Dual-path fusion lithium ion battery short circuit fault multi-level diagnosis method and system

Abstract

The present application belongs to the technical field of battery fault diagnosis, and particularly relates to a lithium ion battery short circuit fault multi-stage diagnosis method and system based on double-path fusion. The method comprises the following steps: calculating an offset spectrum of an electrochemical impedance spectrum relative to a baseline; extracting multi-class complementary features from the electrochemical impedance spectrum and the offset spectrum; inputting the multi-class complementary features into a first diagnosis model to obtain a health grade to which a lithium ion battery to be diagnosed belongs and a probability of each short circuit fault grade, i.e. a first probability vector; inputting the offset spectrum into a second diagnosis model to obtain a health grade to which the lithium ion battery to be diagnosed belongs and a probability of each short circuit fault grade, i.e. a second probability vector; and inputting the first probability vector and the second probability vector into a stacked fusioner to obtain a final short circuit fault grade diagnosis result. The present application adopts two complementary diagnosis paths, the error modes of which are complementary, and can be compensated after fusion, thereby realizing accurate classification of weak faults and serious faults.

Description

Technical Field

[0001] This invention belongs to the field of battery fault diagnosis technology, and particularly relates to a multi-level diagnosis method and system for short-circuit faults in lithium-ion batteries using a dual-path fusion approach. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles and large-scale energy storage systems due to their high energy density and long cycle life. However, with continuously increasing energy density, thermal runaway has become an increasingly serious safety issue. Thermal runaway is typically triggered by a progressively worsening internal defect process, with internal short circuits being a common cause. In the early stages of an internal short circuit, the short circuit is minimal, self-discharge is slow, and changes in macroscopic signals such as voltage and temperature are extremely weak. This provides conditions for early detection but also makes the fault difficult to identify using general diagnostic methods. Therefore, timely detection of early short-circuit faults and their severity classification are crucial for ensuring battery safety.

[0003] Electrochemical impedance spectroscopy (EIS), through small-amplitude sinusoidal excitation, can non-invasively observe charge transfer, solid-electrolyte interfaces, and diffusion processes over a wide frequency range. It can distinguish minute impedance changes caused by weak short circuits, making it suitable for early short-circuit fault detection. Recent advancements in online impedance spectroscopy measurement hardware have made acquiring impedance spectra during operation increasingly feasible. However, current research on applying EIS to short-circuit fault diagnosis is still in its early stages. Common limitations include single-temperature testing, narrow severity ranges, binary outputs, and decreased sensitivity with aging. It has not yet been able to simultaneously cover weak faults as low as approximately C / 6000 and achieve multi-level fault classification under operating conditions where temperature, health status, and state of charge change together. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a multi-level diagnostic method and system for short-circuit faults in lithium-ion batteries using a dual-path fusion approach. Based on multiple complementary features extracted from electrochemical impedance spectroscopy and the deviation spectrum of electrochemical impedance spectroscopy relative to the baseline, two complementary diagnostic paths are designed. The diagnostic results of the two complementary diagnostic paths are stacked and fused at the probabilistic level, thereby achieving accurate classification that takes into account both weak and severe fault boundaries. This effectively addresses the problem of weak and difficult-to-identify changes in early short-circuit fault electrical signals under conditions where temperature, battery state of health (SOH), and battery state of charge (SOC) change together.

[0005] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a multi-level diagnostic method for short-circuit faults in lithium-ion batteries using a dual-path fusion approach.

[0006] A multi-level diagnostic method for short-circuit faults in lithium-ion batteries using a dual-path fusion approach includes the following steps: Electrochemical impedance spectroscopy (EIS) of the lithium-ion battery under diagnosis was collected at multiple frequency points to determine the baseline of the EIS and calculate the deviation spectrum of the EIS relative to the baseline. Multiple complementary features are extracted from electrochemical impedance spectroscopy and deviation spectroscopy. These complementary features include at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectral features. By inputting multiple complementary features into the first diagnostic model, the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level are obtained, i.e., the first probability vector. The deviation spectrum is input into the second diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the second probability vector. The first probability vector and the second probability vector are input into a pre-trained stacked fusion unit, and the first probability vector and the second probability vector are fused to obtain the final short-circuit fault level diagnosis result of the lithium-ion battery to be diagnosed.

[0007] The second aspect of the present invention provides a multi-level diagnostic system for short-circuit faults in lithium-ion batteries with dual-path fusion.

[0008] A dual-path fusion multi-level diagnostic system for short-circuit faults in lithium-ion batteries includes: The deviation spectrum calculation module is configured to: acquire the electrochemical impedance spectrum of the lithium-ion battery to be diagnosed at multiple frequency points, determine the baseline of the electrochemical impedance spectrum, and calculate the deviation spectrum of the electrochemical impedance spectrum relative to the baseline. A multi-complementary feature extraction module is configured to extract multi-complementary features from electrochemical impedance spectroscopy and deviation spectroscopy, wherein the multi-complementary features include at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectral features. The first probability vector calculation module is configured to input multiple complementary features into the first diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the first probability vector. The second probability vector calculation module is configured to input the deviation spectrum into the second diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the second probability vector. The fusion module is configured to input the first probability vector and the second probability vector into a pre-trained stacked fusion unit, fuse the first probability vector and the second probability vector, and obtain the final short-circuit fault level diagnosis result of the lithium-ion battery to be diagnosed.

[0009] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in the first aspect of the present invention.

[0010] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in the first aspect of the present invention.

[0011] The above one or more technical solutions have the following beneficial effects: This invention provides a multi-level diagnostic method and system for short-circuit faults in lithium-ion batteries using a dual-path fusion approach. Based on multiple complementary features extracted from electrochemical impedance spectroscopy and the deviation spectrum of electrochemical impedance spectroscopy relative to the baseline, two complementary diagnostic paths are stacked and fused at the probability level. Since the feature sources and classification structures of the two paths are different, their error modes are complementary and can compensate for each other after fusion, thereby ensuring accurate classification of the boundary between weak and severe faults.

[0012] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is an overall flowchart of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the dual-path diagnostic framework according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the topology of the equivalent circuit model used in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram showing the high-frequency parasitic inductance compensation process and the comparison of the Nyquist spectrum before and after compensation in Embodiment 1 of the present invention. Figure 5 This is a schematic diagram illustrating the four types of complementary features in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the deviation spectrum and relaxation time distribution under different short-circuit severity conditions in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram illustrating the dataset construction process and synthetic impedance spectrum verification of Embodiment 1 of the present invention. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] Example 1 Existing diagnostic methods for early short-circuit faults in lithium-ion batteries mainly include time-domain electrical signal-based methods, model-based methods, and data-driven methods. The time-domain electrical signal-based method can detect faults in seconds to about an hour, but it is usually only effective for moderate to severe short circuits with a high leakage current ratio (e.g., higher than about C / 31), and it is mostly a binary judgment of "fault / no fault". Model-based methods can detect minor faults and estimate their intensity, but they require high model accuracy, are susceptible to interference and noise, and are mostly validated only under a single temperature and a single aging condition. Data-driven methods do not rely on precise models, but they generally suffer from insufficient fault severity discrimination, limited detection range for minor faults, and a lack of verification under combined temperature-SOH-SOC variation conditions. In particular, when the leakage current ratio drops to about C / 100 or even lower, voltage changes are almost indistinguishable from normal self-discharge and are further masked by polarization and operating condition drift, making it difficult for time-domain indicators to effectively identify faults.

[0019] To address the aforementioned issues, there is an urgent need for a diagnostic method that can reliably detect and classify weak early short-circuit faults under combined changing operating conditions, while minimizing the risk of misdiagnosing faults as healthy conditions.

[0020] Based on this, this embodiment discloses a dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries, which can effectively address the problem of weak and difficult-to-identify changes in early short-circuit fault electrical signals under operating conditions where temperature, SOH, and SOC change together.

[0021] The technical concept of this invention is as follows: Short-circuit faults alter the battery's impedance spectrum in two physically different ways: First, the external short-circuit resistance causes the overall impedance to decrease in an approximately first-order relationship, that is, the impedance deviation amplitude is approximately inversely proportional to the external short-circuit resistance. Among them, severe faults cause the entire Nyquist curve to shrink towards the origin, while weak faults hardly cause any observable impedance changes. Second, short circuits alter specific electrochemical processes, leaving subtle changes in certain frequency bands. These changes only become apparent after the impedance spectrum is subtracted from a healthy baseline under the same operating conditions.

[0022] For severe faults, the overall impedance drop is sufficient to distinguish them from a healthy state; for minor faults, the overall impedance change is less than that caused by operating condition drift, but the frequency band change after baseline subtraction can still reveal the fault; conversely, the large wideband impedance drop of severe faults will mask the aforementioned local frequency band characteristics.

[0023] A single classification structure cannot simultaneously address the identification needs of both minor and major faults. Therefore, this embodiment of the invention employs two independent diagnostic paths to capture the two types of changes mentioned above, and fuses their outputs at the probability level, so that the diagnostic results of the two paths compensate for each other and improve the overall diagnostic accuracy.

