Lithium battery internal short circuit fault intelligent detection method based on electrochemical impedance spectroscopy

Abstract

The invention provides a lithium battery internal short circuit fault intelligent detection method based on electrochemical impedance spectroscopy. According to the system, battery data are obtained through electrochemical impedance spectroscopy testing, and innovative feature extraction is carried out. The core technology is to extract four key features from a Nyquist graph: a minimum value point of an intermediate frequency region, a straight slope of a low frequency region, a radius of a circle of a high frequency region and an intersection point of the high frequency region and a real part axis. The characteristics effectively characterize the short circuit state in the battery. And the system inputs the extracted features into a long short-term memory (LSTM) network deep learning model to realize accurate identification and classification of the internal short circuit fault. The model is deployed in a cloud after being trained, and monitors the battery state in real time and performs graded early warning. According to the invention, the accuracy and efficiency of short-circuit fault detection in the lithium ion battery are obviously improved, and powerful support is provided for battery safety management of an electric vehicle and an energy storage system.

Description

Technical Field

[0001] The present invention relates to the technical field of battery safety detection. Specifically, it relates to a method and system platform for detecting internal short - circuit faults of lithium - ion batteries based on Electrochemical Impedance Spectroscopy (EIS) and Long Short - Term Memory Network (LSTM), which is used to detect internal short - circuit faults of lithium - ion batteries in electric vehicles and energy storage systems. Background Art

[0002] With the wide application of lithium - ion batteries in electric vehicles and energy storage systems, the safety issues caused by internal short - circuit (ISC) have attracted increasing attention. Existing ISC detection methods mainly rely on terminal voltage, surface temperature, and leakage gas monitoring, but these methods have problems such as late early detection and difficulty in feature extraction.

[0003] For example, the detection method based on pulse charge - discharge proposed in patent publication number CN118731721A requires the use of a standard battery as a reference, increasing the detection cost and complexity. The detection method based on open - circuit voltage proposed in publication number CN115774200B has low hardware requirements, but the detection process takes a long time and is difficult to meet the real - time monitoring needs.

[0004] The Electrochemical Impedance Spectroscopy (EIS) method shows advantages in early high - sensitivity detection of internal short - circuits, and can provide multi - parameter evaluation and non - invasive measurement. However, existing EIS - based detection methods still have problems. For example, the method based on impedance spectrum and Elman neural network proposed in publication number CN116990716A has a complex feature extraction process, which may affect the real - time detection efficiency, and the neural network used may have the problem of gradient disappearance when dealing with long - term sequence data.

[0005] Therefore, there is an urgent need for a lithium - ion battery internal short - circuit fault detection system that can overcome the limitations of existing technologies and improve the detection accuracy, real - time performance, and reliability. Summary of the Invention

[0006] In order to overcome the deficiencies of existing technologies for detecting internal short - circuit (ISC) of lithium - ion batteries, such as the lag in early detection ability, difficulty in extracting internal short - circuit fault characteristics, and complexity of algorithm cloud application deployment, based on terminal voltage, surface temperature monitoring, leakage characteristic gas monitoring, and battery internal state monitoring, the present invention provides a lithium - ion battery internal short - circuit fault detection system based on Electrochemical Impedance Spectroscopy (EIS) detection and Long Short - Term Memory Network (LSTM) deep learning, which is used to improve the accuracy and efficiency of battery internal short - circuit fault detection. The technical solution adopted by the present invention to solve its technical problems is:

[0007] In a first aspect, the present invention provides a lithium-ion battery internal short circuit fault detection system for use in detecting faults in the lithium-ion batteries of electric vehicles. The system includes:

[0008] Obtain lithium-ion battery samples with different degrees of internal short circuit, perform electrochemical impedance spectroscopy (EIS) detection on them, and output test data;

[0009] Preprocess the test data, including deleting high-frequency data, extracting the real part impedance (Z') and the imaginary part impedance (-Z");

[0010] Draw a Nyquist plot, and extract the minimum value point in the intermediate frequency region, the slope of the straight line in the low frequency region, the radius of the circle in the high frequency region, and the intersection point of the high frequency region and the real axis as characteristic values;

[0011] Fuse the characteristic values with the ohmic resistance labels as the input of a long short-term memory network (LSTM) deep learning model;

[0012] Construct and train an LSTM deep learning model to identify and classify internal short circuit faults in the battery;

[0013] Deploy the trained LSTM model to a cloud system to achieve real-time detection and early warning of internal short circuit faults in lithium-ion batteries.

[0014] In a second aspect, the present invention provides a lithium-ion battery internal short circuit fault detection system based on EIS detection and LSTM deep learning, which includes four modules, specifically:

[0015] An electrochemical impedance test module for obtaining lithium-ion battery samples with different degrees of internal short circuit, performing electrochemical impedance spectroscopy (EIS) detection on them, and outputting test data;

[0016] A data preprocessing and feature extraction module for preprocessing the test data, including deleting high-frequency data, extracting the real part impedance (Z') and the imaginary part impedance (-Z"), drawing a Nyquist plot, and extracting the minimum value point in the intermediate frequency region, the slope of the straight line in the low frequency region, the radius of the circle in the high frequency region, and the intersection point of the high frequency region and the real axis as characteristic values;

[0017] A model construction and training module for fusing the characteristic values with the ohmic resistance labels as the input of a long short-term memory network (LSTM) deep learning model, constructing and training an LSTM deep learning model to identify and classify internal short circuit faults in the battery;

[0018] A cloud deployment and early warning module for deploying the trained LSTM model to a cloud system to achieve real-time detection and early warning of internal short circuit faults in lithium-ion batteries.

[0019] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program which, when executed by a processor, implements the internal short-circuit fault detection system for the lithium-ion battery as described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] The internal short-circuit fault detection method, system and computer-readable storage medium provided by the present invention collect real part impedance and imaginary part impedance data by performing electrochemical impedance tests on different lithium-ion battery samples. Then, through data preprocessing and the drawing of Nyquist diagrams, characteristic values such as the minimum value point in the intermediate frequency region, the slope of the straight line in the low frequency region, the radius of the circle in the high frequency region, and the intersection point of the high frequency region and the real part axis are extracted. Next, using these characteristic values as inputs, an LSTM deep learning model is trained to identify and classify internal short-circuit faults in the battery. Finally, the trained model is deployed in a cloud system to achieve the detection and early warning of internal short-circuit faults in lithium-ion batteries. The present invention aims to improve the accuracy and efficiency of internal short-circuit fault detection in batteries, and is of great significance for battery safety management in fields such as electric vehicles and energy storage systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be further described below in conjunction with the drawings and embodiments.

[0023] Figure 1 It is a schematic diagram of the detection process of the present invention;

[0024] Figure 2 It is a schematic diagram of experimental sample data;

[0025] Figure 3 It is a schematic diagram of Nyquist curves (for batteries with short-circuit resistances of 10 Ω and 200 Ω);

[0026] Figure 4 It is a schematic diagram of Nyquist curves of lithium-ion batteries with different degrees of internal short-circuit faults;

[0027] Figure 5 It is a schematic diagram of the fitting circle characteristics in the high frequency region;

[0028] Figure 6 It is a schematic diagram of feature extraction of the present invention;

[0029] Figure 7 It is a schematic diagram of the internal short-circuit fault detection architecture of the LSTM machine learning model;

[0030] Figure 8 It is a schematic diagram of the LSTM 5-fold cross-validation training and validation curves;

[0031] Figure 9Schematic diagram of the confusion matrix for the validation set. Detailed implementation manners

[0032] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Apparently, the described embodiments are some, but not all, of the embodiments of the present application. Components of the embodiments of the present application generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0033] Therefore, the detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but merely represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0034] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not require further definition and explanation in subsequent drawings.

[0035] In the description of the present application, it should be noted that the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be construed as indicating or implying relative importance. In addition, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0036] In view of this, this embodiment provides a lithium-ion battery internal short circuit fault detection system based on electrochemical impedance spectroscopy (EIS) detection and long short-term memory network (LSTM) deep learning. The overall system flow chart is as Figure 1As shown. In this system, the detection device conducts electrochemical impedance tests on different lithium-ion battery samples to collect real part impedance and imaginary part impedance data. Then, through data preprocessing and the drawing of Nyquist plots, characteristic values such as the minimum value point in the intermediate frequency region, the slope of the straight line in the low frequency region, the radius of the circle in the high frequency region, and the intersection point of the high frequency region and the real part axis are extracted. Next, using these characteristic values as inputs, an LSTM deep learning model is trained to identify and classify internal short circuit faults in the battery. Finally, the trained model is deployed in the cloud system to achieve the detection and early warning of internal short circuit faults in lithium-ion batteries. Through the above detection method, the internal short circuit faults in lithium-ion batteries are monitored in real time until the internal short circuit faults in the battery are detected and an early warning is issued.

[0037] In some embodiments, the detection device can be a data platform, which is communicatively connected to a vehicle equipped with a battery pack to collect the state information of the battery pack during vehicle driving and charging. For example, the mileage of the vehicle, the charging current, the charging voltage, the temperature of each single battery, the voltage of each single battery, and the SOC value during charging. In order to facilitate subsequent use in training a battery state prediction model, the above data also carries a time stamp at the time of collection.

