A battery external short circuit diagnosis method based on multi-scale fusion

Abstract

The application discloses a battery external short circuit diagnosis method based on multi-scale fusion, and belongs to the technical field of lithium ion battery safety monitoring and artificial intelligence fault diagnosis. The method only collects battery terminal voltage time series data, and constructs event level labels after preprocessing; a virtual and real hybrid sample generation module based on real voltage samples is established to generate virtual external short circuit samples and difficult negative samples; further, multi-scale voltage features such as statistics, segmentation, context, transient and morphology are extracted, and through input projection, residual feature enhancement, multi-token reconstruction, multi-head attention fusion and hierarchical prediction head, the output is the external short circuit occurrence probability, fault scene, severity and alarm confidence. The application does not depend on the measured current and temperature signals, and can alleviate the problems of real external short circuit sample scarcity, normal disturbance false alarm and high resistance early fault identification difficulty.

Description

Technical Field

[0001] This invention relates to the fields of lithium-ion battery safety monitoring, external short-circuit fault diagnosis, artificial intelligence models, and battery management systems. Specifically, it relates to a method for diagnosing external short-circuit faults in lithium iron phosphate batteries that uses only battery terminal voltage time-series data as online input, combined with physical information-based virtual-real hybrid sample generation, multi-scale voltage feature construction, hierarchical attention diagnostic models, and continuous window alarm confirmation. The method can be encapsulated as a computer program model and deployed on battery testing platforms, battery management systems, energy storage monitoring terminals, edge computing devices, or cloud diagnostic platforms. It is used to identify, classify, estimate the severity of, and output alarms for external short-circuit faults in lithium iron phosphate batteries under static, constant-current charging, and preset external short-circuit scenarios. Background Technology

[0002] Lithium iron phosphate batteries are widely used in new energy vehicles, energy storage systems, and other fields due to their high safety, long cycle life, and low cost. However, with the expansion of applications, electrical safety anomalies may occur in batteries under conditions such as loose connections, insulation degradation, damaged wiring harnesses, or accidental contact with external conductors. External short circuits, as a typical form of electrical abuse, can create an unexpected low-resistance path between the positive and negative terminals of the battery, leading to a rapid drop in terminal voltage, abnormally increased current, heat accumulation, and intensified polarization, which may, in severe cases, induce thermal runaway.

[0003] Existing battery fault diagnosis technologies have disclosed methods such as multi-source information fusion, micro-short-circuit resistance estimation, and time-series data classification. For example, document CN104714175A uses multi-source signals such as temperature, humidity, current, and voltage combined with a BP neural network for battery system fault diagnosis; document CN108363016B trains a neural network to estimate micro-short-circuit resistance using samples under different temperature, aging, and external resistance conditions; and document CN117388716B converts one-dimensional time-series data into two-dimensional images and uses a network model to complete battery pack fault classification. While these methods can achieve a certain degree of battery fault identification, there is still room for further optimization in addressing the high-resistance early weak characteristics of external short-circuit faults in lithium iron phosphate batteries, transient evolution across multiple operating scenarios, scarcity of real samples, discrepancies in the distribution of virtual and real samples, and the hierarchical output of fault detection, scene recognition, and severity estimation, even when using only battery terminal voltage signals. This invention addresses the external short-circuit voltage response mechanism by unifying physical information (hybrid samples of virtual and real data), multi-scale voltage characteristics, hierarchical diagnostic models, and online alarm confirmation mechanisms into a dedicated diagnostic framework. This framework aims to improve the applicability of early weak fault identification, disturbance rejection under complex operating conditions, and code model deployment. Summary of the Invention

[0004] The purpose of this invention is to provide a battery external short circuit diagnosis method based on multi-scale fusion. Under the conditions of limited real external short circuit samples, weak early voltage changes in external short circuits, and easy confusion between normal operating disturbances and short circuit events, this method uses a hybrid sample of virtual and real samples constrained by voltage response, multi-scale voltage characteristics, and a hierarchical diagnostic model to achieve external short circuit fault detection, fault scenario identification, severity estimation, and alarm confirmation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] Time-series data of lithium iron phosphate batteries under normal operation and external short-circuit conditions are collected. The time-series data includes battery terminal voltage data and selectable operating condition indicators; measured current and temperature signals are not required as inputs during the online diagnostic phase. The external short-circuit conditions include static external short circuit, constant current charging external short circuit, external short circuit with different external short-circuit resistances, different fault durations, and external short-circuit events under different initial charging states.

