Battery self-discharge prediction method and system based on artificial intelligence technology

Abstract

The application discloses a battery self-discharge prediction method and system based on artificial intelligence technology. The battery self-discharge prediction method based on artificial intelligence technology according to the application embodiment comprises the following steps: acquiring battery operation data, incremental capacity data and electrochemical impedance spectrum characteristic data; obtaining a prediction model according to the battery operation data, the incremental capacity data and the electrochemical impedance spectrum characteristic data; and predicting a self-discharge parameter of the battery according to the prediction model. The battery self-discharge prediction method and system based on artificial intelligence technology can greatly improve the sensitivity of battery self-discharge detection.

Description

Technical Field

[0001] This application relates to the field of battery technology, and in particular to a battery self-discharge prediction method and system based on artificial intelligence technology. Background Technology

[0002] With the widespread adoption of new energy vehicles, electrochemical energy storage systems, and portable electronic devices, the performance stability and safety of lithium-ion batteries, as core energy storage components, have become a key focus of the industry's development.

[0003] Self-discharge is the spontaneous loss of capacity caused by inherent processes such as internal electrochemical side reactions and physical micro-short circuits when a battery is left in an open-circuit resting state. This characteristic not only significantly reduces battery energy storage efficiency and exacerbates the degradation of battery pack consistency, but may also induce major safety accidents such as lithium plating and thermal runaway. Therefore, accurate and timely detection and fault classification of battery self-discharge rate is an indispensable core link in the whole life cycle management of batteries.

[0004] Current battery self-discharge detection technologies in the industry all have significant technical limitations, making it difficult to simultaneously achieve real-time performance, detection sensitivity, and adaptability to all operating conditions. These limitations fail to meet the management needs of the entire battery lifecycle. Specific shortcomings are as follows:

[0005] Static voltage method: The battery needs to be left to stand for a long time to obtain voltage decay data. The detection cycle is as long as several days to several weeks. It cannot be used for battery systems in operation. Moreover, the detection results are greatly affected by ambient temperature and battery history. It is not capable of identifying early micro self-discharge.

[0006] Equalization counting method: such as Chinese patent with publication number CN120870913A, which uses statistical information of the battery equalization process to realize self-discharge detection. Although it can meet the real-time requirements, it can only reflect the difference in voltage consistency of the battery pack and cannot characterize the changes in the internal electrochemical state of the battery. It is not sensitive to early micro self-discharge and is prone to missed detection.

[0007] Differential capacity curve method: such as Chinese patent publication number CN118275908A, which evaluates the self-discharge state based on the incremental capacity curve characteristics, but requires the acquisition of complete charge and discharge data of the battery, which places extremely high demands on the sampling accuracy and data integrity of the BMS, and the detection effect is greatly reduced for battery systems with long voltage platforms such as lithium iron phosphate.

[0008] Multi-process screening method: such as the Chinese patent with publication number CN116381518A, which completes the screening of self-discharge batteries through multi-stage static placement and charging / discharging processes, is only applicable to the battery manufacturing and delivery stage and cannot be used for online monitoring during battery use.

[0009] Traditional machine learning methods, such as the Chinese patent with publication number CN117991135A, build a self-discharge prediction model through machine learning models. However, they rely on a large amount of labeled training data, and the generalization ability of the model is limited by the working conditions and battery model, and cannot adapt to the aging state changes of the battery throughout its entire life cycle.

[0010] SOC difference method: such as the Chinese patent with announcement number CN119861297B, the self-discharge rate is calculated by the difference of SOC before and after standing. Although the influence of the equalization process is taken into account, it has extremely high requirements for the accuracy of BMS sensor and SOC estimation, and is difficult to apply in engineering.

[0011] Volume-voltage correlation method: such as Chinese patent application number 202510835206.4, which evaluates self-discharge characteristics by the correlation between battery volume change and voltage, is only applicable to the battery material system development stage and is difficult to use for online monitoring of finished batteries.

[0012] Furthermore, most existing technologies rely on single-dimensional external operating data such as voltage, capacity, and temperature, lacking direct perception of the battery's internal electrochemical state. This makes it impossible to capture early-stage internal state changes during self-discharge, resulting in the difficulty in timely detection of early, weak self-discharge. Electrochemical impedance spectroscopy (EIS) can accurately reflect core electrochemical information such as charge transfer, lithium-ion diffusion, and electrode interface states within the battery. It is an effective tool for diagnosing battery aging, lithium plating, and SEI film abnormalities. However, current technologies have not yet systematically applied it to real-time quantitative assessment and fault tracing of battery self-discharge, nor have they achieved multi-source fusion and adaptive evaluation of EIS characteristics and conventional BMS operating data.

[0013] Therefore, there is a need for a new battery self-discharge prediction method and system based on artificial intelligence technology that can overcome at least one of the above problems. Summary of the Invention

[0014] In view of the above problems, the purpose of this application is to provide a battery self-discharge prediction method and system based on artificial intelligence technology, which aims to significantly improve the sensitivity of battery self-discharge detection.

[0015] According to a first aspect of the embodiments of this application, a battery self-discharge prediction method based on artificial intelligence technology is provided, comprising:

[0016] Acquire battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data;

[0017] A prediction model is obtained based on the battery operation data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data;

[0018] The self-discharge parameters of the battery are predicted based on the prediction model.

[0019] Optionally, the battery operating data includes at least one selected from cell voltage, total current, temperature, and equilibrium state;

[0020] The incremental capacity data includes at least one selected from differential capacity curves, high-voltage inflection point capacity, and low-voltage inflection point capacity.

[0021] The electrochemical impedance spectroscopy characteristic data includes at least one selected from the real part of impedance at the characteristic frequency, the imaginary part of impedance at the characteristic frequency, and the peak value of relaxation time distribution at the characteristic frequency.

[0022] The prediction model derived from the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data includes:

[0023] Based on the battery operation data, extract and construct the operation feature vector;

[0024] Based on the incremental capacity data, extract and construct the incremental capacity feature vector;

[0025] Based on the electrochemical impedance spectroscopy feature data, an electrochemical impedance spectroscopy feature vector is extracted and constructed.

[0026] The operating feature vector, the incremental capacity feature vector, and the electrochemical impedance spectroscopy feature vector are normalized.

[0027] The prediction model is obtained by calculating the normalized operating feature vector, incremental capacity feature vector, and electrochemical impedance spectroscopy feature vector.

[0028] Optionally, the battery self-discharge prediction method further includes:

[0029] Obtain historical battery statistics;

[0030] The prediction model is obtained based on the battery operating data, the incremental capacity data, the electrochemical impedance spectroscopy characteristics, and the battery historical statistical data.

[0031] The historical statistical data includes at least one selected from available balance counts, effective balance counts, and cumulative idle time.

[0032] Optionally, the battery self-discharge prediction method further includes:

[0033] When the battery is in a static state or operating at a constant current less than a set threshold, inject a multi-frequency sinusoidal current disturbance with an amplitude less than the threshold.

[0034] The voltage response under the multi-frequency sinusoidal current perturbation was measured and the impedance spectrum was calculated. The real and imaginary parts of the impedance at different characteristic frequency points were extracted to generate the electrochemical impedance spectrum eigenvector.