[0024] like Figure 1 As shown, the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries includes the following steps: S1. Collect the electrochemical impedance spectrum of the lithium-ion battery to be diagnosed at multiple frequency points, and use the health state impedance spectrum that matches the current operating conditions of the battery as the baseline to calculate the deviation spectrum of the electrochemical impedance spectrum relative to the baseline. S2. Extract multiple complementary features from the electrochemical impedance spectrum and / or the deviation spectrum, wherein the multiple complementary features include at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectrum features. S3. Input the multi-class complementary features into the first diagnostic model; the first diagnostic model first performs binary classification detection of the battery for fault and health to obtain the fault probability, then estimates the conditional probability of the battery belonging to each short-circuit fault level under fault conditions, and combines the fault probability and the conditional probability into a first probability vector. The first probability vector includes the probability that the battery belongs to a health level and the probability of each short-circuit fault level. S4. Input the deviation spectrum into the second diagnostic model. The second diagnostic model directly classifies the deviation spectrum into health levels and various short-circuit fault levels, and outputs a second probability vector. ; S5. Input the first probability vector and the second probability vector into the pre-trained stack fusion unit, and the stack fusion unit fuses the two probability vectors to output the short circuit fault level diagnosis result of the lithium-ion battery.

[0025] This embodiment acquires the electrochemical impedance spectroscopy of the battery to be diagnosed and calculates the deviation spectrum using a healthy baseline under the same operating conditions. It extracts complementary features such as equivalent circuit model parameters, relaxation time distribution, Nyquist geometry, and the deviation spectrum. A first diagnostic model performs binary classification detection of faults and health conditions, then estimates the conditional probabilities of each fault level under fault conditions and combines them into a first probability vector. A second diagnostic model directly classifies the deviation spectrum into different levels and outputs a second probability vector. Finally, a stacked fusion unit fuses the two probability outputs to produce the final diagnostic result. Because the two paths use different input features and classification structures, their error patterns are complementary and can compensate for each other after stacking and fusion. This method achieves reliable detection and multi-level classification of weak early short-circuit faults under various combined operating conditions, without missing any fault samples.

[0026] In this invention, the battery operating condition is determined by three dimensions: battery temperature, battery state of health (SOH), and battery state of charge (SOC). The two electrochemical impedance spectra are considered to be in the same operating condition if and only if they are identical in all three dimensions: temperature, SOH, and SOC.

[0027] In step S1: The baseline is the frequency-point average or median of multiple health spectra under the same operating condition: the baseline is the frequency-point average or median of multiple health state impedance spectra collected under the same temperature, health state, and charge state; when a certain operating condition lacks a health state impedance spectrum, in the temperature-health state-charge state space normalized according to each dimension, several operating conditions with baselines closest to the current operating condition are selected, and the baseline of the current operating condition is obtained by weighting the average of the inverse distance. High-frequency parasitic inductance compensation is performed before calculating the deviation spectrum; the deviation spectrum is subjected to channel-by-channel signed logarithmic compression based on the lowest fault level and channel-by-channel robust normalization, and then resampled to equally spaced logarithmic frequency points.

[0028] Electrochemical impedance spectroscopy was performed using an electrochemical workstation, sampling 150 frequency points at logarithmic intervals within the frequency range of 0.01 Hz to 10 kHz, with a sinusoidal excitation amplitude of 10 mV. Before each measurement, the battery was allowed to stand at the target temperature until the terminal voltage stabilized. After removing the high-frequency range affected by parasitic inductance, the remaining frequency points were resampled to 200 uniformly spaced logarithmic frequency points, serving as a unified frequency grid for subsequent feature extraction and model input.

[0029] Specifically, electrochemical impedance spectroscopy (EIS) of the lithium-ion battery under diagnosis is collected at multiple frequency points. Using the healthy state impedance spectrum, which matches the battery's current operating conditions, as a baseline, the specific formula for calculating the deviation spectrum is as follows: (1) in, The impedance of the battery to be diagnosed. For the same operating condition, the healthy baseline impedance, LV0 represents the angular frequency; the subscript LV0 indicates the health level, and LV1, LV2, and LV3, etc., represent the short-circuit fault levels from minor to severe, respectively.

[0030] Here, "State of Health (SOH)" characterizes the degree of battery aging (e.g., capacity retention) and is independent of whether the battery has experienced a short circuit; "healthy" specifically refers to the state where the battery has not experienced a short circuit, not that it has not aged. The baseline impedance for health is... Impedance spectra are collected from batteries that have been confirmed to be free of short-circuit faults, operating at the same temperature, in the same health condition, and under the same state of charge as the battery under diagnosis. Aged batteries, in their healthy state, still possess a corresponding baseline free of short-circuit faults.

[0031] The health baseline is pre-established during the training / calibration phase before the deployment of the diagnostic model and stored as a health baseline library indexed by operating conditions. Whether a battery is healthy (excluding short-circuit faults) is determined by its charge retention rate: after fully charging the battery and allowing it to stand for a set period, its charge retention rate is measured. If it is not lower than a set threshold, it is determined that no short-circuit fault has occurred (LV0). In one specific embodiment, referring to the requirements for room temperature charge retention capacity in GB / T 31486-2024 (the charge retention capacity after standing at room temperature for a specified number of days should not be less than 90% of the initial capacity, which corresponds to a leakage current ratio on the order of approximately C / 7200), and taking into account the lower bound of the measurement sensitivity of electrochemical impedance spectroscopy, approximately C / 6000, which is on the same order of magnitude as this threshold but slightly higher, is taken as the boundary between fault and health. It should be understood that when the measurement sensitivity allows, this boundary can also be taken as a lower leakage current ratio to obtain a more stringent determination. Since it is difficult to perform such long-term static measurements during online operation, the confirmation of healthy batteries, the collection of healthy baselines, and the collection of fault samples with different short-circuit severity are all completed in advance in the laboratory or at the factory stage.

[0032] Specifically, batteries of the same type that have passed the initial capacity and impedance consistency screening and have been confirmed to have no short-circuit faults according to the above criteria are selected. These batteries are then subjected to cyclic aging at various set temperatures. Multiple impedance spectra are collected at multiple health status checkpoints and at each state of charge level. The median or average of each frequency point is taken as the health baseline for that operating condition and stored in the baseline library. During diagnosis, a matching baseline is retrieved from the baseline library based on the current temperature, health status, and state of charge of the battery to be diagnosed. If a certain operating condition is missing, it is supplemented using the aforementioned distance-inverse weighted interpolation.

[0033] This embodiment classifies the severity of short-circuit faults in lithium-ion batteries into four levels: LV0 represents a healthy state; LV1 represents the weakest (lowest) fault level; LV2 is the second most severe level; LV3 is the highest severity level.

[0034] Before calculating the deviation spectrum, the method further includes performing high-frequency parasitic inductance compensation on the electrochemical impedance spectrum: In the high-frequency range where the impedance spectrum is inductive, the inductance value is estimated by linear regression of the imaginary part of the impedance and the angular frequency, and the estimated inductance contribution is subtracted from each measurement point.

[0035] After calculating the deviation spectrum, the deviation spectrum is subjected to channel-by-channel signed logarithmic compression and robust normalization: The deviation spectrum is organized from the same electrochemical impedance spectroscopy measurement into an input tensor containing 200 frequency points and 7 channels, as follows: The seven channels and their acquisition methods are as follows: Channel 1 is the real part of the deviation. This is equal to the real part of the compensated electrochemical impedance spectrum minus the real part of the healthy baseline impedance under the same operating conditions. Channel 2 is the imaginary part of the deviation. This is equal to the imaginary part of the compensated impedance spectrum minus the imaginary part of the healthy baseline impedance under the same operating conditions. Channel 3 is the deviation amplitude. The impedance magnitude is obtained by taking the real and imaginary parts of the compensated impedance spectrum and by taking the difference between the real and imaginary parts of the healthy baseline impedance. Channel 4 is the offset phase The phase angle is obtained by subtracting the real and imaginary parts of the compensated impedance spectrum from the baseline phase angle obtained from the healthy baseline impedance. Channel 5 is the logarithm of frequency. , where is the logarithm of the frequency at each frequency point; Channel 6 represents the relaxation time distribution deviation. It is equal to the relaxation time distribution obtained by solving the compensated impedance spectrum minus the relaxation time distribution of the healthy baseline under the same operating conditions. Channel 7 is the logarithm of the time constant. , where is the logarithm of each time constant point.