[0038] Among them, the data platform can be a single data platform or a group of data platforms. The group of data platforms can be centralized or distributed (for example, the data platform can be a distributed system). In some embodiments, the data platform can be local or remote relative to the user terminal. In some embodiments, the cloud platform can be implemented on the data platform; by way of example only, the cloud platform can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an inter-cloud, a multi-cloud, etc., or any combination thereof. In some embodiments, the data platform can be implemented on an electronic device having one or more components.

[0039] Based on the above introduction, each step included in this system will be elaborated in detail below. It should be understood that the operations in the flowchart can be implemented out of order, and steps without a logical context relationship can be reversed or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart or remove one or more operations from the flowchart under the guidance of the content of this application.

[0040]

Embodiment 1

[0041] 1. Sample Selection: A total of 100 lithium-ion battery samples were collected, including 50 normal lithium-ion battery samples and 50 internally short-circuited lithium-ion battery samples. All samples must ensure that the specifications and models are exactly the same, and the production batch is the same batch. During storage and transportation, the ambient temperature is controlled at 25±2°C, and the humidity is 45±5%RH, so as to exclude external factors interference to the greatest extent and ensure that the test results focus on the characteristics reflected by the difference between the normal and internally short-circuited states of the battery itself. The test equipment uses a high-precision electrochemical workstation (accuracy ±0.1%).

[0042] 2. Test Conditions: The entire EIS test process was carried out under open-circuit conditions to simulate the battery's static state. The excitation voltage was set to 10mV (far lower than the battery polarization voltage threshold of 30mV) to ensure that the test process would not cause changes in the internal electrochemical state of the battery, thus ensuring non-destructive testing. The frequency range was accurately set to 5mHz - 10kHz, which can comprehensively scan the electrochemical characteristics of the battery from macroscopic to microscopic, while ensuring the integrity and reliability of data acquisition.

[0043] 3. Sample Installation and Initialization: The lithium-ion battery samples were installed one by one into a dedicated four-point probe fixture. This fixture uses a spring pressure method to ensure stable electrode connection, with a contact resistance less than 0.05Ω, effectively eliminating the influence of poor contact on the measurement results. After installation, the samples were placed on a constant temperature (25±0.5°C) test platform and left to stand for 60 minutes to reach thermal equilibrium and electrochemical steady state.

[0044] 4. EIS Test Execution and Raw Data Storage: Start the EIS test program preset in the electrochemical workstation and test the samples in sequence according to the set parameters. The device applies a sine wave excitation signal, collects the current signal, and analyzes the electrochemical impedance data of the battery at different frequencies based on Ohm's law and the principle of complex number operation, including key parameters such as real part impedance, imaginary part impedance, and phase angle. At the same time, the supporting software of the electrochemical workstation automatically records the corresponding data of each sample, outputs it in a standardized data format and imports it into an Excel table. The Excel table headers are set in sequence as "Frequency (Hz)", "Real Part Impedance (Ω)", "Imaginary Part Impedance (Ω)", "Magnitude (Ω)", "Phase Angle (°)", "Detection Time (s)", etc., as Figure 2 shown.

[0045] The EIS test sampling rate settings are as follows: in the frequency band of 0.01 Hz - 1 Hz, 10 points are collected for each frequency band; in the frequency band of 1 Hz - 10 Hz, 8 points are collected for each frequency band; in the frequency band of 10 Hz - 100 Hz, 6 points are collected for each frequency band; in the frequency band of 100 Hz - 1 kHz, 5 points are collected for each frequency band; in the frequency band of 1 kHz - 3 kHz, 4 points are collected for each frequency band. Each sampling point is measured 3 times repeatedly and the average value is taken to reduce random errors. The acquisition accuracy ensures that the measurement errors of the real and imaginary part impedance values do not exceed ±0.1%, and the frequency measurement error does not exceed ±0.05%. For each battery sample, a total of 89 ± 3 valid data points are collected. This sampling density has been statistically verified to be sufficient to capture the key feature changes in the Nyquist plot while maintaining computational efficiency.

[0046]

Example 2

[0047]

Basis for Frequency Interval Division

[0048] Through the experimental analysis of 100 battery samples (50 normal and 50 internal short circuits), combined with Fourier transform and correlation analysis, the optimal segmentation points of the frequency interval are determined. This segmentation scheme is optimized through 3 rounds of cross-validation, and finally 0.35 Hz and 500 Hz are selected as the key demarcation points to achieve the best separation of signal features. The electrochemical processes corresponding to the three intervals are completely different: the low-frequency region (0.01 Hz - 0.35 Hz) mainly reflects the diffusion behavior of lithium ions in the solid phase, and the change in the current path caused by internal short circuits leads to significant changes in this region; the mid-frequency region (0.35 Hz - 500 Hz) is mainly related to the charge transfer process at the electrode / electrolyte interface; the high-frequency region (500 Hz - 3 kHz) mainly reflects the SEI film impedance and electrode reaction kinetic characteristics. Near 0.35 Hz, the electrochemical characteristics of lithium-ion batteries change from a diffusion-controlled process to a charge-transfer-controlled process, and near 500 Hz, it changes from a charge-transfer control to an ohmic control and surface process.

[0049] This frequency interval division method improves the accuracy of internal short circuit detection by 6.8 percentage points (from 90.2% to 97.0%) compared with the traditional 2 kHz and 100 Hz demarcation point (using empirical division) scheme. This improvement is statistically significant (p < 0.01, based on Fisher's exact test). The p-value is a measure in statistics used to determine whether the observed results are statistically significant.

[0050]

Definition of Frequency Interval

[0051] - Low - frequency region: Defined as 0.01 Hz to 0.35 Hz. This frequency range reflects the characteristics of the diffusion process inside the battery and the lithium - ion transport in the solid phase. In the case of internal short - circuit, the change in the impedance characteristics in this region is manifested as a decrease in the slope of the low - frequency straight line in the Nyquist plot. According to the data analysis of 100 battery samples (50 normal and 50 with internal short - circuit) described in the patent, the slope of the internal - short - circuit batteries in the low - frequency region decreases by an average of 52.8% ± 3.2%, and this characteristic change is statistically significant (p < 0.01).

[0052] - Medium - frequency region: Specifically 0.35 Hz to 500 Hz, corresponding to the range where the real - part impedance value in the Nyquist plot is between 0.22 Ω and 0.24 Ω. This region reflects the charge - transfer process at the electrode / electrolyte interface, and internal short - circuit causes a displacement of the position of the minimum value point in this region.

[0053] - High - frequency region: Specifically 500 Hz to 3 kHz, which appears as a semicircle in the Nyquist plot. This region reflects the impedance of the SEI film inside the battery and the characteristics of the electrode reaction kinetics. Internal short - circuit causes a change in the radius of the semicircle and a shift of the intersection point with the real - part axis in this region.

[0054] - Ultra - high - frequency region: Refers to the region above 3 kHz. The data is deleted during the pre - processing stage because: (1) the data in this frequency band is affected by the measurement device accuracy (±0.1%) and external electromagnetic interference, resulting in large data noise; (2) the fast transient response corresponding to this frequency band has no direct relevance to the analysis of internal short - circuit faults.

[0055] 1. Deleting high - frequency data: Considering that in the higher - frequency band, on the one hand, due to factors such as the detection accuracy of the electrochemical workstation itself and external electromagnetic interference, the high - frequency data obtained has a large noise component, resulting in a decrease in data reliability; on the other hand, from the perspective of the electrochemical characteristics of lithium - ion batteries, the electrochemical processes corresponding to too high frequencies are often some fast transient responses, which have little relevance to the long - term and slow - changing electrochemical mechanisms related to the analysis of internal short - circuit faults in the battery, but will increase the data volume and the complexity of subsequent data analysis. Therefore, some high - frequency data is deleted. The specific operation method for deleting high - frequency data is: setting a frequency threshold of 3 kHz, and using the filtering function of Excel to delete the entire rows of high - frequency data outside the corresponding frequency range, and retaining the valid data in the range of 0.01 Hz to 3 kHz to optimize the data quality and focus on the more valuable mid - and low - frequency band data.

[0056] 2. Data reading, pre - processing, and graph plotting

[0057] Use the readmatrix function in MATLAB to read the EIS data stored in an Excel file. Each file contains the EIS data of a battery, and the data includes information such as frequency, real part impedance (Z'), and imaginary part impedance (-Z").

[0058] Extract the real part impedance (Z') and the imaginary part impedance (-Z"), and reverse the sign of the imaginary part impedance to meet the requirements for plotting the Nyquist diagram.

[0059] Use the plot function in MATLAB to plot the Nyquist diagram, and plot the Nyquist curves of two batteries respectively, as Figure 3 shown.

[0060] 3. Preliminary Judgment of Internal Short Circuit Characteristics

[0061] Data Preparation: Extract the values of the real part impedance (Z') and the imaginary part impedance (-Z") corresponding to the frequency range of 0.01 - 3000 Hz from the Excel data table, and export the extracted data as a text file.

[0062] Data Import and Fitting: Import the exported text file into Zview software for fitting processing to generate a Nyquist diagram and a Bode diagram.

[0063] Result Analysis: Through the analysis of the Nyquist diagram, specifically observe the following characteristic changes:

[0064] - Curve slope in the low-frequency region (0.01 Hz - 0.35 Hz): The slope of the 10Ω battery is 0.625092, which is significantly less than the slope of the 200Ω battery, which is 1.265921.

