A method for testing failure of a lithium ion battery

Abstract

The application discloses a fault test method of a lithium ion battery, and relates to the technical field of intelligent testing, which comprises the following steps: acquiring external signals and internal signals of the battery in a test process, and generating a standard label vector of a battery health state based on the internal signals; the standard label vector comprises a solid electrolyte interface film health degree, an active lithium inventory loss, a lithium precipitation risk index and a positive electrode health degree; the time sequence data of the external signals are paired with the standard label vector to construct a training sample database; a benchmark diagnosis model adopting an encoder-multitask head architecture is trained by using the training sample database; the ability of the benchmark diagnosis model is transferred to an online diagnosis model through knowledge distillation; the external signals are input into the online diagnosis model, and an online prediction result of the battery health index is output; and fault early warning is performed according to the online prediction result and a preset threshold value.

Description

Technical Field

[0001] This invention relates to the field of intelligent testing technology, and in particular to a fault testing method for lithium-ion batteries. Background Technology

[0002] As a core component in current energy storage and power applications, the safety, reliability, and lifespan prediction of lithium-ion batteries are of paramount importance. During battery research and development, quality inspection, and use, testing is necessary to assess their performance and health status. However, traditional battery fault testing and health status assessment methods have significant limitations, making it difficult to meet the need for accurate and real-time monitoring of early-stage, latent faults.

[0003] Existing methods typically rely on overall battery performance parameters, such as capacity decay or increased internal resistance, as criteria for assessing health. Changes in these macroscopic parameters are often the result of multiple internal aging mechanisms working together, exhibiting significant hysteresis. When observable and significant changes occur in capacity or internal resistance, irreversible damage may have already accumulated inside the battery, even approaching safety boundaries such as thermal runaway. Therefore, relying solely on macroscopic parameters cannot provide early warning of faults, let alone root cause analysis of performance degradation. To explore the internal aging mechanisms of batteries, researchers need to use sophisticated instruments such as electrochemical impedance spectroscopy, differential voltage analysis, relaxation testing, and calorimetry for offline diagnosis. These methods can acquire internal signals reflecting the battery's internal state, but the testing process is cumbersome, time-consuming, and costly, and usually requires interrupting normal battery operation, making it impossible to integrate into continuous testing or usage processes for real-time online monitoring. With the development of artificial intelligence technology, using readily available external signals during battery operation to predict its health status has become a research hotspot. However, existing methods still face many challenges. Most models only predict single aging patterns or health indicators, lacking the ability to collaboratively diagnose multi-dimensional, coupled aging states. The training of the model relies on data of known aging patterns, and it lacks effective detection and early warning capabilities for unknown failure modes that may suddenly appear during the testing process and are not covered by the training data. This constitutes a potential blind spot in test safety. Summary of the Invention

[0004] This application provides a fault testing method for lithium-ion batteries, addressing the problems of delayed fault warnings, inability to perform real-time online diagnosis, and lack of collaborative diagnosis of multi-dimensional aging states and detection of unknown faults in existing technologies. Based on external signals, it enables early online prediction and assessment of the battery's multi-dimensional health status, further completing the attribution analysis of overall capacity decay, and detecting and warning of unknown or sudden faults, thereby effectively improving the safety, foresight, and diagnostic depth of the battery testing process.

[0005] This application provides a fault testing method for lithium-ion batteries, including: S1: Acquire external and internal signals of the battery during the test process, and generate a standard label vector of the battery health status based on the internal signals; the standard label vector includes: solid electrolyte interface membrane health, active lithium inventory loss, lithium plating risk index, and cathode health. S2: Pair the time-series data of external signals with standard label vectors to construct a training sample database; S3: Using the training sample database, train a benchmark diagnostic model with an encoder-multi-task head architecture. The encoder is used to extract a comprehensive feature representation vector from the time series data of external signals. The multi-task head contains several parallel regressors to predict health indicators in the standard label vector based on the comprehensive feature representation vector. S4: Transfer the capabilities of the baseline diagnostic model to the online diagnostic model through knowledge distillation; input external signals into the online diagnostic model, output online prediction results of battery health indicators, and provide fault warnings based on the online prediction results and preset thresholds.

[0006] Furthermore, the external signal is the time-series data of the battery's total voltage, charging and discharging current, and battery surface temperature, which are collected in real time during the battery test. The internal signals are the electrochemical impedance spectroscopy, capacity-voltage curve, and heat generation power curve collected by measuring instruments during the battery testing process; The standard label vector includes: calculating the solid electrolyte interface membrane health by extracting the change in solid electrolyte interface membrane resistance through equivalent circuit model fitting based on electrochemical impedance spectroscopy; calculating the active lithium inventory loss by locating the shift of characteristic peaks on the capacity axis based on capacity-voltage curves through differential voltage analysis; calculating the lithium plating risk index based on the abnormal exothermic characteristics in the voltage relaxation curve and heat generation power curve after charging; and calculating the positive electrode health based on active lithium inventory loss and the overall battery health status.

[0007] Furthermore, the testing process employs a composite testing protocol that covers typical stress conditions that lead to early latent faults; The composite test protocol includes a baseline cycle phase and a stress cycle phase. The reference cycle stage is constant current and constant voltage charging and constant current discharging at a standard rate of 1C at 25°C. The stress cycling phase introduces one or more of the following stress conditions alternately to accelerate the aging process: charging and discharging at -10°C; charging at a high rate of 2C; leaving the battery at 45°C with a state of charge above 90% for an extended period; and deep discharging to below the cutoff voltage.

[0008] Furthermore, the construction of the training sample database includes: each 5 stress cycles constitute a diagnostic cycle; taking the start of the diagnostic cycle as the time label, extracting the external signal time series data containing the 5 most recent complete stress cycles as runtime time series data; if the number of cycles is less than 5 in the early stage of the test, extracting all data from the start of the test; pairing the data with the standard label vector of the same time label to form training sample pairs; and traversing the entire life cycle diagnostic cycle of the test samples to store the sample pairs in the training sample database.