[0007] The time-series data is preprocessed, including outlier removal, missing value handling, time alignment, unified resampling, standardization, window slicing, and event label construction. The event labels include at least normal state labels, external short-circuit fault labels, fault scenario labels, and fault severity labels.

[0008] A hybrid virtual-real sample generation module is constructed. Based on the initial voltage value, baseline drift trend, voltage drop amplitude, fault duration, recovery behavior, and noise distribution extracted from real voltage samples, and combined with preset external short circuit equivalent severity parameters, the module generates virtual external short circuit samples and difficult negative samples under different external short circuit intensities, trigger times, durations, recovery behaviors, and noise conditions. These samples are then combined with real experimental samples to form a hybrid virtual-real training set.

[0009] Construct a multi-scale physical information feature vector. For each input window, extract statistical features, segmented evolution features, contextual temporal features, transient mutation features, and morphological features to form a feature vector for model input; the features include at least the maximum voltage difference, short-time window slope, voltage drop persistence, recovery degree, local slope mutation, morphological sharpness, and standardized residual features.

[0010] A multi-scale temporal fusion hierarchical diagnostic model is constructed. The model includes an input projection layer, a residual feature enhancement layer, a channel attention layer, a multi-token reconstruction layer, a multi-head self-attention fusion layer, a hierarchical prediction head, and a contrastive learning projection head. Preferably, the input projection layer maps 89-dimensional multi-scale voltage features to a 256-dimensional latent space; the residual squeezing excitation module enhances external short-circuit sensitive channels; the token reconstruction layer reconstructs the latent representation into multiple low-dimensional tokens; and the multi-head self-attention fusion layer learns the correlations between statistical, segmented, contextual, transient, and morphological features. The hierarchical prediction head includes at least a first prediction head for determining whether an external short circuit has occurred, a second prediction head for determining the fault scenario or fault type, and a third prediction head for estimating the fault severity.

[0011] The hierarchical diagnostic model is trained. During training, real voltage samples, virtual external short-circuit samples, and difficult negative samples are used simultaneously. Fault detection loss, fault scenario classification loss, severity estimation loss, and supervised contrastive learning loss are jointly optimized to form distinguishable representations of normal state, different external short-circuit scenarios, and samples of different severity in the feature space.

[0012] Online external short-circuit fault diagnosis is performed. Real-time collected battery time-series data is input into the preprocessing module and feature extraction module to obtain the window to be diagnosed; the window to be diagnosed is then input into the trained hierarchical diagnostic model, which outputs the probability of external short circuit occurrence, fault scenario, fault severity, and alarm confidence level.

[0013] Perform post-processing verification and alarm output. Combine continuous window consistency, fault duration, confidence threshold, scenario constraints, and severity level to verify the model output, suppress false alarms caused by isolated anomalies and normal operation transitions, and output the fault occurrence time, fault type, severity level, and safety control recommendations.

[0014] Furthermore, in the multi-scale temporal fusion hierarchical diagnostic model, the input feature vector can be an 89-dimensional physical information feature, which is first projected to a 256-dimensional latent space via linear mapping, then enhanced by two residual squeezing excitation modules, and subsequently reconstructed into multiple low-dimensional tokens and input into a multi-head self-attention module to capture the correlation between features at different scales. Subsequently, the first prediction head outputs a binary classification result of normal or external short-circuit fault, the second prediction head outputs categories such as stationary external short circuit, constant current charging external short circuit, or other preset scenarios, and the third prediction head outputs the severity level corresponding to the external short-circuit resistance or fault response intensity.