[0065] - Intersection point with the real part axis in the high-frequency region: For the 10Ω battery, it is 0.1899Ω, and for the 200Ω battery, it is 0.1869Ω, with a displacement of 0.0030Ω.

[0066] - Semi-circle radius in the high-frequency region: For the 10Ω battery, it is 0.0228Ω, which is greater than that of the 200Ω battery, which is 0.0203Ω.

[0067] The results of statistical analysis show that: Based on the data of 100 battery samples (50 normal and 50 with internal short circuit), the slope of the internal short circuit batteries in the low-frequency region decreases by an average of 52.8% ± 3.2%, and the semi-circle radius in the high-frequency region increases by an average of 12.3% ± 1.5%. These characteristic changes are statistically significant (p < 0.01).

[0068] Further, extract the real part impedance (Z') and imaginary part impedance (-Z") values corresponding to the frequencies of 0.01 - 3000 Hz from the data table. Export the extracted data as a text file and plot the Nyguist curves of lithium-ion batteries with different degrees of internal short circuit faults, as shown in the appendix Figure 4 As shown. To accurately judge the severity of the internal short circuit fault, the present invention establishes the following sensitivity determination criteria:

[0069]

Quantitative definition of the degree of internal short circuit

[0070] Severe internal short circuit: 0Ω < R ≤ 50Ω, that is, the internal short circuit equivalent resistance is less than or equal to 50Ω, corresponding to the label value 2. In this state, the following specific changes occur inside the battery:

[0071] (1) A metallic lithium dendrite short circuit path with a diameter greater than 0.5 mm is formed

[0072] (2) The current leakage is greater than 200 mA (measured based on the ambient temperature of 25°C and the voltage condition of 3.7V)

[0073] (3) The rising rate of the battery surface temperature exceeds 2°C / hour

[0074] (4) The reduction amplitude of the slope in the low-frequency region of the Nyquist diagram is greater than 40%

[0075] (5) The increase amplitude of the radius of the circle in the high-frequency region is greater than 15%

[0076] In this state, the safety risk rating is level 5 high grade, and the following specific measures must be taken:

[0077] (1) The system automatically disconnects the main battery circuit within 15 seconds after detecting a severe internal short circuit

[0078] (2) Activate the battery pack isolation device through the BMS controller

[0079] (3) Issue the highest level warning and require professional maintenance personnel to arrive at the scene within 2 hours for handling

[0080] (4) Start the temperature monitoring emergency program and increase the sampling frequency to once every 5 seconds

[0081] - Minor internal short circuit: 50Ω < R ≤ 200Ω, that is, the internal short circuit equivalent resistance is greater than 50Ω and less than or equal to 200Ω, corresponding to the label value 1. In this state, the following specific changes occur inside the battery:

[0082] (1) A tiny metallic lithium dendrite short circuit path with a diameter less than 0.5 mm is formed

[0083] (2) The current leakage is between 50 mA and 200 mA (measured based on an ambient temperature of 25 °C and a voltage condition of 3.7 V).

[0084] (3) The rate of increase in the battery surface temperature does not exceed 1 °C / hour.

[0085] (4) The reduction amplitude of the slope in the low-frequency region of the Nyquist plot is between 20% and 40%.

[0086] (5) The increase amplitude of the radius of the circle in the high-frequency region is between 5% and 15%.

[0087] In this state, the safety risk rating is medium, and the system needs to take the following specific measures:

[0088] (1) The system generates a yellow warning signal and notifies the maintenance personnel through the designated channel.

[0089] (2) Arrange for inspection and maintenance within 48 hours.

[0090] (3) Increase the battery status monitoring frequency from the normal once every 10 minutes to once every 5 minutes.

[0091] (4) Record the abnormal status data for subsequent analysis and model optimization.

[0092] - Normal state: R > 200 Ω (the equivalent internal short-circuit resistance is greater than 200 Ω or there is no internal short-circuit phenomenon), corresponding to the label value 0. In this state, the battery leakage current is less than 50 mA, and the impedance characteristics are within the normal parameters (±5%) specified by the manufacturer, and the system can operate normally.

[0093] 4. Determination of the internal short-circuit characteristic points

[0094] Characteristic 1. Determination of the minimum point: Screen all data points in the Nyquist plot where the real part impedance value is between 0.22 Ω and 0.24 Ω. This interval is determined based on the analysis of the Nyquist plots of 50 normal batteries and 50 internally short-circuited batteries. Use a traversal algorithm for the screening results to find the point with the minimum imaginary part impedance and its corresponding real part impedance value. The calculation formula is as follows.

[0095] Z 10ohm ″ = min{Z1″, Z2″, Z3″…, Z k ′} = 0.002586 Ω, at this time Z 10ohm ′ = 0.229653 Ω (1)

[0096] Z 200ohm ″ = min{Z1″, Z2″, Z3″…, Z j ′} = 0.002967 Ω, at this time Z 200ohm ′ = 0.222378 Ω (2)

[0097] Output (Z 10ohm ′, Z 10ohm ″) is (0.229653, 0.002586)

[0098] Output (Z 200ohm ′, Z 200ohm ″) is (0.222378, 0.002967)

[0099] Displacement Δ = |Z 10ohm ′ - Z 200ohm ′| = 0.229653 - 0.222378 = 0.007275 Ω (3)

[0100] It is calculated that the minimum point of the 10 Ω battery is (0.229653 Ω, 0.002586 Ω), the minimum point of the 200 Ω battery is (0.222378 Ω, 0.002967 Ω), and the displacement is 0.007275 Ω. The exact coordinates of this point are an important part of the eigenvector.

[0101] Characteristic 2. Determination of the slope of the straight line in the low-frequency region: Execute according to the following steps:

[0102] (1) Screen all data points within the frequency range of 0.01 Hz to 0.35 Hz

[0103] (2) Take the imaginary part impedance value as the dependent variable (y-axis), and the real part impedance value as the independent variable (x-axis)

[0104] (3) Use the least squares method for linear fitting, and the specific calculation results:

[0105] Output the fitting straight line equation:

[0106] y1 = 0.625092x1 - 0.140312 (4)

[0107] y2 = 1.265921x2 - 0.279096 (5)

[0108] Extract the slopes of the two straight lines:

[0109] k1 = 0.625092 (6)

[0110] k2 = 1.265921 (7)

[0111] Among them, 1 is the 10 Ω battery sample; 2 is the 200 Ω battery sample.

[0112] It can be seen that the slope corresponding to the straight line in the low-frequency region of the lithium-ion battery with an internal resistance of 10 Ω (severe internal short circuit) is smaller than the slope corresponding to the straight line in the high-frequency region of the lithium-ion battery with an internal resistance of 200 Ω (mild internal short circuit), and the slope corresponding to the straight line in the low-frequency region of the lithium-ion battery with an internal resistance of 10 Ω (severe internal short circuit) is 0.5 times that of the lithium-ion battery with an internal resistance of 200 Ω (mild internal short circuit). From this, it can be concluded that the low-frequency region is more sensitive to the internal short-circuit phenomenon. Therefore, the slope of the straight line in the low-frequency region can be used as a characteristic value for detecting whether an internal short circuit occurs.

[0113] Feature 3. Fitting of the semicircle in the high-frequency region: Execute according to the following steps:

[0114] (1) Screen all data points within the frequency range of 500 Hz to 3 kHz

[0115] (2) Use the least squares method for circle fitting, based on the equation (x - a) 2 +(y - b) 2 =r 2

[0116] (3) Specific fitting results:

[0117] Use the least squares method in MATLAB for circle fitting, and extract the radius of the circle as a feature point, as Figure 5 shown.

[0118] Then, according to the equation of the circle

[0119] (x - a) 2 +(y - b) 2 =r 2 (8)

[0120] It can be transformed into

[0121] x 2 +y 2 =2ax + 2by - c (9)

[0122] where c=r 2 -a 2 -b 2 , and a and b are the horizontal and vertical coordinates of the center of the circle respectively.

[0123] Write it in matrix form as Q = M·s, where Q is the vector containing the coordinates of the measurement points, M is the coefficient matrix composed of the measurement points, and s = [-2a; -2b; c] is the parameter vector to be solved. By solving the linear equation system Q / Y, the estimated value of s is obtained, and thus the center coordinates (a, b) and the radius r are deduced.

[0124] Finally, according to the obtained s, calculate the coordinates x fit and y fitand radius r fit These parameters are obtained by fitting the actually measured impedance data points and can accurately reflect the impedance characteristics of the battery in the high-frequency region.

[0125] The following are the fitting results applied to the actual battery samples:

[0126] Battery sample with 10 Ω:

[0127] Equation of the circle (x - 0.2103) 2 +(y + 0.0104) 2 = 0.02282 (10)

[0128] Center coordinates C1(0.2103, -0.0104), radius R1 = 0.0228

[0129] Battery sample with 200 Ω:

[0130] Equation of the circle (x - 0.2053) 2 +(y + 0.0084) 2 = 0.02032 (11)

[0131] Center coordinates C2(0.2053, -0.0084), radius R2 = 0.0203

[0132] Feature 4. Determination of the intersection point of the high-frequency region and the real-axis: Execute according to the following steps:

[0133] (1) Use the circle equation obtained by fitting and let y = 0

[0134] (2) Solve the equation to obtain two x values:

[0135] y1 = 0, x1 = 0.1899, x1' = 0.2306 (rejected) (12)

[0136] y2 = 0, x2 = 0.1869, x2' = 0.2237 (rejected) (13)

[0137] Among them, 1 is the 10 Ω battery sample; 2 is the 200 Ω battery sample.