[0036] Channels 1, 2, 3, 4, and 6 are the deviation value channels that carry fault deviation information. Each channel is independently determined and subjected to signed logarithmic compression parameters and robust normalized statistics. The parameters are not shared between different channels. The compression parameters for each channel are determined based on the set low quantile of the non-zero absolute value of the fault sample of the lowest fault level on that channel, ensuring that the compressed deviation of the lowest fault level falls within a comparable numerical range to that of higher fault levels. Robust normalization is performed channel by channel, with all frequency points within the same channel sharing a set of medians and interquartile ranges to preserve the relative amplitude relationships between frequency bands. Channels 5 and 7 are coordinate channels marking the positions of each data point on the frequency axis and time constant axis, used only for position encoding and not involved in signed logarithmic compression and robust normalization. After removing the high-frequency range affected by parasitic inductance, all seven channels are resampled to uniformly spaced logarithmic frequency points, collectively forming a 200×7 deviation spectrum input tensor. Among them, channels 1 to 5, provided by the electrochemical impedance spectral deviation, constitute a 200×5 electrochemical impedance spectral deviation spectrum, and channels 6 and 7, provided by the relaxation time distribution deviation, constitute a 200×2 relaxation time distribution deviation spectrum.

[0037] Furthermore, given that the external short-circuit resistance is much greater than the magnitude of the battery's own impedance, a first-order approximation is made of the parallel relationship between the external short-circuit resistance and the battery impedance. The deviation spectrum approximately satisfies: (2) in, The impedance is the equivalent circuit model impedance for the battery's healthy state. This is the external short-circuit resistor.

[0038] S2. Extract multiple complementary features from the electrochemical impedance spectrum and / or the deviation spectrum, including at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectrum features.

[0039] All of the above complementary features are derived from different calculation transformations of the same electrochemical impedance spectroscopy measurement: the equivalent circuit model parameters, relaxation time distribution features and Nyquist geometric features are obtained by fitting, transforming or geometrically labeling the electrochemical impedance spectrum, the deviation spectrum features are calculated by the deviation spectrum, and the deviation spectrum itself is the difference between the electrochemical impedance spectrum and the healthy baseline impedance under the same operating conditions (see Equation (1)).

[0040] Various features are fed into different encoding structures according to their data formats: In the first diagnostic model, the deviation spectrum is encoded by a sequence encoder, the relaxation time distribution deviation is encoded by a multi-scale convolutional network, and the equivalent circuit model parameter deviation, Nyquist geometric features, relaxation time distribution peak features, and operating condition variables are encoded by a multilayer perceptron. Each encoded feature is fused by a multi-head self-attention mechanism. The equivalent circuit model parameter deviation, Nyquist geometric features, frequency band deviation statistics, spectral line shape features, operating condition variables, and cross-charge state features form a statistical feature vector that is fed into a gradient boosting decision tree. In the second diagnostic model, only the deviation spectrum tensor is used as input.

[0041] The deviation spectrum characteristics in step S2 include the normalized deviation characteristics obtained by dividing the deviation spectrum by the square of the impedance magnitude of the healthy baseline under the same operating conditions, and its amplitude is approximately proportional to... When extracting Nyquist geometric features, candidate arcs are sorted and aligned first by impedance amplitude and then by frequency position, so that impedance spectra with different numbers of arcs produce feature vectors with consistent positions. It can also include cross-charge state features calculated under the same temperature and health conditions and across multiple charge states.

[0042] Specifically: (1) Equivalent circuit model parameters: Fit each impedance spectrum as Equivalent circuit model of the structure, such as Figure 3 As shown, the impedances for ohmic, SEI film, charge transfer, and solid-state diffusion processes are characterized, respectively. The total impedance and finite-length diffusion (Warburg) impedance are as follows: (3) (4) The Levenberg–Marquardt fitting with physical constraints was used to obtain the results. , , , , , , , , , There are ten parameters in total. The changes in these parameters driven by operating conditions are on a much larger scale than the impedance disturbances introduced by minor faults. Therefore, what enters the model is the difference between the parameters and the healthy parameters under the same operating conditions, which constitutes the parameter deviation of the ten-dimensional equivalent circuit model.

[0043] like Figure 3 As shown, the symbol "||" in the structure indicates parallel connection, and the whole structure is a series branch, that is, it consists of a series ohmic resistor. The first parallel component and the second parallel component are connected in series to form the following structure: (i) The series ohmic resistance characterizes the internal ohmic resistance of the electrolyte, diaphragm, and current collector, corresponding to the intersection of the high-frequency end of the spectral line and the real axis on the Nyquist plot (the starting point of the spectral line); (ii) First parallel link From the first constant phase element With the first resistor Parallel configurations are used to characterize solid electrolyte interphase (SEI) membrane processes, wherein For SEI film resistance, By pseudo capacitor With diffusion index The description corresponds to the first downward-pressing arc in the high-mid frequency region of the Nyquist plot; (iii) Second parallel link By the second constant phase element With "second resistor" Series finite-length Warburg diffusion impedance The parallel connection of branches forms a characterizing process for charge transfer and solid-state diffusion. For charge transfer resistor, By pseudo capacitor With diffusion index describe, For finite-length Warburg diffusion impedance, derived from diffusion resistance diffusion time constant With diffusion index The description corresponds to the second charge transfer arc in the middle region of the Nyquist plot and the subsequent Warburg diffusion tail in the low-frequency region. The impedance of the constant-phase element is... Its total impedance and finite-length Warburg impedance are shown in equations (3) and (4), respectively.

[0044] Using Levenberg-Marquardt fitting with physical constraints to obtain the equivalent circuit model parameters is a mature practice in the existing technology, and will not be elaborated further here.

[0045] (2) Characteristics of relaxation time distribution: The relaxation time distribution is obtained from the electrochemical impedance spectroscopy as follows: The cell polarization impedance (i.e., the total impedance) is subtracted from the series ohmic resistance. The latter part is modeled as a relaxation time constant. The superposition of infinitely many fundamental relaxation processes continuously distributed above, located at the time constant The fundamental relaxation process at that point contributes a fundamental impedance. Its weight is Integrating over all time constants yields the relationship between impedance and distribution function (5).

[0046] By solving the first kind of Fredholm integral equation, the impedance is decomposed into a relaxation time constant. continuous distribution on : (5) In the formula, the measured impedance Given a known quantity, the kernel function Given a fixed kernel, the distribution function Let be the unknown function to be solved, then this expression is given by . This is a first-kind Fredholm integral equation with unknowns. When solving it, we will... Discretizing the equation at 200 equally spaced logarithmic time constant points transforms it into an ill-conditioned linear system of equations. This system is then solved using Tikhonov regularization, with regularization parameters selected through generalized cross-validation. The relaxation time distribution is then obtained from the measured electrochemical impedance spectroscopy. .

[0047] Each of the main peaks identified above represents an electrochemical process with a characteristic time constant inside the battery (such as charge transfer, SEI film relaxation, solid-state diffusion, etc.). The peaks and processes may not correspond strictly one-to-one. The peak height, peak area and time constant of each peak are extracted, arranged in a fixed order and missing peaks are filled with zeros to form a relaxation time distribution peak feature vector (21-dimensional in a specific embodiment). The value entered into the model is the difference between the relaxation time distribution peak feature and the healthy baseline under the same operating conditions.

[0048] (3) Nyquist geometric features: The Nyquist geometric features include at least one of the following: high-frequency real intercept, center, radius, chord length and vertex coordinates of the charge transfer arc, and inflection point coordinates, regression slope and intercept of the low-frequency diffusion (Warburg) tail. The specific process of extracting geometric features from electrochemical impedance spectroscopy is as follows: (i) High-frequency real intercept: After high-frequency parasitic inductance compensation in the electrochemical impedance spectroscopy, the imaginary part of the high-frequency end of the spectrum is taken. The real part coordinates corresponding to the zero-crossing (or closest to zero) points are used as the high-frequency real intercept. And record its corresponding frequency; (ii) Charge transfer arc characteristics: In the mid-frequency region, least-squares circle fitting is performed on the data points on the charge transfer arc, which is a downward-pressing semicircle. The center coordinates and radius of the fitted circle are obtained; the straight-line distance between the start and end points of the arc is taken as the chord length; the imaginary part of the arc is taken as the chord length. Take the coordinates of the point where the maximum value is located as the vertex coordinates; and record the frequency corresponding to the start and end points of the arc and the vertex. (iii) Low-frequency diffusion (Warburg) tail characteristics: In the low-frequency region, first determine the boundary between the charge transfer arc and the diffusion tail, take the data point where the slope of the spectral line changes as the inflection point and record its coordinates; then perform linear regression on the low-frequency diffusion tail data points after the inflection point to obtain the slope and intercept of the regression line.

[0049] When multiple arcs can be distinguished from the same spectral line, the candidate arcs are first sorted by impedance amplitude, ensuring that the arc with the main charge transfer occupies the first position. Then, they are sorted by frequency position to eliminate juxtapositions, thereby generating geometric feature vectors with consistent positions for impedance spectra with different numbers of arcs. After extraction and alignment with the arcs, the above features constitute a 27-dimensional geometric feature vector in a specific embodiment.