[0015] The method also includes a model package generation step, which encapsulates the trained model parameters, feature normalization parameters, label mapping relationships, threshold configuration files, and inference programs into a callable code model package for use by the battery management system, host computer, or edge computing terminal.

[0016] To enhance feasibility, the code model package may also include a model structure configuration file, an input / output interface specification file, and a runtime environment dependency file; wherein, the model structure configuration file is used to record the input feature dimension, latent space dimension, number of tokens, number of attention heads, and number of labels for each prediction head, and the input / output interface specification file is used to define the data fields, window length, sampling interval, and return result format during online inference.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects:

[0018] First, this invention models the external short-circuit fault mechanism, which not only determines whether a fault exists, but also outputs the severity of the fault and the alarm confidence level, making it easier for the battery management system to execute a graded protection strategy.

[0019] Second, the present invention utilizes physical information to generate virtual and real mixed samples to alleviate the problem of scarcity of real external short-circuit samples, and can cover the continuous response range from high-resistance early faults to low-resistance severe faults.

[0020] Third, this invention improves the comprehensive characterization ability of weak pressure drops, transient changes, persistent anomalies, and recovery behaviors by using multi-scale physical information features and attention fusion models.

[0021] Fourth, the present invention can reduce false alarms caused by normal operating condition switching, sampling noise and short-term disturbances through hierarchical prediction and post-processing confirmation mechanisms.

[0022] Fifth, the present invention can form independent code model packages, electronic devices and storage media solutions, which are easy to deploy on battery testing platforms, battery management systems, energy storage monitoring terminals or edge computing devices. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the consistency verification between the real voltage trajectory and the virtual voltage trajectory in this invention.

[0024] Figure 2 This is a schematic diagram of the multi-scale temporal fusion hierarchical diagnostic model structure of the present invention;

[0025] Figure 3 This is a schematic diagram illustrating the training process and changes in validation metrics of the hierarchical diagnostic model of the present invention.

[0026] Figure 4 This is a schematic diagram illustrating the external short-circuit fault diagnosis performance and feature space separability of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0028] The complete steps of the method of this invention are as follows:

[0029] First, data acquisition and sample construction were performed. Lithium iron phosphate batteries were selected as the test object, and voltage data were collected under normal resting, normal charging, different external short-circuit resistances, different trigger times, and different durations. Event-level labels were constructed based on the external short-circuit trigger time, end time, operating scenario, and short-circuit resistance level. These event-level labels included at least a normal state label, an external short-circuit fault label, a fault scenario label, and a severity label. Simultaneously, the influence of parasitic resistance in the short-circuit loop on the external short-circuit strength was considered, and the parasitic resistance of the short-circuit loop was set as R. p The nominal value of the external short-circuit resistor is R. nom The effective short-circuit resistance R applied to the battery terminal is then... eff Determine using the following formula:

[0030]

[0031] Among them, R eff The actual electrical load used to characterize external short-circuit faults is used to avoid inaccurate estimates of fault severity based solely on nominal short-circuit resistance.

[0032] Next, the data is preprocessed and input is standardized. The acquired battery time-series data undergoes outlier removal, missing value handling, time alignment, unified resampling, standardization, and window slicing. The voltage sequence after unified resampling is denoted as X, and is expressed by the following formula:

[0033]

[0034] Where, x i Let N represent the voltage value at the i-th sampling time, and N represent the input window length. By standardizing the window length and sampling interval, the impact of different sampling frequencies, different durations of experiments, and individual differences in different batteries on the model input can be reduced.

[0035] Then, a physical information virtual-real hybrid sample generation module is constructed. Since this invention only uses battery terminal voltage timing data as input during the online diagnostic phase, the virtual sample generator does not rely on measured current and temperature signals. Instead, it establishes a voltage-type virtual-real hybrid sample generator for external short-circuit fault diagnosis based on the initial voltage value, static drift trend, voltage change trend during charging, short-circuit trigger time, voltage drop amplitude, continuous abnormal characteristics, recovery behavior, and sampling noise distribution extracted from the real voltage samples. The generator generates virtual external short-circuit samples under different operating conditions and difficult negative samples similar to the external short-circuit waveform by adjusting the equivalent severity of the external short-circuit, fault trigger time, fault duration, voltage drop amplitude, recovery strength, contact disturbance, and sampling noise. The consistency verification between the real voltage trajectory and the physical information virtual voltage trajectory is as follows: Figure 1 As shown, the difficult negative samples include normal voltage change samples caused by static charging, constant current charging disturbance samples, short-time spike noise samples, and local fluctuation samples, which are used to improve the model's ability to distinguish between normal disturbances and real external short circuits.