[0035] The electrochemical impedance spectroscopy eigenvectors are used to obtain the prediction model.

[0036] Optionally, the battery self-discharge prediction method further includes:

[0037] The prediction model is adjusted according to the dynamic weight matrix;

[0038] The generation of the dynamic weight matrix includes:

[0039] The operating characteristic weights are calculated based on the current state of charge of the battery.

[0040] The incremental capacity feature weights are calculated based on the battery's current temperature and the current rate of the most recent complete cycle.

[0041] The characteristic weights of the electrochemical impedance spectroscopy are calculated based on the current resting time and operating current of the battery.

[0042] The dynamic weight matrix is ​​obtained by calculating the operating characteristic weight, the incremental capacity characteristic weight, and the electrochemical impedance spectroscopy characteristic weight.

[0043] Optionally, the battery self-discharge prediction method further includes:

[0044] Obtain the baseline feature vector corresponding to a healthy battery;

[0045] The weighted Mahalanobis distance formula is used to calculate the multi-dimensional deviation between the feature vectors corresponding to the current battery operating data, incremental capacity data, and electrochemical impedance spectroscopy feature data and the benchmark feature vector;

[0046] The weighted Mahalanobis distance is mapped to a self-discharge assessment value using an exponential mapping function.

[0047] Optionally, the battery self-discharge prediction method further includes:

[0048] Obtain the self-discharge assessment value of the nearest neighboring number of times;

[0049] Calculate the average slope of change based on the nearest self-discharge assessment values ​​of the set number of times.

[0050] Specifically, when the average slope of change is greater than a preset slope threshold, the battery is determined to be in a state of accelerated self-discharge.

[0051] Optionally, the battery self-discharge prediction method further includes:

[0052] Establish a modified model, which is used to achieve at least one of the following steps:

[0053] Receive the feature vector, the weighted Mahalanobis distance, the self-discharge assessment value, and the average change slope;

[0054] The system performs abnormal data verification on the input data, identifies sensor drift, data loss, and disturbances in the measurement of electrochemical impedance spectroscopy characteristic data, and completes data repair or invalidation marking.

[0055] The self-discharge assessment value and the calculated average change slope are corrected based on the current state of charge, temperature, health status, and application scenario of the battery.

[0056] For abnormal states, the cause of abnormal self-discharge is determined based on the fitting characteristics and operating characteristics of the electrochemical impedance spectroscopy data.

[0057] Optionally, the modified model includes a convolutional neural network and a long short-term memory network;

[0058] The modified model is also used to achieve at least one of the following steps:

[0059] Construct a battery self-discharge data sample library, and collect and store the benchmark feature vector and the self-discharge evaluation value;

[0060] When the number of new valid samples in the battery self-discharge data sample library reaches a preset number or the judgment accuracy is lower than a preset threshold, online incremental learning is triggered.

[0061] The revised model is optimized based on the newly added valid samples;

[0062] Clustering algorithm is used to update the baseline feature vector of healthy batteries of the same model and under the same working conditions;

[0063] Using the prediction error of the self-discharge assessment value as the objective function, the parameters in the dynamic weight formula are adjusted by the gradient descent method.

[0064] The scale parameter in the self-discharge evaluation value calculation formula is adjusted according to at least one of the battery model, material system, and application scenario.

[0065] According to a second aspect of this application, a battery self-discharge prediction system based on artificial intelligence technology is provided, comprising:

[0066] The acquisition unit is used to acquire battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data;

[0067] The computing unit is used to obtain a prediction model based on the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data.

[0068] A prediction unit is used to predict the self-discharge parameters of the battery based on the prediction model.

[0069] According to the battery self-discharge prediction method and system based on artificial intelligence technology in the embodiments of this application, a prediction model is obtained based on battery operation data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data; then, the self-discharge parameters of the battery are predicted based on the prediction model; during the prediction process, battery operation data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data are fused to achieve deep fusion of multi-source information of electrochemical mechanism and operation data, and to evaluate the self-discharge state from the dual dimensions of internal electrochemical mechanism and external operation performance of the battery, so as to achieve a comprehensive characterization of the self-discharge state, break through the limitations of traditional single-dimensional detection and evaluation, improve the sensitivity of early micro self-discharge identification by more than an order of magnitude, and can capture early signals of self-discharge anomalies in advance.

[0070] Furthermore, when the battery is in a static state or operating at a constant current less than a set threshold, a multi-frequency sinusoidal current disturbance with an amplitude less than the threshold is injected to achieve impedance spectrum measurement without stopping the battery or performing a full charge and discharge. Without interfering with the normal operation of the battery, key impedance characteristics reflecting the internal state can be accurately obtained, truly realizing real-time online monitoring of self-discharge, and adapting to dynamic operating scenarios such as vehicle and energy storage.

[0071] Furthermore, the weight allocation of different features is adaptively adjusted according to the actual operating state of the battery, which greatly improves the detection robustness under all operating conditions and throughout the entire life cycle. A dynamic weight matrix based on battery SOC, temperature, SOH, and operating conditions is constructed, which solves the problem of decreased detection accuracy of traditional fixed weight models under high and low temperatures, high and low SOC, and different aging stages.

[0072] Furthermore, a self-discharge comprehensive score based on weighted Mahalanobis distance is proposed, which can accurately characterize the severity of self-discharge; combined with grading thresholds and trend judgment, a full-process tiered control of "monitoring-early warning-intervention" is realized, taking into account both safety and user experience.

[0073] Furthermore, the self-discharge comprehensive score based on characteristic deviation and weighted Mahalanobis distance transforms the originally discrete anomaly judgment into a (continuous) quantitative assessment, which can clearly distinguish the severity level of self-discharge. Attached Figure Description

[0074] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0075] Figure 1 A flowchart of a battery self-discharge prediction method according to Embodiment 1 of this application is shown;

[0076] Figure 2A flowchart of the battery self-discharge prediction method according to Embodiment 2 of this application is shown;

[0077] Figure 3 A schematic diagram of the system architecture for implementing the battery self-discharge prediction method according to Embodiment 3 of this application is shown;

[0078] Figure 4 A schematic diagram of data acquisition and preprocessing according to Embodiment 4 of this application is shown;

[0079] Figure 5 A schematic diagram illustrating the construction of multi-source feature vectors according to Embodiment Six of this application is shown;

[0080] Figure 6 A flowchart of the battery self-discharge prediction method according to Embodiment 5 of this application is shown;

[0081] Figure 7 A schematic diagram illustrating the algorithm principle according to Embodiment Seven of this application is shown;

[0082] Figure 8 A schematic diagram of the structure of a battery self-discharge prediction system according to an embodiment of this application is shown. Detailed Implementation

[0083] The present application will now be described in more detail with reference to the accompanying drawings. In the various drawings, the same elements are indicated by similar reference numerals. For clarity, the various parts in the drawings are not drawn to scale. Furthermore, some well-known parts may not be shown. The present application may be presented in various forms; only some examples are described below.