[0138] (3) Select the smaller real part value (0.1899 and 0.1869) as the feature value and discard the larger real part value

[0139] (4) Calculate the displacement: Δx = x1 - x2 = 0.1899 - 0.1869 = 0.0030

[0140] This is based on the following electrochemical principle: In the Nyquist plot analysis of lithium-ion batteries, the left intersection point (smaller real part value) of the high-frequency region circle and the real axis accurately corresponds to the ohmic internal resistance (R) of the batterys ) This is the pure resistance component composed of electrolyte resistance, current collector resistance, and electrode contact resistance; while the right intersection point (with a larger real part value) usually falls in the mid-frequency region of the battery, including the contribution of charge transfer resistance, and cannot accurately reflect the change in ohmic internal resistance under the condition of internal short circuit. Since internal short circuit will directly affect the equivalent circuit structure of the battery, especially its ohmic internal resistance part, the left intersection point is a more accurate feature for evaluating the internal short circuit state.

[0141] Calculations show that the intersection point of the 10Ω battery curve and the real part axis is (0.1899, 0), and the intersection point of the 200Ω battery curve and the real part axis is (0.1869, 0). It can be seen that the occurrence of internal short circuit shifts the Nyquist curve to the right, and the displacement amount is:

[0142] Δx = x1 - x2 = 0.1899 - 0.1869 = 0.0030 (14)

[0143] According to the above feature extraction steps, the minimum value point in the mid-frequency region, the slope of the straight line in the low-frequency region, the radius of the circle in the high-frequency region, and the intersection point of the high-frequency region and the real part axis are obtained, as Figure 6 shown.

[0144]

Specific implementation method of feature fusion

[0145] In the present invention, the feature fusion method is to organize the four extracted feature values into a structured feature vector, and fuse it with the battery ohmic resistance value label and then input it into the LSTM deep learning model. The specific implementation steps are as follows:

[0146] The first step, feature vector construction: The feature vector X is further specifically defined as:

[0147] X = [Δ, k, r, x]

[0148] Where:

[0149] Δ = |Z' 10 -Z' 200 | is the displacement amount of the minimum value point in the mid-frequency region, with the unit of Ω;

[0150] k is the slope of the straight line in the low-frequency region, dimensionless;

[0151] r is the radius of the circle in the high-frequency region, with the unit of Ω;

[0152] x is the coordinate value of the intersection point of the high-frequency region and the real part axis, with the unit of Ω.

[0153] Corresponding to the sample data, the specific values of the feature vector are X = [0.007275, 0.625092, 0.0228, 0.1899] (short - circuit sample within 10Ω) and X = [0, 1.265921, 0.0203, 0.1869] (short - circuit sample within 200Ω). For normal samples (without internal short - circuit), the reference value of the feature vector is X normal = [0, 1.45 ± 0.15, 0.018 ± 0.002, 0.184 ± 0.005].

[0154] Second step, feature standardization processing: Since the physical meanings and dimensions of the four features are different and there are large differences in their numerical ranges, to eliminate the influence of dimensions and ensure the stability of model training, each element in the feature vector X is standardized as follows:

[0155] x' i = (x i - μ i ) / σ i (15)

[0156] where μ i and σ i are calculated based on the mean and standard deviation of the i - th feature in the training set, respectively, using 100 samples (50 normal and 50 with internal short - circuit).

[0157] Third step, feature and label fusion: The standardized feature vector X' is fused with the Ohmic resistance label R of the battery to form an augmented feature vector F = [X', R]. Among them, the R value is processed as follows:

[0158] For normal batteries, R > 200Ω, and the label value is set to 0

[0159] For slightly internally short - circuited batteries, 50Ω < R ≤ 200Ω, and the label value is set to 1

[0160] For severely internally short - circuited batteries, 0Ω < R ≤ 50Ω, and the label value is set to 2

[0161] Fourth step, construction of time - series features: Considering the characteristics of the LSTM model in processing time - series data, the feature fusion vector F collected at different time points is organized into a time - series feature matrix T, where each row represents the feature vector at a time point, and the matrix dimension is [n, 5], n is the time - series length, and the default value is set to 10, representing 10 consecutive time - point data collections.

[0162] Fifth step, construction of batch processing: A batch - processing data set is constructed for model training. Each batch contains 32 samples, and each sample is a time - series feature matrix, which constitutes the input data of the LSTM model.

[0163] Through the above steps, the present invention effectively organizes and integrates four key features into structured time-series feature data, providing high-quality input for the LSTM deep learning model, thereby achieving accurate identification and classification of internal short-circuit faults in lithium batteries. Statistical analysis shows that the contribution degrees of the features are in the following order: the slope of the straight line in the low-frequency region (contribution rate of about 40%), the radius of the circle in the high-frequency region (contribution rate of about 30%), the displacement of the minimum point in the middle-frequency region (contribution rate of about 20%), and the intersection point of the high-frequency region and the real-axis (contribution rate of about 10%), which is consistent with the feature sensitivity ranking observed in the patent.

[0164] In the present invention, the feature fusion method is to combine the four extracted feature values into a feature vector for subsequent analysis by the LSTM deep learning model. Specifically, the displacement of the minimum point in the middle-frequency region, the slope of the straight line in the low-frequency region, the radius of the circle in the high-frequency region, and the coordinate value of the intersection point of the high-frequency region and the real-axis are used to form a feature vector, which is combined with the battery ohmic resistance value label. According to the data analysis described in the patent, the slope in the low-frequency region is the most sensitive to the internal short-circuit state, showing the most obvious change characteristics. Secondly, it is the radius of the circle in the high-frequency region, with the average radius of the semicircle in the high-frequency region of the internal short-circuit battery increasing by 12.3% ± 1.5%.

[0165]

Example 3

[0166] Principle of the LSTM model: LSTM (Long Short-Term Memory network) is a special type of Recurrent Neural Network (RNN) with unique memory units, which can effectively process sequence data, especially suitable for prediction and classification tasks of time-series data. Its principle architecture is as Figure 7 shown. In the present invention, the LSTM model is used to process the EIS (Electrochemical Impedance Spectroscopy) data of the battery. By learning the feature points in the EIS data and using them as input, the internal short-circuit fault category of the battery is finally output, which can be divided into two categories: mild internal short-circuit and severe internal short-circuit. Through its internal gating mechanism, including the input gate, forget gate, and output gate, the model can selectively remember and forget information, thus solving the problem of gradient disappearance in traditional RNNs and enabling the network to long-term remember and process long time-series information. The LSTM deep learning model adopted in the present invention includes the following clearly defined structures and parameters:

[0167] Model structure:

[0168] Input layer, which receives a 5-dimensional feature vector, specifically including: the coordinates of the minimum point in the middle-frequency region (Z', Z”), the slope k value of the straight line in the low-frequency region, the radius r value of the circle in the high-frequency region, the coordinate x value of the intersection point of the high-frequency region and the real-axis, and the battery ohmic resistance R. Each neuron in the input layer corresponds to a feature dimension and receives the feature values after being normalized. The normalization formula is: x' i =(xi -μ i ) / σ i , where μ i and σ i are the mean and standard deviation of the i-th feature on the training set, respectively. The correlation between features is calculated by the Pearson correlation coefficient. The result shows that the correlation coefficient between the slope in the low-frequency region and the severity of internal short circuit reaches -0.87, which is the highest among all features. The slope value of the battery ohmic resistance in the low-frequency region is calculated by linearly fitting the low-frequency part (0.01 Hz - 0.35 Hz) of the EIS data, reflecting the characteristics of the internal charge transfer and diffusion process of the battery in the low-frequency region; the radius of the circle in the high-frequency region is obtained by circular fitting of specific data points in the high-frequency region (3 kHz - 10 kHz) of the EIS data, reflecting the kinetic characteristics of the battery electrode reaction in the high-frequency region; the minimum value of the imaginary part corresponding to the intermediate-frequency region (the real part is between 0.22 and 0.24) is obtained by analyzing the corresponding region; the real part value corresponding to when the imaginary part is 0 is obtained by interpolation calculation. These features together constitute the input of the input layer, providing a rich information source for subsequent model processing. In addition, it also includes normalizing the input data to eliminate the influence of features with different magnitudes on model training, ensuring the equal importance of each feature in training, and enabling the model to learn more stably.

[0169] Hidden layer, the first LSTM layer of the present invention, contains 256 LSTM units, the activation function is tanh, and the recurrent activation function is sigmoid; Dropout layer: The dropout rate is set to 0.3 to prevent overfitting. The second LSTM layer: contains 128 LSTM units, configured the same as the first layer, for further feature extraction; Dropout layer: The dropout rate is set to 0.3. Each LSTM unit selectively remembers and forgets the input data through the coordinated work of the input gate, forget gate, and output gate, thus effectively processing the long-term dependence relationship in the sequence data. Each LSTM layer further abstracts and extracts higher-level feature information based on the previous layer, enabling the network to gradually learn the temporal dependence relationship and complex patterns in the data.

