A method for analyzing and rapidly evaluating self-discharge characteristics of a battery

Abstract

This invention relates to a method for analyzing and rapidly evaluating battery self-discharge characteristics, aiming to improve the accuracy and efficiency of battery self-discharge performance testing. Within a defined environmental stability range, this method acquires the instantaneous voltage change sequence and current perturbation response sequence of the target battery through micro-load perturbation, constructing a battery internal resistance dynamic evolution tensor characterizing the battery's internal dynamic response characteristics. A multi-scale time window segmentation strategy is applied to extract local features, forming a feature vector set containing a voltage nonlinear rebound factor and a delay response exponent. This feature vector is then input into a trained Bayesian weighted adaptive mapping network, outputting an estimated battery self-discharge rate that incorporates manufacturing batch residuals and temperature offset compensation. Trend fitting and consistency checks are performed on the estimated values ​​to obtain a battery self-discharge level assessment result, generating a structured test report including initial offset, rebound amplitude, and estimation confidence level. This method can be widely applied in battery production, quality control, and performance screening scenarios.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a method for analyzing and rapidly evaluating the self-discharge characteristics of batteries. Background Technology

[0002] In existing technologies, the assessment of battery self-discharge characteristics mainly relies on long-term static observation, which involves monitoring the voltage decay of the battery for several days to several weeks under constant temperature conditions to estimate its self-discharge rate. In addition, some methods indirectly infer the self-discharge trend through static impedance measurement or open-circuit voltage comparison. These methods are mostly used in production line sampling or laboratory verification, and have certain theoretical basis and engineering feasibility.

[0003] However, the aforementioned existing technologies generally suffer from problems such as long detection cycles, poor real-time performance, and strong dependence on environmental conditions, making them difficult to meet the needs of rapid screening and refined evaluation in modern battery mass production. Furthermore, some indirect inference methods lack the ability to model dynamic response behavior, resulting in limited accuracy and stability of evaluation results, particularly making it difficult to distinguish performance fluctuations caused by manufacturing process differences or temperature drift.

[0004] Therefore, there is an urgent need to propose a new method that can achieve high-precision analysis and evaluation of battery self-discharge characteristics in a short time. Summary of the Invention

[0005] This application provides a method for analyzing and rapidly evaluating the self-discharge characteristics of batteries, thereby improving the accuracy and efficiency of battery self-discharge performance testing.

[0006] This application provides a method for analyzing and rapidly evaluating the self-discharge characteristics of batteries, including:

[0007] Within a set environmental stability range, acquire the instantaneous voltage change sequence and current perturbation response sequence of multiple short time segments of the target battery under micro-load perturbation;

[0008] Based on the instantaneous voltage change sequence and current perturbation response sequence, a battery internal resistance dynamic evolution tensor is constructed to represent the dynamic response characteristics of the battery.

[0009] For the battery internal resistance dynamic evolution tensor, a multi-scale time window segmentation strategy is applied to extract local features, and a feature vector set containing voltage nonlinear rebound factor and delay response exponent is obtained.

[0010] The feature vector set is input into the trained Bayesian weighted adaptive mapping network to obtain the battery self-discharge rate estimate considering manufacturing batch residual modeling and temperature offset compensation mechanism.

[0011] The estimated self-discharge rate of the battery for at least three consecutive cycles is fitted with a trend curve, and the consistency of the curve is checked to obtain the battery self-discharge level assessment result.

[0012] Based on the battery self-discharge level assessment results and the corresponding battery self-discharge rate estimate, a structured test report is generated, which includes the initial offset, rebound amplitude, and estimated confidence level.

[0013] Furthermore, the micro-load disturbance refers to applying a short-duration pulse current signal with a current amplitude less than 1% of the rated capacity to the target battery during the test, and the duration of each disturbance is less than 2 seconds, in order to avoid significantly changing the battery's state of charge.

[0014] Furthermore, the voltage nonlinear rebound factor represents the degree of deviation of the voltage recovery process from the linear trend after the disturbance ends. It is obtained by fitting the voltage sequence within the disturbance recovery segment and calculating the average squared residual. The delay response index is used to reflect the voltage response time delay after the current disturbance is applied. Its value is the time interval between the current transition moment and the moment of maximum voltage change, and is normalized to the proportion of the disturbance time window length.

[0015] Furthermore, the Bayesian weighted adaptive mapping network includes a feature encoding unit, a joint modeling unit for manufacturing batches and temperature conditions, and an output fusion and confidence inference unit;

[0016] The feature encoding unit is used to receive a normalized feature vector group, which includes multiple voltage nonlinear rebound factors and delay response exponents. The feature vector group is processed by a multi-layer feedforward network to extract and compress features, and output a fixed-length encoded vector to represent the characteristics of disturbance response behavior.

[0017] The manufacturing batch and temperature condition joint modeling unit is used to receive the manufacturing batch and ambient temperature information of the target battery, and concatenate it with the fixed-length encoding vector to generate a joint input vector; it includes multiple sub-mapping paths, each of which is optimized based on training data of a specific manufacturing batch and temperature range, and Bayesian posterior weights are assigned during inference based on the similarity between the current input conditions and the central condition vector of each sub-path.

[0018] The output fusion and confidence inference unit is used to receive the candidate battery self-discharge rate estimates output by all sub-mapping paths, and to perform weighted fusion of each candidate value in combination with the Bayesian posterior weight to generate the final battery self-discharge rate estimate; and to calculate the confidence score based on the degree of dispersion between candidate values ​​to characterize the reliability of the final battery self-discharge rate estimate.

[0019] Furthermore, the construction of a battery internal resistance dynamic evolution tensor to represent the battery's internal dynamic response characteristics based on the instantaneous voltage change sequence and current perturbation response sequence includes:

[0020] Within the set environmental stability range, the instantaneous voltage change sequence and current perturbation response sequence of each group of short time segments are divided into a pre-perturbation resting segment, a perturbation excitation segment, and a perturbation end recovery segment, respectively. The segment division is based on the initial slope and amplitude threshold of the current change, and the boundary time is automatically determined.

[0021] In each perturbation excitation segment, the time relationship between the voltage drop start point, the voltage minimum point and the current transition point corresponding to the current perturbation is extracted and marked as the voltage response window index field, which serves as the time positioning reference for subsequent tensor construction.

[0022] In each disturbance end recovery segment, the recovery velocity vector between continuous voltage sampling points within the segment is calculated, and combined with the initial recovery offset value, maximum rebound velocity and rebound amplitude, a three-dimensional battery internal resistance dynamic evolution tensor with time step, ternary physical characteristics and segment number is constructed.

[0023] The maximum rebound velocity and the time difference between the delayed start-up time corresponding to each segment of the battery internal resistance dynamic evolution tensor are used as local dynamic feature fields to participate in the subsequent operation of local feature extraction using a multi-scale time window segmentation strategy.

[0024] The beneficial effects of the technical solution provided in this application include:

[0025] (1) By acquiring instantaneous voltage and current response sequences under micro-load disturbances and combining them with Bayesian weighted mapping networks for modeling and estimation, the battery self-discharge rate can be quickly obtained without long-term static placement, greatly improving detection efficiency and making it suitable for rapid screening in the production process. (2) By introducing the battery internal resistance dynamic evolution tensor, multi-scale time window segmentation strategy and feature vector group construction, key parameters reflecting the micro-changes inside the battery can be effectively extracted, enhancing the modeling ability for complex dynamic behaviors and improving the accuracy and stability of self-discharge evaluation results. (3) By considering the modeling of manufacturing batch residuals and temperature offset compensation mechanism, the system can effectively eliminate evaluation bias caused by batch differences and environmental factors, ensuring the universality and reliability of the algorithm under multiple scenarios and battery specifications. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for analyzing and rapidly evaluating the self-discharge characteristics of a battery, provided in the first embodiment of this application. Detailed Implementation

[0027] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0028] The first embodiment of this application provides a method for analyzing and rapidly evaluating the self-discharge characteristics of a battery. Please refer to... Figure 1 This figure is a schematic diagram of the first embodiment of this application. The following is in conjunction with... Figure 1 The first embodiment of this application provides a detailed description of a method for analyzing and rapidly evaluating the self-discharge characteristics of a battery.

[0029] Step S101: Within the set environmental stability range, acquire the instantaneous voltage change sequence and current perturbation response sequence of multiple short time segments of the target battery under micro-load disturbance.

[0030] In step S101, the system first needs to establish an environmental stability zone. This zone refers to a physical space where external conditions such as temperature, humidity, electromagnetic interference, and airflow are essentially constant and do not significantly disturb battery behavior. Preferably, this environment can be achieved through a constant temperature chamber or a controlled laboratory test chamber, with the temperature maintained at 25±1℃ and the relative humidity controlled between 40% and 60%. There should be no significant airflow vibrations or electromagnetic interference externally to ensure the stability and repeatability of data during battery testing.

[0031] In this environment, the target battery is connected to an electronic load module with milliampere-level resolution and programmable control capabilities. This electronic load should support constant current mode and be able to stably apply current perturbations within milliseconds. Preferably, the electronic load is controlled by a microcontroller or host computer (e.g., STM32 or LabVIEW platform), and is set to apply a small-amplitude current perturbation, such as 0.5 mA, 1.0 mA, or 2.0 mA, within a specified time window. To avoid activating the battery's macroscopic response or affecting its state of charge, the perturbation current should be much smaller than the operating current corresponding to the battery's rated capacity, and the perturbation duration is recommended to be between 0.5 seconds and 2 seconds. An idle period of at least 0.2 seconds should also be set before and after the perturbation to observe the natural voltage change and recovery characteristics after the perturbation, thus forming a complete perturbation cycle.