[0009] Furthermore, the benchmark diagnostic model is trained by minimizing a multi-task weighted loss function: , in, For the total loss, The prediction error for SEI health status. This is the prediction error for active lithium inventory loss. The prediction error of the lithium plating risk index is... This represents the prediction error for the health of the positive electrode. , , and These are the corresponding weighting coefficients.

[0010] Furthermore, the online diagnostic model employs a lightweight temporal network with a two-layer bidirectional gated recurrent unit as its backbone, and training uses a composite loss function: , in, The total loss of the online diagnostic model. The error between the predicted output of the online diagnostic model and the standard label vector. The mean squared error between the predicted output of the online diagnostic model and the predicted output of the benchmark diagnostic model on the same input data is given. This is the balance coefficient.

[0011] Furthermore, the fault warning based on the online prediction results and preset thresholds includes: maintaining a fixed-length sliding time window to cache the timing data of external signals from the most recent complete stress cycles; the window slides and updates as a new cycle is completed; the updated window data is input into the online diagnostic model, and the predicted values ​​of health indicators are output in real time. For each health indicator, multiple warning thresholds are preset. When the predicted value exceeds the first-level warning threshold, a visual prompt is given on the human-computer interaction interface; when the predicted value exceeds the second-level warning threshold, an audible and visual alarm is triggered and a command is sent to the test equipment to pause the test.

[0012] Furthermore, the method also includes attribution analysis of the overall battery capacity degradation: constructing a battery simulation model of an aging mechanism sub-model, performing simulation using a composite test protocol, and generating simulation training samples containing external signal timing data and corresponding decoupling label vectors, wherein the decoupling label vectors are the percentage of capacity loss due to known aging mechanisms; Based on the benchmark diagnostic model, a multi-task decoupling model is constructed and trained, which can predict the percentage of capacity loss due to known aging mechanisms based on external signals. The capabilities of the multi-task decoupling model are transferred to an online decoupling diagnostic model through knowledge distillation, which is used to output an aging decoupling report.

[0013] Furthermore, the method also includes the detection of unknown faults: a synthetic anomaly perturbation module simulating unknown fault modes is added to the battery simulation model and randomly activated in the simulation to generate an extended decoupling label vector containing contributions from unknown anomalies. An unknown anomaly detection head is added to the multi-task decoupling model to form an enhanced multi-task decoupling model, which is then trained to enable the model to identify unknown faults. Through knowledge distillation, the capabilities of the enhanced multi-task decoupling model are transferred to an online anomaly diagnosis model, which outputs an unknown anomaly diagnosis report.

[0014] Furthermore, the fault warning also includes: setting independent multi-level warning thresholds for the percentage of unknown abnormal capacity loss predicted by the online anomaly diagnosis model; When the percentage of unknown abnormal capacity loss is greater than zero for the first time but lower than the first-level abnormal warning threshold, a first-level abnormal warning is triggered, the test continues, but the tester is prompted to pay attention; when the percentage of unknown abnormal capacity loss exceeds the second-level abnormal warning threshold, or continues to rise for several consecutive diagnostic cycles, a second-level abnormal warning is triggered, an audible and visual alarm is immediately triggered, a command is sent to the test equipment, and the current test cycle is terminated.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages: By constructing a multi-dimensional health indicator system and benchmark diagnostic model, a comprehensive quantitative assessment of the battery's internal state and accurate identification of early latent faults are achieved. Knowledge distillation technology is used to compress complex models into lightweight online diagnostic models. Combined with real-time data input via a sliding window and multi-level early warning threshold linkage device control, online health monitoring and proactive safety protection throughout the battery's entire lifespan are completed. Battery simulation and multi-task decoupling models incorporating aging mechanism sub-models are used to perform quantitative attribution analysis of capacity decay and identify dominant failure mechanisms. By introducing a synthetic abnormal disturbance module into the simulation and training an enhanced multi-task decoupling model, the detection and quantification of unknown or sudden fault modes are achieved. Attached Figure Description

[0016] Figure 1This is a flowchart of a fault testing method for a lithium-ion battery according to an embodiment of the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0018] Unless otherwise defined, 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; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1: As Figure 1 As shown, a fault testing method for lithium-ion batteries is described.

[0020] S1: Acquire external and internal signals of the battery during the test process, and generate a standard label vector of the battery health status based on the internal signals; the standard label vector includes: solid electrolyte interface membrane health, active lithium inventory loss, lithium plating risk index, and cathode health. The testing process employs a composite testing protocol that covers typical stress conditions that lead to early latent faults. Specifically, lithium-ion batteries from the same production batch with consistent initial performance were selected as test samples. A composite test protocol covering typical stress conditions that lead to early latent faults was adopted, including a baseline cycling stage and a stress cycling stage. The baseline cycling stage involved constant current and constant voltage charging and discharging at a standard 1C rate at 25°C, serving as a reference for performance degradation. The stress cycling stage introduced one or more of the following stress conditions alternately to directionally accelerate specific aging modes: charging and discharging at -10°C to induce lithium plating risk and related interface side reactions; charging at a high 2C rate to exacerbate electrode polarization and potentially trigger lithium plating and unstable growth of the interface film; allowing the battery to remain stationary at 45°C with a state of charge exceeding 90% for an extended period to accelerate electrolyte oxidation and decomposition and positive electrode interface side reactions; and deep discharging to below the cutoff voltage to induce abnormal reactions such as dissolution of the copper current collector at the negative electrode.