[0050] (4) Deviation spectrum characteristics and transcharge state characteristics: All impedance spectra and relaxation time distribution spectra are subtracted relative to a healthy baseline under the same operating condition to suppress operating condition drift. Deviation spectrum characteristics include normalized deviation characteristics, obtained by dividing the deviation spectrum by the square of the impedance magnitude of the healthy baseline under the same operating condition. The normalized deviation characteristics are approximately proportional to the reciprocal of the external short-circuit resistance to suppress impedance amplitude determined by the operating condition and highlight components related to short-circuit severity. Since healthy-state impedance varies with state of charge, the same short-circuit resistance produces different spectral deviations under different states of charge. To capture the changes in fault characteristics with state of charge, under the same temperature and healthy condition, the deviation spectra at multiple state of charge levels are calculated across several frequency bands for dispersion, trend of change, and deviation from the healthy baseline, and then concatenated into cross-state of charge characteristics.

[0051] S3. Input the aforementioned complementary features into the first diagnostic model; first perform binary classification detection of fault and health to obtain the fault probability. Then, the conditional probability of each short-circuit fault level under fault conditions is estimated by a cascaded classifier, and the first probability vector is constructed according to the following formula. : (6) in, For the probability of health level, Let k be the probability of the fault level. These are the conditional probabilities for the corresponding levels, and the sum of all conditional probabilities is 1.

[0052] That is, the first probability vector Construct it as follows: The probability that the battery is in good condition is equal to 1 minus the probability of failure. The probability of belonging to the k-th short-circuit fault level is equal to The product of the conditional probability of the level and the conditional probability of the level, and the sum of the conditional probabilities of all levels is 1.

[0053] The first diagnostic model includes a detection module and a cascaded fault classification module; In the detection module, the deviation spectrum is encoded by a sequence encoder, the relaxation time distribution deviation is encoded by a multi-scale convolutional network, and the equivalent circuit model parameters, Nyquist geometric features, relaxation time distribution features (i.e., peak features extracted from the relaxation time distribution), and operating condition variables are encoded by a multilayer perceptron. The operating condition information is modulated by a feature modulation layer to modulate each encoded feature. After each encoded feature is fused by a multi-head self-attention mechanism, it is input into a binary classifier along with the frequency band root mean square features, and the output is the fault probability. ; To prevent the frequency band energy information from being weakened during the attention fusion process, the root mean square features of each frequency band are input into the classification head separately without going through the attention layer.

[0054] The cascaded fault classification module consists of two cascaded gradient boosting decision tree binary classifiers. The input to each classifier is the same statistical feature vector, which includes equivalent circuit model parameter deviation, Nyquist geometric features, frequency band deviation statistics, spectral line shape features, operating condition variables, and cross-charge state features.

[0055] The specific data processing flow for the two cascaded classification stages is as follows: Phase 1: Using the statistical feature vector as input, a first-level binary classifier distinguishes the lowest short-circuit fault level LV1 from combinations of other higher short-circuit fault levels (i.e., the union of LV2 and LV3), and outputs the conditional probability that the battery belongs to LV1 under the fault condition. Conditional probability of belonging to a higher-level combination The sum of the two is 1; during training, this level uses all fault samples as the training set and labels them as "LV1" and "non-LV1". The second stage: Using the same statistical feature vector as input, a second-level binary classifier further distinguishes between LV2 and LV3, outputting the conditional probability of belonging to LV2 under higher-level combinations. Conditional probability of belonging to LV3 The sum of the two is 1; during training, this level only uses LV2 and LV3 fault samples as the training set, and labels them as "LV2" and "LV3"; the two stages above are calculated in parallel and independently of each other for the same input sample during inference. The second stage does not take the hard decision result of the first stage as a premise, and its output is the conditional probability of "under the condition that the battery belongs to a higher level combination"; the sequential relationship between the two stages is reflected in the subsequent chain multiplication of conditional probabilities, rather than the serial dependency in the calculation process.

[0056] Subsequently, the conditional probabilities of each short-circuit fault level under fault conditions are obtained by combining the probabilities: the conditional probability of belonging to LV1 is the output of the first stage. The conditional probability of belonging to LV2 is: and The product; the conditional probability of belonging to LV3 is and The product of the three conditional probabilities is 1. Finally, the failure probability output by the detection module will be... The conditional probabilities at each level mentioned above are combined according to equation (6) to form the first probability vector. That is, the probability that the battery belongs to health level LV0 is 1 minus the probability of failure. The probability of belonging to the k-th short-circuit fault level is the fault probability. The product of the conditional probability of the corresponding level.

[0057] In the above cascaded design, the first stage targets the severity boundary between LV1 and LV2, and the second stage targets the severity boundary between LV2 and LV3. Each boundary is handled by a dedicated binary classifier, which is trained using only the samples on both sides of the boundary. This concentrates all modeling capabilities on the distinction of the boundary, avoiding the risk of a single quad classifier spreading the probability across too many categories.

[0058] The decision threshold used by the detection module for binary classification of faults and health is determined through a constrained search on the validation data. Among the candidate thresholds, only those thresholds that simultaneously satisfy the minimum failure level (LV1) with a false negative rate lower than the first set upper limit and a health level (LV0) recall rate higher than the second set lower limit are retained. Among the thresholds that satisfy the constraints, the threshold that optimizes the comprehensive evaluation index of each level classification is selected. When there is no threshold that satisfies the constraints, the first set upper limit is relaxed and the search is repeated.

[0059] In summary, the detection module of the first diagnostic model integrates multimodal coding features through a multi-head self-attention mechanism and outputs the fault probability. Its judgment threshold is determined by constrained search to maximize the detection of weak faults while ensuring that faults are not misjudged as healthy. The cascaded fault classification module provides targeted two-class training at each level boundary to avoid a single multi-class classifier spreading the probability across too many categories.

[0060] S4. Input the deviation spectrum into the second diagnostic model, directly classify it into health level and various fault levels, and output the second probability vector. .

[0061] The second diagnostic model does not use the aforementioned failure probabilities, but instead directly predicts the four levels. This approach uses a multi-scale one-dimensional convolutional neural network to process the bias spectrum tensor.

[0062] S5. Input the two probability vectors into the pre-trained stacked fusion classifier and output the final diagnostic result. In this invention, the stacked fusion classifier refers to a fusion classifier trained in a stacking ensemble manner, that is, an upper-level classifier that takes the probability vectors output by the two diagnostic paths and the differences derived from them as input features and the diagnostic level as the training objective.

[0063] The input features of the stacked fusion unit consist of fifteen scalars, which are composed of five parts: the first probability vector. (4-dimensional), second probability vector (4-dimensional) Difference between two probability vectors (4-dimensional), the component with the highest probability value for each of the two paths (2-dimensional in total), and a flag indicating whether the predicted categories indicated by the two probability vectors are consistent (1-dimensional; 1 if the predicted categories are the same, 0 otherwise).

[0064] The data processing flow of the stacked fusion unit includes: During training, the first and second diagnostic models first generate out-of-division predicted probabilities on samples that were not part of their training. Then, they construct 15-dimensional input features using these out-of-division predicted probabilities in the manner described above, and train a stacked fusion engine using the actual short-circuit fault level of the sample as the label. During diagnosis, 15-dimensional input features are constructed for the same sample to be diagnosed and then input into the stacked fusion engine. The level with the highest probability value after fusion is used as the final diagnostic result. Since the input features simultaneously contain two probability vectors, their difference, their respective highest predicted probabilities, and a consistency indicator of the predicted categories, the stacked fusion engine can make targeted judgments when the predictions of the two paths are inconsistent. In particular, when one path is judged as faulty and the other as healthy, it can make more accurate decisions based on the difference in confidence between the two paths.

[0065] Furthermore, before training, the first diagnostic model, the second diagnostic model, and the stacked fusion unit augment the training data using at least one of the following methods: The parameters of the equivalent circuit model obtained by fitting the measured impedance spectrum under healthy conditions are perturbed within a preset physical range to generate enhanced samples. By combining the impedance of the equivalent circuit model in the healthy state with the target short-circuit resistance according to the parallel short-circuit impedance model, a synthetic impedance spectrum corresponding to different short-circuit severity is generated.

[0066] The parallel short-circuit impedance model is as follows: the fault impedance is equal to the product of the impedance of the equivalent circuit model in the healthy state and the target short-circuit resistance, divided by the sum of the two; the synthetic samples of intermediate fault levels without actual measured data are only used for training.

[0067] The parallel short-circuit impedance model is as follows: (7) Wherein, the target short-circuit resistance is the external short-circuit resistance value corresponding to the short-circuit fault level to be characterized by this synthetic sample. Its corresponding leakage current ratio is , This is the battery's nominal voltage. This is the rated capacity. For intermediate fault levels where no measured data is available, the impedance of the equivalent circuit model of the healthy state under this operating condition is used. With the target short-circuit resistance By combining the equations (7), the composite impedance spectrum of the intermediate fault level under the corresponding operating conditions can be obtained. These intermediate-level synthetic samples are only used for training.

