network, a sliding scan is performed on the 2D plane of the 2D image data to extract high-dimensional feature vectors corresponding to the 2D image data as fault features. For example, a generated 64x64 pixel GAF image (2D image data) is used as input. A CNN backbone network containing four convolutional blocks is constructed. Each convolutional block consists of two consecutive 3x3 convolutional layers, batch normalization (BatchNorm), ReLU activation function, and a 2x2 max pooling layer. The number of convolutional kernels increases progressively in each block (e.g., 32, 64, 128, 256), and the number of convolutional kernels in the convolutional layers within each block is consistent to progressively extract hierarchical features from low-level edge textures to high-level semantic patterns. The first convolutional layer learns to capture local correlations in the GAF image (e.g., correlation patterns at short temporal scales); deeper convolutional layers combine these local patterns to form an abstract representation of the global temporal structure (e.g., the overall shape of the entire relaxation process). After the last pooling layer, the feature map is flattened into a high-dimensional feature vector. , as the corresponding fault characteristics.

[0063] The convolutional kernel performs sliding convolution operations on the input GAF image, automatically learning and extracting multi-level spatial features from local subtle patterns (such as unique textures during a specific relaxation time period) to global macroscopic structures (such as the overall shape pattern of the entire relaxation process). After several layers of convolution and pooling, the extracted high-dimensional feature maps are flattened to form a set of deep feature vectors that can comprehensively represent the internal short-circuit state.

[0064] S105: The fault features are projected into multiple different representation subspaces through the multi-head attention mechanism, and the estimated internal short-circuit resistance is output.

[0065] like Figure 4 As shown, the input sequence is projected into multiple different representation subspaces through multi-head attention (MHA), allowing the 2D CNN model to collectively focus on information from different representation subspaces at different locations. This enhances the model's expressive power and enables high-accuracy and high-reliability estimation of the core temperature of lithium-ion batteries, making it a preferred solution for efficient sequence modeling. The main computational process of the multi-head attention mechanism is as follows: The fault features are reshaped into corresponding feature sequences; the flattened feature vectors extracted by the CNN are then processed. Reshape it into a feature sequence (e.g., consider it as a sequence of length L, with a feature dimension of D at each position).

[0066] The feature sequence is processed through a multi-head attention mechanism, which linearly projects the feature sequence into the corresponding query matrix, key matrix, and value matrix. The scaled dot product attention is computed in parallel on multiple heads and transformed into the corresponding probability distribution.

[0067] For input features (feature sequences corresponding to fault features) (N is the number of nodes, d is the feature dimension), design independent multi-head attention heads: ;in, The multi-head attention function for the m-th variable is defined as follows: The calculation method for the i-th attention head is as follows: ;in, These are the projection matrices for the query, key, and value, respectively. This is for outputting the projection matrix.

[0068] The attention function is in the form of a scaled dot product: Where Q, K, and V are the query matrix, key matrix, and value matrix (also called query vector, key vector, and value vector), respectively; d is the dimension; and softmax is the normalization function that transforms the dot product result into a probability distribution between [0,1].

[0069] The probability distributions of the outputs from each head are concatenated and linearly projected to obtain the final feature enhanced with global context information. This final feature is then input into a regressor to obtain the estimated internal short-circuit resistance. The outputs of all heads are concatenated and linearly projected again to obtain the final feature representation enhanced with global context information. .Will The input is a regressor consisting of two fully connected layers, and the final output is a scalar value, namely the estimated internal short-circuit resistance. Using the constructed dataset, the full 2D-CNN-MHA model (i.e., a deep learning model) was trained end-to-end using the Adam optimizer with Mean Squared Error (MSE) as the loss function.

[0070] Multi-head attention mechanisms, by calculating attention weights among query, key, and value vectors, enable the model to dynamically and selectively focus on historical features most relevant to the current internal resistance estimation task, rather than treating all features equally. The attention mechanism operates in parallel across multiple heads, each projecting input features into a different representation subspace. This allows the model to focus on complex dependencies between features from different perspectives and subspaces, such as simultaneously capturing long-range associations between features in the early, middle, and late stages of relaxation. The outputs of each attention head are concatenated and linearly projected to form a final feature representation enhanced with global relational enhancement. This representation integrates local spatial features extracted by CNNs and global contextual dependencies captured by the attention mechanism. Then, through fully connected layers, it ultimately outputs an accurate estimate of the internal short-circuit resistance, enabling quantitative fault diagnosis.

[0071] 1. A data-driven internal short-circuit fault diagnosis method is proposed, which elevates the fault discrimination mechanism of the battery management system from over-limit alarms based on fixed thresholds to a methodological architecture capable of proactively sensing early, subtle fault characteristics. This not only expands the fault diagnosis functional boundaries of the battery management system but also provides key state input parameters for redefining system-level safety boundaries, predicting lifespan degradation trajectories, and optimizing energy management strategies. It serves as an important technical support for building the next-generation intelligent battery management system.

[0072] 2. A simulation model of internal short-circuit fault in lithium-ion batteries is constructed. By introducing battery temperature and state of charge as control variables for model parameters, the mapping relationship between key parameters and operating conditions is established. By adjusting the parameter values ​​of the short-circuit resistance branch in the model, the external response characteristics of the battery under different fault levels are simulated, realizing data augmentation under fault conditions and providing a large-scale labeled sample set covering multiple operating conditions and multiple fault levels for subsequent data-driven model training and generalization.