[0084] Figure 1 A flowchart illustrating a battery self-discharge prediction method based on artificial intelligence technology according to an embodiment of this application is shown. Figure 1 As shown, the battery self-discharge prediction method according to an embodiment of this application includes the following steps:

[0085] In step S101, battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data are acquired;

[0086] Acquire battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data. Optionally, the battery operating data includes at least one selected from cell voltage, total current, temperature, equilibrium state, etc.; the incremental capacity data includes at least one selected from differential capacity curve, high voltage inflection point capacity, low voltage inflection point capacity, etc.; the electrochemical impedance spectroscopy characteristic data includes at least one selected from the real part of impedance at the characteristic frequency, the imaginary part of impedance at the characteristic frequency, the peak value of relaxation time distribution at the characteristic frequency, etc.

[0087] In step S102, a prediction model is obtained based on the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data;

[0088] Based on the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data, a prediction model (prediction feature) is obtained. Optionally, the prediction model is a model based on artificial intelligence technology, i.e., an artificial intelligence prediction model.

[0089] In step S103, the self-discharge parameters of the battery are predicted according to the prediction model.

[0090] The self-discharge parameters of the battery are predicted based on the prediction model, and the obtained self-discharge parameters include, but are not limited to, the self-discharge rate.

[0091] In some embodiments, the prediction model obtained based on the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data includes:

[0092] Based on the battery operation data, extract and construct the operation feature vector;

[0093] Based on the incremental capacity data, extract and construct the incremental capacity feature vector;

[0094] Based on the electrochemical impedance spectroscopy feature data, an electrochemical impedance spectroscopy feature vector is extracted and constructed.

[0095] The operating feature vector, the incremental capacity feature vector, and the electrochemical impedance spectroscopy feature vector are normalized.

[0096] The prediction model is obtained by calculating the normalized operating feature vector, incremental capacity feature vector, and electrochemical impedance spectroscopy feature vector.

[0097] In some embodiments, the battery self-discharge prediction method further includes:

[0098] Obtain historical battery statistics;

[0099] The prediction model is obtained based on the battery operating data, the incremental capacity data, the electrochemical impedance spectroscopy characteristics, and the battery historical statistical data.

[0100] The historical statistical data includes at least one selected from available balance counts, effective balance counts, and cumulative idle time.

[0101] In some embodiments, the battery self-discharge prediction method further includes:

[0102] When the battery is in a static state or operating at a constant current less than a set threshold, inject a multi-frequency sinusoidal current disturbance with an amplitude less than the threshold.

[0103] The voltage response under the multi-frequency sinusoidal current perturbation was measured and the impedance spectrum was calculated. The real and imaginary parts of the impedance at different characteristic frequency points were extracted to generate the electrochemical impedance spectrum eigenvector.

[0104] The electrochemical impedance spectroscopy eigenvectors are used to obtain the prediction model.

[0105] In some embodiments, the battery self-discharge prediction method further includes:

[0106] The prediction model is adjusted according to the dynamic weight matrix;

[0107] The generation of the dynamic weight matrix includes:

[0108] The operating characteristic weights are calculated based on the current state of charge of the battery.

[0109] The incremental capacity feature weights are calculated based on the battery's current temperature and the current rate of the most recent complete cycle.

[0110] The characteristic weights of the electrochemical impedance spectroscopy are calculated based on the current resting time and operating current of the battery.

[0111] The dynamic weight matrix is ​​obtained by calculating the operating characteristic weight, the incremental capacity characteristic weight, and the electrochemical impedance spectroscopy characteristic weight.

[0112] In some embodiments, the battery self-discharge prediction method further includes:

[0113] Obtain the baseline feature vector corresponding to a healthy battery;

[0114] The weighted Mahalanobis distance formula is used to calculate the multi-dimensional deviation between the feature vectors corresponding to the current battery operating data, incremental capacity data, and electrochemical impedance spectroscopy feature data and the benchmark feature vector;

[0115] The weighted Mahalanobis distance is mapped to a self-discharge assessment value using an exponential mapping function.

[0116] Optionally, the battery self-discharge prediction method also includes:

[0117] Obtain the self-discharge assessment value of the nearest neighboring number of times;

[0118] Calculate the average slope of change based on the nearest self-discharge assessment values ​​of the set number of times.

[0119] Specifically, when the average slope of change is greater than a preset slope threshold, the battery is determined to be in a state of accelerated self-discharge.

[0120] Optionally, the battery self-discharge prediction method also includes:

[0121] Establish a modified model, which is used to achieve at least one of the following steps:

[0122] Receive the feature vector, the weighted Mahalanobis distance, the self-discharge assessment value, and the average change slope;

[0123] The system performs abnormal data verification on the input data, identifies sensor drift, data loss, and disturbances in the measurement of electrochemical impedance spectroscopy characteristic data, and completes data repair or invalidation marking.

[0124] The self-discharge assessment value and the calculated average change slope are corrected based on the current state of charge, temperature, health status, and application scenario of the battery.

[0125] For abnormal states, the cause of abnormal self-discharge is determined based on the fitting characteristics and operating characteristics of the electrochemical impedance spectroscopy data.

[0126] Optionally, the modified model includes convolutional neural networks and long short-term memory networks;

[0127] The modified model is also used to achieve at least one of the following steps:

[0128] Construct a battery self-discharge data sample library, and collect and store the benchmark feature vector and the self-discharge evaluation value;

[0129] When the number of new valid samples in the battery self-discharge data sample library reaches a preset number or the judgment accuracy is lower than a preset threshold, online incremental learning is triggered.

[0130] The revised model is optimized based on the newly added valid samples;

[0131] Clustering algorithm is used to update the baseline feature vector of healthy batteries of the same model and under the same working conditions;

[0132] Using the prediction error of the self-discharge assessment value as the objective function, the parameters in the dynamic weight formula are adjusted by the gradient descent method.

[0133] The scale parameter in the self-discharge evaluation value calculation formula is adjusted according to at least one of the battery model, material system, and application scenario.

[0134] Figure 2 A flowchart of a battery self-discharge prediction method according to Embodiment 2 of this application is shown, specifically a battery self-discharge rate prediction and fault classification method based on AI-fused static features. Figure 2 As shown, the battery self-discharge prediction method according to Embodiment 2 of this application includes the following steps:

[0135] In step S201, data acquisition and preprocessing are performed;

[0136] Simultaneously, battery operation data, incremental capacity data, and EIS (Electrochemical Impedance Spectroscopy) characteristic data are collected, and the battery data is cleaned and standardized preprocessed. Optionally, the collected and preprocessed data also includes historical battery statistics. This comprehensive data collection and standardized preprocessing lays a solid data foundation for subsequent feature fusion and model evaluation; the collected data covers four core dimensions, taking into account both the external performance and internal electrochemical state of the battery.

[0137] In step S202, feature extraction and normalization are performed.

[0138] Extract and construct running feature vectors respectively Capacity increment feature vector EIS feature vectors Complete the normalization process for all features.

[0139] In step S203, the severity score of self-discharge is calculated;

[0140] Multi-feature concatenation and self-discharge score calculation: Three types of features are concatenated into the total feature set. A dynamic weight matrix is ​​obtained based on the actual battery operating conditions. The deviation between the current battery state and a healthy battery state is obtained using weighted Mahalanobis distance, which is then mapped to obtain a self-discharge severity score in the range of 0 to 100. .