[0170] Output layer, finally outputting the internal short circuit fault categories of the battery, including slight internal short circuit and severe internal short circuit. Through feature extraction and processing of the input data, the output layer maps the final learning result of the network to these two categories to complete the classification task. Specifically, a combination of a fully connected layer and a Softmax layer is used. Fully connected layer: The number of neurons is the number of classifications (in this example, 3: normal, slight internal short circuit, severe internal short circuit); Softmax activation function: Converts the output into a probability distribution. The internal short circuit fault detection architecture of the LSTM machine learning model is as Figure 7 shown.

[0171] Dataset construction: First, collect EIS data files from multiple sources. These files contain measurement data of the battery under different conditions, including but not limited to different states of charge (SOC), different battery characteristics (BD), and measurement results under different load resistances (such as 10 ohm and 200 ohm). For each EIS data file, perform data processing through a series of feature extraction algorithms. Specifically, use a custom feature extraction function to operate on the read data, such as performing linear fitting on the data in the low-frequency part, circular fitting on the data points in the high-frequency region, finding the minimum value points in the intermediate-frequency region, and calculating the real part value when the imaginary part is 0. Integrate the extracted feature values such as the slope value in the low-frequency region, the radius of the circle in the high-frequency region, the minimum value in the intermediate-frequency region, and the real part value when the imaginary part is 0 with the corresponding labels of the battery internal short-circuit fault categories (slight internal short circuit / serious internal short circuit) to form a complete dataset. The setting of these labels is based on prior knowledge or obtained through the analysis of experimental data. For example, according to the statistical characteristics of the experimental data and the experience of professionals, certain eigenvalue ranges are corresponding to the slight internal short-circuit category, while other eigenvalue ranges are corresponding to the serious internal short-circuit category. To increase the diversity of the data and the generalization ability of the model, perform data augmentation operations on the original dataset. When adding noise, select Gaussian white noise within a certain standard deviation range according to the noise characteristics of the EIS data and the requirements of the model robustness; the scaling range is set between 0.95 and 1.05 to increase the diversity of the data while maintaining the data characteristics and avoiding feature distortion caused by excessive data transformation. Repeat adding different noises to each original sample 5 times and scale the original samples within this range. Combine the augmented data with the original data to form the final dataset. In this way, construct a dataset with rich information and clear labels, providing a basis for subsequent model training.

[0172] Model training: Use the constructed dataset and divide it into a training set and a test set. Specifically, use the 5-fold cross-validation method to divide the dataset into 5 parts. Each time, select one part as the test set, and the remaining 4 parts as the training set. Repeat the training and testing process 5 times to make full use of the data and evaluate the stability of the model. Its training and validation curves are as Figure 8 shown. The training set data is used to train the LSTM model. During the training process, optimize the model performance by adjusting various parameters of the model, such as the number of units in the LSTM layer, the learning rate, the optimizer, etc. Use the Adam optimizer, and its key parameters are:

[0173] Initial learning rate: 0.005;

[0174] β1: 0.9, which is used for the first - moment estimation (first - order momentum) of the gradient to control the influence degree of historical gradients;

[0175] β2: 0.999, which is used for the second - moment estimation (second - order momentum) of the gradient to control the influence degree of the squares of historical gradients;

[0176] ε: 1e -8 , a numerical stability constant used to prevent division - by - zero errors

[0177] Gradient clipping threshold: 1.0, which is used to prevent gradient explosion. These default values are determined based on a large number of experiments and studies and can effectively optimize the model training process in most cases. This algorithm dynamically adjusts the learning rate of each parameter according to the first - moment estimation and second - moment estimation of the gradient to accelerate the convergence of the model and improve the training effect.

[0178] During the training process, the maximum number of training epochs is set to 800, the initial learning rate is 0.005, and a piece - wise learning rate adjustment strategy is adopted. The learning rate is decreased by 30% every 200 epochs to avoid falling into local optimal solutions in the later stage of training. At the same time, the gradient threshold is set to 1 to prevent gradient explosion problems.

[0179]

Specific methods for adjusting training parameters

[0180] Learning rate adjustment: The initial learning rate is set to 0.005; a step - decay learning rate attenuation strategy (StepDecay) is adopted; the learning rate is reduced by 30% every 200 epochs of training. The calculation formula is:

[0181]

[0182] where, initial lr represents the initial learning rate, new lr represents the new learning rate, epoch represents the current training epoch, and n represents the total number of training epochs.

[0183] The minimum learning rate threshold is set to 1e -6 , and the learning rate will not be decreased below this value; Batch Size adjustment: The initial batch size is set to 32. If the training is unstable, it can be adjusted to 16; if memory permits and the convergence speed needs to be improved, it can be adjusted to 64; the batch size adjustment is based on the fluctuation degree of the validation set loss function. When the volatility > 10%, the batch size is reduced

[0184] LSTM Layer Parameter Adjustment: Number of LSTM Layers: The basic configuration is 2 layers, which can be adjusted to 1 - 3 layers according to the data complexity; Number of LSTM Units: 256 units in the first layer and 128 units in the second layer, which can be adjusted within the range of [64, 128, 256, 512]; According to the performance of the validation set, when the validation accuracy improvement < 0.5% and the training accuracy > 98%, reduce the number of units to prevent overfitting.

[0185] Dropout Rate Adjustment: The initial Dropout rate is set to 0.3; Observe the accuracy gap between the training set and the validation set. When the gap > 5%, increase the Dropout rate to 0.4 - 0.5; When the gap < 2% and the validation set accuracy does not reach the target, the Dropout rate can be reduced to 0.2.

[0186] Adjustment of Other Hyperparameters:

[0187] Activation Function: By default, tanh is used, and ReLU or LeakyReLU can be tried; Sequence Length: Adjusted according to the time window, and by default, 10 time steps are used; Regularization Strength: Initially 0, L2 regularization (weight decay) can be added, and the strength range is [1e -5 , 1e -3 .

[0188] During the training process, a custom output function is used to store the training information in the training progress information manager for monitoring and subsequent analysis of the training process. The trained model is evaluated using the test set data. The test set data is input into the trained model to obtain the prediction results of the model. The prediction results are compared with the actual battery internal short - circuit fault categories to calculate the accuracy rate of the model. In addition, to more intuitively show the performance of the model in different category predictions, a confusion matrix is drawn, as Figure 9 shown.

[0189] The confusion matrix clearly shows the statistical information of various prediction situations, including the number of normal batteries correctly predicted as normal, the number of normal batteries misjudged as internal short - circuit, the number of internal short - circuit batteries correctly predicted as internal short - circuit, the number of internal short - circuit batteries misjudged as normal, etc.

[0190] Figure 9 The confusion matrix of... shows the classification performance of the model on the test set. In the matrix, the vertical axis represents the actual category and the horizontal axis represents the predicted category. From the matrix data, it can be seen that for severe internal short - circuit samples, the recognition accuracy rate of the model reaches 97.8% (45 / 46), for mild internal short - circuit samples, the recognition accuracy rate is 94.3% (33 / 35), and for normal battery samples, the recognition accuracy rate is 96.5% (55 / 57). Especially for severe internal short - circuit samples, the false negative rate (i.e., misjudged as normal or mild internal short - circuit) is only 2.2%, which is of great significance for ensuring battery safety.

[0191] Based on the above confusion matrix, the present invention calculates the following complete model performance evaluation metrics:

[0192] 1. Severe internal short circuit category:

[0193] Precision: 97.8% (45 / 46)

[0194] Recall: 97.8% (45 / 46)

[0195] F1 score: 97.8%

[0196] False Negative Rate: 2.2%

[0197] 2. Minor internal short circuit category:

[0198] Precision: 94.3% (33 / 35)

[0199] Recall: 94.3% (33 / 35)

[0200] F1 score: 94.3%

[0201] False Negative Rate: 5.7%

[0202] 3. Normal battery category:

[0203] Precision: 96.5% (55 / 57)

[0204] Recall: 96.5% (55 / 57)

[0205] F1 score: 96.5%

[0206] False Positive Rate: 3.5%

[0207] 4. Overall performance:

[0208] Average Precision: 96.5%

[0209] Weighted F1 score (assigning 1.5 times the weight to the severe internal short circuit category): 97.0%

[0210] Model convergence time: average 378 ± 42 rounds (about 6.3 hours, using NVIDIA T4 GPU)

[0211] Inference time: single prediction < 15ms (CPU implementation)

[0212] In the actual application scenario test, the time advance of the model for detecting internal short - circuit faults (i.e., the early warning time before the traditional method detects the fault) is on average 12.5 ± 3.2 hours, which has important value for preventing potential safety accidents.

[0213] According to the evaluation results, including the information reflected by the accuracy rate and the confusion matrix, further adjust the model structure or parameters, such as adding or reducing the number of LSTM layers, adjusting the Dropout rate, etc., to improve the model performance, so that the model can better adapt to different battery EIS data and accurately classify the internal short - circuit fault categories of the battery.

[0214] During the entire training process, use the validation set to verify the training effect of the model in real - time. The verification frequency is once every 50 epochs. Adjust the number of LSTM layers, Dropout rate, and learning rate to adjust the training parameters according to the validation set accuracy rate and the volatility of the loss function, ensuring that the model is continuously optimized during the training process.