[0032] During the application of a current disturbance, a set of high-resolution, high-sampling-rate data acquisition modules (e.g., a 16-bit ADC module, a differential amplifier, and a high-speed sampling system) simultaneously measures the voltage change across the battery and the actual current flowing through it. This acquisition system should support a sampling rate of at least 1 kHz to ensure accurate capture of the transient process. The resulting data includes voltage and current signal sequences marked with uniform timestamps, referred to as the instantaneous voltage change sequence and the current perturbation response sequence, respectively. Both should be strictly aligned on the time axis and have an equal number of data points.

[0033] To improve the signal-to-noise ratio and ensure data quality, it is recommended to perform preliminary cleaning of the raw voltage and current data after acquisition using signal processing algorithms such as wavelet transform denoising, Savitzky-Golay filtering, or mean smoothing. During processing, key abrupt changes and gradual transitions in the signal should be preserved to avoid feature loss due to excessive smoothing.

[0034] In addition, to achieve high-reliability data acquisition, it is recommended that each disturbance test should collect no fewer than 5 independent short time segments (which can be completed across several minute intervals), and each segment should include a complete cycle of "rest-disturbance-recovery". The system needs to evaluate the acquisition quality of each cycle. If discontinuous voltage changes, sampling jitter, or abnormal disturbance response are detected, the data should be marked as invalid and automatically discarded to prevent erroneous data from entering the subsequent analysis process.

[0035] Finally, the output of step S101 is a set of short-time data segments with a unified format, time synchronization, and high signal-to-noise ratio. Specifically, these include instantaneous voltage change sequences and current perturbation response sequences, which serve as the basic input data for subsequent steps to construct the dynamic evolution tensor of the battery's internal resistance. This step achieves high-precision capture of the battery's microscopic electrochemical response behavior by constructing a non-destructive, low-perturbation, and standardized data acquisition mechanism, balancing testing efficiency and accuracy, and ensuring the repeatability and industrial applicability of the entire evaluation process.

[0036] Furthermore, the micro-load disturbance refers to applying a short-duration pulse current signal with a current amplitude less than 1% of the rated capacity to the target battery during the test, and the duration of each disturbance is less than 2 seconds, in order to avoid significantly changing the battery's state of charge.

[0037] In this invention, the "micro-load perturbation" is used to excite the short-time voltage response characteristics of the battery in order to extract the dynamic characteristic parameters of the battery's internal resistance without disrupting its state of charge. To ensure that this perturbation does not significantly change the state of charge of the target battery or trigger irreversible electrochemical changes, this embodiment limits the amplitude and duration of the perturbation process to simultaneously satisfy strict boundary conditions.

[0038] Specifically, in actual testing, the system control current source or electronic load device applies short-duration current pulses to the battery, and the current amplitude of each pulse should be less than 1% of the battery's rated discharge capacity. For example, for a 2000 mAh lithium-ion battery, the maximum allowable perturbation current is 20 mA. This current can be implemented through a constant current electronic load, a precision programmable DC source, or a digital-analog control circuit, ensuring control accuracy within ±0.1 mA.

[0039] Furthermore, to further reduce the impact of disturbances on the battery polarization state and electrochemical stability, the duration of each pulse application should not exceed 2 seconds. A preferred range is 0.5 to 1.5 seconds to ensure that the disturbance induces a measurable voltage change while avoiding the activation of internal charge migration or interface reactions. This timing control can be achieved by setting trigger and stop logic via a microcontroller or by automatic timing control executed by the test program, ensuring high repeatability of each disturbance.

[0040] The entire micro-load disturbance process must be performed within the environmental stability range, meaning that environmental conditions such as temperature, humidity, and electromagnetic interference must remain constant during the test to eliminate interference from external factors on the disturbance response data. A rest window of at least 200 milliseconds should be allowed before and after the disturbance to collect the baseline voltage before the disturbance and the voltage recovery behavior after the disturbance ends.

[0041] By employing the aforementioned dual constraints of amplitude and time, existing precision electronic loads or testing systems can be used to perform non-destructive, repeatable perturbation tests on any type of battery, thereby acquiring sufficient response information for subsequent analysis without affecting battery performance. This type of perturbation control mechanism is widely applicable to electrochemical energy storage devices of lithium-ion, nickel-metal hydride, and lead-acid types.

[0042] Step S102: Based on the instantaneous voltage change sequence and current perturbation response sequence, construct a battery internal resistance dynamic evolution tensor to represent the battery's internal dynamic response characteristics.

[0043] In step S102, the system needs to construct a tensor structure to characterize the internal electrochemical dynamic changes of the battery, based on the instantaneous voltage change sequence and current perturbation response sequence acquired in step S101. This structure is called the battery internal resistance dynamic evolution tensor. The construction process of this tensor not only includes the integration of the basic voltage and current correspondence, but also needs to be expanded through the time dimension, perturbation period dimension, and signal feature dimension, thereby forming a time-series tensor that can bear the multidimensional characteristics of perturbation response behavior.

[0044] Specifically, each voltage and current sequence contains multiple sampling points. The system first aligns the voltage and current signals point by point to ensure that the two signals have corresponding physical responses at the same time point. Based on this, the system divides a complete disturbance cycle into several time periods, such as a pre-disturbance rest period, a disturbance excitation period, and a disturbance end recovery period. Within each time period, key physical parameters such as local voltage slope, current response amplitude, and recovery speed are extracted. For each disturbance segment, the system can encode it into a three-dimensional array, where the first dimension is the time step, the second dimension is the signal type (voltage, current, slope, rebound rate, etc.), and the third dimension is the disturbance sequence number or test batch information. By concatenating multiple disturbance segments into a unified tensor structure, the system can establish a multi-dimensional data volume containing multiple samples, multiple physical variables, and their temporal evolution relationships.

[0045] In the actual construction of the dynamic evolution tensor of battery internal resistance, it is necessary to first segment the complete perturbation cycle in each perturbation test segment to reflect the dynamic response process of the battery to small perturbations. Specifically, each perturbation cycle can be divided into three periods based on the application time of the current perturbation: the pre-perturbation resting period, the perturbation excitation period, and the perturbation end recovery period. The pre-perturbation resting period refers to the stable state interval before the current perturbation is applied, which usually lasts for 200 to 500 milliseconds and is used to capture the baseline voltage level. The perturbation excitation period refers to the high response region after the micro-load current is formally applied, with a duration consistent with the perturbation duration, usually 500 milliseconds to 2 seconds. At this time, the voltage drops or changes significantly, which is an important segment for observing the battery's equivalent series internal resistance (ESR) behavior. The perturbation end recovery period is the stage where the voltage naturally rebounds after the current is disconnected, lasting for 500 milliseconds to 1 second, and is used to characterize polarization recovery and interface charge migration effects.

[0046] After segmentation, the system will extract features from the voltage and current data for each time period. The pre-disturbance static segment is used to calculate the average voltage as a reference voltage baseline; the disturbance excitation segment calculates the instantaneous voltage drop amplitude, instantaneous current amplitude, and voltage slope (ΔU / Δt) as indicators of internal resistance and polarization rate; the disturbance-end recovery segment analyzes the maximum rebound rate, rebound amplitude, and the fitting index of the voltage recovery curve during the voltage recovery process. These features reflect the hysteresis and recovery capability of the battery's internal electrochemical system. All extracted features will be embedded as signal dimensions into the channels used to construct the tensor structure, providing high-dimensional, time-defined microscopic behavioral data for subsequent models.

[0047] Maximum rebound rate refers to the maximum rate of voltage change per unit time during the voltage recovery process after the disturbance current terminates. Its calculation method is as follows: During the period from the end of the disturbance to the voltage stabilizing, let the sampling period be... For every two adjacent voltage sampling points and Calculate its first difference After iterating through all recovery segment sampling points, the maximum value is taken as the maximum rebound velocity, usually expressed in volts per second (V / s). This value characterizes the battery's instantaneous recovery capability during the initial stage of polarization release.

[0048] The rebound amplitude refers to the total voltage rise from the lowest point of the disturbance to the end of the recovery phase. Let the lowest voltage at the end of the disturbance be... The voltage at the end of the recovery is Then the rebound amplitude Defined as:

[0049] ;

[0050] The unit is volt (V). This value reflects the battery's total ability to recover from a polarized state to a steady state.

[0051] The fitting exponent of the voltage recovery curve is used to characterize the dynamic process of voltage recovering to a steady state over time. This fitting typically takes the form of an exponential recovery function:

[0052] ;

[0053] in To recover the voltage at any point in time within the segment, This represents the final stable voltage (which can be approximated as the voltage value at the end of the recovery segment).

[0054] To restore the initial offset, indicating the time when the disturbance ends. At that time, the voltage relative to the final stable voltage The initial deviation. That is to say, This is the voltage difference between the voltage when it begins to recover from its lowest point after the disturbance and the final recovered voltage. This value reflects the degree of polarization or transient voltage drop of the battery at the moment the disturbance ends. The larger the value, the greater the recovery required after the disturbance, and the more significant the internal impedance or polarization response. The unit is volts, consistent with the voltage signal itself.