[0141] In step S204, the adaptive threshold state is determined;

[0142] Based on self-discharge severity score By setting three-level thresholds and combining the slope of the scoring trend to determine the self-discharge development rate, the basic state of self-discharge is determined.

[0143] In step S205, the AI ​​performs self-assessment and outputs a warning.

[0144] The AI ​​self-judgment module self-corrects abnormal data, self-adjusts operating conditions, and self-diagnoses the driving force of self-discharge. It outputs qualitative status indicators and graded handling suggestions, which together play a role in BMS / EMS graded early warning and targeted intervention.

[0145] In step S206, AI performs self-iterative optimization.

[0146] The AI ​​self-iterative module self-adjusts model accuracy, self-updates baseline feature vectors, self-adjusts weight coefficients, and self-adjusts scale coefficients, enabling self-iteration throughout the system's entire lifecycle.

[0147] Figure 3A schematic diagram of the system architecture for implementing the battery self-discharge prediction method according to Embodiment 3 of this application is shown. (Combined with...) Figure 3 As shown, the system architecture of the battery self-discharge prediction method includes the following parts: data acquisition, feature extraction, self-discharge score calculation, AI status determination, graded early warning and intervention, vehicle / energy storage / cloud platform implementation, and AI self-iterative optimization.

[0148] Optionally, the system architecture also includes fault tolerance and security mechanisms, including a fallback mechanism for AI model failures, a parameter update rollback mechanism, a hierarchical operation permission mechanism, and a real-time data backup mechanism, to ensure stable system operation under complex working conditions. Specifically:

[0149] AI model failure fallback: If the AI ​​self-judgment / self-iteration module fails, the system automatically switches to the traditional fixed threshold judgment mode to ensure the normal operation of the basic early warning function, while triggering the AI ​​module failure early warning.

[0150] Parameter update rollback: If the AI ​​self-iterative optimization of parameters / models causes a drop in judgment accuracy of more than 10%, the system will automatically roll back to the previous version of valid parameters / models to avoid optimization failure;

[0151] Access control is tiered: AI's self-iterative parameter updates and model optimizations are all executed automatically. Humans only have the permissions to view / pause / rollback to prevent human error.

[0152] Data backup: The sample library, model parameters, and benchmark feature vectors are all backed up in real time to the cloud to prevent data loss and support rapid recovery after system failure.

[0153] This solution can be integrated into vehicle BMS, energy storage EMS, or battery cloud management platforms. At the hardware level, only a low-frequency micro-disturbance EIS measurement module needs to be adapted, without interfering with normal battery operation. At the software level, standardized API interfaces are provided for rapid integration with existing battery management systems. The entire system can achieve millisecond-level real-time status assessment, hourly light parameter iteration, and day-level deep model optimization, fully meeting the real-time and long-term optimization requirements of online monitoring throughout the battery's entire lifecycle, providing full-cycle technical support for the safe and stable operation of lithium-ion batteries.

[0154] Figure 4 A schematic diagram illustrating data acquisition and preprocessing according to Embodiment 4 of this application is shown. (Combined with...) Figure 4 As shown, four main categories of data were collected: operational data, capacity characteristics, EIS electrochemical characteristics, and historical statistical data. The collected data underwent standardization preprocessing to obtain effective data output.

[0155] Specifically, the operating data includes data selected from individual cell voltages. Total current ,temperature At least one of the following: balanced state, etc. Optionally, the normal sampling frequency is 1Hz, and event details are recorded synchronously when a balanced event is triggered.

[0156] (Incremental) capacity characteristics include those selected from charge / discharge segments. Curve (differential capacity curve), high voltage inflection point capacity Low voltage inflection point capacity At least one of the following. Incremental capacity characteristics are acquired and extracted after each complete charge-discharge cycle.

[0157] EIS electrochemical characteristics include the real part of impedance selected from the characteristic frequency. virtual part At least one of the following: Distribution of Relaxation Times (DRT) peak value, etc. EIS electrochemical characteristics are collected daily or per cycle during battery rest or low-current operation. Optionally, EIS measurements employ a low-frequency micro-perturbation strategy, specifically: when the battery meets... In the static state or constant current low current operation state, a multi-frequency sinusoidal current disturbance with an amplitude of less than 0.5%C is injected, covering a frequency range of 0.1Hz-10kHz, preferably five characteristic frequency points: 0.1Hz, 1Hz, 10Hz, 100Hz, and 1000Hz. The impedance spectrum is calculated by measuring the voltage response, and each characteristic frequency point is extracted. and Constructing EIS feature vectors not only ensures the effectiveness of the features and enables online acquisition of EIS features, but also significantly reduces the computational load of BMS.

[0158] Extract Zre and Zim for each frequency point to construct the EIS feature vector:

[0159] .

[0160] Historical statistics include data selected from available equilibrium counts. Effective balanced counting Cumulative settling time At least one of the following, historical statistical data is continuously updated and stored as the battery operates.

[0161] Optionally, corresponding collection frequencies and triggering conditions can be set for different data types. The collected raw data undergoes standardized preprocessing such as filtering, noise reduction, outlier removal, and timestamp alignment to ensure data validity. For data of different dimensions, feature extraction and normalization are performed separately to eliminate the impact of dimensional differences on the fusion evaluation, ultimately forming three categories of standardized feature vectors.

[0162] Figure 5 A schematic diagram illustrating the construction of multi-source feature vectors according to Embodiment Six of this application is shown. Combined with... Figure 5 As shown, the constructed total feature vector For example, it can be 10-dimensional, including EIS eigenvectors representing core electrochemical characteristics. , representing the capacity increment eigenvector of the internal material state , representing the operational feature vector of external operational performance .

[0163] Specifically, in combination Figure 4 Related description, EIS feature vector Including the real part of the impedance at the characteristic frequency virtual part EIS feature vectors are composed of three dimensions: relaxation time distribution peak value, relaxation time distribution peak value, and other features. Charge transfer resistance, interfacial reaction characteristics, diffusion impedance, and lithium-ion transport characteristics. Capacity increment eigenvector. Including the charge and discharge segments Curve, high voltage inflection point capacity Low voltage inflection point capacity These three dimensions, capacity increment feature vector The voltage inflection point capacity difference reflects the loss of active material. Operating characteristic vector. Including individual unit voltage Total current ,temperature The four dimensions of equilibrium state are used to run feature vectors. The normalized voltage difference reflects consistency, the equilibrium count reflects historical equilibrium status, and the temperature correction parameter eliminates environmental influences. The specific construction methods for these three types of feature vectors are as follows:

[0164] Running feature vectors The core coverage includes real-time operating parameters such as battery voltage consistency, equalization state, and temperature conditions, reflecting the battery's external operating status and consistency performance. Through normalization with the initial state and rated range, interference from individual battery differences and operating condition fluctuations is eliminated, ultimately forming a 4-dimensional standardized feature:

[0165]

[0166] in, This represents the current maximum voltage difference. This represents the initial maximum voltage difference of the battery (initial value). To balance the count values ​​and normalize them according to the total number of battery packs; This is the average temperature. Optionally, ; .

[0167] Capacity Increment Feature Vector Based on the dQ / dV curves of the charge-discharge segments, the capacity difference characteristics at the high and low voltage inflection points are extracted. After normalization by the rated capacity, a 3D feature is formed, which accurately reflects the capacity decay (loss of active materials inside the battery) and changes in electrochemical reactions.