[0036] Difficult negative samples and virtual external short-circuit samples are input into the model using the same preprocessing and feature extraction process, so that the model can learn the characteristics of sudden voltage drop during external short circuit and the characteristics of normal operating condition switching during the training phase, thereby reducing the risk of misjudging normal disturbances as external short-circuit faults.

[0037] The above-mentioned physical information virtual-real mixed sample generation process is used for offline training and model robustness enhancement. The external short circuit equivalent severity parameter, fall time constant, recovery ratio and noise standard deviation can all be obtained by statistical fitting of real voltage samples, offline calibration or preset level mapping. In the online diagnosis stage, only the battery terminal voltage window and the trained model parameters are called, and real-time measurement of short circuit current or battery temperature is not required.

[0038] To eliminate the influence of different initial battery voltage levels and sampling biases, the voltage window is relatively normalized, as expressed by the following formula:

[0039]

[0040] Where, μ pre and σ pre These are the mean and standard deviation of the voltage in the reference window before the fault is triggered, respectively, and ε is a minimal constant to prevent the denominator from being zero.

[0041] Furthermore, a voltage baseline trend function is established using normal samples, expressed as follows:

[0042]

[0043] Among them, f base (·) represents the baseline function obtained by fitting real normal voltage samples, Θb This represents the baseline parameters. The baseline function can be a linear function, a piecewise linear function, a polynomial function, an exponential smoothing function, or a moving average function.

[0044] When an external short-circuit fault is triggered, the fault voltage disturbance term is superimposed on the normal voltage baseline to generate a virtual external short-circuit voltage sample, expressed by the following formula:

[0045]

[0046] x i sim This represents the generated virtual voltage sample. ξ represents the abnormal voltage drop caused by an external short circuit. i The sampling noise term is represented by the following formula: The abnormal drop in external short-circuit voltage is expressed as follows:

[0047]

[0048] A s The voltage drop magnitude is represented by i0, the fault trigger sampling point is represented by Δt, and the sampling time interval is represented by τ. s This represents the voltage drop time constant.

[0049] To reflect the impact of different external short circuit severity levels on the voltage drop amplitude, the voltage drop amplitude can be set using the following formula:

[0050]

[0051] Among them, R eff The resistance parameters represent the severity of the equivalent external short circuit, where α and β are proportional parameters fitted from real voltage samples. It should be noted that R... eff In this embodiment, it is only used to control the severity of the fault in the virtual sample, and is not an input quantity that must be measured in the online diagnostic stage.

[0052] For the voltage sample recovered after a short-term external short circuit, the recovery term is expressed by the following formula:

[0053]

[0054] Where i1 represents the fault termination sampling point, ρ represents the recovery ratio, and τ r This represents the recovery time constant. By adjusting ρ and τ... r It can generate different types of voltage samples, such as those with complete recovery, those with incomplete recovery, and those with persistent abnormalities.

[0055] The sampling noise term is expressed by the following formula:

[0056]

[0057] Where, σ n This represents the noise standard deviation estimated from real normal voltage samples. Through the above method, this embodiment can generate virtual external short-circuit samples with different short-circuit severity, trigger times, durations, and recovery behaviors using only voltage data, and simultaneously generate normal disturbance samples similar to external short circuits, thereby expanding the model training data and improving the model's ability to distinguish highly confused samples. The consistency verification results between the real voltage trajectory and the virtual voltage trajectory based on physical information are as follows: Figure 1 As shown.