[0055] The fitting exponent (also called the recovery rate constant) of the voltage recovery curve is expressed in units of 1000 kJ / m². By fitting the voltage recovery segment sampling points to the above functional form using the nonlinear least squares method, the fitting exponent can be obtained. The larger this value, the faster the voltage recovers and the smaller the hysteresis.

[0056] During tensor construction, normalization is preferably used to make all signal dimensions dimensionless to eliminate the offset effect caused by differences in sampling amplitude. For example, voltage values ​​can be normalized to the percentage change relative to the initial perturbation value, and current signals can be normalized to the relative response under the perturbation amplitude unit. Furthermore, to enhance the expressive power of the tensor in subsequent modeling tasks, the system can also introduce a set of derived feature dimensions, such as parameters like the location of local extrema, the time interval between peaks and valleys, and the mutation rate of the derivative at the reversal point, and embed them into the tensor as extended channels.

[0057] The constructed battery internal resistance dynamic evolution tensor not only records the instantaneous response of the battery to external stimuli under each perturbation, but also implicitly reflects the dynamic trajectory of complex mechanisms such as its internal polarization characteristics, charge migration rate, and interface impedance changes. This tensor will serve as the main input data structure for subsequent steps to further extract representative features and build a prediction model.

[0058] To improve construction efficiency and stability, the system should encapsulate the tensor generation process as an automated program module, and ensure that all segment processing uses a unified time alignment strategy, signal processing algorithm, and feature encoding format to facilitate consistent training and evaluation of subsequent models. The core of this step lies in using structured, multi-dimensional tensor representations to transform short-term battery perturbation response data from time series into an information carrier with spatial patterns and statistical characteristics, thereby providing a solid data foundation for accurately characterizing battery self-discharge behavior.

[0059] Furthermore, the construction of a battery internal resistance dynamic evolution tensor to represent the battery's internal dynamic response characteristics based on the instantaneous voltage change sequence and current perturbation response sequence includes:

[0060] Within the set environmental stability range, the instantaneous voltage change sequence and current perturbation response sequence of each group of short time segments are divided into a pre-perturbation resting segment, a perturbation excitation segment, and a perturbation end recovery segment, respectively. The segment division is based on the initial slope and amplitude threshold of the current change, and the boundary time is automatically determined.

[0061] In each perturbation excitation segment, the time relationship between the voltage drop start point, the voltage minimum point and the current transition point corresponding to the current perturbation is extracted and marked as the voltage response window index field, which serves as the time positioning reference for subsequent tensor construction.

[0062] In each disturbance end recovery segment, the recovery velocity vector between continuous voltage sampling points within the segment is calculated, and combined with the initial recovery offset value, maximum rebound velocity and rebound amplitude, a three-dimensional battery internal resistance dynamic evolution tensor with time step, ternary physical characteristics and segment number is constructed.

[0063] The maximum rebound velocity and the time difference between the delayed start-up time corresponding to each segment of the battery internal resistance dynamic evolution tensor are used as local dynamic feature fields to participate in the subsequent operation of local feature extraction using a multi-scale time window segmentation strategy.

[0064] In the battery self-discharge characteristic analysis and rapid evaluation method described in this invention, the process of constructing a battery internal resistance dynamic evolution tensor to represent the battery's internal dynamic response characteristics is the core step in extracting key perturbation behaviors from instantaneous voltage change sequences and current perturbation response sequences. To accurately model the detailed response mechanism of the battery under perturbation conditions, the system uses multiple short time segments as analysis units, sequentially performing operations such as segment identification, behavioral index extraction, structured organization, and feature field injection within a set environmental stability interval, ultimately forming a three-dimensional tensor representation with temporal structure, behavioral features, and segment recognition capabilities.

[0065] First, the system divides the instantaneous voltage change sequence and current perturbation response sequence in each short time segment into perturbation stages. This division process is completed automatically by the program, without relying on manual annotation, and is based on the initial slope and amplitude threshold of the current signal change. When the current waveform rapidly transitions from a steady state to the perturbation current level, the system detects that the rising or falling slope exceeds a set threshold, which is then determined as the start of the perturbation excitation segment; when the current waveform falls back and stabilizes near the baseline, and the change amplitude is below the set threshold, the system marks this position as the end of the perturbation. After the division is completed, each segment is divided into a pre-perturbation resting segment, a perturbation excitation segment, and a perturbation end recovery segment, corresponding to the static steady state before the perturbation, the response stage during current loading, and the gradual voltage recovery process after the perturbation is removed, respectively.

[0066] Within each defined disturbance excitation segment, the system further locates the starting point of voltage drop, the lowest voltage point, and the peak point of current change rate, measures the relative time interval between them, and records this time relationship as the voltage response window index field. This field can be regarded as a characteristic time marker of the disturbance response, which is subsequently used to assist in the relative positioning and window alignment control of the time step sequence in the tensor, so as to ensure the comparability of the voltage dynamic behavior within the tensor during the tensor construction process.

[0067] Subsequently, the system focuses on the recovery segment after each disturbance, that is, the time-domain interval in which the voltage rises back to a stable level after the disturbance is removed. Within this segment, the system extracts the difference between all consecutive voltage sampling points and divides it by the sampling interval to construct a recovery velocity vector, which reflects the trajectory of the voltage recovery rate within this segment. Simultaneously, the system further calculates three key physical characteristics: first, the initial recovery offset value, i.e., the magnitude of the voltage drop relative to the pre-disturbance steady-state voltage at the moment the disturbance ends; second, the maximum rebound velocity, i.e., the maximum rate of rise in the recovery velocity vector; and third, the rebound amplitude, i.e., the amplitude of the voltage recovery from the lowest point to the final stable value. These three physical indicators, together with the recovery velocity vector, constitute a ternary physical characteristic dimension.

[0068] After completing the above steps, the system constructs tensors in a unified format. The first dimension of the tensor is the continuous time step, reflecting the high-frequency response trajectory within the perturbation segment; the second dimension is the ternary physical feature, including the recovered velocity vector, the recovered initial offset value, the rebound amplitude, and the maximum rebound velocity; the third dimension is the segment number dimension, used to identify which group of perturbation segments it belongs to, ensuring that different segment data can be processed independently or batch-classified in subsequent analysis. The tensor adopts a standard structured memory layout, has backward-propagating index information, and can support layer-by-layer input to neural network models.

[0069] Furthermore, to ensure that the subsequent multi-scale time window segmentation strategy can fully utilize the key dynamic patterns of the perturbation recovery features, the system extracts the maximum rebound velocity and the delayed start-up time difference corresponding to each segment of the aforementioned tensor, and appends them as independent fields to the local dynamic feature set in the tensor. The delayed start-up time difference refers to the time delay between the end of the current perturbation and the start of a significant voltage recovery, reflecting the time lag in the response of the battery polarization or diffusion mechanism. These local dynamic feature fields not only enhance the tensor's ability to represent the dynamics of the perturbation response but also provide a structured foundation for subsequent time window selection, feature extraction, and classification aggregation operations.

[0070] Through the above processing, the present invention can accurately model the multi-segment response behavior of the target battery under micro-perturbation conditions without long-term placement or large external disturbances. It constructs a three-dimensional battery internal resistance dynamic evolution tensor with clear physical meaning, stable time alignment, and diverse feature granularity, which serves as the core input structure for subsequent analysis processes, ensuring that the self-discharge characteristic evaluation has responsiveness, stability, and engineering feasibility.

[0071] Step S103: Apply a multi-scale time window segmentation strategy to extract local features from the battery internal resistance dynamic evolution tensor to obtain a feature vector set containing the voltage nonlinear rebound factor and the delay response exponent.

[0072] In step S103, the system takes the battery internal resistance dynamic evolution tensor as input, performs structured processing on the disturbance response data contained therein, and extracts the voltage nonlinear rebound behavior and voltage response delay characteristics. To accurately capture the local response characteristics of the battery at different time scales, the system employs a multi-scale time window segmentation strategy. "Multi-scale" refers to setting multiple time window lengths for the same disturbance data segment, such as 50 milliseconds, 100 milliseconds, 200 milliseconds, and 500 milliseconds, each length representing a different granular temporal resolution. At each scale, the system sequentially slides the time windows to extract segments from the data, typically using a partial overlap method to ensure that the overall disturbance process is completely covered.

[0073] For each extracted time window segment, the system performs two types of feature analysis. The first type is the extraction of the voltage nonlinear rebound factor, which aims to identify whether the voltage exhibits a regular, uniform change within that time period, or whether there are sudden rebounds, jitters, or nonlinear distortions. This is achieved by constructing a linear fitting curve within the current window to simulate the voltage's trend over time. Then, the actual sampled voltage values ​​are compared point-by-point with the expected values ​​of the fitting curve, and the sum of the squares of all deviations is calculated and divided by the total number of sampling points in the window, resulting in a value representing the "average degree of deviation." If this value is close to zero, it indicates that the voltage change within that segment is nearly linear; if the value is large, it indicates that the voltage response contains significant nonlinear components, such as rebound peaks, hysteresis breakpoints, or transient fluctuations. This value is defined in this system as the "voltage nonlinear rebound factor," measured in square volts, and is used to quantify the intensity of local nonlinearity.