[0168]

[0169] in, This refers to the battery's rated capacity. This represents the high-voltage inflection point capacity of the current cycle. The high-voltage inflection point capacity of the initial cycle; This represents the low-voltage inflection point capacity of the current cycle. This represents the theoretical capacity difference at the high-voltage inflection point.

[0170] By fitting the equivalent circuit ( Three core electrochemical parameters were extracted from impedance spectroscopy: charge transfer resistance, which reflects interfacial reactivity. Double-layer capacitance, which reflects changes in interface area. Diffusion resistance coefficient, which reflects the diffusion ability of lithium ions. Then, by normalizing with the initial state of battery health, the internal state of the battery can be directly quantified. EIS feature vector for:

[0171]

[0172] in, It is a charge transfer resistor; It is a double-layer capacitor; This represents the diffusion impedance coefficient; the subscript "0" indicates the initial state of battery health.

[0173] Furthermore, the steps for multi-dimensional feature fusion are as follows:

[0174] The three types of features are concatenated into a total feature vector. .

[0175] Figure 6 A flowchart of a battery self-discharge prediction method according to Embodiment 5 of this application is shown. Figure 6As shown, after feature extraction and normalization, multi-dimensional feature fusion and self-discharge (severity) scoring are performed. Unlike traditional fixed-weight fusion schemes, we first construct a dynamic weight matrix based on the current operating state of the battery. The weight allocation is not static but is adjusted in real time according to the battery's SOC, temperature, SOH, resting time, and current rate: the weight of operating features adapts to the degree of battery aging and SOC deviation, the weight of capacity increment features is adjusted according to temperature conditions and charge / discharge rate, and the weight of EIS electrochemical features is dynamically optimized according to resting time and current magnitude, ensuring that the most effective feature dimensions receive the highest weight allocation under different operating conditions, maximizing the model's evaluation accuracy.

[0176] The three types of features are concatenated into a total feature vector. .

[0177] Perform dynamic weight matrix calculation: Calculate the weight vector based on the battery's current SOC, temperature, and SOH. ,in, , , In the formula, These are adjustable parameters; The current multiplier for the most recent complete cycle; The battery is in its initial health state of 100%. This is the cumulative settling time; This refers to the battery's rated capacity.

[0178] Construct a diagonal weight matrix based on the weight vector. ,in Indicates length is A vector of all 1s.

[0179] A weighted Mahalanobis distance is calculated to measure the deviation of the current state from the healthy baseline state. The formula is as follows:

[0180]

[0181] in, The baseline feature vector of a healthy battery at the same SOC and temperature (which can be obtained from historical data or calibration); It is the covariance matrix of the eigenvectors (which can be obtained from statistics of healthy battery clusters).

[0182] To enhance interpretability, a self-discharge severity score mapping was performed, mapping the Mahalanobis distance to a self-discharge severity score in the range of 0 to 100. The formula is:

[0183]

[0184] in, These are scale parameters, adjustable depending on the battery type (e.g., ...). ).

[0185] Combination Figure 6 As shown, the specific rules for adaptive threshold determination include:

[0186] Tiered threshold setting: This is the normal state; It is in a state of slight self-discharge; This indicates a severe self-discharge state.

[0187] Trend judgment: Calculate the most recent Average slope of sub-scores .

[0188]

[0189] Among them, if (For example ( / day), indicating that self-discharge is in an accelerated development state; The preset number of sampling times; This is a preset slope threshold.

[0190] Optionally, the AI ​​self-judgment module adopts an architecture combining a CNN (Convolutional Neural Network) + LSTM (Long Short-Term Memory) deep learning classification model with a rule engine. The AI ​​self-judgment module is the intelligent central hub of the entire early warning system, possessing independent judgment logic and fault-tolerance mechanisms. It can complete the entire process of intelligent judgment, anomaly tracing, and decision suggestion generation without manual intervention. It also features a full-lifecycle AI self-iteration mechanism, enabling the system to continuously self-optimize. The core execution process of the AI ​​self-judgment module includes:

[0191] Receive the feature vector, weighted Mahalanobis distance, self-discharge score, and trend slope standardized data output by the fusion evaluation layer;

[0192] Abnormal data self-verification: Identifies abnormal data such as sensor drift, missing data, and measurement disturbance, and completes data repair (automatically marks invalid data and completes repair through interpolation, replacement with similar historical data, etc.) or invalid marking. When the proportion of invalid data exceeds 30%, a data acquisition fault warning is triggered; the fault warning and self-discharge warning are completely decoupled to ensure the validity of the judgment basis;

[0193] Adaptive threshold correction based on operating conditions: Dynamically correct the graded threshold and slope threshold based on the current SOC, temperature, SOH and application scenario to avoid false alarms and missed alarms;

[0194] Basic state determination: Based on the corrected threshold, output the final state label, including normal, slight self-discharge (stable), slight self-discharge (accelerated), severe self-discharge (stable), and severe self-discharge (accelerated).

[0195] Anomaly Root Cause Tracing: For abnormal states, the root cause of self-discharge anomalies is located based on EIS fitting features, operational characteristics, and an electrochemical fault correlation model; optionally, if An abnormal increase indicates an abnormal charge transfer impedance at the battery interface; abnormal voltage difference and equalization count indicate deterioration in battery pack consistency; if... An abnormal increase indicates a decrease in lithium-ion diffusion capacity. Optionally, the root cause of the self-discharge anomaly can be precisely located, such as damage to the SEI film at the interface, increased charge transfer impedance, decreased lithium-ion diffusion capacity, deterioration of battery pack consistency, or malfunction of the equalization system, completely overcoming the limitations of traditional solutions that "only report anomalies without finding the cause."

[0196] Tiered decision-making and early warning output: Based on status tags, traceability results and application scenarios, four-level processing suggestions are generated, corresponding to five-level tiered early warnings, and the BMS (Battery Management System) / EMS (Energy Management System) are linked to execute corresponding intervention actions, realizing integrated management and control from status monitoring and early warning prompts to automatic intervention execution.

[0197] Optionally, the tiered early warning and coordinated execution specifically include:

[0198] Level 0 has no warning, corresponding to the normal state. There are no prompts, and basic operating parameters are displayed (self-discharge score, current SOC, temperature, SOH, etc.).

[0199] Level 1 alert corresponds to a slight self-discharge (stable) state. A green alert box is silent, and the score and status are displayed, with a recommendation for close monitoring.

[0200] Level 2 warning corresponds to a slight self-discharge (acceleration) state. A yellow prompt box will sound a short alarm, display the score and status, and output a mild intervention suggestion, such as automatically triggering passive equalization of the battery pack for 1 hour.

[0201] Level 3 alarm corresponds to severe self-discharge (stable) state. An orange prompt box indicates an intermittent audible alarm. The alarm displays the scoring and tracing results, outputs a moderate intervention suggestion and abnormal cell number, automatically triggers active balancing of the battery pack, and limits the charge / discharge rate to ≤0.5C.