[0215] Clear end criteria for the training process:

[0216] The present invention adopts the following precisely defined training end conditions:

[0217] Main end criteria:

[0218] Early Stopping strategy. When the validation set accuracy rate does not improve for 15 consecutive epochs (the improvement threshold is 0.1%), the training automatically stops; accuracy rate threshold. When the validation set accuracy rate reaches or exceeds 98.5%, the training stops; maximum number of training epochs: Regardless of the performance, the training is forced to stop when the number of training epochs reaches 800.

[0219] Auxiliary end indicators:

[0220] The loss function converges, and the change rate of the validation set loss function value is less than 0.01% for 10 consecutive epochs; confusion matrix indicator. When the recall rate of the severe internal short - circuit category reaches 99.5% and the precision rate reaches 98%, the training is preferentially ended; calculation resource constraint: The training stops when the training time exceeds the preset maximum allowable time (12 hours).

[0221] Model selection criteria:

[0222] Adopt the model with the best performance on the validation set as the final model.

[0223] The performance evaluation uses the weighted F1 - score, and the calculation formula is:

[0224]

[0225] Among them, precision represents precision rate, and recall represents recall rate.

[0226] Assign a weight of 1.5 times to the severe internal short - circuit category to improve the ability to identify high - risk states.

[0227] The specific implementation of the above training end conditions: Adopt a round - by - round evaluation scheme. After each round of training, conduct an evaluation and check the end conditions in the following order of priorities:

[0228] 1. First, check whether the accuracy of the validation set reaches the 98.5% threshold. If it reaches, immediately end the training;

[0229] 2. If the first condition is not met, check whether the accuracy of the validation set has not increased significantly (increase < 0.1%) for 15 consecutive rounds. If so, end the training;

[0230] 3. If the first two conditions are not met, check whether the recall rate of the severe internal short - circuit category reaches 99.5% and the accuracy rate reaches 98%. If so, end the training;

[0231] 4. If the first three conditions are not met, check whether the change rate of the loss function value of the validation set is less than 0.01% for 10 consecutive rounds. If so, end the training;

[0232] 5. If the first four conditions are not met, check whether the number of training rounds reaches 800 rounds or the training time exceeds 12 hours. If so, forcefully end the training.

[0233] This training end strategy with multiple conditions and multiple priorities ensures that the model reaches better performance while avoiding over - training, improves the training efficiency and guarantees the model quality.

[0234] Finally, through multiple trainings and validations, an LSTM model that can effectively classify internal short - circuit faults in battery EIS data is obtained, and the calculated average accuracy rate can reflect the comprehensive performance of the model on different data.

[0235] Through the collaborative work of the above LSTM model construction, dataset construction, model training, and auxiliary functions, the effective processing of battery EIS data and the classification of internal short - circuit faults are realized, improving the accuracy and efficiency of battery fault diagnosis, and providing important technical support for the safe use and maintenance of batteries.

[0236]

Example 4

[0237] 1. Model Deployment: Import the trained LSTM model into the cloud system (digital twin model). This cloud system, as a complex information integration and processing platform, has powerful data storage, processing, and transmission capabilities. To achieve this deployment, first package the trained LSTM model to ensure it can exist in a stable, efficient, and callable form in the cloud environment. During the import process, consider the model's compatibility and scalability to adapt to different cloud architectures and potential future model update requirements.

[0238] The cloud system of the present invention obtains the EIS data and environmental information of the battery in real time through the digital twin model interface. The system supports multiple data collection methods, including but not limited to the direct communication interface with the battery management system (BMS). The collected data includes the complete EIS test results of the battery (real part impedance and imaginary part impedance data), battery temperature, voltage, current, and key information such as charge and discharge status. The system preprocesses the collected data, including data cleaning, standardization, and feature extraction, to provide high-quality input data for the fault detection of the LSTM model. The data transmission process uses encryption technology to ensure data security. The system also backs up and stores historical data to support subsequent analysis and model optimization. The data collection interface obtains data from the battery management system (BMS) or other sensors at a certain frequency (such as per second, per minute, or a period set according to specific application scenarios) and preprocesses it, such as data cleaning, denoising, normalization, etc., to ensure the data quality input into the LSTM model.

[0239] In addition, to ensure data security and integrity, encryption technologies such as the SSL / TLS protocol are used during the data transmission process to prevent data from being tampered with or leaked during transmission. At the same time, the cloud system backs up and stores the data for subsequent analysis and retraining of the model. The stored data can be stored in a distributed file system or object storage service, such as HDFS or S3, etc., to ensure data reliability and accessibility.

[0240] 2. Fault Detection and Warning: Input the preprocessed EIS data and related environmental information into the LSTM fault detection model in the digital twin model to achieve the detection and warning of internal short circuit faults in lithium-ion batteries. The LSTM model will judge the state of the battery based on the characteristics of the input data, using the patterns and rules learned during the training process, and output the possibility of internal short circuit faults in the battery and the corresponding fault categories (minor internal short circuit / severe internal short circuit).

[0241] Generate a warning signal according to the detection result.

[0242] The fault warning system of the present invention realizes hierarchical warning according to the severity of internal short - circuit. For a slight internal short - circuit (the equivalent resistance of the internal short - circuit is between 50Ω and 200Ω), the system generates a routine maintenance prompt to notify relevant personnel to conduct inspections and maintenance within a specified time (such as within 48 hours). For a severe internal short - circuit (the equivalent resistance of the internal short - circuit is less than or equal to 50Ω), the system triggers a higher - level warning, including automatically cutting off the battery power supply and sending an emergency repair notice to professional maintenance personnel. These warning signals are transmitted to the monitoring platform through a reliable message queue protocol to ensure the real - time and reliability of information. The system also records the sending time and receipt confirmation of each warning signal for subsequent tracking and analysis.

[0243] Specific implementation of internal short - circuit fault warning:

[0244] 1. Trigger conditions for slight internal short - circuit warning (triggered when any of the following conditions is met):

[0245] The probability of the LSTM model outputting the category of slight internal short - circuit is greater than 70% and less than 90%;

[0246] The slope of the straight line in the low - frequency region is between 0.60 and 0.80 (41.4% to 55.2% of the normal value);

[0247] The radius value of the circle in the high - frequency region is between 0.0205Ω and 0.0230Ω (an increase of 13.9% to 27.8% compared to the normal value);

[0248] The displacement of the minimum point in the middle - frequency region is between 0.003Ω and 0.007Ω.

[0249] Trigger action: The system generates a yellow warning signal (priority P3), sends it to the monitoring platform through the Kafka message queue (QoS level is set to 1 to ensure at least one delivery), and at the same time sends an email to the maintenance personnel through the SMTP protocol, requiring inspections and maintenance within 48 hours. The battery data collection frequency is increased from once every 10 minutes to once every 5 minutes.

[0250] 2. Trigger conditions for severe internal short - circuit warning (triggered when any of the following conditions is met):

[0251] The probability of the LSTM model outputting the category of severe internal short - circuit is greater than 90%;

[0252] The slope of the straight line in the low - frequency region is less than 0.60 (less than 41.4% of the normal value);

[0253] The radius of the circle in the high - frequency region is greater than 0.0230Ω (an increase of more than 27.8% compared to the normal value);

[0254] The displacement of the minimum point in the middle - frequency region is greater than 0.007Ω.

[0255] Trigger action: The system generates a red warning signal (priority P1), which is sent to the monitoring platform through the Kafka message queue (QoS level set to 2 to ensure only one delivery), and at the same time, a text message is sent to the emergency repair team through the SMS gateway, and the power cut operation is immediately triggered through a control instruction (by calling the emergency control interface of the battery management system through the REST API). The emergency repair personnel must arrive at the scene within 2 hours to handle it, and the system starts the temperature monitoring emergency procedure with the sampling frequency increased to once every 5 seconds.

[0256] This warning signal not only contains information about the internal short - circuit fault in the battery, but also detailed information such as the time of the fault occurrence, the identification of the battery, and the severity of the fault. The generation of the warning signal follows certain rules. For example, when the probability predicted by the model that the battery has an internal short - circuit fault exceeds a set threshold (such as 70%), the corresponding warning level will be triggered. Different warning levels and handling measures are set for different internal short - circuit fault categories and different fault probabilities.

[0257] It is transmitted to the relevant monitoring platform in real - time through the cloud system, and this transmission process will follow a reliable message queue protocol, such as RabbitMQ or Kafka, to ensure the real - time and reliability of the message.

[0258] The warning system of the present invention adopts a distributed architecture to ensure high availability and reliability. The system transmits warning signals through a message queue middleware, and sets dedicated communication channels for warnings of different levels. The system supports multiple notification methods, including but not limited to graphical interface display, email notification, SMS notification, etc., to ensure that warning information can be delivered to relevant personnel in a timely manner. The system also sets clear performance indicators, including service availability targets, warning delay time control, etc., to ensure a quick response when an internal short - circuit fault occurs. The system monitors and records the transmission of warning signals, and continuously evaluates the detection performance of the model, triggering the model update process when necessary to keep the system in the best state all the time.

[0259] The specific technical implementation of the warning system includes the following aspects: First, in terms of the warning signal transmission architecture, the system uses Kafka as the message queue middleware, and configures 3 proxy nodes to ensure high availability. The system sets dedicated topics for warnings of different levels, including "battery.alert.minor" and "battery.alert.severe". The system partitions according to the battery ID to ensure that warning messages of the same battery can be processed in sequence. The message retention policy of the system sets the message retention period to 7 days, and the upper limit of the size of a single partition is 1GB.