[0021] The external signals are time-series data of the battery's total voltage, charging and discharging current, and battery surface temperature, which are collected in real time during the battery test. The internal signals are the electrochemical impedance spectroscopy, capacity-voltage curve, and heat generation power curve collected by measuring instruments during the battery testing process; Specifically, during the testing process, external and internal signals are acquired in real time. External signals include the total battery voltage, charge / discharge current, and temperatures at multiple points on the battery surface (at least including the positive electrode, negative electrode, and the middle of the casing). Internal signals include: applying a small AC signal to the battery and measuring its impedance response to obtain an electrochemical impedance spectroscopy; recording the capacity-voltage curve during constant current charge / discharge using high-precision data acquisition equipment (accuracy better than 0.1mV); and monitoring the heat generation power curve of the battery during charge / discharge and resting processes using an isothermal calorimeter.

[0022] The standard label vector includes: Specifically, after every 5 stress cycles, the sampling time at the end of the constant current discharge in the 5th stress cycle is used as the time stamp for this diagnostic cycle. The stress test is paused, and the diagnostic mode is entered to analyze the collected data and generate corresponding standard label vectors. Through equivalent circuit model fitting, the changes in charge transfer resistance and membrane resistance due to the growth of the negative electrode SEI film (solid electrolyte interface film) are extracted and normalized to represent the SEI health. , in, For SEI health, The initial SEI film resistance was obtained through electrochemical impedance spectroscopy. The SEI film resistance after aging is obtained by performing electrochemical impedance spectroscopy under the exact same conditions as the initial test when the battery capacity decays to 80% of the initial capacity. The SEI film resistance threshold represents the end-of-life threshold, a preset engineering threshold calibrated experimentally. It indicates that the battery's performance degradation due to SEI film issues has reached the end-of-life value. 100% is used to convert the result to a percentage. A value of 100% indicates consistency with the initial state, while a decrease in the value indicates that membrane growth has led to performance degradation.

[0023] Differential voltage analysis of the capacity-voltage curves yielded the active lithium inventory loss: , in, Losses due to active lithium inventory This refers to the capacity position of the characteristic peak corresponding to the negative electrode-dominated phase transition with the most significant amplitude on the capacity-voltage curve in the initial state of the battery. The new position of the same characteristic peak on the capacity axis when the capacity of the same battery decays to 80% of its initial capacity; This represents the total capacity of all recyclable active lithium within the battery in its initial state.

[0024] By comprehensively evaluating the shape of the voltage relaxation curve after low temperature and fast charging, and the abnormal exothermic peak in the micro-calorimetry data, the lithium plating risk index is calculated: , in, This is a lithium plating risk index; the closer it is to 1, the greater the risk. This is the voltage relaxation quantity. , This is the voltage at the charging cutoff point. The stable voltage after standing for 30 minutes; The maximum abnormal heat release peak power is extracted from the heat generation power curve during the resting period. The normalization coefficients were determined by testing a large number of battery samples known to have experienced severe lithium plating, and statistically analyzing the extremely high-risk scenarios. Determined by the product value; Used to limit the upper limit of the calculation result to 1.

[0025] Calculate the cathode health based on active lithium inventory loss and overall battery health status: , in, The positive electrode health rating indicates the percentage of effective retention of the positive electrode active material. For the overall health of the battery, , This represents the average actual discharge capacity of the battery in the current test. This refers to the initial rated capacity of the battery. This represents losses from active lithium inventory.

[0026] The construction of the training sample database includes: Specifically, the starting point of each diagnostic cycle in the diagnostic mode is used as a time stamp, and the SEI health, active lithium inventory loss, lithium plating risk index, and cathode health calculated at that moment are used as standard label vectors. Using the time stamp as the endpoint, all external signal timing data corresponding to the entire operation process, including the most recent five complete stress cycles, are extracted as runtime timing data. If the number of cycles in the initial stage of the test is less than five, all data since the start of the test are extracted. The runtime timing data consists of voltage, current, and temperature sequences obtained from external signals.

[0027] S2: Pair the time-series data of external signals with standard label vectors to construct a training sample database; Specifically, the runtime sequence data is normalized using a minimum-maximum method, scaling the data to the [0, 1] interval, and then paired with standard label vectors generated at the same time label to form a training sample pair. For all test sample batteries, all diagnostic cycles throughout their entire test life are traversed, and the resulting training sample pairs are stored in the training sample database. Each training sample pair includes: a unique identifier for the battery sample, the cumulative number of cycles, and an encoding vector of the stress conditions experienced. The encoding vector of the stress conditions experienced is the truncated 5 stress cycles, recording the stress type applied in each cycle in sequence. For example: [low temperature, high rate, high temperature rest, deep discharge, baseline cycle].

[0028] S3: Using the training sample database, train a benchmark diagnostic model with an encoder-multi-task head architecture. The encoder is used to extract a comprehensive feature representation vector from the time series data of external signals. The multi-task head contains several parallel regressors to predict health indicators in the standard label vector based on the comprehensive feature representation vector. Specifically, a benchmark diagnostic model employing an encoder-multi-task-head architecture is trained to automatically learn the complex mapping relationship between routine runtime sequence data and the microscopic health state of the battery. The encoder, as the core processing unit of the model, receives runtime sequence data and performs in-depth analysis and abstraction. A deep neural network extracts temporal patterns and deep features related to early battery aging from the original sequence, such as subtle deformations of the voltage plateau, coordinated changes in electro-thermal signals, and cross-cycle decay trends. This information is then aggregated into a fixed-dimensional comprehensive feature representation vector. After the encoder outputs the comprehensive feature representation vector, it is sent to the parallel multi-task head. The multi-task head consists of four independent fully connected neural network branches, each acting as a specialized regressor. Its function is to predict a specific health indicator from the standard label vector based on the shared comprehensive feature representation vector: the first task head predicts SEI health; the second task head predicts active lithium inventory loss; the third task head predicts the lithium plating risk index; and the fourth task head predicts cathode health.