[0058] For the construction of multi-scale voltage features, statistical features, segmented evolution features, contextual time series features, transient change features, and morphological features should be extracted from the voltage time series for each voltage window to be diagnosed, forming a multi-scale voltage feature vector F, expressed by the following formula:

[0059]

[0060] Among them, F stat Indicates statistical characteristics used to describe the overall distribution of the voltage window, including the voltage mean, minimum, maximum, standard deviation, kurtosis, and range of variation; F seg This represents the piecewise evolution characteristic, used to describe the local variation process of voltage within different sub-intervals; F ctam This represents the contextual time window feature, used to characterize the voltage trend difference before and after a candidate event; F trans This indicates transient change characteristics, used to capture rapid voltage drops and differential changes at the moment of external short-circuit triggering; F morph It represents morphological characteristics and is used to describe the concentration, steepness, and recovery characteristics of voltage waveform changes.

[0061] In the specific calculation, the voltage window is first uniformly resampled and normalized to obtain a voltage sequence of consistent length. Then, the first-order difference of the voltage is calculated to obtain the voltage change between adjacent sampling points. Preferably, the transient difference standard deviation D... std It can be used to quantify the intensity of voltage surges, expressed by the following formula:

[0062]

[0063] Among them, a i a represents the voltage difference between adjacent sampling points. avg This represents the mean of the differential sequence. This feature can reflect whether the voltage undergoes a concentrated sudden change in a short period of time, which helps to distinguish between external short-circuit voltage collapse and normal slow operating conditions.

[0064] Furthermore, the segmented evolution features are obtained by dividing the voltage window into multiple continuous sub-intervals; the context time window features are obtained by comparing the voltage slope, mean, and fluctuation level before and after the candidate event; and the morphological features are obtained by statistically analyzing the concentration of voltage changes, the completeness of recovery, and the steepness of the waveform. Through these feature construction methods, stable and interpretable input can be provided for subsequent hierarchical diagnostic models using only voltage time-series data.

[0065] Next, a hierarchical diagnostic model is constructed by inputting the aforementioned multi-scale voltage feature vector F into the model. The model first maps the 89-dimensional feature vector to a high-dimensional latent space through an input projection layer, as shown in the following equation:

[0066]

[0067] Among them, W p and b p These are the trainable weight matrix and the bias term, respectively. Preferably, the input feature vector is 89-dimensional, and the mapped latent space is 256-dimensional, so as to unify voltage features of different types and dimensions into the same latent space.

[0068] Subsequently, the model employs a residual squeezing activation module to recalibrate channels and enhance nonlinearity in latent features. This residual squeezing activation module includes fully connected layers, nonlinear activation functions, dropout layers, channel weight generation structures, and skip connection structures, used to enhance external short-circuit sensitive features while maintaining training stability.

[0069] After completing the residual squeezing excitation process, the 256-dimensional latent features are reconstructed into four 64-dimensional tokens, which are then fused using a four-head multi-head self-attention module to learn the correlation between statistical features, segmented features, contextual features, transient features, and morphological features.

[0070] The model output has three levels of prediction heads: the first level is a fault detection head, used to determine whether an external short circuit has occurred; the second level is a fault type head, used to distinguish different external short circuit scenarios; and the third level is a severity estimation head, used to output the severity level of the external short circuit. Through this hierarchical output method, the model can simultaneously provide fault alarm, fault scenario, and severity information, facilitating the battery management system to execute graded protection strategies. The structure of the multi-scale time-series fusion hierarchical diagnostic model is as follows: Figure 2 As shown.