[0202] Level 4 emergency alarm corresponds to severe self-discharge (acceleration) state. A red prompt box with continuous audible alarm and a pop-up window at the top displays the score, trend, and source tracing results. It outputs emergency handling suggestions and abnormal cell number, automatically limits battery charging and discharging (limits power output in vehicle scenarios and cuts off the cell / module in energy storage scenarios), and sends emergency alarm information (including location, battery status, and handling suggestions) to the cloud platform.

[0203] Optionally, it also includes an AI self-iteration module. This module enables the early warning system to learn itself, optimize its model, and update its parameters. Based on an online incremental learning algorithm, it can continuously optimize model performance using massive amounts of real-world data without offline retraining, solving the problems of poor generalization and limited adaptability of traditional machine learning methods, and achieving full lifecycle self-evolution of the system. The AI ​​self-iteration module implements at least one of the following functions based on the online incremental learning algorithm:

[0204] Model accuracy self-optimization: When the number of new effective samples reaches 1000, or when the system's judgment accuracy (the matching degree between the actual state and the AI's judgment result) is lower than 90%, online incremental learning is automatically started. The new samples are used to fine-tune the weights of the CNN+LSTM classification model to optimize the accuracy of state judgment and root cause tracing. At the same time, the model is tested for generalization, and the optimization effect is verified using unseen working condition samples to ensure the model's adaptability under new working conditions.

[0205] Benchmark eigenvector Self-updating: Regularly analyzes the health status of batteries of the same model and under the same operating conditions (e.g., The feature vectors of the battery (which are stable) are used to generate new benchmark feature vectors through clustering algorithms to replace the fixed benchmark values, making the calculation of weighted Mahalanobis distance more consistent with the actual aging pattern of the battery and improving the accuracy of the assessment.

[0206] Dynamic weight parameter self-tuning: Using the prediction error of self-discharge score (the deviation between the score and the actual self-discharge rate of the battery) as the objective function, the adjustable parameters are adjusted in real time using the gradient descent method. Optimize the weight allocation under different working conditions (high SOC / low SOC, high temperature / low temperature, high SOH / low SOH) to improve the effectiveness of feature fusion.

[0207] Scale parameters Self-adaptive: Automatically adjusts the value of the scale parameter τ based on the battery model, material system, and application scenario. For example, for lithium iron phosphate batteries (with a long voltage plateau). The value was optimized from AI to 3 and from ternary lithium battery to 2, so that the score can more accurately reflect the differences in self-discharge of different battery systems.

[0208] Optionally, the self-iterative closed-loop mechanism includes:

[0209] Data acquisition and automatic labeling: Real-time acquisition of full battery operation data, automatic sample labeling and cleaning based on subsequent battery state changes, and storage in a full life cycle sample library;

[0210] Trigger condition verification: The decision-making layer monitors the iteration trigger conditions in real time. If the conditions are met, the iteration task is locked, the iteration level is divided, computing resources are allocated, and all parameters and models of the current stable version are archived.

[0211] Iterative task execution: According to the iteration level, the optimization tasks of the corresponding modules are executed to generate the optimized benchmark system, parameters, models, and decision rules, forming iterative candidate versions;

[0212] Offline performance verification: The candidate version is tested offline using the memory replay dataset and the newly added validation set. The core metrics (accuracy ≥ 92%, false negative rate = 0, false positive rate ≤ 8%) must be better than the previous stable version; otherwise, the iteration is terminated.

[0213] Gray-scale release and online verification: Candidate versions that pass offline verification undergo a 72-hour gray-scale trial run on 10% of the target battery cells, continuously monitoring the operating status and performance indicators. If there are no abnormalities or performance degradation, the next step is initiated.

[0214] Full update and version archiving: After the gray-scale verification is passed, the iterative version is released to all target battery units, and the parameters, models and benchmarks on the edge and cloud are updated synchronously. At the same time, version information is archived, recording the iteration content, verification results, release time, and generating a unique version number.

[0215] Continuous performance monitoring and rollback: After the version update, the system performance is monitored for 7 consecutive days. If any abnormal situation occurs, such as a drop in accuracy of more than 10%, a significant increase in false positive rate, or missed events, the system will be automatically rolled back to the previous stable version immediately, and an emergency iteration will be triggered to locate the problem.

[0216] An AI self-iterative module based on an online incremental learning algorithm was built, a battery self-discharge big data sample library was constructed, and iteration trigger conditions were set. It realized four core functions: model accuracy self-optimization, benchmark feature vector self-updating, dynamic weight parameter self-tuning, and scale parameter self-adaptation. It completed the closed-loop self-evolution of "data collection-judgment output-effect verification-model optimization", which solved the problems of poor generalization ability and inability to adapt to the aging of the entire battery life cycle of traditional machine learning models.

[0217] Optionally, a battery self-discharge big data sample library can be built to collect and store all data during system operation in real time (as the core data foundation for self-iteration), including: standardized feature vectors, weighted Mahalanobis distance, self-discharge scores, AI judgment results, actual treatment measures, subsequent changes in battery state (to verify the accuracy of judgment), operating condition information (SOC, temperature, SOH, application scenario), etc., to provide labeled real samples for iteration; the sample library supports retrieval by battery model, application scenario, and aging state, which is convenient for incremental learning.

[0218] The sample database is stored using a three-level index system: "single battery cell - single cycle - timestamp" (a comprehensive data storage specification), covering four core data categories:

[0219] Basic characteristic data: standardized feature vector, weighted Mahalanobis distance, self-discharge score, dynamic weight parameters, operating condition information (SOC, temperature, SOH, current ratio);

[0220] AI-powered data analysis includes: status labels, fault tracing results, warning levels, and generated decision-making suggestions.

[0221] Closed-loop verification data: actual intervention measures implemented, changes in battery state after intervention, true self-discharge rate verified by post-disassembly / testing, and records of false alarms / missed alarms;

[0222] Metadata: Battery model, material system, batch number, application scenario (vehicle / energy storage / portable device), cumulative cycle count, and service life.

[0223] Abandoning manual labeling, automatic labeling is achieved based on the state changes throughout the battery's entire life cycle, with a labeling accuracy of ≥95%.

[0224] Positive samples (normal state): Battery data with a self-discharge score consistently ≤20, stable state without warning, no abnormal capacity decay in the subsequent 30 cycles, and no frequent equalization triggers;

[0225] Negative samples (abnormal self-discharge): triggering level 2 or higher warnings, subsequent verification confirming excessive self-discharge rate, disassembly confirming electrochemical faults, and genuine abnormal data verified manually;

[0226] Sample classification: Based on the level of abnormality, samples are divided into minor self-discharge and severe self-discharge; based on the root cause of the fault, samples are divided into 6 types of source tracing samples; based on the operating conditions, samples are divided into scenarios such as low temperature, high temperature, high SOC, low SOC, static, and running, so as to achieve iterative optimization by scenario.

[0227] Streaming incremental storage: Stores only valid samples, automatically removes duplicates and noise, eliminates abnormal and interfering data, and avoids sample redundancy;

[0228] Memory replay dataset: 10% of the classic full-condition samples are retained to prevent catastrophic forgetting during incremental learning;

[0229] Access control: The sample database only supports appending and reading, and prohibits modification / deletion of historical data to ensure data traceability;

[0230] Hot and cold data separation: Hot data from the past 3 months is stored at the edge / local, while historical cold data is archived to the cloud, balancing access efficiency and storage costs.