[0068] Compared with the prior art, the technical advantages of this embodiment are as follows: First, two complementary diagnostic paths are employed and stacked and fused at the probabilistic level. The stepwise classification path, based on circuit model parameters and global spectral features, has a strong ability to identify large impedance changes caused by severe faults; the one-step classification path, based on local frequency band patterns of the deviation spectrum, is more sensitive to weak faults. Because the two paths have different feature sources and classification structures, their error patterns are complementary and can compensate for each other after fusion, thus ensuring accurate classification of both weak and severe fault boundaries.

[0069] Secondly, under various operating conditions with combined changes in temperature, health status, and state of charge, it has achieved reliable detection and multi-level classification of weak early short-circuit faults as low as approximately C / 6000. In experimental verification, none of the fault samples were misjudged as healthy, significantly reducing the safety risks caused by missed fault detection.

[0070] Third, by expanding the training data through parameter perturbation based on the equivalent circuit model and the parallel short-circuit impedance model, the problem of insufficient measured samples is alleviated, and the stability of the model under multiple operating conditions is improved.

[0071] Fourth, by using channel-by-channel signed logarithmic compression based on the lowest fault level, normalized deviation characteristics based on first-order physical approximation, and fault judgment threshold search with safety constraints, the detection sensitivity of weak faults is further improved, and the safety requirement that faults are not misjudged as healthy is met.

[0072] like Figure 1 and Figure 2 As shown in the figure, this invention provides a multi-level diagnostic method for short-circuit faults in lithium-ion batteries based on dual-path fusion of electrochemical impedance spectroscopy. The steps are described in detail below with reference to a specific embodiment.

[0073] I. Experimental Platform and Data Acquisition In this embodiment, three NCM811 soft-pack lithium-ion batteries (nominal capacity 10 Ah, nominal voltage 3.6 V) from the same production batch and screened for initial capacity and impedance consistency were selected. They were placed in independent constant temperature chambers at 15℃, 25℃ and 35℃ for cycle aging. Cycling was paused at six checkpoints: 100%, 96%, 92%, 88%, 84% and 80% of the healthy state. At each checkpoint, electrochemical impedance spectroscopy measurements were performed at five levels of state of charge: 5%, 20%, 50%, 70% and 95%, covering a total of 3×6×5=90 operating conditions.

[0074] The severity of the short circuit was simulated by connecting a precision metal film resistor in parallel between the battery terminals, determined by the leakage current ratio. : (8) in, For external short-circuit resistance, Nominal voltage, This is the rated capacity. This embodiment sets 15 assessment fault levels. The value ranges from 7.2 Ω to 2160 Ω, corresponding to leakage current ratios C / 20 to C / 6000. Approximately C / 6000 is used as the boundary between fault and health (this boundary is determined with reference to the requirements for the charge retention capacity of power batteries at room temperature: after a specified number of days of static storage at room temperature, the charge retention capacity should not be less than 90% of the initial capacity, which corresponds to a leakage current ratio on the order of approximately C / 7200; in this embodiment, at the same order of magnitude as this threshold, while considering the lower bound of the measurement sensitivity of electrochemical impedance spectroscopy, a value slightly higher than this threshold of approximately C / 6000 is used; minor leakage current ratios below this boundary fall within the range of normal self-discharge and measurement uncertainty, and are not considered short-circuit faults). The severity of the fault is divided into four levels: LV0 is a healthy state without external resistance; LV1 corresponds to... The range is 720~2160 Ω (approximately C / 2000 to C / 6000), which is the weakest fault range; LV2 corresponds to 72~720 Ω (approximately C / 200 to C / 2000); LV3 corresponds to The impedance range is 7.2~72 Ω (approximately C / 20 to C / 200), representing the highest severity level in this embodiment. The boundaries of these levels are defined based on the relationship that the amplitude of the deviation spectrum is approximately proportional to the reciprocal of the external short-circuit resistance (see Equation (2)). This ensures that LV2 and LV3 have similar range widths in the impedance deviation space, while LV1 has a narrower range due to the upper limit being limited by the fault and health boundaries. It should be understood that the above number of levels, resistance range, and boundary divisions are merely examples and can be set to other numbers of levels and division methods as needed.

[0075] Electrochemical impedance spectroscopy was performed using an electrochemical workstation, sampling 150 frequency points at logarithmic intervals within the frequency range of 0.01 Hz to 10 kHz, with a sinusoidal excitation amplitude of 10 mV. Before each measurement, the battery was allowed to stand at the target temperature until the terminal voltage stabilized. After removing the high-frequency range affected by parasitic inductance, the remaining frequency points were resampled to 200 uniformly spaced logarithmic frequency points, serving as a unified frequency grid for subsequent feature extraction and model input.

[0076] II. Impedance Spectrum Preprocessing The original impedance spectrum exhibits a positive imaginary part in the high-frequency region above approximately 1 kHz due to the parasitic inductance of the cable and connections. For example... Figure 4 As shown, in this embodiment, in the high-frequency range where the impedance spectrum is inductive, the imaginary part of the impedance and the angular frequency are linearly regressed to estimate the inductance value L, and the inductance contribution is subtracted from each measurement point so that the compensated impedance spectrum starts from the real axis intercept R0 and does not contain the high-frequency inductive tail. Then, the linear Kramers-Kronig test is used to verify the causality, linearity and stationarity of the compensated impedance spectrum.

[0077] For each operating condition, the median or average value of multiple health-state impedance spectra collected under the same temperature, health state, and charge state is used as the health baseline for that operating condition. Calculate the deviation spectrum according to formula (1) When a certain operating condition lacks a healthy state impedance spectrum, the three nearest operating conditions with baselines are selected in the temperature-health-charged state space that is normalized by the full range, and the healthy baseline of the operating condition is obtained by weighting by the inverse distance.

[0078] To avoid compressing weak faults into numerical noise through direct global normalization, this embodiment performs signed logarithmic compression on the deviation spectrum channel by channel: (9) Among them, the compression parameters of each channel The 10th percentile of the absolute value of the non-zero deviation of the weakest fault level (LV1) fault sample on the channel is taken (and a lower limit is set). After the above compression, the magnitude of the deviation of the weakest fault is approximately 1. The large deviation of the severe fault is suppressed after logarithmic compression, thus enabling the three fault levels to be effectively distinguished within the same numerical range. Subsequent robust normalization is performed on a channel-by-channel basis. All frequency points within the same channel share a set of median and interquartile range scaling, which suppresses outlier amplitudes while preserving the relative amplitude relationships between frequency bands and not destroying the spectral shape. After removing the high-frequency range affected by parasitic inductance, the deviation spectrum is resampled to uniformly spaced logarithmic frequency points (e.g., 200 points).

[0079] Specifically, the seven channels constituting the 200×7 input tensor and their acquisition methods from the original impedance spectrum are as follows: real part of the deviation The deviation is obtained by subtracting the real part of the healthy baseline impedance from the real part of the compensated impedance; the imaginary part is also obtained. The deviation amplitude is obtained by subtracting the imaginary part of the baseline from the imaginary part of the compensated impedance. The deviation phase is obtained by subtracting the magnitudes of the compensated impedance and the baseline impedance (obtained from their real and imaginary parts); The relaxation time distribution deviation is obtained by subtracting the phase angles of the compensated impedance and the baseline impedance (obtained from their real and imaginary parts); The frequency logarithm is obtained by subtracting the baseline relaxation time distribution from the relaxation time distribution obtained from the compensated impedance solution. Logarithm of time constant The coordinate channels are used to mark the positions of each point on the frequency axis and the time constant axis. The real part of the deviation, the imaginary part of the deviation, the deviation amplitude, the deviation phase angle and the relaxation time distribution deviation are the deviation numerical channels that carry the deviation information. Each channel is subjected to signed logarithmic compression and robust normalization by equation (9). The frequency logarithm and the time constant logarithm are coordinate channels and do not participate in the above compression and normalization.

[0080] Furthermore, this embodiment calculates a set of normalized deviation characteristics: the real and imaginary parts of the deviation spectrum are divided by the square of the impedance modulus of the healthy baseline under the same operating conditions (i.e., This makes the amplitude of the characteristic approximately proportional to the reciprocal of the external short-circuit resistance, thereby removing the impedance amplitude determined by the operating conditions from the fault amplitude.

[0081] Ultimately, the electrochemical impedance spectroscopy bias provides five channels: the real part of the bias, the imaginary part of the bias, the magnitude of the bias, the phase angle of the bias, and the logarithm of the frequency. The relaxation time distribution bias provides two channels: the bias and the logarithm of the time constant. Together, they constitute an input tensor with 200 frequency points and 7 channels.

[0082] III. Extraction of Complementary Features from Multiple Classes To separate fault-related impedance changes from operating condition drift, this embodiment extracts four complementary features from each impedance spectrum, such as... Figure 5 As shown in the figure, the equivalent circuit model parameters and geometric annotations are obtained by fitting the measured impedance spectrum; the synthetic samples of non-measured spectra are directly generated using the equivalent circuit model parameters on which they are based.