[0074] The second category is the voltage response delay index to current disturbances, which aims to assess whether a battery can quickly generate a voltage response after a minor disturbance is applied. The system first marks the start of the disturbance in the current sequence, typically the point in time when the current abruptly changes from zero to a set disturbance value. Then, in the subsequent voltage data, it searches for the point in time where the voltage change rate is fastest, usually the initial inflection point where the voltage drops sharply or rises sharply. This point can be found by analyzing the rate of voltage change, i.e., dividing the voltage difference between consecutive sampling points by the time interval, to find the point where the maximum value is located. Next, the system calculates the time interval between this point and the start of the current disturbance, obtaining a response delay value in seconds. To eliminate the influence of different time window scales on the absolute magnitude of this value, the system divides this delay value by the total duration of the current time window, obtaining a proportional value, which is defined as the "delay response index". For example, in a 200-millisecond window, if the voltage response begins 40 milliseconds later than the current disturbance, the index is 0.2.

[0075] The two aforementioned features, the voltage nonlinear rebound factor and the delay response exponent, are combined into a single local feature vector for a time window. The system repeats this process across multiple scales and sliding time windows, yielding multiple local feature vectors. All feature vectors are combined in temporal and scale order to form a complete set of feature vectors, representing the time-domain and scale-domain behavioral distribution of the battery's disturbance response. These vectors are directly input into the subsequent self-discharge rate estimation model to identify the self-discharge characteristics of different batteries.

[0076] The entire feature extraction process is automatically executed by a pre-set analysis program. The data interface is unified, the sliding window strategy is standardized, and it has high repeatability and good batch processing capabilities.

[0077] Furthermore, the voltage nonlinear rebound factor represents the degree of deviation of the voltage recovery process from the linear trend after the disturbance ends. It is obtained by fitting the voltage sequence within the disturbance recovery segment and calculating the average squared residual. The delay response index is used to reflect the voltage response time delay after the current disturbance is applied. Its value is the time interval between the current transition moment and the moment of maximum voltage change, and is normalized to the proportion of the disturbance time window length.

[0078] In the battery self-discharge characteristic analysis and rapid evaluation method described in this invention, in order to extract key characteristic parameters that characterize the internal dynamic behavior of the battery from the disturbance response process, the system introduces two core indicators within a multi-scale time window: the voltage nonlinear rebound factor and the delay response index. These two indicators are used to capture the nonlinear behavior during the voltage recovery process after the disturbance ends and the time lag of the voltage response to the current disturbance, respectively, which can effectively enhance the discrimination accuracy of subsequent self-discharge rate estimation.

[0079] In the extraction of the voltage nonlinear rebound factor, the system first locates the end time of the current disturbance in each time window, i.e., the instant the micro-load application terminates. The system then extracts a recovery period voltage sequence after this moment, typically ranging from several hundred milliseconds to one second in length. Within this time period, the system uses a first-order linear trend as a reference baseline, fitting the voltage signal of the entire recovery segment to an optimal straight line. The fitting process can be achieved using the least squares method, ensuring that the linear trend line closely approximates the actual voltage change trend. After fitting, the system squares the deviation between the actual value and the fitted value at each voltage sampling point and calculates its average or total value. The average of this deviation is the voltage nonlinear rebound factor. The unit of this factor is the square of volts; a larger value indicates a more significant nonlinear fluctuation during the voltage recovery process, which may be related to interface capacitance delay, electrolyte migration hindrance, or local polarization reconstruction.

[0080] The extraction of the delay response index focuses on measuring the time lag behavior of the voltage response. Within each disturbance time window, the system first determines the moment when the micro-load begins to be applied based on the current signal, typically the time point when the current transitions from a near-zero steady state to the specified micro-disturbance current value. The system then analyzes the rate of change of the corresponding voltage sequence, constructing a voltage change rate sequence by dividing the voltage difference between consecutive time points by the time interval. The system then searches for the time point corresponding to the maximum rate of change in this sequence, i.e., the moment when the voltage changes most rapidly, which is usually also the response point where the voltage begins to drop or rise rapidly. The system uses the time interval between this moment and the start of the current disturbance as the delay value of the voltage response. To maintain comparability across different time window scales, this delay value is also divided by the total duration of the current time window, transforming it into a dimensionless relative value, the delay response index. This index typically ranges between 0 and 1; a larger value indicates a slower battery response to the disturbance, potentially characterizing physical phenomena such as increased diffusion impedance or sluggish interface layer charge response.

[0081] The extraction process of these two features can be automated on a digital signal processing platform without relying on manual judgment. It is suitable for standardized batch evaluation processes and can be integrated into battery testing systems, controllers, or embedded analysis platforms through programming.

[0082] Furthermore, the dynamic evolution tensor of the battery internal resistance is subjected to a multi-scale time window segmentation strategy for local feature extraction to obtain a feature vector set containing the voltage nonlinear rebound factor and the delay response exponent, including:

[0083] Based on the difference between the maximum rebound speed of the recovery segment and the delayed start-up time for each segment in the dynamic evolution tensor of the battery internal resistance, a local perturbation behavior index table is constructed to mark the relative time positions of the most drastic voltage change segment and the response start-up hysteresis segment in the tensor.

[0084] Guided by the local disturbance behavior index table, multiple time windows of different lengths are applied to slide and divide the area centered on the peak rebound speed segment. The time window lengths include at least 50 milliseconds, 100 milliseconds, 200 milliseconds, and 500 milliseconds, and each time window coverage area must include at least one marked position.

[0085] In the tensor recovery segment covered by each time window, a first-order linear trend fitting is performed, and the sum of squared residuals between the actual voltage sampled values ​​and the fitted trend line within the time window is normalized as the voltage nonlinear rebound factor corresponding to the time window.

[0086] Within each time window, the time difference between the transition moment of the disturbance current and the time point corresponding to the maximum voltage change rate in the tensor is calculated, and the time difference is normalized to serve as the delay response index corresponding to the time window.

[0087] The voltage nonlinear rebound factor and delay response exponent calculated in all time windows are combined into feature pairs, and all feature pairs are organized into feature vector groups according to the perturbation segment index and time window order, which are used as perturbation behavior input features for the subsequently trained Bayesian weighted adaptive mapping network.

[0088] In the battery self-discharge characteristic analysis and rapid evaluation method described in this invention, in order to effectively extract representative perturbation behavior features from the battery internal resistance dynamic evolution tensor to support subsequent accurate modeling and classification of battery self-discharge rate, the system further performs multi-scale time window segmentation and local physical feature extraction operations based on the tensor construction, and finally generates a set of feature vectors with engineering interpretability and model usability. This process not only combines the dynamic structure of multi-segment perturbation response within the tensor, but also introduces calibration reference indicators closely related to perturbation intensity and response hysteresis to ensure the focus and discriminativeness of local feature extraction.

[0089] First, in the constructed dynamic evolution tensor of battery internal resistance, the system extracts two important behavioral features from the recovery segment of each perturbation fragment: the maximum rebound velocity, representing the steepest voltage recovery rate after the perturbation is removed; and the delayed start-up time difference, representing the time delay from the end of the perturbation current to the point where the voltage begins to recover significantly. These two physical quantities reflect the speed of battery polarization release and diffusion reactions, possessing a highly individualized behavioral description capability. Based on these two features, the system constructs a local perturbation behavior index table to record the location of the most drastic voltage change segment and the start-up hysteresis region within the tensor in each fragment. This index table provides a structured basis for localization, which is subsequently used to guide the selection of the starting point and scale adaptation of the sliding time window.

[0090] Based on the aforementioned perturbation behavior index table, the system performs a multi-scale sliding time window segmentation operation within each perturbation segment, centered on the time point corresponding to the maximum rebound velocity. The selected time window lengths cover multiple typical scales, including 50 milliseconds, 100 milliseconds, 200 milliseconds, and 500 milliseconds, covering the two common behavioral characteristics of battery dynamic response: fast recovery and slow diffusion. To ensure that the segmented windows can capture the most critical physical behavioral characteristics, the tensor region covered by each time window must contain at least one position marked in the index table, i.e., ensuring that the center of the time window is aligned with the behavioral inflection point.

[0091] Within each segmented time window, the system first extracts the complete sampling sequence of the voltage recovery segment and performs a first-order linear trend fitting on this sequence. The fitting method can employ least-squares linear fitting, constructing an equally spaced slope baseline on the time axis to simulate an ideal linear rebound path. Subsequently, the system squares the point-to-point residuals between the actual voltage sampling values ​​within the time window and the fitted trend line, then normalizes the sum of squared residuals by dividing by the number of sampling points and the voltage amplitude range, yielding a dimensionless index, namely the voltage nonlinear rebound factor corresponding to that time window. This factor reflects the degree of deviation between the voltage recovery process and the ideal linear model; a larger value indicates a more irregular or severe recovery process, commonly seen in high-impedance degraded cells or severely polarized cells.

[0092] Simultaneously, the system extracts another key indicator within each time window: the delay response index. This index is calculated based on the time difference between the moment of the disturbance current transition and the time point corresponding to the maximum voltage change rate in the tensor. Specifically, the system first identifies the initial moment of the perturbation application through the current signal and locates this moment in the tensor; then, it analyzes the first-order difference of the voltage sequence and finds the sampling point corresponding to the maximum rate of change. The time interval between the two is the delay time. This time difference is then divided by the current time window length to convert it into a proportional form, forming the delay response index. This index characterizes the degree of lag in the voltage response relative to the current disturbance and is an important measure of dynamic response speed.