[0231] Figure 7 A schematic diagram illustrating the algorithm principle according to Embodiment Seven of this application is shown. Combined with... Figure 7 As shown, the battery self-discharge prediction method according to the embodiments of this application is implemented through some core algorithms.

[0232] Specifically, the system takes real-time operating parameters such as SOC, temperature T, SOH, and resting time as input, calculates and adaptively assigns weights to three types of features using a dynamic weight matrix, quantifies the deviation from the health benchmark using weighted Mahalanobis distance, converts the results into a 0-100 self-discharge severity score using a scoring mapping, and outputs a self-discharge severity score.

[0233] In one specific embodiment of this application, a method for predicting battery self-discharge rate and classifying faults based on AI-integrated static features is provided. This method is applied to the management system of a ternary lithium battery pack in a new energy vehicle. The battery pack has a rated capacity of 100Ah and a rated voltage of 355V, and is composed of 96 individual cells connected in series.

[0234] Data is collected synchronously according to preset rules via the onboard BMS acquisition module. The collected operational data includes individual cell voltage, total current, individual battery cell temperature, equalization switch status, and equalization duration. For EIS feature acquisition, data is collected daily after the vehicle has been parked and turned off, when the battery meets the specified requirements. When the conditions are at rest, start the low-frequency micro-perturbation EIS measurement.

[0235] The raw data collected is preprocessed as follows: voltage and current sampling noise is removed by moving average filtering, abnormal sampling values ​​are eliminated by the 3σ criterion, missing data is filled by linear interpolation, and timestamps are aligned for multi-source data.

[0236] During feature extraction and normalization, the feature vector is run. During the construction process, This represents the battery's initial maximum voltage difference, for example, 5mV. Capacity increment feature vector. of The battery's rated capacity is 100Ah. In the dynamic weight matrix calculation, the initial values ​​of the adjustable parameters are set to... , .

[0237] Construct a diagonal weight matrix based on the weighting coefficients:

[0238]

[0239] in, , These are two vectors of length 4 and 3, both consisting of only 1s.

[0240] In the self-discharge severity rating mapping process, for ternary lithium battery systems, the scale parameters are... The initial value is set to 2.

[0241] When the AI ​​self-judgment module performs adaptive threshold correction under working conditions, when the battery temperature is below 0℃, the AI ​​automatically raises the scoring threshold by 10% to avoid false alarms caused by low temperature; when the battery SOH is below 80%, the slope threshold is adaptively adjusted to adapt to the self-discharge change pattern of aging batteries.

[0242] When a Level 4 emergency alarm occurs, corresponding to a severe self-discharge (acceleration) state, a red pop-up window will appear on the vehicle's instrument panel, accompanied by a continuous audible alarm. The alarm will display a score, trend, source tracing results, abnormal cell number, and emergency handling suggestions: "Stop the vehicle immediately, cease use of the vehicle, and contact after-sales service for emergency handling." At the same time, the BMS will automatically limit the battery discharge power and send emergency alarm information to the cloud platform and the owner's app, including vehicle location, battery status, and handling suggestions.

[0243] The AI ​​self-iterative module is deployed on a cloud-based battery management platform and is implemented based on an online incremental learning algorithm. The specific closed-loop mechanism is as follows:

[0244] A big data sample library for battery self-discharge was built. All data uploaded by the vehicle BMS was collected and stored in real time, including standardized feature vectors, weighted Mahalanobis distance, self-discharge score, AI judgment results, actual treatment measures, subsequent changes in battery state, and operating condition information. Data annotation was completed, and an annotated sample library that can be used for incremental learning was constructed.

[0245] Benchmark eigenvector Self-updating: Monthly statistics on healthy batteries of the same model and under the same operating conditions. The feature vectors of batteries with a value of ≤20 and stable state are used to generate new benchmark feature vectors using the K-means clustering algorithm to replace the original fixed benchmark values ​​and adapt to the natural aging pattern of the battery.

[0246] Self-adaptive scale parameter τ: The value of scale parameter τ is adaptively adjusted according to the battery aging state and actual test results to ensure that the score can accurately reflect the actual self-discharge state of the battery.

[0247] The optimized model and parameters are then updated in reverse to the vehicle BMS and cloud-based fusion evaluation layer and AI self-judgment layer, completing a self-iterative closed loop and enabling the system to continuously evolve.

[0248] Multiple fault-tolerance and security mechanisms are included:

[0249] AI model failure fallback: If the vehicle-mounted AI self-judgment module fails, the system automatically switches to the traditional fixed threshold judgment mode to ensure the normal operation of the basic warning function, while triggering the AI ​​module failure warning.

[0250] Parameter update rollback: If the AI's self-iterative optimization of parameters / models causes a drop in judgment accuracy of more than 10%, the system will automatically roll back to the previous version of valid parameters / models;

[0251] Access control is tiered: AI's self-iterative parameter updates and model optimizations are all executed automatically in the cloud. Humans only have the permissions to view, pause, and roll back, to prevent human error.

[0252] Data backup: The sample library, model parameters, and benchmark feature vectors are all backed up in real time in the cloud to prevent data loss and support rapid recovery after system failure.

[0253] This specific embodiment enables real-time online monitoring of self-discharge throughout the entire lifecycle of vehicle-mounted ternary lithium battery packs. The sensitivity for identifying early micro-self-discharge is more than 10 times that of the traditional static voltage method, and the detection accuracy under all operating conditions is ≥95%. It can effectively provide early warning of abnormal self-discharge risks and ensure battery safety and lifespan.

[0254] According to another aspect of this application, a battery self-discharge prediction system is provided for implementing the battery self-discharge prediction method as described above. Figure 8 A schematic diagram of the structure of a battery self-discharge prediction system according to an embodiment of this application is shown. Figure 8 As shown, the battery self-discharge prediction system according to an embodiment of this application includes an acquisition unit 100, a calculation unit 200, and a prediction unit 300.

[0255] Specifically, the acquisition unit 100 is used to acquire battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data.

[0256] The computing unit 200 is used to obtain a prediction model based on the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data.

[0257] The prediction unit 300 is used to predict the self-discharge parameters of the battery based on the prediction model.

[0258] The battery self-discharge prediction method and system provided in this application simultaneously collect battery operating data, incremental capacity characteristics, and electrochemical impedance spectroscopy (EIS) characteristics. After separating, extracting, and dimensionality-reducing the collected features, a dynamic weighted adaptive matrix based on battery SOC, temperature, and SOH is introduced. A weighted Mahalanobis distance is used to estimate the comprehensive self-discharge score, achieving fuzzy quantitative analysis of the self-discharge level. An AI self-judgment module is applied to identify abnormal data, automatically correct operating condition adaptation parameters, trace and classify the root causes of self-discharge, and an AI self-iteration module enables online self-improvement of the model, benchmark, and weight coefficients. This achieves a closed loop of data acquisition, feature processing, quantitative evaluation, state determination, intelligent early warning, and autonomous optimization, enabling accurate online real-time self-discharge prediction throughout the battery's entire lifespan. This system monitors and manages safety, demonstrating highly accurate identification of early-stage micro-self-discharge. It is adaptable to various scenarios and can be seamlessly integrated into vehicle BMS, energy storage EMS, and battery cloud systems. For comprehensive self-discharge scoring, it establishes a dual-judgment mechanism of graded early warning and trend analysis, avoiding false alarms and missed alarms caused by a single fixed threshold. The AI ​​self-judgment module can perform self-verification of abnormal data, adaptive correction of operating condition thresholds, and intelligent tracing of the root cause of self-discharge, achieving precise location and personalized decision-making for self-discharge anomalies. The accompanying AI self-iteration module, based on online incremental learning, enables continuous optimization of the model, baseline parameters, and weight coefficients without offline retraining, perfectly adapting to the aging patterns and operating condition changes throughout the battery's entire lifecycle, addressing the industry pain point of insufficient generalization ability in traditional models. (Self-discharge scoring is used as a basis for further analysis.) Using the trend slope k as the core input, combined with an AI-trained working condition correction model, the warning threshold can be automatically adjusted according to extreme working conditions such as low temperature and high aging, avoiding false alarms and missed alarms caused by fixed thresholds.