[0083] 1) Equivalent circuit model characteristics (10-dimensional): Fit each measured impedance spectrum as... Equivalent circuit model of the structure (e.g.) Figure 3 As shown, the structure and the physical meaning of each component are explained in step S2. Figure 3The detailed descriptions of the impedances (Equations 3 and 4) characterize the ohmic, SEI film, charge transfer, and solid-state diffusion processes, respectively. The total impedance and finite-length diffusion (Warburg) impedance are shown in Equations (3) and (4), respectively. The Levenberg-Marquardt fitting with physical constraints is used to obtain the impedances. , , , , , , , , , There are ten parameters in total. The changes in these parameters driven by operating conditions are on a much larger scale than the impedance disturbances introduced by minor faults. Therefore, what is entered into the model are the differences from the healthy parameters under the same operating conditions.

[0084] 2) Relaxation time distribution characteristics (21 dimensions): By solving the first kind of Fredholm integral equation as shown in equation (5), the impedance is decomposed into relaxation time constants. continuous distribution on In this embodiment, the discretization is performed at 200 equally spaced logarithmic time constant points, and the ill-conditioned inversion is stabilized using Tikhonov regularization, which selects the regularization parameter by generalized cross-validation. Each of the main peaks identified above represents an electrochemical process with a characteristic time constant inside the battery (such as charge transfer, SEI film relaxation, solid-state diffusion, etc.). The peaks and processes may not correspond strictly one-to-one. The peak height, peak area and time constant of each peak are extracted, arranged in a fixed order, and missing peaks are filled with zeros. In a specific embodiment, a 21-dimensional relaxation time distribution peak feature vector is constructed. The value entered into the model is the difference between the peak and the healthy baseline under the same operating conditions.

[0085] 3) Geometric Features Aligned with Arcs (27-Dimensional): Geometric features such as high-frequency real intercept, charge transfer arcs (center, radius, chord length, vertex coordinates, and start and end points), and low-frequency diffusion (Warburg) tails (inflection point coordinates, regression slope, and intercept) are extracted from each Nyquist plot. Since the number of distinguishable arcs varies with health status and temperature (new batteries often exhibit a single charge transfer arc, while aged batteries may exhibit a second arc related to interface aging), extracting them sequentially would cause overall misalignment of the arc-related features due to the presence of a second arc. Therefore, candidate arcs are first sorted by impedance amplitude to ensure the main charge transfer arc always occupies the first position, and then sorted by frequency to eliminate juxtapositions, resulting in geometrically consistent feature vectors for single and double arc impedance spectra. After extraction and arc alignment, these features constitute a 27-dimensional geometric feature vector in a specific embodiment. For the specific extraction methods of each geometric feature (zero-crossing point of the high-frequency real axis intercept, least-squares circle fitting of the charge transfer arc, determination of the inflection point of the low-frequency tail and linear regression), please refer to the detailed description in step S2.

[0086] 4) Deviation spectrum and trans-charge state characteristics under the same operating condition: All impedance spectra and relaxation time distribution spectra are subtracted from the healthy baseline under the same operating condition to suppress operating condition drift, such as... Figure 6 As shown, since the healthy state impedance changes with the state of charge, the same short-circuit resistance produces different spectral deviations under different states of charge. In order to capture the changes in fault characteristics with the state of charge, under the same temperature and healthy state, the deviation spectra of multiple state of charge levels (5 in this embodiment) are divided into several frequency bands. Within each frequency band, a set of statistics is calculated across each state of charge level and spliced ​​together to form cross-state of charge characteristics. In one specific embodiment, the frequency bands include the real part, imaginary part, and phase angle of the deviation in the low-frequency, mid-frequency, and high-frequency regions of the electrochemical impedance spectroscopy, as well as the low-frequency and high-frequency time constant regions of the relaxation time distribution, totaling 11 frequency bands. Within each frequency band, nine statistics are calculated across five states of charge levels, including the difference between two specified states of charge levels, standard deviation, concentration ratio (the ratio of the maximum absolute value to the average absolute value), slope at the high state of charge end, full-range difference (the difference between the highest and lowest states of charge levels), kurtosis, rank correlation coefficient, deviation from the mean of the same group, and deviation from the healthy baseline. The 11 frequency bands and nine statistics together constitute a 99-dimensional cross-state of charge feature. Since this feature requires the acquisition of spectra from multiple states of charge under the same temperature and healthy state, this framework is suitable for periodic or offline diagnosis. When only a single state of charge spectrum is available, this feature and cascaded classification can be omitted, degenerating into a diagnosis based on direct classification of the deviation spectrum.

[0087] In one specific embodiment, the statistical feature vector used by the cascaded fault classification module is 201-dimensional, which is composed of 102-dimensional planar statistical features and 99-dimensional transcharge state features. The planar statistical features include 10-dimensional equivalent circuit model parameter deviation, 27-dimensional Nyquist geometric features, 6-dimensional root mean square of frequency bands, 24-dimensional mean / maximum / 90th percentile / standard deviation of electrochemical impedance spectroscopy in several channels in the low-frequency and mid-frequency bands, 8-dimensional statistics of relaxation time distribution in the low-frequency and mid-frequency bands, 6-dimensional normalized deviation frequency band mean, 10-dimensional electrochemical impedance spectral line shape features (peak value, slope, curvature and ratio), 8-dimensional relaxation time distribution peak shape features (peak position, peak height, peak spacing and relative spacing), and 3-dimensional operating condition features (normalized temperature, normalized health state and their interaction terms). The 99-dimensional transcharge state features are constructed as described in "4) above with operating condition deviation spectrum and transcharge state features".

[0088] The frequency bands are divided as follows: the resampled logarithmic frequency grid is divided into three segments—low-frequency, mid-frequency, and high-frequency—according to frequency from low to high; the resampled logarithmic time constant grid is divided into three time constant segments—high-frequency, mid-frequency, and low-frequency—according to time constant from small to large. The root mean square (RMS) characteristics of the frequency bands are calculated on the deviation spectrum after signed logarithmic compression: for the low-frequency, mid-frequency, and high-frequency segments of the electrochemical impedance spectral deviation, the RMS values ​​of all four channels—real part, imaginary part, amplitude, and phase angle of the deviation—are taken to obtain three dimensions; for the three time constant segments of the relaxation time distribution deviation, the RMS values ​​of the relaxation time distribution deviation values ​​are taken to obtain three dimensions; a total of six dimensions are used to characterize the total energy of the deviation signal in each frequency band.

[0089] IV. First Diagnostic Model: Hierarchical Probability Path The first diagnostic model first estimates the probability of a fault occurring, and then estimates the conditional distribution of severity under fault conditions, such as... Figure 2 The path is shown in path A.

[0090] Since different types of features have different data formats and time scales, the detection module adopts a modal coding and fusion strategy. Its data processing flow includes three stages: modal coding, multi-head self-attention fusion, and detection.

[0091] In the modal coding stage, the six inputs are mapped to a unified embedding dimension by corresponding encoders: the electrochemical impedance spectral deviation spectrum (200×5) is linearly projected point by point, and then affinely modulated by the scaling and offset parameters generated by the feature modulation layer (FiLM) of the operating condition embedding. It is then modeled by the multi-layer multi-head Transformer encoder to model the dependence between frequency points and compressed into a modal vector by mean pooling; the relaxation time distribution deviation spectrum (200×2) is modulated by the operating condition, and then multi-scale features are extracted by multi-path parallel one-dimensional convolution branches with different kernel sizes. After being concatenated, it is compressed into a modal vector by convolution block and global mean pooling; the equivalent circuit model parameter deviation, Nyquist geometric features, relaxation time distribution peak features and operating condition variables are projected into modal vectors by their respective multi-layer perceptrons.

[0092] In the multi-head self-attention fusion stage, the above modal vectors and a learnable fusion token are combined to form a token sequence, which is input into the fusion Transformer. The modal tokens exchange information with each other through multi-head self-attention, and the output of the fusion token is taken as the multimodal fusion representation.

[0093] During the detection phase, the frequency-segment root mean square features fused from bypass multi-head self-attention are concatenated with the multimodal fusion representation, and then processed by a multilayer perceptron with Sigmoid activation to output the fault probability. To prevent the frequency band energy information from being weakened during the attention fusion process, the root mean square features of each frequency band are input to the detection head separately without passing through the attention layer. The detection head is trained using a focal loss as the training loss and employs the AdamW optimizer combined with a cosine annealing strategy. After training, the detection head is evaluated on the validation set using a constrained threshold search strategy described later. The fault determination threshold is set. The number of layers, number of attention heads, embedding dimension, convolution kernel size, loss function, and optimization hyperparameters mentioned above are only examples.

[0094] The cascaded fault classification module consists of a two-stage gradient boosting decision tree classifier (e.g., XGBoost), whose input is the aforementioned 201-dimensional statistical feature vector.