[0029] After constructing the training sample database, the benchmark diagnostic model is trained and its performance validated using the database under supervision. The training objective is to make the four predicted values ​​output by the model as close as possible to the true values ​​in the standard label vector. A multi-task weighted loss function is defined as the optimization objective of the model. , in, For the total loss, The prediction error for SEI health status. This is the prediction error for active lithium inventory loss. The prediction error of the lithium plating risk index is... To calculate the prediction error of the positive electrode health, the prediction errors of all four tasks were calculated using the mean square error. , , and These are the corresponding weight coefficients, calibrated experimentally. The database is divided into training, validation, and test sets. A grid search method is used to traverse different weight combinations within a predefined set of typical values ​​{0.01, 0.1, 1, 10, 100}. For each set of candidate weight parameters, the model is trained using the training set, and the model's total loss on the validation set is used as the benchmark. As the core evaluation metric, a set of weight coefficients that minimizes the total loss and provides relatively balanced prediction performance across tasks is selected as the final configuration on the validation set.

[0030] All training samples in the database are randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. A batch of samples is drawn from the training set at a set batch size (e.g., 64). The runtime sequence data from these samples is input into the benchmark diagnostic model. The model performs forward propagation, extracts features through the encoder, and outputs predicted values ​​for four health indicators via the multi-task head. The total loss between the predicted values ​​and the standard label vector is calculated. Using the backpropagation algorithm, the gradient of the model parameters is calculated based on the loss value, and the Adam optimizer is used to update the model parameters to minimize the loss function. This process iterates through all samples in the training set within one training cycle and is repeated for multiple cycles. After each training cycle, the current model is evaluated using the entire validation set, and the total loss on the validation set is calculated. The validation set loss is continuously monitored. If the value does not decrease within N consecutive training cycles (e.g., 5), the model performance is considered to have reached its limit, training is immediately terminated, and the model parameters corresponding to the cycle with the lowest validation set loss are saved as the final model. By running the training process multiple times and comparing the model's performance on the validation set, the model is re-evaluated. , , and These weighting coefficients are adjusted and optimized to determine a set of configurations that optimize model performance.

[0031] S4: Transfer the capabilities of the baseline diagnostic model to the online diagnostic model through knowledge distillation; input external signals into the online diagnostic model, output online prediction results of battery health indicators, and provide fault warnings based on the online prediction results and preset thresholds.

[0032] Specifically, to enable the diagnostic capabilities of the benchmark diagnostic model to be applied in real-time on-site, knowledge distillation technology is employed to transfer the decision-making knowledge of the benchmark diagnostic model to an online diagnostic model. The online diagnostic model uses a lightweight temporal network as its backbone, with inputs identical to the benchmark model. A two-layer bidirectional gated recurrent unit replaces the encoder in the benchmark model. The first-layer bidirectional gated recurrent unit receives the original sequence and learns the underlying temporal patterns; the second-layer bidirectional gated recurrent unit receives the output of the first layer and learns temporal feature abstractions. This bidirectional structure allows the model to simultaneously consider past and future contextual information at each time point. The forward and backward hidden states generated by the second-layer bidirectional gated recurrent unit at the final time step are concatenated to form a comprehensive feature vector. Feature compression and integration are performed through a shared fully connected layer, connecting to four linear output layers, corresponding to the predictions of SEI health, active lithium inventory loss, lithium plating risk index, and cathode health, respectively.

[0033] The online diagnostic model is trained using a composite loss function: , in, The total loss of the online diagnostic model. For hard label loss, the error between the predicted output of the online diagnostic model and the standard label vector is calculated. To determine the soft knowledge loss, the mean squared error between the predicted output of the online diagnostic model and the predicted output of the benchmark diagnostic model on the same input data is calculated. This is a balancing coefficient, with a value range of [0, 1]. It is used to balance the relative importance of hard labels and soft knowledge during training and can be adjusted based on the validation set results.

[0034] During training, the parameters of the benchmark diagnostic model remain frozen, while the online diagnostic model uses the exact same pre-split training and validation sets as the benchmark model. For each training batch, runtime sequence data is simultaneously input into both the benchmark and online diagnostic models. The total loss is calculated using a composite loss function, and the parameters of the online diagnostic model are updated only through backpropagation; the parameters of the benchmark model are not updated. The Adam optimizer is used to minimize the loss. Repeat the iteration until the performance of the online diagnostic model converges on the validation set.

[0035] The online diagnostic model is embedded into the host computer or industrial control computer of the standard battery testing system. Continuous online diagnostics are performed on the battery under test, continuously filling a sliding time window buffer with real-time acquired and preprocessed time-series data. The window length is fixed at the data volume corresponding to the last five complete stress cycles. Each time a new stress cycle is completed, the window automatically slides and updates, discarding the oldest cycle data and incorporating the latest data. After each sliding window data update, the complete time-series data within the current window is input into the online diagnostic model, and a forward propagation calculation is performed, simultaneously outputting the predicted values ​​of four health indicators. The diagnostic results are displayed in real-time on the host computer's human-machine interface in the form of dashboards, trend graphs, and numerical panels, facilitating observation of the battery's microscopic health status evolution.

[0036] The fault early warning based on online prediction results and preset thresholds includes: Specifically, multiple warning thresholds are preset for each health indicator. The model output value is compared with each threshold in real time, and the corresponding warning status is triggered. When any indicator deviates from the normal range but does not yet pose a direct risk, a Level 1 warning is triggered. The Level 1 warning thresholds are: Lithium plating risk index > 0.3, SEI health < 95%, active lithium inventory loss > 2%, and cathode health < 95%. When a Level 1 warning is triggered, the corresponding indicator value in the interactive interface turns yellow. The test is not interrupted, but the tester is prompted to pay attention to the trend of the indicator. When a key indicator deteriorates to the point that it may affect battery safety or is about to cause serious performance degradation, a Level 2 warning is triggered. The Level 2 warning thresholds are: Lithium plating risk index > 0.7, SEI health decreasing by more than 5% within three consecutive diagnostic cycles, and active lithium inventory loss > 5%. When a Level 2 warning is triggered, an audible and visual alarm is activated, a red warning box pops up on the interface, clearly indicating the exceeding indicator and its severity, and a command is sent to the testing equipment to automatically pause the current test cycle.