[0093] Finally, the system pairs the voltage nonlinear rebound factor and the delay response exponent calculated in each time window to form a feature pair. After processing all time windows of all perturbation segments in the entire tensor, the system organizes all feature pairs according to segment index and time window order to form the final feature vector group. This feature vector group not only preserves the temporal local structure and physical change logic of the perturbation behavior, but also provides a unified numerical encoding of the response morphology at different scales, exhibiting good model input adaptability.

[0094] The aforementioned feature vector set will serve as the perturbation behavior input features for the subsequently trained Bayesian weighted adaptive mapping network, enabling the system to generate high-precision and high-confidence estimates of battery self-discharge rates, taking into account manufacturing batch residual modeling and temperature offset compensation mechanisms, thus providing a reliable basis for structured evaluation and classification.

[0095] Step S104: Input the feature vector group into the trained Bayesian weighted adaptive mapping network to obtain the battery self-discharge rate estimate considering manufacturing batch residual modeling and temperature offset compensation mechanism.

[0096] Furthermore, the Bayesian weighted adaptive mapping network includes a feature encoding unit, a joint modeling unit for manufacturing batches and temperature conditions, and an output fusion and confidence inference unit;

[0097] The feature encoding unit is used to receive a normalized feature vector group, which includes multiple voltage nonlinear rebound factors and delay response exponents. The feature vector group is processed by a multi-layer feedforward network to extract and compress features, and output a fixed-length encoded vector to represent the characteristics of disturbance response behavior.

[0098] The manufacturing batch and temperature condition joint modeling unit is used to receive the manufacturing batch and ambient temperature information of the target battery, and concatenate it with the fixed-length encoding vector to generate a joint input vector; it includes multiple sub-mapping paths, each of which is optimized based on training data of a specific manufacturing batch and temperature range, and Bayesian posterior weights are assigned during inference based on the similarity between the current input conditions and the central condition vector of each sub-path.

[0099] The output fusion and confidence inference unit is used to receive the candidate battery self-discharge rate estimates output by all sub-mapping paths, and to perform weighted fusion of each candidate value in combination with the Bayesian posterior weight to generate the final battery self-discharge rate estimate; and to calculate the confidence score based on the degree of dispersion between candidate values ​​to characterize the reliability of the final estimate.

[0100] In step S104, the feature vector set obtained in step S103 is input into the trained Bayesian weighted adaptive mapping network to obtain an estimate of the battery self-discharge rate considering manufacturing batch residual modeling and temperature offset compensation mechanisms. The Bayesian weighted adaptive mapping network includes a feature encoding unit, a joint modeling unit for manufacturing batch and temperature conditions, and an output fusion and confidence inference unit, which work together to form a complete mapping prediction path.

[0101] First, the system receives a normalized feature vector set, which consists of voltage nonlinear rebound factors and delay response exponents extracted at multiple time window scales. This feature vector set is input to a feature encoding unit, which performs feature structure mapping and compression on the original input to generate a more discriminative feature representation vector, serving as the basis for subsequent modeling steps. The feature encoding unit can be implemented using a multilayer feedforward network, a fully connected structure, or other methods capable of performing linear / nonlinear mappings. Its output is a fixed-length feature encoding vector used to characterize the perturbation response behavior of the target battery.

[0102] Subsequently, the system reads the manufacturing batch information and test environment temperature information of the current target battery, and encodes them into a manufacturing batch and temperature condition vector. The manufacturing batch can be represented using one-thermal encoding, and the temperature information can be normalized. The system concatenates the manufacturing batch and temperature condition vector with the aforementioned feature encoding vector to form the joint input of the model.

[0103] In the joint modeling unit for manufacturing batches and temperature conditions, the system aims to improve the overall stability and adaptability of the prediction results by modeling and compensating for the interference of manufacturing origin differences and ambient temperature on the prediction results of battery self-discharge rate. The key technical structure of this unit consists of a set of several sub-mapping paths, each corresponding to a sample distribution characteristic represented by a combination of manufacturing batches and temperature ranges. All sub-mapping paths have been overfitted and optimized using sample data of the corresponding category during the training phase, thus possessing targeted modeling capabilities.

[0104] In the actual inference process, the system first encodes the manufacturing batch and temperature condition vector of the current target battery. The manufacturing batch information comes from production labels or a test database, representing the batch number to which the battery belongs. The system uses one-hot encoding to convert this into a vector representation of equal length; for example, "Batch 1" is represented as [1,0,0,0], "Batch 2" as [0,1,0,0], and so on. Temperature information is obtained through a measurement system, in degrees Celsius. To ensure model compatibility, the system normalizes the current temperature using the historical temperature range recorded in the training dataset, converting it into a real value between [0,1]. This process ensures that inputs under different temperature conditions do not cause model imbalance due to numerical differences.

[0105] The system concatenates the two variables to form a unified manufacturing batch and temperature condition vector. Then, based on this vector, the system calculates the similarity between each path and the input in the already trained set of sub-mapping paths, and dynamically assigns participation weights to each path. Each sub-mapping path records the distribution of manufacturing batches and temperature statistics it covers during training, such as mean and variance. The current input condition vector is matched against the historical distribution of each sub-path, for example, by calculating a matching score using Euclidean distance, cosine similarity, Mahalanobis distance, or a Gaussian kernel function. These matching scores, after normalization, are used as Bayesian posterior weight estimates for the current input on each path.

[0106] For example, if the current battery comes from "batch 3" with an ambient temperature of 33°C, and the training data for sub-path A mainly comes from samples of "batch 3" at 30-35°C, whose distribution characteristics highly overlap with the current input conditions, the system will assign a higher weight to sub-path A. Conversely, if sub-path B mainly comes from "batch 1" and its training temperature is around 20°C, the system will assign it a lower weight, even close to zero. This weight allocation mechanism essentially models the manufacturing batch residuals of the input samples, allowing predictions to be based on data sub-distributions with similar manufacturing backgrounds, while using temperature to compensate for temperature shifts, ensuring that the prediction results are not misjudged due to cross-environment data.

[0107] After weight allocation, the system inputs the current feature vector set into each of the sub-mapping paths. Each sub-path typically consists of a small set of feedforward networks or regressors, optimized during training using samples from its batch and temperature conditions. While maintaining a consistent structure, the weight parameters differ. Each sub-path outputs a candidate estimate of the battery self-discharge rate, in units of volts per hour, millivolts per day, or other custom units, configured by the system.

[0108] All candidate battery self-discharge rate estimates are sequentially input into the output fusion and confidence inference unit. In this unit, the system first obtains the Bayesian weight corresponding to each candidate value. This weight is derived from the weight allocation results of each sub-mapping path in the joint modeling unit of manufacturing batch and temperature conditions, and has been normalized to a probability form with a sum of 1.

[0109] The system pairs each candidate battery self-discharge rate estimate with its corresponding Bayesian weight, then performs a weighted fusion operation, multiplying each candidate value by its corresponding Bayesian weight and summing all the results. The value obtained from this operation is the final battery self-discharge rate estimate for the current target battery. The physical meaning of this value is: after comprehensively considering factors such as disturbance response behavior characteristics, manufacturing batch origin, and test environment temperature conditions, the system infers the natural voltage decay rate per unit time. The unit can be "millivolts per day," "volts per hour," etc., as set by the system calibration.

[0110] In addition to outputting the final result, the output fusion and confidence inference unit also needs to quantify the credibility of the current prediction. To this end, the system treats the set of battery self-discharge rate estimates output by all sub-mapping paths participating in this round of fusion as a set of numerical samples. The system calculates the dispersion of this set, which can be achieved through conventional statistical indicators, such as standard deviation, maximum difference, or distribution variance. If the estimates of all sub-paths are nearly consistent, meaning the dispersion between candidate values ​​is small, the system considers the current input to be within the range supported by the training data, indicating that the model prediction has high consistency and stability, and thus assigns a high confidence score. Conversely, if the prediction results of each sub-path differ significantly, it indicates that the current sample may have out-of-distribution characteristics, be from uncovered batches, or be subject to extreme temperature conditions, and the system will lower the confidence score.

[0111] The confidence score ranges from 0 to 1, and the system can set confidence level thresholds according to actual business needs for risk warning or manual review. Ultimately, the estimated battery self-discharge rate and its corresponding confidence score will be combined as structured result data and provided to the subsequent structured inspection report generation module for report output and evaluation visualization. This mechanism ensures that the system not only has numerical prediction capabilities during large-scale automatic evaluation but also provides criteria for prediction reliability, helping engineers make rapid decisions in quality control or anomaly screening scenarios.

[0112] With the above structure, the system can output a battery self-discharge rate estimate with manufacturing batch residual modeling and temperature offset compensation mechanism based on the feature vector set of disturbance response behavior and combined with the known manufacturing batch and temperature condition vectors. This method is suitable for rapid evaluation of batteries from different sources under various test environments and can effectively improve the accuracy, stability and industrial applicability of the prediction.