[0260] Secondly, in terms of the early warning display system, the Web front-end of the system is developed using the React.js framework and adopts the Material UI component library. The mobile end of the system is developed using Flutter and supports Android 8.0+ and iOS 12.0+. The system uses the WebSocket protocol to push early warnings, ensuring that the latency is less than 200 ms. The notification methods of the system include email (SMTP), SMS (SMS API), mobile app push (FCM / APNS), and web browser notification (Web Push API).

[0261] Thirdly, in terms of operation and maintenance monitoring metrics, the service availability target of the system is 99.95%, with no more than 4.38 hours of downtime throughout the year. The end-to-end latency of the system's early warning, from detecting an anomaly to issuing an early warning, is no more than 3 seconds at P95. The system uses Prometheus to collect key performance indicators and Grafana for visualization. When the service availability of the system is lower than 99.5% or the latency exceeds 5 seconds, an alarm is automatically sent to the operation and maintenance team.

[0262] The relevant monitoring platform can be a web application, a mobile application, or a dedicated monitoring system. It receives early warning signals from the cloud system and presents them to relevant personnel in an intuitive way, such as through interface display, SMS notification, email notification, or push notification.

[0263] For example, when the model detects a risk of internal short circuit fault in the battery, it issues an early warning signal in a timely manner to notify relevant personnel to take actions, such as checking the battery status, taking repair measures, or adjusting the battery usage strategy. For minor internal short circuit faults, the monitoring platform can remind the operator to perform routine inspections and maintenance operations, such as checking whether the battery connections are loose and conducting simple performance tests on the battery; for severe internal short circuit faults, more advanced operations may be triggered, such as immediately cutting off the battery power supply, starting an emergency handling procedure, and notifying professional maintenance personnel to conduct on-site repairs, etc., to ensure the safe operation of the battery.

[0264] To further improve the reliability and availability of the system, the cloud system monitors and records the transmission of early warning signals, including information such as the sending time, receiving time, and whether the reception is successful for each early warning signal. At the same time, it also continuously monitors the detection performance of the model. By collecting actual fault data and early warning data, it evaluates the performance of the model. When the model performance deteriorates, it triggers a retraining process to ensure that the model is always in the best state.

Claims

1. An intelligent detection method for internal short - circuit faults in lithium - ion batteries based on electrochemical impedance spectroscopy, characterized in that: Obtain lithium - ion battery samples with different degrees of internal short - circuit, perform electrochemical impedance spectroscopy (EIS) detection on them, and output test data; Pre - process the test data, including deleting high - frequency data, extracting the real - part impedance (Z') and the imaginary - part impedance (-Z"); Draw a Nyquist plot, and extract the minimum value point in the intermediate - frequency region, the slope of the straight line in the low - frequency region, the radius of the circle in the high - frequency region, and the intersection point of the high - frequency region and the real - axis as characteristic values; Fuse the characteristic values with the ohmic - value labels as the input of the long short - term memory network (LSTM) deep - learning model; Construct and train the LSTM deep - learning model to identify and classify internal short - circuit faults in batteries; Deploy the trained LSTM model to the cloud system to achieve real - time detection and early warning of internal short - circuit faults in lithium - ion batteries.

2. The method according to claim 1, characterized in that: 1) Sample selection: Collect and obtain n lithium - ion battery samples, including m normal lithium - ion battery samples and multiple lithium - ion battery samples with internal short - circuit, where 1 < m < n; all sample specifications and models are exactly the same, the production batch is the same batch, and the environmental temperature during storage and transportation is controlled at 25 ± 2 °C, and the humidity is 45 ± 5% RH; the test equipment uses a high - precision electrochemical workstation with an accuracy of ±0.1%; 2) Test conditions: The entire EIS test process is carried out under open - circuit conditions to simulate the battery's static state; the excitation voltage range is set to 1 mV - 10 mV, and the frequency range is accurately set to 5 mHz - 10 kHz; 3) Sample installation and initialization: Install lithium - ion battery samples one by one into the fixture. After installation, place the samples on a test platform at 25 ± 0.5 °C and let them stand for more than 60 minutes to achieve thermal equilibrium and electrochemical steady state; 4) EIS test execution and original data storage: Start the EIS test program preset by the electrochemical workstation, and sequentially test the samples according to the set parameters; the device applies a sine - wave excitation signal, collects the current signal, and analyzes the electrochemical impedance data of the battery at different frequencies; at the same time, the supporting software of the electrochemical workstation automatically records the corresponding data of each sample, outputs it in a standardized data format and imports it into an Excel table. The Excel table header is set to include "Frequency (Hz)", "Real - part impedance (Ω)", "Imaginary - part impedance (Ω)", "Magnitude (Ω)", "Phase angle (°)", "Detection time (s)".

3. The method according to claim 1, characterized in that Including: 1). EIS test sampling rate setting Specifically as follows: In the frequency band from 0.01 Hz to 1 Hz, 10 points are collected for each frequency band; in the frequency band from 1 Hz to 10 Hz, 8 points are collected for each frequency band; in the frequency band from 10 Hz to 100 Hz, 6 points are collected for each frequency band; in the frequency band from 100 Hz to 1 kHz, 5 points are collected for each frequency band; in the frequency band from 1 kHz to 3 kHz, 4 points are collected for each frequency band; each sampling point is measured more than 3 times and the average value is taken; The frequency range of the EIS test for lithium - ion batteries is specifically divided as follows: - Low - frequency region: Defined as 0.01 Hz to 0.35 Hz, this frequency range reflects the characteristics of the diffusion process inside the battery and the lithium - ion transport in the solid phase; - Medium - frequency region: Specifically from 0.35 Hz to 500 Hz, corresponding to the range where the real - part impedance value in the Nyquist plot is between 0.22 Ω and 0.24 Ω; This region reflects the charge - transfer process at the electrode / electrolyte interface, and internal short - circuit causes the displacement of the position of the minimum point in this region; - High - frequency region: Specifically from 500 Hz to 3 kHz, which appears as a semicircle shape in the Nyquist plot; This region reflects the impedance of the SEI film inside the battery and the characteristics of the electrode reaction kinetics, and internal short - circuit causes the change of the semicircle radius and the shift of the intersection point with the real - part axis in this region; - Ultra - high - frequency region: Refers to the region above 3 kHz; The data is deleted during the pre - processing stage. 2). Data reading, pre - processing, and graph plotting Use the readmatrix function in MATLAB to read the EIS data stored in the Excel file; Each file contains the EIS data of one battery, and the data includes frequency, real - part impedance (Z'), and imaginary - part impedance (-Z"); Extract the real - part impedance (Z') and the imaginary - part impedance (-Z"), and reverse the sign of the imaginary - part impedance. Use the plot function in MATLAB to plot the Nyquist plot, and plot the Nyquist curves of the two batteries respectively. 3). Preliminary judgment of internal short - circuit characteristics Data preparation: Extract the values of the real - part impedance (Z') and the imaginary - part impedance (-Z") corresponding to the frequency range of 0.01 - 3000 Hz from the Excel data table, and export the extracted data as a text file; Data import and fitting: Import the exported text file into Zview software for fitting processing to generate a Nyquist plot and a Bode plot; Result analysis: Through the analysis of the Nyquist plot, specifically observe the following characteristic changes: - Slope of the curve in the low - frequency region (0.01 Hz - 0.35 Hz): - Intersection point of the high - frequency region with the real - part axis: - Semicircle radius in the high - frequency region: Extract the values of the real - part impedance (Z') and the imaginary - part impedance (-Z") corresponding to the frequency range of 0.01 - 3000 Hz from the data table again; Export the extracted data as a text file and plot the Nyguist curves of lithium - ion batteries with different degrees of internal short - circuit faults. The following sensitivity judgment criteria are established: According to the magnitude of the internal short - circuit equivalent resistance (R) value, the internal short - circuit faults of lithium - ion batteries are classified as follows: Severe internal short - circuit: 0 Ω < R ≤ 50 Ω, that is, the internal short - circuit equivalent resistance is less than or equal to 50 Ω, corresponding to the label value 2; In this state, the safety risk rating is at the high level of 5, and the following specific measures must be taken: (1) Automatically disconnect the main battery circuit within 15 seconds after detecting a severe internal short - circuit; (2) Activate the battery pack isolation device through the BMS controller; (3) Issue the highest - level warning; (4) Start the temperature monitoring emergency program, and increase the sampling frequency to once every 5 seconds; Minor internal short - circuit: 50 Ω < R ≤ 200 Ω, that is, the internal short - circuit equivalent resistance is greater than 50 Ω and less than or equal to 200 Ω; Corresponding to the label value 1; In this state, the security risk rating is medium, and the system needs to take the following specific measures: (1) The system generates a yellow warning signal; (2) Arrange inspection and maintenance within 48 hours (3) Increase the frequency of battery status monitoring from once every 10 minutes to once every 5 minutes (4) Record abnormal status data; Normal state: R>200Ω, that is, the internal short circuit equivalent resistance is greater than 200Ω or there is no internal short circuit phenomenon, and the corresponding label value is 0; Determination of internal short circuit characteristic points Feature 1. Determination of the minimum point: Screen all data points with real impedance values ​​between 0.22Ω and 0.24Ω in the Nyquist diagram, and use the traversal algorithm to find the point with the minimum imaginary impedance and its corresponding real impedance value Feature 2. Determination of the slope of the straight line in the low frequency region: Follow the steps below: (1) Filter all data points with a frequency range of 0.01 Hz to 0.35 Hz (2) The imaginary impedance value is used as the dependent variable (y-axis) and the real impedance value is used as the independent variable (x-axis) (3) Use the least squares method to perform linear fitting, output the fitted straight line equation, and extract the slope of the straight line; Feature 3. Fitting of the semicircle in the high frequency region: Follow the steps below: (1) Filter all data points with a frequency range of 500 Hz to 3 kHz (2) Use the least squares method to fit the circle based on equation (xa) 2 +(yb) 2 =r 2 (3) Transform the equation of the circle into: x 2 +y 2 =2ax+2by-c Where c = r 2 -a 2 -b 2 , a and b are the horizontal and vertical coordinates of the center of the circle respectively; (4) It is written in matrix form as Q = M·s, where Q is a vector containing the coordinates of the measurement points, M is a coefficient matrix consisting of the measurement points, and s = [-2a; -2b; c] is the parameter vector to be solved; by solving the linear equation system M / Q, the estimated value of s is obtained, and the coordinates of the center of the circle (a, b) and the radius r are calculated; (5) Specific fitting results: Use the least squares method in MATLAB to fit the circle, extract the radius of the circle as the feature point, and calculate the coordinate x of the center of the circle. fit and fit and the radius r fit ; Feature 4. Determination of the intersection of the high frequency region and the real axis: Follow the steps below: (1) Using the fitted circle equation, let y = 0 (2) Solve the equation to get two x values: (3) Select the smaller real part value as the eigenvalue and discard the larger real part value (4) Calculate the displacement.