[0037] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application constructs a multi-dimensional health index, including SEI health, active lithium inventory loss, lithium plating risk index, and cathode health, by simultaneously collecting external and internal signals. This enables a comprehensive quantitative assessment of the battery's internal state, overcoming the limitations of traditional methods that rely solely on capacity decay. A benchmark diagnostic model employing an encoder-multi-task-head architecture automatically learns the complex mapping relationship from conventional time-series data to health indicators. This allows for accurate prediction of the battery's microscopic health state based solely on external signals, without relying on complex internal measurement methods, significantly reducing diagnostic costs and hardware barriers. Knowledge distillation technology compresses the benchmark model into a lightweight online diagnostic model, embedding it into the test system's host computer. Combined with real-time sliding window inference, online health monitoring throughout the battery's entire lifespan is achieved, solving the problem of deploying complex models in real-world testing scenarios. Multiple warning thresholds are preset for each health indicator and linked to the human-machine interface and device control commands. This enables timely alerts, warnings, and even automatic test pauses when indicators are abnormal, effectively preventing safety accidents such as battery thermal runaway and improving the safety of the battery testing process.

[0038] Example 2: Example 1 achieved online monitoring and evaluation of battery health status, but it still has the limitation of not being able to perform attribution analysis on overall performance degradation. This example further supplements and explains the content of Example 1.

[0039] The method also includes attribution analysis of the overall battery capacity degradation: constructing a battery simulation model of an aging mechanism sub-model, performing simulation using a composite test protocol, and generating simulation training samples containing external signal time series data and corresponding decoupling label vectors, wherein the decoupling label vectors are the percentage of capacity loss due to known aging mechanisms; Specifically, a battery simulation model is constructed, integrating multiple independent and adjustable aging mechanism sub-models. These sub-models include an SEI growth sub-model, a lithium plating and dead lithium formation sub-model, and a positive electrode active material loss sub-model. The SEI growth sub-model simulates the thickening process of the solid electrolyte interfacial film during cycling; the lithium plating and dead lithium formation sub-model simulates the deposition reaction of metallic lithium on the negative electrode surface under conditions such as low temperature and fast charging, as well as the subsequent deactivation and conversion of deposited lithium into dead lithium; and the positive electrode active material loss sub-model simulates the deactivation of the positive electrode material due to structural degradation, phase transition, or interfacial side reactions.

[0040] Simulations were performed using the exact same stress conditions and sequence as the composite testing protocol, recording the time-varying sequences of the battery's total voltage, total current, and average temperature. During simulation, the battery simulation model recorded the contribution of each aging sub-model to the battery's current overall state. For example, at the end of the simulation, the capacity loss caused by SEI growth, lithium plating, and cathode degradation in the total capacity decay could be calculated. A multi-dimensional decoupled label vector was generated for each segment of simulation data.

[0041] Five stress cycles are completed as a diagnostic cycle, and the simulation data is sliced. The external signal sequence corresponding to each diagnostic cycle is used as the input feature, and the decoupling label vector calculated at the end of the cycle is used as the output target. These are paired to form a simulation training sample. By randomizing or fully permuting the stress condition combinations of the battery simulation model, a simulation training sample database can be generated.

[0042] Based on the benchmark diagnostic model, a multi-task decoupling model is constructed and trained, which can predict the percentage of capacity loss due to known aging mechanisms based on external signals. The capabilities of the multi-task decoupling model are transferred to an online decoupling diagnostic model through knowledge distillation, which is used to output an aging decoupling report.

[0043] Specifically, based on the benchmark diagnostic model, a multi-task output head is constructed, consisting of multiple parallel and structurally independent regression sub-networks. Each sub-network is responsible for predicting a specific dimension in the decoupling label vector. Specifically, decoupling task head A predicts the percentage of capacity loss caused by SEI growth, decoupling task head B predicts the percentage of capacity loss caused by lithium plating and dead lithium formation, and decoupling task head C predicts the percentage of capacity loss caused by loss of positive electrode active material.

[0044] Define a multi-task total loss function as the training objective: , in, The total loss function for multiple tasks. , and These are the mean square errors between the predicted values ​​and the true simulation values ​​of task headers A, B, and C, respectively. , and These are the weighting coefficients for each task, used to balance the importance of different aging modes' contributions to the total loss, and determined on the validation set through grid search.

[0045] During training, the simulation training sample database is randomly divided into training, validation, and test sets in a 70%:15%:15% ratio. Batch data from the training set is input into the model, the encoder extracts features, each decoupling task head outputs predicted values, and the total multi-task loss function is calculated. Gradients are calculated using backpropagation, updating only the parameters of the multi-task decoupling output heads, and the Adam optimizer is used to minimize the loss function. After each training epoch, the model performance is evaluated using the validation set. Training is terminated early when the validation set loss no longer decreases over N consecutive epochs (e.g., 5). The model parameters corresponding to the epoch with the lowest validation set loss are selected as the multi-task decoupling model.

[0046] Physical battery samples consistent with the simulation model were selected, and the composite testing protocol and diagnostic mode were reproduced for accelerated aging testing. External signals and four comprehensive health indicators were simultaneously collected during each diagnostic cycle. At the end of the battery's lifespan (e.g., capacity decay to 80%), the battery was disassembled, and physicochemical analysis methods were used to assess the relative severity of three aging modes: SEI growth, lithium plating, and cathode decay. The main stress types experienced during periods of significant capacity decay were analyzed, establishing a correlation between aging phenomena and stress conditions; for example, capacity jumps mainly occur after continuous low-temperature cycling. The total capacity loss (100%) was considered... Based on the assessed relative proportions, a decoupling contribution reference vector is assigned. For example: [SEI loss: 65%, lithium plating loss: 25%, cathode loss: 10%].