[0095] The first level distinguishes between LV1 and combinations of LV2 and LV3. During training, all fault samples are used as the training set, and the labels are set to "LV1" and "non-LV1". The conditional probability is output. and Finally, the failure probability and the conditional probabilities at each level are combined into a first probability vector according to equation (6); The second level further distinguishes between LV2 and LV3. During training, only LV2 and LV3 fault samples are used as the training set, and the labels are set to "LV2" and "LV3", outputting conditional probabilities. and .

[0096] The two levels perform parallel computation on the same statistical feature vector, with no sequential dependency. The resulting combination yields the conditional probabilities of each level under fault conditions: , , The sum of the three is 1; finally, the failure probability and the conditional probabilities at each level are combined into the first probability vector according to equation (6). .

[0097] In this embodiment, the three fault levels are arranged according to severity, and there are two severity boundaries between adjacent levels: the boundary between LV1 and LV2 is located at approximately C / 2000 ( ≈720Ω), the boundary between LV2 and LV3 is located at approximately C / 200 ( (≈72Ω). The two cascaded classifiers described above are trained for these two boundaries respectively: The first-level classifier targets the boundary between LV1 and LV2, using all fault samples as the training set and reclassifying the labels into "LV1" and "non-LV1". Its decision surface specifically learns to distinguish samples on both sides of this boundary. During training, healthy samples (LV0) are not included, so it is not affected by the healthy and faulty boundary. The second-level classifier targets the boundary between LV2 and LV3, using only LV2 and LV3 samples as the training set (excluding LV1) and classifying the labels into "LV2" and "LV3". Its decision surface specifically learns to distinguish samples on both sides of this boundary. The entire modeling capability of each classifier is concentrated on the single boundary it is responsible for. In contrast, a single four-class classifier needs to learn all the boundaries between the four categories simultaneously, distributing the probability across too many categories, thus weakening the discriminative ability at each boundary.

[0098] To improve the detection of weak faults and reduce the risk of them being misclassified as healthy, a constrained search is performed on the fault determination threshold of the detection module: Candidate thresholds are examined point-by-point within a set interval on the validation set. Only thresholds that simultaneously satisfy the condition that the false negative rate of the weakest fault level (LV1) is lower than the first upper limit and the recall rate of the healthy level is higher than the second lower limit are retained. From these constrained thresholds, the one that optimizes the comprehensive evaluation index for the four categories is selected. If no constrained threshold is found, the first upper limit is relaxed and the search is repeated. In this embodiment, the candidate interval is set to 0.2 to 0.8, the first upper limit is set to 30%, the second lower limit is set to 85%, and the comprehensive evaluation index is the macro-average F1 score. If no constrained threshold is found, the first upper limit is relaxed to 60% and the search is repeated.

[0099] V. Second Diagnostic Model: Sequence Probability Path The second diagnostic model does not use the aforementioned failure probabilities, but instead directly predicts four levels, such as... Figure 2Path B is shown. The model takes a 200×7 bias spectral tensor as input, with 7 channels as the feature dimension and 200 frequency points as the sequence dimension. It employs three sets of parallel convolutional branches at different scales (using small, medium, and large kernel sizes respectively to capture local, mid-range, and slowly varying spectral patterns). Each branch contains one-dimensional convolution, batch normalization, and activation, while maintaining a constant output length. The three feature paths are merged along the channel dimension and then processed by convolutional blocks and global average pooling to extract the global representation. Finally, a fully connected classifier with Softmax activation outputs four probability vectors, i.e., the second probability vector. The sum of the four terms is 1. This model uses cross-entropy loss as the training objective and is trained using the AdamW optimizer in conjunction with a cosine annealing strategy. The kernel size, number of layers, loss function, and optimization hyperparameters mentioned above are merely examples.

[0100] VI. Stacking and Fusion The two paths converge only at the stacked fusion unit. During diagnosis, each path outputs a four-class probability vector for the same sample; a gradient-boosting decision tree fusion unit comprehensively evaluates the two probability vectors, their difference, and prediction consistency information, thus enabling it to make a judgment distinct from when the predictions of the two paths are consistent, even when they are inconsistent. In a specific implementation, the fusion unit's input features consist of 15 scalars: the four-class probabilities of the two paths, their difference, the highest predicted probability of each path, and a flag indicating whether the predicted classes are consistent. During training, each base model first generates outlier predictions on samples not used in its training, and then uses these outlier predictions to train the fusion unit; finally, the highest probability level after fusion is used as the diagnostic result, without the need to set an additional decision threshold.

[0101] VII. Training Data Expansion and Training Settings To alleviate the problem of insufficient measured samples, this embodiment expands the measured impedance spectrum, such as... Figure 7 As shown: For each healthy operating condition, the parameters of the fitted equivalent circuit model are perturbed within the physical range calibrated by the degree of dispersion of repeated measurements of multiple healthy states under that operating condition to generate enhanced healthy samples. For each measured fault level, the fitted parameters are perturbed to generate enhanced samples while keeping its nominal short-circuit resistance unchanged. For intermediate fault levels without measured data, the target short-circuit resistance is introduced based on the parameters of the equivalent circuit model of the healthy state under each operating condition, according to the parallel short-circuit impedance model shown in Equation (7). Synthetic impedance spectra are generated, and these intermediate-level synthetic samples are used only for training. In one specific embodiment, there are 1710 measured spectra under 90 operating conditions, and after the above expansion, the dataset contains 15390 samples.

[0102] Model evaluation employs a six-fold cross-validation approach grouped by operating condition: each fold excludes all evaluation samples (including their enhancement samples) for the specified operating condition from base model training, hyperparameter selection, and fusion set fitting, ensuring strict isolation of evaluation samples from training. In a specific implementation, each neural network component is trained using the AdamW optimizer and cosine annealing strategy, the detection head uses Focal Loss, the sequence path uses cross-entropy loss, and the hyperparameters of the gradient boosting decision tree are selected through Bayesian optimization. The above network structure, loss function, and hyperparameters are merely examples and do not constitute a limitation of the present invention.

[0103] VIII. Diagnostic Results Under six-fold cross-validation grouped by operating condition, the framework described in this embodiment achieves an overall accuracy of approximately 96.9% and a macro F1 score of approximately 0.9704 on a dataset covering 90 operating conditions and fault severity ranging from C / 20 to C / 6000. In experimental verification, no fault samples were misclassified as healthy. In comparison, the overall accuracy using only the first diagnostic model (hierarchical probability path) is approximately 86.8%. This path correctly identifies all samples with the highest severity but misclassifies some minor faults as healthy. The overall accuracy using only the second diagnostic model (sequential probability path) is approximately 89.1%. This path performs well with minor faults but misclassifies some severe faults as healthy. After stacking and fusing the two paths, the overall accuracy improves to approximately 96.9%, and the misclassification of faults as healthy is completely eliminated. This indicates that the fused framework simultaneously utilizes the advantages of the hierarchical path at the severe fault boundary and the sequential path at the minor fault boundary. When evaluated on a subset consisting only of measured spectra, the framework still achieved an accuracy of approximately 98.2% and a macro F1 score of approximately 0.9760, indicating that its diagnostic performance is not dependent on synthetic data. The above values ​​are results from this embodiment under specific datasets and settings, and may vary depending on implementation conditions.

[0104] Example 2 This embodiment discloses a dual-path fusion multi-level diagnostic system for short-circuit faults in lithium-ion batteries.

[0105] A dual-path fusion multi-level diagnostic system for short-circuit faults in lithium-ion batteries includes: The deviation spectrum calculation module is configured to: acquire the electrochemical impedance spectrum of the lithium-ion battery to be diagnosed at multiple frequency points, determine the baseline of the electrochemical impedance spectrum, and calculate the deviation spectrum of the electrochemical impedance spectrum relative to the baseline. A multi-complementary feature extraction module is configured to extract multi-complementary features from electrochemical impedance spectroscopy and deviation spectroscopy, wherein the multi-complementary features include at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectral features. The first probability vector calculation module is configured to input multiple complementary features into the first diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the first probability vector. The second probability vector calculation module is configured to input the deviation spectrum into the second diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the second probability vector. The fusion module is configured to input the first probability vector and the second probability vector into a pre-trained stacked fusion unit, fuse the first probability vector and the second probability vector, and obtain the final short-circuit fault level diagnosis result of the lithium-ion battery to be diagnosed.

[0106] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0107] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in Embodiment 1 of the present invention.

[0108] Example 4 The purpose of this embodiment is to provide an electronic device.

[0109] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in Embodiment 1 of the present invention.

[0110] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0111] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0112] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-level diagnostic method for short-circuit faults in lithium-ion batteries using a dual-path fusion approach, characterized in that: Includes the following steps: Electrochemical impedance spectroscopy (EIS) of the lithium-ion battery under diagnosis was collected at multiple frequency points to determine the baseline of the EIS and calculate the deviation spectrum of the EIS relative to the baseline. Multiple complementary features are extracted from electrochemical impedance spectroscopy and deviation spectroscopy. These complementary features include at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectral features. By inputting multiple complementary features into the first diagnostic model, the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level are obtained, i.e., the first probability vector. The deviation spectrum is input into the second diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the second probability vector. The first probability vector and the second probability vector are input into a pre-trained stacked fusion unit, and the first probability vector and the second probability vector are fused to obtain the final short-circuit fault level diagnosis result of the lithium-ion battery to be diagnosed.