[0113] The following is a reference implementation of the Bayesian weighted adaptive mapping network:

[0114] import numpy as np

[0115] # Input data structure definition

[0116] # Input feature vector set: for example, 20-dimensional (10 voltage nonlinear rebound factors + 10 delay response exponents)

[0117] feature_vector = np.random.rand(20) # Simulate a set of perturbation feature vectors (normalized)

[0118] # Current battery manufacturing batch (assuming a total of 4 batches, using unique thermal coding)

[0119] batch_id = 2

[0120] batch_one_hot = np.eye(4)[batch_id] # [0, 0, 1, 0]

[0121] # Current ambient temperature of the battery (degrees Celsius, normalized to [0, 1], assuming a temperature range of [20, 40] degrees Celsius)

[0122] temperature_celsius = 33.0

[0123] temperature_norm = (temperature_celsius - 20.0) / (40.0 - 20.0) # Normalize to [0,1]

[0124] # Splicing manufacturing batch and temperature condition vector

[0125] condition_vector = np.concatenate([batch_one_hot, [temperature_norm]])

[0126] # Feature coding unit

[0127] def encode_features(input_vector):

[0128] # Multi-layer feedforward structure (which can be considered part of a shallow neural network) simulates the extraction of behavioral feature embeddings

[0129] # Here, linear mapping with ReLU activation is used.

[0130] W1 = np.random.randn(32, len(input_vector)) 0.1

[0131] b1 = np.zeros(32)

[0132] hidden = np.dot(W1, input_vector) + b1

[0133] encoded = np.maximum(hidden, 0) # ReLU activation

[0134] return encoded # Output 32-dimensional encoded vector

[0135] encoded_vector = encode_features(feature_vector)

[0136] # Concatenate the feature encoding vector and the conditional vector as joint input

[0137] joint_input = np.concatenate([encoded_vector, condition_vector])

[0138] # Sub-mapping path definition (assuming there are N sub-paths in the system)

[0139] # Each sub-path has been optimized for a specific batch-temperature range, and its central condition vector has been recorded.

[0140] N_PATHS = 5

[0141] subpaths = []

[0142] for i in range(N_PATHS):

[0143] path = {

[0144] "center_condition": np.random.rand(len(condition_vector)), # Batch temperature distribution center of the subpath

[0145] "model_weights": np.random.randn(1, len(encoded_vector)), # Simulate sub-model weights

[0146] "model_bias": np.random.randn(1)[0] # Bias

[0147] }

[0148] subpaths.append(path)

[0149] # Calculate similarity & Bayesian posterior weight assignment (using cosine similarity + softmax normalization)

[0150] def compute_cosine_similarity(a, b):

[0151] return np.dot(a, b) / (np.linalg.norm(a) np.linalg.norm(b) +1e-8)

[0152] similarities = []

[0153] for path in subpaths:

[0154] similarity = compute_cosine_similarity(condition_vector, path["center_condition"])

[0155] similarities.append(similarity)

[0156] # Softmax normalization yields Bayesian posterior weights (ensuring the sum is 1).

[0157] exp_sim = np.exp(similarities - np.max(similarities))

[0158] bayesian_weights = exp_sim / np.sum(exp_sim)

[0159] # Output candidate self-discharge rate estimates for each sub-path

[0160] candidate_outputs = []

[0161] for i, path in enumerate(subpaths):

[0162] # Use fixed weights in the sub-path for linear regression output (or use a pre-trained regressor).

[0163] weights = path["model_weights"]

[0164] bias = path["model_bias"]

[0165] candidate_value = float(np.dot(weights, encoded_vector) + bias)

[0166] candidate_outputs.append(candidate_value)

[0167] # Output fusion and confidence inference units, weighted fusion to obtain the final estimate.

[0168] final_estimate = sum(w v for w, v in zip(bayesian_weights,candidate_outputs))

[0169] # Calculate confidence level (using inverse standard deviation logic)

[0170] std_dev = np.std(candidate_outputs)

[0171] confidence_score = np.exp(-std_dev) # The smaller the dispersion, the higher the confidence level (between 0 and 1).

[0172] The training process of this Bayesian weighted adaptive mapping network consists of three stages. First, the system divides the samples into several subsets based on the battery's perturbation response feature vector set, manufacturing batch label, and test temperature from historical data. Each subset corresponds to a combination of manufacturing batch and temperature range, and a sub-mapping path is established for each subset. Second, the system trains each sub-mapping path independently, enabling it to predict the battery self-discharge rate under corresponding conditions from the input feature vector. The training process uses a standard regression loss function, such as mean squared error, and cross-validation is used to prevent overfitting. Finally, the system calculates the mean of the condition vector of each sub-path in the training samples, which serves as the condition center for the path. This is used for similarity matching and Bayesian weight allocation in the inference stage, thereby achieving distribution awareness and condition adaptation capabilities. This training mechanism retains the specificity of local modeling while supporting the dynamic fusion of input conditions, ensuring stable predictive performance of the model under multiple batches and environmental data.

[0173] Step S105: Fit a trend curve to the estimated battery self-discharge rate for at least three consecutive cycles, and verify its consistency to obtain the battery self-discharge level evaluation result.

[0174] In step S105, the system uses the battery self-discharge rate estimate obtained in step S104 as a basis to perform time-series trend analysis and consistency judgment on multiple estimation results collected from the same target battery at different time periods, thereby obtaining the battery self-discharge level evaluation result for level classification. To ensure the reliability of the evaluation result, this step requires that it be based on self-discharge rate estimates from at least three independent time segments. These time segments can be obtained under different disturbance periods or different sampling windows, provided that they correspond to the same physical battery and the test environment remains consistent.

[0175] In the specific implementation process, the system first timestamps all estimated self-discharge rates of the same target battery, records the test order in which each value was generated, and organizes them into an estimated sequence in chronological order. Each value, at the time of generation, already possesses complete input features, model structure, manufacturing batch, and temperature conditions. Therefore, the system can confirm the internal consistency of the sequence and determine whether it can be used for subsequent trend fitting processing. If the number of samples in the sequence is less than three, the system will terminate this step and return a "data insufficient" flag to avoid errors in grade judgment.

[0176] After the estimated sequence meets the minimum sample size requirement, the system uses this set of estimates as input to perform a trend fitting process. Trend fitting refers to the system performing numerical analysis on multiple self-discharge rate estimates to identify whether their trend is stable, such as whether it continuously rises or falls within a short period, or fluctuates drastically. The system can use methods such as moving averages or piecewise linear fitting to model the overall trend of the estimated sequence. The system comprehensively analyzes the trend by judging the following three indicators:

[0177] First, the system analyzes the maximum difference between all estimates, that is, it finds the absolute difference between the highest and lowest estimates. If this difference is less than the absolute tolerance set by the system (e.g., 1 millivolt per day), it means that the range of variation of each value is acceptable.

[0178] Second, the system determines whether the changes in valuation values ​​are consistent, that is, whether the changes between adjacent valuation values ​​are increasing, decreasing, or fluctuating. For example, if the value shows a pattern of first rising and then falling, it can be considered as having high volatility; if the change always shows a decreasing trend, it indicates that the self-discharge rate tends to weaken, which may be a recovery battery.

[0179] Third, the system evaluates the relative standard deviation among multiple estimates, which is the ratio of the magnitude by which multiple estimates deviate from the mean to the mean itself, to quantify volatility. The system sets a threshold for this indicator, for example, allowing the standard deviation to not exceed 15% of the mean.

[0180] Once all three indicators meet the system's preset stability conditions, the system determines that the self-discharge rate estimate of the target battery has good consistency and possesses a statistical basis for grade assessment. Subsequently, the system classifies the battery based on the average estimate. The grade classification criteria can be preset by the user; for example, a daily self-discharge rate of less than 1 millivolt can be classified as Grade 1 (Excellent), 1-2 millivolts as Grade 2 (Qualified), and more than 2 millivolts as Grade 3 (Risk). This grade classification logic should be consistent with product quality control requirements and can be adjusted through the system parameter configuration interface.

[0181] If the trend fitting shows significant inconsistencies in the valuation, such as the difference between multiple values ​​exceeding a set threshold, or the appearance of abrupt trends or non-monotonic changes, the system will output a "valuation inconsistency" prompt and refuse to output the grade result, prompting the user to re-collect data or review the model input settings.

[0182] Ultimately, the battery self-discharge level assessment result output in this step is a standardized label, such as "Level 1", "Level 2", "Level 3" or "Result Invalid", which can be directly used for the summarization, judgment or display of structured test reports, and also provides the input basis for step S106.

[0183] Furthermore, the process of fitting a trend curve to the estimated battery self-discharge rate over at least three consecutive cycles and verifying its consistency to obtain a battery self-discharge level assessment result includes:

[0184] The estimated self-discharge rate of the battery for at least three consecutive cycles is constructed into an ordered estimation sequence in chronological order, and the rebound amplitude corresponding to each cycle is recorded together to form a pairing sequence of self-discharge rate and rebound amplitude.

[0185] The self-discharge rate estimate between any two adjacent periods in the estimated sequence is calculated by difference, and its direction of change is determined. The number of times the direction of change reverses in the entire sequence is further counted. If the number of reversals is greater than or equal to two, the trend fluctuation indicator of the sequence is output.

[0186] Based on the trend fluctuation indicator, the paired sequence is traversed, the period with the largest increase in rebound amplitude is extracted, and the change in the self-discharge rate estimate of the period relative to the previous period is calculated as the offset sensitive difference.

[0187] When the trend fluctuation indicator exists and the offset sensitivity difference is greater than the preset offset expansion threshold, the judgment result that does not meet the consistency condition is output, and the aforementioned operation of obtaining the instantaneous voltage change sequence and current micro-perturbation response sequence of multiple short time segments of the target battery under micro-load disturbance is triggered; otherwise, the ordered estimation sequence is output as the basis for generating the battery self-discharge level evaluation result.