[0047] Fine-tuning the baseline diagnostic model: , in, To fine-tune the loss function, The mean squared error between the four health indicators and the comprehensive health indicator measured by physical examination is given. The mean square error between the percentage of capacity loss output by the multi-task output head and the decoupling contribution reference vector; and For the corresponding weighting coefficients, > In fine-tuning, direct measurement results dominate, and the optimal value is determined experimentally on the validation set through grid search.

[0048] A lightweight time-series network (with a two-layer bidirectional gated recurrent cell structure), identical to the online diagnostic model, is used as the backbone to construct the online decoupling diagnostic model, with external signal time-series data as input. Three parallel linear output layers are designed to predict the percentage capacity loss caused by SEI growth, lithium plating and dead lithium formation, and loss of positive electrode active material, respectively. Training employs a knowledge distillation framework, using the multi-task decoupling model as the knowledge source and the online decoupling diagnostic model to be trained as the knowledge receiver. Training is performed using a simulated training sample database. The composite loss function for the decoupling task is defined as follows: , in, The total loss of the online decoupled diagnostic model. The mean square error between the predicted output of the online decoupled diagnostic model and the decoupled label vector is given. The mean square error between the predicted output of the online decoupled diagnostic model and the predicted output of the multi-task decoupled model is given. The balance coefficient is used for weighing... and The relative importance of the two losses during training, and the optimal values, are determined by grid search of performance on the validation set.

[0049] During training, the parameters of the multi-task decoupling model remain frozen, while the online decoupling diagnostic model uses the same pre-split training and validation sets as the training model. For each training batch, runtime sequence data is simultaneously input into both the multi-task decoupling model and the online decoupling diagnostic model. The total loss of the online decoupling diagnostic model is calculated based on the composite loss function of the decoupling tasks. Using backpropagation, only the parameters of the online decoupling diagnostic model are updated; the parameters of the multi-task decoupling model are not updated. The Adam optimizer is used to minimize... Repeat the iteration until the performance of the online decoupled diagnostic model converges on the validation set.

[0050] The online decoupled diagnostic model is embedded into the host computer or industrial control computer of a standard battery testing system. While outputting predicted values ​​for four health indicators, it simultaneously outputs an aging decoupled report comprising the percentage capacity loss due to SEI growth, lithium plating and dead lithium formation, and loss of positive electrode active material. This allows for the quantitative decomposition of overall battery performance degradation into specific failure mechanisms, thereby accurately identifying the dominant causes of capacity loss.

[0051] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application uses a battery simulation model to generate decoupled label vectors that include the contributions of aging mechanisms such as SEI growth, lithium plating and dead lithium formation, and loss of positive electrode active material. A multi-task decoupled model is then trained to achieve quantitative attribution analysis of overall battery capacity degradation. By fine-tuning the model using real battery test data and employing knowledge distillation technology to compress the model into a lightweight online decoupled diagnostic model, a decoupled report identifying the dominant causes of battery aging can be output in real-time and online, relying solely on external signals. This not only assesses the battery's health status but also identifies the specific failure mechanisms leading to performance degradation, providing in-depth decision-making support for battery failure analysis, lifespan prediction, and targeted improvements.

[0052] Example 3: Example 2 achieved quantitative attribution analysis of overall battery capacity degradation, but it still has the limitation of not being able to detect sudden or unknown faults. This example further supplements and explains the content of Example 2.

[0053] The method also includes the detection of unknown faults: a synthetic anomaly perturbation module simulating unknown fault modes is added to the battery simulation model and randomly activated in the simulation to generate an extended decoupling label vector containing contributions from unknown anomalies. Specifically, multiple independent synthetic anomalous perturbation modules are employed to simulate sudden or unknown failures that may occur in real-world testing but are not covered by existing known aging models. The mechanisms of these modules differ from conventional aging modes, providing diagnostic models with the ability to identify unknown characteristics. Among them, the anomalous electrolyte decomposition module simulates an atypical side reaction pathway. Unlike conventional side reactions that form stable SEI films or positive electrode interface films, by setting differentiated reaction kinetic parameters, it simulates an anomalous decomposition process that may produce gaseous or heterogeneous solid products. Its reaction heat generation rate and its impact on battery voltage differ from known reactions. The electrode material local phase transition module simulates sudden defects in the microstructure of active materials. In the simulation, local regions of the positive or negative electrode are randomly selected, instantly reducing their lithium-ion diffusion coefficient to 10%-50% of its original value, or causing some active sites to permanently fail. Simulating local lattice distortion, phase separation, or particle microcracks that may occur during cycling directly affects the ion migration and reaction kinetics of the region. The internal contact impedance step module simulates sudden degradation of the battery's internal mechanical connections. At random time points in the simulation, the contact resistance of a connection inside the battery (such as a tab weld) is instantaneously increased by a fixed value. The increment is set in the range of 0.2 milliohms to 2 milliohms to simulate a sudden increase in ohmic impedance caused by loosening or deformation of the weld; this effect is instantaneous and independent of the electrochemical reaction process. A metal foreign object micro-short circuit module simulates the hidden dangers introduced by manufacturing process defects. A random, dynamically changing bypass leakage current is introduced between the positive and negative electrodes. The leakage current ranges from 10 microamps to 100 microamps and can fluctuate with the battery's state of charge or simulated mechanical vibration, thereby simulating intermittent micro-short circuits caused by metal impurities, leading to unpredictable capacity loss and minute voltage drops.

[0054] During simulation, the same composite testing protocol as the benchmark diagnostic model is employed, including a benchmark cycle and alternating stress conditions. Within a randomly selected stress cycle, one or more synthetic anomalous disturbance modules are activated. The activation start time, duration, and strength parameters of each module are randomly set within their respective defined ranges to simulate the randomness and diversity of fault occurrence. The simulation runs continuously, with each five stress cycles defined as a diagnostic cycle, maintaining consistency with the diagnostic rhythm of physical testing.