2. The dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in claim 1, characterized in that, The baseline is obtained as follows: When acquiring electrochemical impedance spectroscopy, the temperature, health status, and state of charge of the lithium-ion battery to be diagnosed should be determined. During the acquisition of electrochemical impedance spectroscopy, multiple impedance spectra in the same healthy state as the lithium-ion battery to be diagnosed were collected at the same temperature, in the same healthy state, and in the same charged state. Calculate the frequency-by-frequency average or frequency-by-frequency median of multiple health-state impedance spectra as a baseline; When a certain operating condition lacks a healthy state impedance spectrum, in the temperature-health state-charged state space normalized according to each dimension, select several operating conditions with baselines that are closest to the current operating condition, and obtain the baseline of the current operating condition by weighting the inverse distance. or, Before calculating the deviation spectrum, high-frequency parasitic inductance compensation is also performed on the electrochemical impedance spectroscopy. In the high-frequency range where the impedance spectrum is inductive, the inductance value is estimated by linear regression of the imaginary part of the impedance and the angular frequency, and the estimated inductance contribution is subtracted from each measurement point.

3. The dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in claim 1, characterized in that, This also includes channel-by-channel signed logarithmic compression and robust normalization of the bias spectrum: The deviation spectrum and the relaxation time distribution deviation are jointly organized into an input tensor with seven channels. The relaxation time distribution deviation is equal to the relaxation time distribution obtained by the electrochemical impedance spectroscopy minus the relaxation time distribution obtained by the baseline. The seven channels are, in order, the real part of the deviation, the imaginary part of the deviation, the deviation amplitude, the deviation phase angle, the logarithm of the frequency, the relaxation time distribution deviation, and the logarithm of the time constant; Among them, the real part of the deviation, the imaginary part of the deviation, the magnitude of the deviation, the phase angle of the deviation, and the deviation of the relaxation time distribution are the deviation numerical channels, and the logarithm of the frequency and the logarithm of the time constant are the coordinate channels; The compression parameters for each deviation numerical channel are determined based on the set low quantile of the non-zero absolute value of the fault sample with the lowest fault level in the current channel. The coordinate channels are used only for position encoding and do not participate in signed logarithmic compression transformation and robust normalization; Robust normalization is performed on a channel basis, with all frequency points within the same channel sharing a set of median and interquartile range, preserving the relative amplitude relationship between each frequency band; After removing the high-frequency range affected by parasitic inductance, the channels arranged along the frequency axis are resampled to logarithmic frequency points with uniform spacing, and the channels arranged along the time constant axis are resampled to logarithmic time constant points with uniform spacing.

4. The dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in claim 1, characterized in that, The specific process for determining the characteristics of the relaxation time distribution includes: The impedance is decomposed into a continuous distribution over the relaxation time constant; For each peak identified on the continuous distribution, the peak height, peak area and peak position time constant are extracted, arranged in a fixed order, and missing peaks are filled with zeros to form a relaxation time distribution feature vector. The specific process for determining the Nyquist geometric features includes: Extract high-frequency real intercept features, charge transfer arc features, and low-frequency diffusion tail features from each Nyquist plot; the charge transfer arc features include at least one of the center, radius, chord length, and vertex coordinates; the low-frequency diffusion tail features include at least one of the inflection point coordinates, regression slope, and intercept. The deviation spectrum characteristics include normalized deviation characteristics, which are obtained by dividing the deviation spectrum by the square of the impedance modulus of the healthy baseline under the same operating conditions. or, When extracting the features of charge transfer arcs, candidate arcs are first sorted by impedance amplitude so that the main charge transfer arc occupies the first position. Then, they are sorted by frequency position to eliminate parallelism, so that impedance spectra with different numbers of arcs produce geometric feature vectors with consistent positions.

5. The dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in claim 1, characterized in that, Multiple complementary features are input into the first diagnostic model. The specific data processing flow includes: The first diagnostic model first performs a binary classification detection of the battery as faulty or healthy to obtain the fault probability; Then estimate the conditional probability that the battery belongs to each short-circuit fault level under the fault conditions; The failure probability and conditional probability are combined into a first probability vector. ; in: The first diagnostic model includes a detection module and a cascaded fault classification module; The inputs to the detection module include the deviation spectrum, relaxation time distribution deviation, equivalent circuit model parameters, Nyquist geometric features, relaxation time distribution features, operating condition variables, and root mean square features of the frequency band. The relaxation time distribution deviation is equal to the relaxation time distribution obtained by the electrochemical impedance spectroscopy minus the relaxation time distribution obtained by the baseline. The operating condition variables consist of the battery's temperature, health status, and state of charge; the frequency band root mean square characteristics consist of the deviation spectrum and the root mean square of the relaxation time distribution deviation in each set frequency band. In the detection module: The bias spectrum is encoded by a sequence encoder, the relaxation time distribution bias is encoded by a multi-scale convolutional network, and the equivalent circuit model parameters, Nyquist geometric features, relaxation time distribution features, and operating condition variables are encoded by a multilayer perceptron. The operating condition information is modulated by a feature modulation layer to modulate each encoded feature. After being fused by a multi-head self-attention mechanism, each encoded feature is input into a binary classifier along with the frequency band root mean square features, and the output is the fault probability. ; The cascaded fault classification module includes at least two cascaded classification stages: First, in the first phase, the lowest short-circuit fault level is distinguished from the other short-circuit fault levels; The second stage further distinguishes the remaining short-circuit fault levels to obtain the conditional probability of each fault level. The first probability vector Construct it as follows: The probability that a battery is in good condition is equal to 1 minus the probability of failure. ; The probability that the battery belongs to the kth short-circuit fault level is equal to The product of the conditional probability of the k-th short-circuit fault level and the conditional probability of each level is 1.

6. The dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in claim 5, characterized in that, The decision threshold used by the detection module when performing binary classification of faults and health is determined through a constrained search on the validation data: Among the candidate thresholds, only those thresholds that simultaneously satisfy the minimum failure level with a false negative rate lower than the first set upper limit and a health level recall rate higher than the second set lower limit are retained, and the threshold that makes the comprehensive evaluation index of each level classification optimal is selected from the thresholds that satisfy the constraints. If no threshold that satisfies the constraints is found, relax the first upper limit and search again.

7. The dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in claim 1, characterized in that, The input features of the stacked fusion unit include: a first probability vector, a second probability vector, the difference between the first probability vector and the second probability vector, the component with the highest probability value in each of the first probability vector and the second probability vector, and a flag indicating whether the predicted categories indicated by the two probability vectors are consistent. During training, the first diagnostic model and the second diagnostic model first generate out-of-bounds prediction probabilities on samples that were not involved in training, and then use the out-of-bounds prediction probabilities to construct input features to train the stacked fusion engine. During diagnosis, the highest probability value after fusion is used as the final diagnostic result; or, Before training, the first diagnostic model, the second diagnostic model, and the stacked fusion unit augment the training data using at least one of the following methods: The parameters of the equivalent circuit model obtained by fitting the measured impedance spectrum under healthy conditions are perturbed within a preset physical range to generate enhanced samples. The impedance of the equivalent circuit model in the healthy state and the target short-circuit resistance are combined according to the parallel short-circuit impedance model to generate a synthetic impedance spectrum corresponding to different short-circuit severity. The parallel short-circuit impedance model is as follows: The fault impedance is equal to the product of the impedance of the equivalent circuit model in the healthy state and the target short-circuit resistance, divided by the sum of the two.

8. A multi-level diagnostic system for short-circuit faults in lithium-ion batteries with dual-path fusion, characterized in that, include: The deviation spectrum calculation module is configured to: acquire the electrochemical impedance spectrum of the lithium-ion battery to be diagnosed at multiple frequency points, determine the baseline of the electrochemical impedance spectrum, and calculate the deviation spectrum of the electrochemical impedance spectrum relative to the baseline. A multi-complementary feature extraction module is configured to extract multi-complementary features from electrochemical impedance spectroscopy and deviation spectroscopy, wherein the multi-complementary features include at least two of the following: equivalent circuit model parameters, relaxation time distribution features, Nyquist geometric features, and deviation spectral features. The first probability vector calculation module is configured to input multiple complementary features into the first diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the first probability vector. The second probability vector calculation module is configured to input the deviation spectrum into the second diagnostic model to obtain the health level of the lithium-ion battery to be diagnosed and the probability of each short-circuit fault level, i.e., the second probability vector. The fusion module is configured to input the first probability vector and the second probability vector into a pre-trained stacked fusion unit, fuse the first probability vector and the second probability vector, and obtain the final short-circuit fault level diagnosis result of the lithium-ion battery to be diagnosed.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the dual-path fusion multi-level diagnosis method for short-circuit faults in lithium-ion batteries as described in any one of claims 1-7.

Images