[0071] During training, both real and virtual voltage samples are used simultaneously, and the model parameters are jointly optimized using fault detection loss, fault scenario classification loss, severity estimation loss, and supervised contrastive learning loss. The changes in the objective function, validation metrics, and learning rate scheduling during model training are described below. Figure 3As shown. The total loss function is expressed by the following formula:

[0072]

[0073] Where L1 is the cross-entropy loss for binary fault classification, L2 is the focus loss for fault scenario classification, and L3 is the cross-entropy loss on the effective severity labels. c λ2, λ3, and λ4 are the supervised contrastive learning loss; λ2, λ3, and λ4 are weight coefficients, which can be set according to the distribution of the training set, preferably 2.0, 0.3, and 0.10 respectively. The supervised contrastive learning loss is expressed by the following formula:

[0074]

[0075] Among them, z i Let be the normalized embedding vector of the i-th sample, P(i) be the set of positive samples with the same class label as the i-th sample, A(i) be the set of candidate samples other than the i-th sample, and τ be the temperature coefficient, preferably 0.07. This loss is used to enhance the clustering of similar samples in the feature space and improve the separability between different fault types.

[0076] The trained model parameters, feature normalization parameters, label mapping files, and threshold configuration files are packaged into a code model package. During online operation, the battery management system or host computer continuously collects battery voltage time-series data and calls the preprocessing module and multi-scale feature extraction module according to a fixed time window to obtain the input to be diagnosed; then, the hierarchical diagnostic model is called to output the probability of external short circuit occurrence, fault scenario, fault severity, and alarm confidence.

[0077] In a preferred embodiment, the post-processing verification module verifies the model output based on continuous window consistency and an alarm threshold. Let p be the probability of an external short circuit occurring in the t-th window. t Let the alarm threshold be θ and the number of consecutive confirmation windows be M. Then, when p is satisfied in M ​​consecutive windows... t When the number of windows ≥ θ reaches a preset proportion, an external short-circuit alarm is output; otherwise, it is judged as a normal disturbance or a pending confirmation state. This method can suppress false alarms caused by isolated anomalies, sampling noise, and normal operating condition switching. When the model is validated using a real sample test set isolated by battery number, the model can output fault detection results, fault scenario categories, and fault severity. Its diagnostic performance and feature space separability are as follows: Figure 4 As shown.

[0078] The invention includes at least the following aspects: First, it incorporates baseline drift, sudden drop amplitude, duration, recovery behavior, and noise distribution from the real external short-circuit voltage trajectory into virtual sample generation, making the virtual samples match the natural response of the battery to an external short circuit; Second, it incorporates difficult negative samples into the training to specifically suppress false alarms caused by static charging, normal charging disturbances, and short-term noise; Third, it constructs statistical, segmented, contextual, transient, and morphological multi-scale features around the terminal voltage, making the algorithm features correspond to the external short-circuit voltage drop, continuous anomalies, and recovery behavior; Fourth, it sets up hierarchical prediction heads for fault detection, scene recognition, severity estimation, and alarm confidence; Fifth, it completes online alarm confirmation through continuous window consistency, forming a deployable code model package and system architecture.

[0079] The various embodiments in this specification are described in a related manner, and the same or similar parts between the embodiments can be referred to each other. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively brief, and the relevant parts can be referred to the description of the method embodiments.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made to sample generation parameters, feature combinations, number of model layers, number of attention heads, loss function weights, alarm thresholds, program module division, or deployment methods within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A battery external short-circuit diagnosis method based on multi-scale fusion, characterized in that, Includes the following steps: The battery terminal voltage time series data under normal operating conditions and external short circuit conditions are collected as necessary inputs for the online diagnostic stage. The measured current signal and measured temperature signal are not necessary inputs for the model. Event-level labels are constructed based on the fault trigger time, fault end time, operating scenario and short circuit severity. The battery terminal voltage time series data is subjected to outlier removal, missing value processing, time alignment, unified resampling, standardization, window slicing and label construction to form a fixed-length voltage window to be diagnosed. A virtual-real hybrid sample generation module is constructed based on real voltage samples to constrain voltage response. Virtual external short-circuit samples and difficult negative samples are generated according to the initial voltage value, baseline drift trend, voltage drop amplitude, fault duration, recovery behavior and noise distribution. Statistical features, segmented evolution features, contextual temporal features, transient change features, and morphological features are extracted from the voltage window to be diagnosed to form multi-scale voltage features. These multi-scale voltage features are used to characterize the local voltage drop, persistent anomaly, recovery degree, and morphological concentration after an external short circuit is triggered. A multi-scale temporal fusion hierarchical diagnostic model is constructed. The hierarchical diagnostic model includes feature projection, residual feature enhancement, channel recalibration, multi-token reconstruction, multi-head attention fusion, and hierarchical prediction output. It is jointly trained using real voltage samples and virtual voltage samples. The real-time collected battery terminal voltage time series data is input into the trained hierarchical diagnostic model, which outputs the probability of external short circuit occurrence, fault scenario, fault severity and alarm confidence. The model output is post-processed and confirmed by combining continuous window consistency, alarm threshold, fault duration and severity level, and the external short circuit alarm result is output.