4. The method according to claim 1, characterized in that: The feature fusion method is to organize the four extracted eigenvalues ​​into a structured feature vector, and then fuse it with the battery ohm resistance label and input it into the LSTM deep learning model; the specific implementation steps are as follows: The first step is to construct the feature vector: the feature vector X is further specified as: X=[Δ,k,r,x] in: Δ=|Z' 10 -Z' 200 | is the displacement of the minimum point in the mid-frequency region, in Ω; where Z' 10 Represents the real impedance value of a 10Ω internal short-circuit battery, Z' 200 It represents the real impedance value of a 200Ω internal short-circuit battery; k is the slope of the straight line in the low-frequency region, dimensionless; r is the radius of the high-frequency region circle, in Ω; x is the coordinate value of the intersection of the high frequency region and the real axis, in Ω; The second step is feature standardization: Since the four features have different physical meanings and dimensions, their numerical ranges vary greatly. In order to eliminate the impact of the dimension and ensure the stability of model training, each element in the feature vector X is standardized as follows to obtain X': x' i =(x i -m i ) / s i (15) where μ i and σ i They are the mean and standard deviation of the i-th feature in the training set respectively; Step 3: Feature and label fusion: The standardized feature vector X' is fused with the battery's ohmic resistance label R to form an augmented feature vector F = [X', R]; the R value is processed as follows: For normal status: R>200Ω, that is, the internal short circuit equivalent resistance is greater than 200Ω or there is no internal short circuit phenomenon, the label value is set to 0 For a slight internal short circuit: 50Ω < R ≤ 200Ω, that is, the equivalent resistance of the internal short circuit is greater than 50Ω and less than or equal to 200Ω, and the label value is set to 1 For a severe internal short circuit: 0Ω < R ≤ 50Ω, that is, the equivalent resistance of the internal short circuit is less than or equal to 50Ω, and the label value is set to 2 Fourth step, construction of temporal features: Organize the feature fusion vector F collected at different time points into a temporal feature matrix T, where each row represents the feature vector at a time point, and the matrix dimension is [n, 5], n is the temporal length, and the default setting is 10, representing the data of 10 consecutive time points collected; Fifth step, construction of batch processing: Construct a batch processing dataset for model training. Each batch contains 32 samples, and each sample is a temporal feature matrix, which constitutes the input data of the LSTM model.

5. The method according to claim 4, wherein: Construction and training of the LSTM model The adopted LSTM deep learning model includes the following clearly defined structures and parameters: Model structure: Input layer, receiving a 5-dimensional feature vector, specifically including: the coordinates (Z', Z”) of the minimum value point in the intermediate frequency region, the slope k value of the straight line in the low frequency region, the radius r value of the circle in the high frequency region, the intersection coordinate x value of the high frequency region and the real part axis, and the battery ohmic resistance R; Each neuron in the input layer corresponds to a feature dimension and receives the feature values after normalization processing; Hidden layer, the first LSTM layer, including 256 LSTM units, with the activation function tanh and the recurrent activation function sigmoid; Dropout layer: The dropout rate is set to 0.3; The second LSTM layer: including 128 LSTM units, with the same configuration as the first layer; Dropout layer: The dropout rate is set to 0.3; Output layer, finally outputting the battery internal short circuit fault categories, including slight internal short circuit and severe internal short circuit; Through feature extraction and processing of the input data, the output layer maps the final learning result of the network to these two categories to complete the classification task; Specifically, a combination of a fully connected layer and a Softmax layer is adopted. Fully connected layer: The number of neurons is the number of classifications, which is 3: normal, slight internal short circuit, severe internal short circuit; Softmax activation function: Converts the output into a probability distribution; Dataset construction: First, collect EIS data files from multiple sources, and these files contain the measurement data of the battery in different states, Model training: Use the constructed dataset and divide it into a training set and a test set; During the training process, adopt the Adam optimizer, and its key parameters are: Initial learning rate: 0.005; β1: 0.9, used to calculate the first-order moment estimate of the gradient, that is, the first-order momentum; β2: 0.999, used to calculate the second-order moment estimate of the gradient, that is, the second-order momentum; ε: 1e -8 , a numerical stability constant used to prevent division by zero errors Gradient clipping threshold: 1.0, to prevent gradient explosion; During the training process, set the maximum number of training epochs to 800, the initial learning rate to 0.005, adopt a piecewise learning rate adjustment strategy, reduce the learning rate by 30% every 200 epochs, and set the gradient threshold to 1; The minimum learning rate threshold is set to 1e -6 , below which it will no longer decrease; Batch size adjustment: The initial batch size is set to 32. If the training is unstable, it can be adjusted to 16; if the memory allows and the convergence speed needs to be improved, it can be adjusted to 64; the batch size adjustment is based on the volatility of the validation set loss function. When the volatility is >10%, it is considered unstable and the batch size is reduced; Number of LSTM units: 256 units in the first layer and 128 units in the second layer.

6. The method according to claim 1, characterized in that: The training end condition is defined as follows: End criteria: Early stopping strategy: when the accuracy of the validation set does not improve for 15 consecutive rounds of training, that is, the improvement threshold is 0.1%, the training stops automatically; accuracy threshold: when the accuracy of the validation set reaches or exceeds 98.5%, the training stops; maximum training rounds: forced to stop when the training rounds reach 800 rounds. Auxiliary end indicators: The loss function converges, and the change rate of the validation set loss function value for 10 consecutive rounds is less than 0.01%; Confusion matrix indicator, when the recall rate of the severe internal short circuit category reaches 99.5% and the accuracy rate reaches more than 98%, the training is terminated first; Computing resource constraints: training is terminated when the training time exceeds the preset maximum allowed time of 12 hours; The model with the best performance on the validation set is used as the final model; The specific implementation of the above training end conditions: adopt a round-by-round evaluation scheme, conduct an evaluation after each round of training, and check the end conditions in the following priority order: 1). First check whether the accuracy of the validation set reaches the 98.5% threshold. If so, end the training immediately. 2) If the first condition is not met, check whether the accuracy of the validation set has not improved significantly for 15 consecutive rounds (improvement < 0.1%). If so, terminate the training; 3). If the first two conditions are not met, check whether the recall rate of the severe internal short circuit category reaches 99.5% and the accuracy reaches 98%. If so, end the training; 4). If the first three conditions are not met, check whether the loss function value of the validation set has a change rate of less than 0.01% for 10 consecutive rounds. If so, terminate the training; 5). If the first four conditions are not met, check whether the training rounds have reached 800 or the training time has exceeded 12 hours. If so, the training is forced to end.

7. A lithium-ion battery internal short circuit fault detection system based on EIS detection and LSTM deep learning, comprising: The electrochemical impedance test module is used to obtain lithium-ion battery samples with different internal short circuit degrees, perform electrochemical impedance spectrum detection on them, and output test data; A data preprocessing and feature extraction module, used to preprocess the test data, including deleting high-frequency data, extracting real impedance (Z') and imaginary impedance (-Z"), drawing a Nyquist diagram, and extracting the minimum point in the intermediate frequency area, the slope of the straight line in the low frequency area, the radius of the circle in the high frequency area, and the intersection of the high frequency area and the real axis as characteristic values; A model building and training module, for fusing the characteristic value with the ohmic resistance label as an input to a long short-term memory network (LSTM) deep learning model, building and training the LSTM deep learning model to identify and classify short circuit faults in the battery; The cloud deployment and early warning module is used to deploy the trained LSTM model on the cloud system to achieve real-time detection and early warning of short-circuit faults in lithium-ion batteries.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the lithium-ion battery internal short circuit fault detection system is implemented.

Images