[0055] During the simulation, the input signal stream, the known aging contribution stream, and the synthetic anomaly contribution stream are recorded simultaneously. The input signal stream is a high-sampling-rate sequence of the battery's total voltage, total current, and volume average temperature changing over time. The known aging contribution stream represents the battery capacity loss caused to battery performance by three known aging sub-models at each simulation moment: SEI growth, lithium plating and dead lithium formation, and loss of positive electrode active material. The synthetic anomaly contribution stream represents the battery capacity loss caused to battery performance by the synthetic anomaly perturbation module.

[0056] At the end of each diagnostic cycle, the percentage of capacity loss caused by the three known aging mechanisms—SEI growth, lithium plating and dead lithium formation, and loss of positive electrode active material—is calculated separately for that diagnostic cycle. For example, the percentage of capacity loss due to SEI growth is calculated as follows: , in, This represents the percentage of capacity loss due to the continued growth of the SEI during this diagnostic period, relative to the total capacity loss during this period. The battery capacity loss value is simulated and calculated by the SEI growth submodel. This represents the total capacity loss of the battery during this diagnostic cycle.

[0057] The percentage of additional capacity loss caused by the activated synthetic anomaly perturbation module during this diagnostic cycle is statistically analyzed. Based on the battery capacity at the end of the simulation, the percentage of the current capacity relative to the initial capacity is calculated. The data is assembled into an extended decoupling label vector, for example: [Known SEI: 2.1%, Known Lithium Plating: 0.8%, Known Cathode: 1.5%, Unknown Anomaly: 0.6%, Battery Capacity: 95.0%].

[0058] The process involves adding an unknown anomaly detection head to the multi-task decoupling model to form an enhanced multi-task decoupling model, which is then trained to enable the model to identify unknown faults. Through knowledge distillation, the capabilities of the enhanced multi-task decoupling model are transferred to an online anomaly diagnosis model, which outputs an unknown anomaly diagnosis report.

[0059] Specifically, a multi-task decoupling model is used as the basic backbone network, with a new unknown anomaly detection head added in parallel. This unknown anomaly detection head is an independent regression sub-network with the same structure as the known decoupling heads, specifically responsible for predicting the percentage of capacity loss due to unknown anomalies in the extended decoupling label vector. The model is upgraded to an enhanced multi-task decoupling model with four parallel output heads, corresponding to: SEI growth, lithium plating and dead lithium generation, loss of positive electrode active material, and capacity loss prediction due to unknown anomalies. During training, a simulated training sample database containing the extended decoupling label vector is used. Only the parameters of the newly added unknown anomaly detection head are trained and optimized, while the parameters of the encoder and the known decoupling task heads (A, B, C) remain completely frozen. The training objective at this stage is to enable the unknown anomaly detection head to accurately identify and quantify the anomalous patterns introduced by the synthetic anomaly perturbation module. The loss function is only for unknown anomaly items. , in, For unknown abnormal losses, Mean square error, The predicted value is the one output by the unknown anomaly detection head, representing the predicted percentage of battery capacity loss. The true value of the label is the known true value calculated during the simulation based on the activated synthetic anomaly disturbance module.

[0060] An enhanced multi-task decoupling model is migrated to an online anomaly diagnosis model using a knowledge distillation framework and embedded into the host computer or industrial control computer of a standard battery testing system. Time-series data of the total voltage, current, and surface temperature of the battery under test are collected, and a sliding time window of the length of the five most recent complete stress cycles is maintained. Each time a new stress cycle is completed, the window is updated, and the complete time-series data sequence within the window is input into the online anomaly diagnosis model to perform a forward inference. The online anomaly diagnosis model simultaneously outputs four results, which are assembled into two core reports displayed on the human-computer interface: an aging decoupling report and an unknown anomaly diagnosis report. The unknown anomaly diagnosis report declares the detection of unidentified anomalies and provides a specific numerical value of their contribution. For example, if an unidentified anomaly is detected, it is estimated to cause an additional capacity loss of approximately 0.6% in the current period.

[0061] The fault warning also includes: Specifically, based on the multi-level early warning system, a new independent early warning threshold is added for the percentage of unknown abnormal capacity loss. When the percentage of unknown abnormal capacity loss first exceeds zero but falls below a preset threshold (e.g., 0.3%), a Level 1 abnormality warning is triggered. The test continues, but testers are alerted to the situation. When the percentage of unknown abnormal capacity loss exceeds the upper limit threshold (e.g., 0.5%), or continues to rise for multiple consecutive diagnostic cycles (e.g., 3), a Level 2 abnormality warning is triggered. This immediately activates an audible and visual alarm, sends a command to the test equipment, and pauses or terminates the current test cycle to prevent potential safety incidents caused by continued operation under unknown risks.

[0062] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application constructs an extended decoupled label vector containing unknown fault contributions by designing multiple independent synthetic anomaly perturbation modules and randomly activating them in simulation. This vector is then used to train an enhanced multi-task decoupled model capable of identifying unknown or sudden anomaly features, enabling the detection of potential faults exceeding conventional aging modes in battery testing. The detection capabilities of this model are transferred to a lightweight online anomaly diagnosis model using knowledge distillation technology. This allows for simultaneous online evaluation and early warning of unknown anomaly contributions while relying solely on external signals to monitor known aging states in real time. Independent multi-level early warning thresholds are set for unknown anomaly contributions and linked to the test equipment control system. This enables timely alarms and even automatic test suspension when anomaly risks exceed safe limits, effectively preventing safety accidents such as thermal runaway that may result from the evolution of unknown faults. This enhances the forward-looking safety protection capabilities and overall reliability of battery testing.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault testing method for lithium-ion batteries, characterized in that, include: S1: Acquire external and internal signals of the battery during the test process, and generate a standard label vector of the battery health status based on the internal signals; The standard label vector includes: solid electrolyte interface membrane health, active lithium inventory loss, lithium plating risk index, and cathode health. S2: Pair the time-series data of external signals with standard label vectors to construct a training sample database; S3: Using the training sample database, train a benchmark diagnostic model with an encoder-multi-task head architecture. The encoder is used to extract a comprehensive feature representation vector from the time series data of external signals. The multi-task head contains several parallel regressors to predict health indicators in the standard label vector based on the comprehensive feature representation vector. S4: Transfer the capabilities of the baseline diagnostic model to the online diagnostic model through knowledge distillation; input external signals into the online diagnostic model, output online prediction results of battery health indicators, and provide fault warnings based on the online prediction results and preset thresholds.