2. The battery external short-circuit diagnosis method based on multi-scale fusion according to claim 1, characterized in that, The battery is a lithium iron phosphate battery; the battery terminal voltage time series data does not use measured current signals and measured temperature signals as model inputs during the online diagnostic stage; the standardization in the preprocessing includes relative normalization processing based on the reference window mean and reference window standard deviation.

3. The battery external short-circuit diagnosis method based on multi-scale fusion according to claim 1, characterized in that, The virtual-real hybrid sample generation module generates virtual external short-circuit voltage samples by superimposing a fault voltage disturbance term onto the normal voltage baseline. The fault voltage disturbance term is jointly determined by the fault trigger time, voltage drop amplitude, drop time constant, recovery ratio, recovery time constant, and sampling noise. The difficult negative samples include at least one of normal operating condition switching, short-time voltage spike noise, and local fluctuation samples. The virtual external short-circuit samples are used to supplement the real external short-circuit experimental samples, and the difficult negative samples are used to constrain the discrimination boundary between normal disturbances and weak external short-circuit windows.

4. The battery external short-circuit diagnosis method based on multi-scale fusion according to claim 1, characterized in that, The multi-scale voltage features include at least the voltage mean, voltage standard deviation, voltage kurtosis, minimum voltage, maximum voltage difference, segmented standard deviation, short-time window slope, context window slope, transient difference standard deviation, voltage drop persistence, recovery degree, morphological sharpness, and morphological concentration; the multi-scale voltage features are preferably 89-dimensional feature vectors.

5. The battery external short-circuit diagnosis method based on multi-scale fusion according to claim 1, characterized in that, The multi-scale temporal fusion hierarchical diagnostic model includes an input projection layer, a residual feature enhancement layer, a channel attention layer, a multi-token reconstruction layer, a multi-head self-attention fusion layer, a hierarchical prediction head, and a contrastive learning projection head. The input projection layer maps 89-dimensional feature vectors to a 256-dimensional latent space. The multi-token reconstruction layer reconstructs the 256-dimensional latent representation into multiple low-dimensional tokens, preferably four 64-dimensional tokens, and inputs them into the multi-head self-attention fusion layer.

6. The hierarchical prediction head includes at least a fault detection head for outputting normal or external short circuit, a fault scene identification head for outputting external short circuit scenarios, and a severity estimation head for outputting severity levels; the total loss function for model training is obtained by weighting the fault detection loss, fault scene classification loss, severity estimation loss, and supervised contrastive learning loss according to preset weights.

7. A battery external short-circuit diagnostic system based on multi-scale fusion, characterized in that, It includes a data acquisition module, a preprocessing module, a virtual-real mixed sample generation module, a multi-scale feature extraction module, a hierarchical diagnostic model training module, an online inference module, a post-processing confirmation module, and an alarm output module; wherein, the online inference module is used to output the probability of external short circuit occurrence, fault scenario, fault severity, and alarm confidence, and the post-processing confirmation module is used to confirm the alarm result based on continuous window consistency and alarm threshold.

8. An electronic device, computer-readable storage medium, or code model package, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores a computer program; the computer-readable storage medium stores the computer program; the code model package includes a model parameter file, a feature normalization parameter file, a label mapping file, an alarm threshold configuration file, and an inference program; when the computer program or the inference program is executed by the processor, the battery external short circuit diagnosis method based on multi-scale fusion as described in any one of claims 1 to 5 is implemented.

Images