[0188] In the battery self-discharge characteristic analysis and rapid evaluation method described in this invention, in order to improve the accuracy of the judgment of battery self-discharge behavior and the stability of the results, the variation trend of the estimated battery self-discharge rate in multiple cycles is analyzed and its consistency is checked, so as to determine whether the battery self-discharge level evaluation result can be directly generated based on this, or whether it is necessary to re-collect data and perform a re-evaluation process.

[0189] First, the estimated battery self-discharge rates obtained over at least three consecutive cycles are arranged chronologically to construct an ordered estimation sequence. To introduce a coupling constraint on changes in the battery's physical state, the rebound amplitude for each cycle is also extracted and paired one-to-one with the corresponding self-discharge rate estimate, forming a paired data structure that includes behavioral characteristics. This structure is used to subsequently determine whether changes in the self-discharge rate are affected by fluctuations in rebound behavior, providing a foundation for trend identification and offset analysis.

[0190] Subsequently, a difference analysis is performed on the aforementioned estimation sequence. This involves calculating the difference between the estimated battery self-discharge rates of any two adjacent periods and determining the direction of change based on the sign of the difference. By statistically analyzing the direction of change across consecutive periods, if the rate change direction is found to reverse two or more times throughout the entire estimation sequence (i.e., a shift from increasing to decreasing or vice versa), the sequence is considered to exhibit unstable trend behavior. A trend fluctuation indicator is then generated to identify repeated oscillations or unexplained reverse fluctuations in the overall estimation trend of the sequence, suggesting potential risks such as local disturbances, measurement errors, or external temperature interference.

[0191] After the trend fluctuation indicator is generated, the paired sequence is traversed again based on the change in rebound amplitude. The rebound amplitude values ​​in all periods are compared, and the period with the most significant increase in rebound amplitude is extracted. Using this period as a reference, the numerical change between the estimated battery self-discharge rate at that point and the estimate of the previous period is calculated, forming a bias-sensitive difference. This difference is used to assess whether the estimated self-discharge rate also shows a significant synchronous increase against the background of a sharp increase in rebound behavior, thereby determining whether the estimate truly reflects the changes in the internal state of the battery or is a local abnormal amplification under the interference of external factors.

[0192] Finally, a consistency determination is performed based on the combined judgment results of the trend fluctuation indicator and the offset sensitivity difference. If the estimated sequence has a trend fluctuation indicator and the offset sensitivity difference exceeds the preset offset expansion threshold, it means that the estimated sequence does not meet the consistency requirements in terms of behavioral continuity and amplitude stability, and cannot be directly used for evaluation. The instantaneous voltage change sequence and current perturbation response sequence of multiple short time segments of the target battery under micro-load disturbance should be obtained again, and feature extraction and self-discharge rate estimation should be performed again. Conversely, if the above two conditions are not met simultaneously, the estimated sequence is considered to have an acceptable level of consistency and can be directly used as the basis for generating the battery self-discharge level evaluation result for subsequent report generation and quality classification processes.

[0193] This consistency verification strategy is based on the structural analysis and physical behavior coupling judgment of multi-cycle features. It does not rely on machine learning models or complex statistical curves, but is completed through the coordinated operation of directional changes, amplitude jumps and judgment thresholds of limited data points. It is suitable for large-scale, periodic battery evaluation scenarios and can be implemented by existing battery testing platforms through logic control at the software layer.

[0194] Step S106: Based on the battery self-discharge level assessment results and the corresponding battery self-discharge rate estimate, generate a structured test report including the initial offset, rebound amplitude, and estimated confidence level.

[0195] In step S106, based on the battery self-discharge level assessment results and the corresponding battery self-discharge rate estimates output in step S105, the system generates a test report with a complete structured format. This test report includes at least key indicators such as initial offset, rebound amplitude, and estimated confidence level. This report supports battery quality management, factory screening, and subsequent process optimization; its structure and content should ensure that it can be automatically read, archived, and analyzed.

[0196] First, the system extracts the original segments of the disturbance response used to generate estimates of the battery self-discharge rate for each group and traces back to the instantaneous voltage change sequence corresponding to each estimate. In these sequences, the system identifies the lowest voltage point after the disturbance ends and the subsequent voltage recovery process. By analyzing the voltage change trend during this recovery process, two core physical behavior parameters are further extracted: the initial offset and the rebound amplitude.

[0197] The initial offset, measured in volts, refers to the voltage drop at the moment the disturbance ends relative to the stable voltage before the disturbance. This metric reflects the instantaneous voltage drop of the battery under a minor load disturbance; a larger value may indicate a more significant internal resistance or polarization effect. The system calculates this metric by extracting the static voltage difference between the disturbance excitation period and the moment the disturbance ends.

[0198] Rebound amplitude refers to the voltage rise from its lowest point to its final stable value during the recovery process after a disturbance ends, also measured in volts. This indicator reflects the battery's polarization mitigation capability and the recovery efficiency of its internal diffusion mechanism, and is often used as an additional reference for assessing battery response hysteresis.

[0199] Next, the system will use the confidence score output in step S104 as the estimated confidence score in this report. This confidence score is generated by the output fusion and confidence inference unit based on the dispersion between the output results of all sub-mapping paths, and is used to reflect the credibility of the current prediction result. This confidence score is a standardized score, ranging from 0 to 1. The system can set a threshold for it to indicate whether manual review is required.

[0200] The report structure uses standardized data fields, and each report includes the following information fields: unique identifier of the target battery, sampling time, test environment temperature, manufacturing batch identifier, complete self-discharge rate estimation sequence, final self-discharge rate estimation, corresponding confidence score, battery self-discharge level assessment result generated by the system according to preset rules, initial offset, voltage rebound amplitude, and whether a re-inspection recommendation has been triggered. The report can be exported as JSON, CSV, or a structured database record, facilitating its use in automated production lines, battery sorting systems, or experimental data platforms.

[0201] Furthermore, the system supports further sorting, filtering, and alarming based on this structured test report. For example, if a battery is assessed as having a "Level 3" self-discharge rating and its confidence level is below a set threshold, the system can insert a "Retest Recommended" label into the report and mark it as requiring manual re-inspection. If all indicators perform well, the system can automatically identify it as "Qualified" and label the corresponding level.

[0202] Ultimately, the generated structured test report not only reflects quantitative estimates but also supplements the qualitative assessment dimension of voltage response behavior in an engineering manner. It has good traceability, interpretability, and batch scalability, making it suitable for large-scale deployment and operation by battery manufacturers, quality inspection agencies, and back-end system integrators.

[0203] Furthermore, based on the battery self-discharge level assessment results and the corresponding estimated battery self-discharge rate, a structured detection report is generated, including the initial offset, rebound amplitude, and estimated confidence level, comprising:

[0204] Based on the battery self-discharge level assessment results and the corresponding battery self-discharge rate estimate, the starting offset, rebound amplitude and estimated confidence of the current cycle are extracted and arranged in the order of calculation when the Bayesian weighted adaptive mapping network generates the battery self-discharge rate estimate. The three are then constructed into a list of detection fields whose structural field order corresponds to the calculation path logic.

[0205] Based on the disturbance segment index of the feature vector group corresponding to the detection field list, determine the data segment to which the corresponding original instantaneous voltage change sequence and current perturbation response sequence belong. Based on the disturbance segment index, time window number and recovery segment start voltage, generate a disturbance segment encoded fingerprint. This disturbance segment encoded fingerprint is attached to the structured detection report as a unique identifier for data tracking and report tracing.

[0206] Combining the battery internal resistance dynamic evolution tensor, the forward rebound velocity and reverse voltage fall velocity of each disturbance segment recovery segment are extracted, and the maximum difference between the two is calculated as the time period rebound asymmetry index. This time period rebound asymmetry index is added to the detection field list as a supplementary field to assist battery quality rating.

[0207] A structured detection report is generated, which includes the initial offset, the rebound amplitude, the estimated confidence level, the perturbation segment encoded fingerprint, and the time-period rebound asymmetry index. The structured detection report is accompanied by a field-level hash verification value generated based on the field content and records the current data version identifier to support the content integrity verification of the report results and the comparative analysis of historical versions.

[0208] In the battery self-discharge characteristic analysis and rapid evaluation method described in this invention, the generation of a structured test report is not merely a result display process, but a deep data encapsulation step encompassing self-discharge behavior extraction, path tracing, and quality attribution. This test report, through a rigorous data structure organization, fully inherits the key computational outputs of the aforementioned evaluation process, while enhancing the report's engineering application value, credibility, and traceability through information traceability and anomaly analysis mechanisms.

[0209] In this process, based on the obtained battery self-discharge level assessment results and their corresponding battery self-discharge rate estimates, three key feature values ​​are extracted from the current cycle: initial offset, rebound amplitude, and estimated confidence level. Initial offset refers to the voltage drop relative to the pre-disturbance steady-state voltage at the end of the disturbance, reflecting the degree of residual polarization after discharge; rebound amplitude describes the voltage rise from its lowest point to steady state, reflecting recovery capability; and estimated confidence level comes from the confidence inference results in the aforementioned Bayesian weighted adaptive mapping network, representing the credibility level of the current self-discharge rate estimate. These three data points are arranged in order, based on the derivation path order corresponding to them during the generation of the battery self-discharge rate estimate in the Bayesian weighted adaptive mapping network. This ensures that the field order in the detection report is consistent with the calculation logic of the assessment model, thus forming a detection field list with a structural field order consistent with the calculation path logic.