2. The fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The external signals are time-series data of the battery's total voltage, charging and discharging current, and battery surface temperature, which are collected in real time during the battery test. The internal signals are the electrochemical impedance spectroscopy, capacity-voltage curve, and heat generation power curve collected by measuring instruments during the battery testing process; The standard label vector includes: calculating the solid electrolyte interface membrane health by extracting the change in solid electrolyte interface membrane resistance through equivalent circuit model fitting based on electrochemical impedance spectroscopy; calculating the active lithium inventory loss by locating the shift of characteristic peaks on the capacity axis based on capacity-voltage curves through differential voltage analysis; calculating the lithium plating risk index based on the abnormal exothermic characteristics in the voltage relaxation curve and heat generation power curve after charging; and calculating the positive electrode health based on active lithium inventory loss and the overall battery health status.

3. The fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The testing process employs a composite testing protocol that covers typical stress conditions that lead to early latent faults. The composite test protocol includes a baseline cycle phase and a stress cycle phase. The reference cycle stage is constant current and constant voltage charging and constant current discharging at a standard rate of 1C at 25°C. The stress cycling phase accelerates the aging process by alternately introducing one or more of the following stress conditions: charging and discharging at -10°C. Charge the battery at a high rate of 2C; allow the battery to remain stationary at 45°C with a state of charge exceeding 90% for an extended period; and perform deep discharge to below the cutoff voltage.

4. The fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The construction of the training sample database includes: each 5 stress cycles constitute a diagnostic cycle; taking the start of the diagnostic cycle as the time label, extracting the external signal time series data containing the 5 most recent complete stress cycles as runtime time series data; if the number of cycles is less than 5 in the early stage of the test, extracting all data from the start of the test; pairing the data with the standard label vector of the same time label to form training sample pairs; and traversing the entire life cycle diagnostic cycle of the test samples to store the sample pairs in the training sample database.

5. A fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The benchmark diagnostic model is trained by minimizing a multi-task weighted loss function: ， in, For the total loss, The prediction error for SEI health status. This is the prediction error for active lithium inventory loss. The prediction error of the lithium plating risk index is... This represents the prediction error for the health of the positive electrode. , , and These are the corresponding weighting coefficients.

6. The fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The online diagnostic model employs a lightweight temporal network with a two-layer bidirectional gated recurrent unit as its backbone, and training uses a composite loss function: ， in, The total loss of the online diagnostic model. The error between the predicted output of the online diagnostic model and the standard label vector. The mean squared error between the predicted output of the online diagnostic model and the predicted output of the benchmark diagnostic model on the same input data is given. This is the balance coefficient.

7. A fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The fault warning based on online prediction results and preset thresholds includes: maintaining a fixed-length sliding time window to cache the timing data of external signals from the most recent complete stress cycles; the window slides and updates as a new cycle is completed; the updated window data is input into the online diagnostic model, and the predicted values ​​of health indicators are output in real time. For each health indicator, multiple warning thresholds are preset. When the predicted value exceeds the first-level warning threshold, a visual prompt is given on the human-computer interaction interface; when the predicted value exceeds the second-level warning threshold, an audible and visual alarm is triggered and a command is sent to the test equipment to pause the test.

8. A fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The method also includes attribution analysis of the overall battery capacity degradation: constructing a battery simulation model of an aging mechanism sub-model, performing simulation using a composite test protocol, and generating simulation training samples containing external signal time series data and corresponding decoupling label vectors, wherein the decoupling label vectors are the percentage of capacity loss due to known aging mechanisms; Based on the benchmark diagnostic model, a multi-task decoupling model is constructed and trained, which can predict the percentage of capacity loss due to known aging mechanisms based on external signals. The capabilities of the multi-task decoupling model are transferred to an online decoupling diagnostic model through knowledge distillation, which is used to output an aging decoupling report.

9. A fault testing method for a lithium-ion battery as described in claim 1, characterized in that, The method also includes the detection of unknown faults: a synthetic anomaly perturbation module simulating unknown fault modes is added to the battery simulation model and randomly activated in the simulation to generate an extended decoupling label vector containing contributions from unknown anomalies. An unknown anomaly detection head is added to the multi-task decoupling model to form an enhanced multi-task decoupling model, which is then trained to enable the model to identify unknown faults. Through knowledge distillation, the capabilities of the enhanced multi-task decoupling model are transferred to an online anomaly diagnosis model, which outputs an unknown anomaly diagnosis report.

10. A fault testing method for a lithium-ion battery as described in claim 9, characterized in that, The fault warning also includes: setting independent multi-level warning thresholds for the percentage of unknown abnormal capacity loss predicted by the online anomaly diagnosis model; When the percentage of unknown abnormal capacity loss is greater than zero for the first time but lower than the first-level abnormal warning threshold, a first-level abnormal warning is triggered, the test continues, but the tester is prompted to pay attention; when the percentage of unknown abnormal capacity loss exceeds the second-level abnormal warning threshold, or continues to rise for several consecutive diagnostic cycles, a second-level abnormal warning is triggered, an audible and visual alarm is immediately triggered, a command is sent to the test equipment, and the current test cycle is terminated.

Images