[0073] 3. The terminal voltage response curve during the battery relaxation stage is selected as the analysis object. Leveraging the fact that this stage's signal is unaffected by charging protocols and load disturbances, and is easily obtained from the battery management system, manual statistical features with physical correlation to internal short-circuit faults are extracted. Simultaneously, the Gram angle field algorithm is used to encode the relaxation voltage time series into a two-dimensional image. While fully preserving temporal dependencies, image domain features with high-dimensional, multi-level expressive capabilities are constructed, providing structured input for the deep learning model.

[0074] 4. The constructed deep learning model is composed of a two-dimensional convolutional neural network and a multi-head attention mechanism. The two-dimensional convolutional neural network is responsible for extracting local spatial hierarchical features in the image, while the multi-head attention mechanism is responsible for modeling the long-range dependencies of feature sequences and selectively enhancing key information. The sequential fusion of the two achieves complementary functional coupling between feature extraction and global context modeling, thereby improving the network's quantitative regression accuracy for fault resistance and enabling the model to differentiate important historical information in the feature sequence.

[0075] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A data-driven method for diagnosing internal short-circuit faults in lithium-ion batteries, characterized in that, include: A simulation model of internal short-circuit fault in lithium-ion battery was constructed, and the corresponding model parameters were determined based on battery temperature and state of charge. The lithium-ion battery internal short-circuit fault simulation model generates external characteristic data of the battery containing internal short-circuit faults, and a training dataset is formed based on the external characteristic data of the battery and the actual battery operation data; wherein, the training dataset is used to train a deep learning model, the deep learning model including a two-dimensional convolutional neural network and a multi-head attention mechanism. Obtain the voltage time series data of the battery during the relaxation phase, and extract the initial features related to the internal short circuit fault based on the voltage time series data; The voltage time series data is converted into two-dimensional image data using the Gram angle field algorithm, and the two-dimensional image data is processed by the two-dimensional convolutional neural network. Multiple learnable convolutional kernels are used to slide and scan on the two-dimensional plane of the two-dimensional image data to extract the fault features corresponding to the two-dimensional image data. The fault features are projected into multiple different representation subspaces using the multi-head attention mechanism, and the estimated internal short-circuit resistance is output.

2. The method according to claim 1, characterized in that, A simulation model of internal short-circuit fault in a lithium-ion battery is constructed, and the corresponding model parameters are determined based on battery temperature and state of charge, specifically including: Based on the principles of battery electrochemistry and equivalent circuit theory, a coupled model describing the normal state and internal short-circuit state of a battery is constructed as a simulation model for internal short-circuit faults in lithium-ion batteries. The simulation model for internal short-circuit faults in lithium-ion batteries includes a parallel equivalent circuit model and an ISC branch. The model parameters in the simulation model of internal short circuit fault of the lithium-ion battery are set based on the function related to battery temperature and state of charge.

3. The method according to claim 2, characterized in that, The lithium-ion battery internal short-circuit fault simulation model generates external characteristic data of the battery that includes internal short-circuit faults, specifically including: In the simulation environment, different degrees of internal short-circuit faults are simulated by dynamically adjusting the corresponding short-circuit resistance parameters. Based on the internal short-circuit fault, the simulation model of the internal short-circuit fault of the lithium-ion battery is run to output the external characteristic data of the battery under various standard operating conditions, different ambient temperatures, and initial states of charge; the external characteristic data of the battery includes the time series of voltage, current, and battery temperature.

4. The method according to claim 1, characterized in that, Acquire voltage time-series data of the battery during the relaxation phase, and extract initial features related to internal short-circuit faults based on the voltage time-series data, specifically including: Based on the battery management system, if the load current drops to a preset threshold and continues for more than a preset duration, it is determined that the system has entered the relaxation phase, and the corresponding voltage time series data is collected. The voltage time series data is subjected to moving average filtering, and initial features related to internal short-circuit faults are extracted; the initial features include initial voltage drop, relaxation time constant, energy at a specific time period, and inflection point curvature of the voltage curve.

5. The method according to claim 1, characterized in that, The voltage time series data is converted into two-dimensional image data using the Gram angle field algorithm, specifically including: For the voltage time series data, perform minimum-maximum normalization; Each point in the normalized voltage time series data is treated as a single angle, and the point is mapped to polar coordinate space using the inverse cosine function; The cosine of the angle sum between two points is determined using the Gram angle field algorithm, and an image matrix is ​​generated.

6. The method according to claim 1, characterized in that, The two-dimensional image data is processed by the two-dimensional convolutional neural network. Multiple learnable convolutional kernels slide and scan across the two-dimensional plane of the image data to extract fault features corresponding to the two-dimensional image data. Specifically, this includes: The two-dimensional convolutional neural network is defined to include four convolutional blocks; wherein, a single convolutional block includes two consecutive convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer; the number of convolutional kernels corresponding to each convolutional block increases sequentially; The high-dimensional feature vector corresponding to the two-dimensional image data is extracted by sliding scan on the two-dimensional plane of the two-dimensional image data through the convolution kernel in the two-dimensional convolutional neural network, and is used as the fault feature.

7. The method according to claim 1, characterized in that, The fault features are projected into multiple different representation subspaces using the multi-head attention mechanism to output the estimated internal short-circuit resistance, specifically including: The fault characteristics are reshaped into corresponding feature sequences; The feature sequence is processed through the multi-head attention mechanism, which linearly projects the feature sequence into the corresponding query matrix, key matrix, and value matrix, and performs parallel scaling dot product attention on multiple heads, and transforms it into the corresponding probability distribution. The probability distributions output by each head are concatenated and linearly projected to obtain the final feature enhanced with global context information. The final feature is input into the regressor to obtain the estimated internal short-circuit resistance.

Images