[0259] Finally, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The embodiments described above, as per the implementation of this application, do not exhaustively describe all details, nor do they limit the application to only the specific embodiments described. Clearly, many modifications and variations can be made based on the above description. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of this application, thereby enabling those skilled in the art to make good use of this application and modifications based on it. This application is limited only by the claims and their full scope and equivalents.

Claims

1. A battery self-discharge prediction method based on artificial intelligence technology, comprising: Acquire battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data; A prediction model is obtained based on the battery operation data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data; The self-discharge parameters of the battery are predicted based on the prediction model.

2. The battery self-discharge prediction method based on artificial intelligence technology according to claim 1, wherein, The battery operating data includes at least one selected from single cell voltage, total current, temperature, and equilibrium state; The incremental capacity data includes at least one selected from differential capacity curves, high-voltage inflection point capacity, and low-voltage inflection point capacity. The electrochemical impedance spectroscopy characteristic data includes at least one selected from the real part of impedance at the characteristic frequency, the imaginary part of impedance at the characteristic frequency, and the peak value of relaxation time distribution at the characteristic frequency. The prediction model derived from the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data includes: Based on the battery operation data, extract and construct the operation feature vector; Based on the incremental capacity data, extract and construct the incremental capacity feature vector; Based on the electrochemical impedance spectroscopy feature data, an electrochemical impedance spectroscopy feature vector is extracted and constructed. The operating feature vector, the incremental capacity feature vector, and the electrochemical impedance spectroscopy feature vector are normalized. The prediction model is obtained by calculating the normalized operating feature vector, incremental capacity feature vector, and electrochemical impedance spectroscopy feature vector.

3. The battery self-discharge prediction method based on artificial intelligence technology according to claim 1, wherein, The battery self-discharge prediction method further includes: Obtain historical battery statistics; The prediction model is obtained based on the battery operating data, the incremental capacity data, the electrochemical impedance spectroscopy characteristics, and the battery historical statistical data. The historical statistical data includes at least one selected from available balance counts, effective balance counts, and cumulative idle time.

4. The battery self-discharge prediction method based on artificial intelligence technology according to claim 1, wherein, The battery self-discharge prediction method further includes: When the battery is in a static state or operating at a constant current less than a set threshold, inject a multi-frequency sinusoidal current disturbance with an amplitude less than the threshold. The voltage response under the multi-frequency sinusoidal current perturbation was measured and the impedance spectrum was calculated. The real and imaginary parts of the impedance at different characteristic frequency points were extracted to generate the electrochemical impedance spectrum eigenvector. The electrochemical impedance spectroscopy eigenvectors are used to obtain the prediction model.

5. The battery self-discharge prediction method based on artificial intelligence technology according to claim 1, wherein, The battery self-discharge prediction method further includes: The prediction model is adjusted according to the dynamic weight matrix; The generation of the dynamic weight matrix includes: The operating characteristic weights are calculated based on the current state of charge of the battery. The incremental capacity feature weights are calculated based on the battery's current temperature and the current rate of the most recent complete cycle. The characteristic weights of the electrochemical impedance spectroscopy are calculated based on the current resting time and operating current of the battery. The dynamic weight matrix is ​​obtained by calculating the operating characteristic weight, the incremental capacity characteristic weight, and the electrochemical impedance spectroscopy characteristic weight.

6. The battery self-discharge prediction method based on artificial intelligence technology according to claim 1, wherein, The battery self-discharge prediction method further includes: Obtain the baseline feature vector corresponding to a healthy battery; The weighted Mahalanobis distance formula is used to calculate the multi-dimensional deviation between the feature vectors corresponding to the current battery operating data, incremental capacity data, and electrochemical impedance spectroscopy feature data and the benchmark feature vector; The weighted Mahalanobis distance is mapped to a self-discharge assessment value using an exponential mapping function.

7. The battery self-discharge prediction method based on artificial intelligence technology according to claim 6, wherein, The battery self-discharge prediction method further includes: Obtain the self-discharge assessment value of the nearest neighboring number of times; Calculate the average slope of change based on the nearest self-discharge assessment values ​​of the set number of times. Specifically, when the average slope of change is greater than a preset slope threshold, the battery is determined to be in a state of accelerated self-discharge.

8. The battery self-discharge prediction method based on artificial intelligence technology according to claim 7, wherein, The battery self-discharge prediction method further includes: Establish a modified model, which is used to achieve at least one of the following steps: Receive the feature vector, the weighted Mahalanobis distance, the self-discharge assessment value, and the average change slope; The system performs abnormal data verification on the input data, identifies sensor drift, data loss, and disturbances in the measurement of electrochemical impedance spectroscopy characteristic data, and completes data repair or invalidation marking. The self-discharge assessment value and the calculated average change slope are corrected based on the current state of charge, temperature, health status, and application scenario of the battery. For abnormal states, the cause of abnormal self-discharge is determined based on the fitting characteristics and operating characteristics of the electrochemical impedance spectroscopy data.

9. The battery self-discharge prediction method based on artificial intelligence technology according to claim 8, wherein, The modified model includes convolutional neural networks and long short-term memory networks; The modified model is also used to achieve at least one of the following steps: Construct a battery self-discharge data sample library, and collect and store the benchmark feature vector and the self-discharge evaluation value; When the number of new valid samples in the battery self-discharge data sample library reaches a preset number or the judgment accuracy is lower than a preset threshold, online incremental learning is triggered. The revised model is optimized based on the newly added valid samples; Clustering algorithm is used to update the baseline feature vector of healthy batteries of the same model and under the same working conditions; Using the prediction error of the self-discharge assessment value as the objective function, the parameters in the dynamic weight formula are adjusted by the gradient descent method. The scale parameter in the self-discharge evaluation value calculation formula is adjusted according to at least one of the battery model, material system, and application scenario.

10. A battery self-discharge prediction system based on artificial intelligence technology, comprising: The acquisition unit is used to acquire battery operating data, incremental capacity data, and electrochemical impedance spectroscopy characteristic data; The computing unit is used to obtain a prediction model based on the battery operating data, the incremental capacity data, and the electrochemical impedance spectroscopy characteristic data. A prediction unit is used to predict the self-discharge parameters of the battery based on the prediction model.

Images