[0210] Subsequently, to enhance the traceability of the report content, the system traces the source of the original perturbation data based on the feature vector group segment index corresponding to each record in the detection field list. During the aforementioned feature vector group generation process, each vector group originates from a specific perturbation segment in the battery internal resistance dynamic evolution tensor. The system uses the index number of this perturbation segment, the time window number it falls within, and the starting voltage value of the recovery segment to construct a perturbation segment encoded fingerprint. This encoded fingerprint is generated in the form of a hash function, providing a unique identifier without revealing the original measurement data content. This fingerprint is bound and written to a fixed field position in the structured detection report, serving as the basis for subsequent verification of the accuracy of report fields, comparison of original data, and anomaly analysis.

[0211] In terms of supplementary feature dimensions, the structured inspection report also includes a time-bound rebound asymmetry index for enhanced quality rating. This index is calculated from the recovery segment data in the battery internal resistance dynamic evolution tensor. Specifically, during the disturbance recovery phase, the system extracts the maximum rebound velocity during the voltage rise and the maximum negative recovery velocity during the subsequent potential voltage drop, and calculates the maximum difference between the two. This difference reflects whether there is a symmetry deviation in the shape of the recovery curve, and can reveal characteristics such as delayed partial polarization release, slow instability, or localized drop-off. In engineering, it is often used to identify suboptimal cells, edge-aged batteries, or signs of compound faults. This asymmetry index is added as a supplementary field to the inspection field list, and together with the aforementioned initial offset, rebound amplitude, and estimated confidence level, it constitutes the core index set of the structured report.

[0212] Finally, the system combines all the above fields to generate a structured detection report. This report includes not only the initial offset, the rebound amplitude, and the estimated confidence level, but also the perturbation fragment encoded fingerprint and the time-period rebound asymmetry index, ensuring comprehensive structure, clear source, and complementary indicators. To prevent report fields from being tampered with or mismatched, the system performs field-level hash verification on all field content and attaches a digest signature formed by the field order, field values, and verification value set corresponding to that version. It also records the version number and configuration identifier used by the current data processing logic. This not only ensures the integrity verification and engineering audit of each structured detection report but also facilitates subsequent algorithm iterations, data batch comparisons, and detection platform version backtracking management.

[0213] The structured test reports generated in the above manner have visual friendliness, data verification capabilities, content credibility, and behavioral interpretation paths, which can significantly improve the efficiency and standardization of large-scale automatic testing and graded quality analysis of batteries.

[0214] The second embodiment of the application provides an electronic device, the electronic device comprising:

[0215] processor;

[0216] The memory is used to store a program, which, when read and executed by the processor, executes a battery self-discharge characteristic analysis and rapid evaluation method provided in the first embodiment of this application.

[0217] The third embodiment of this application provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it performs a battery self-discharge characteristic analysis and rapid evaluation method provided in the first embodiment of this application.

[0218] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for analyzing and rapidly evaluating the self-discharge characteristics of a battery, characterized in that, include: Within a set environmental stability range, acquire the instantaneous voltage change sequence and current perturbation response sequence of multiple short time segments of the target battery under micro-load perturbation; Based on the instantaneous voltage change sequence and current perturbation response sequence, a battery internal resistance dynamic evolution tensor is constructed to represent the dynamic response characteristics of the battery. The first dimension of the battery internal resistance dynamic evolution tensor is a continuous time step, reflecting the high-frequency response trajectory within the perturbation segment; the second dimension is a ternary physical feature, including the recovery velocity vector, initial recovery offset value, rebound amplitude, and maximum rebound velocity; the third dimension is a segment number dimension, used to identify which perturbation segment it belongs to, ensuring that different segment data can be processed independently or batch-classified in subsequent analysis. For the battery internal resistance dynamic evolution tensor, a multi-scale time window segmentation strategy is applied to extract local features, obtaining a feature vector set containing a voltage nonlinear rebound factor and a delay response index. The voltage nonlinear rebound factor represents the degree of deviation of the voltage recovery process from the linear trend after a disturbance ends, and is obtained by fitting the voltage sequence within the disturbance recovery segment and calculating the average squared residual. The delay response index reflects the voltage response time delay after a current disturbance is applied; its value is the time interval between the current transition moment and the moment of maximum voltage change, normalized to the proportion of the disturbance time window length. The feature vector set is input into the trained Bayesian weighted adaptive mapping network to obtain the battery self-discharge rate estimate considering manufacturing batch residual modeling and temperature offset compensation mechanism. The estimated self-discharge rate of the battery for at least three consecutive cycles is fitted with a trend curve, and the consistency of the curve is checked to obtain the battery self-discharge level assessment result. Based on the battery self-discharge level assessment results and the corresponding battery self-discharge rate estimates, a structured test report is generated, including initial offset, rebound amplitude, and estimated confidence level. The initial offset refers to the voltage drop relative to the stable voltage before the disturbance at the moment the disturbance ends, in volts, and is used to reflect the instantaneous voltage drop when the battery is subjected to a micro-load disturbance. The rebound amplitude refers to the voltage rise from the lowest point to the final stable value during the recovery process after the disturbance ends, in volts, and is used to reflect the battery's polarization mitigation capability and the recovery efficiency of the internal diffusion mechanism.

2. The method for analyzing and rapidly evaluating battery self-discharge characteristics according to claim 1, characterized in that, The micro-load disturbance refers to applying a short-time pulse current signal with a current amplitude less than 1% of the rated capacity to the target battery during the test, and the duration of each disturbance is less than 2 seconds, so as to avoid significantly changing the battery's state of charge.

3. The method for analyzing and rapidly evaluating battery self-discharge characteristics according to claim 1, characterized in that, The Bayesian weighted adaptive mapping network includes a feature encoding unit, a joint modeling unit for manufacturing batches and temperature conditions, and an output fusion and confidence inference unit. The feature encoding unit is used to receive a normalized feature vector group, which includes multiple voltage nonlinear rebound factors and delay response exponents. The feature vector group is processed by a multi-layer feedforward network to extract and compress features, and output a fixed-length encoded vector to represent the characteristics of disturbance response behavior. The manufacturing batch and temperature condition joint modeling unit is used to receive the manufacturing batch and ambient temperature information of the target battery, and concatenate it with the fixed-length encoding vector to generate a joint input vector; it includes multiple sub-mapping paths, each of which is optimized based on training data of a specific manufacturing batch and temperature range, and Bayesian posterior weights are assigned during inference based on the similarity between the current input conditions and the central condition vector of each sub-path. The output fusion and confidence inference unit is used to receive the candidate battery self-discharge rate estimates output by all sub-mapping paths, and to perform weighted fusion of each candidate value in combination with the Bayesian posterior weight to generate the final battery self-discharge rate estimate; and to calculate the confidence score based on the degree of dispersion between candidate values ​​to characterize the reliability of the final battery self-discharge rate estimate.

4. The method for analyzing and rapidly evaluating battery self-discharge characteristics according to claim 1, characterized in that, The construction of a battery internal resistance dynamic evolution tensor to represent the battery's internal dynamic response characteristics, based on the instantaneous voltage change sequence and current perturbation response sequence, includes: Within the set environmental stability range, the instantaneous voltage change sequence and current perturbation response sequence of each group of short time segments are divided into a pre-perturbation resting segment, a perturbation excitation segment, and a perturbation end recovery segment, respectively. The segment division is based on the initial slope and amplitude threshold of the current change, and the boundary time is automatically determined. In each perturbation excitation segment, the time relationship between the voltage drop start point, the voltage minimum point and the current transition point corresponding to the current perturbation is extracted and marked as the voltage response window index field, which serves as the time positioning reference for subsequent tensor construction. In each disturbance end recovery segment, the recovery velocity vector between continuous voltage sampling points within the segment is calculated, and combined with the initial recovery offset value, maximum rebound velocity and rebound amplitude, a three-dimensional battery internal resistance dynamic evolution tensor with time step, ternary physical characteristics and segment number is constructed. The maximum rebound velocity and the time difference between the delayed start-up time corresponding to each segment of the battery internal resistance dynamic evolution tensor are used as local dynamic feature fields to participate in the subsequent operation of local feature extraction using a multi-scale time window segmentation strategy.

5. The method for analyzing and rapidly evaluating battery self-discharge characteristics according to claim 4, characterized in that, The dynamic evolution tensor of the battery internal resistance is subjected to a multi-scale time window segmentation strategy for local feature extraction, resulting in a feature vector set containing the voltage nonlinear rebound factor and the delay response exponent, including: Based on the difference between the maximum rebound speed of the recovery segment and the delayed start-up time for each segment in the dynamic evolution tensor of the battery internal resistance, a local perturbation behavior index table is constructed to mark the relative time positions of the most drastic voltage change segment and the response start-up hysteresis segment in the tensor. Guided by the local disturbance behavior index table, multiple time windows of different lengths are applied to slide and divide the area centered on the peak rebound speed segment. The time window lengths include at least 50 milliseconds, 100 milliseconds, 200 milliseconds, and 500 milliseconds, and each time window coverage area must include at least one marked position. In the tensor recovery segment covered by each time window, a first-order linear trend fitting is performed, and the sum of squared residuals between the actual voltage sampled values ​​and the fitted trend line within the time window is normalized as the voltage nonlinear rebound factor corresponding to the time window. Within each time window, the time difference between the transition moment of the disturbance current and the time point corresponding to the maximum voltage change rate in the tensor is calculated, and the time difference is normalized to serve as the delay response index corresponding to the time window. The voltage nonlinear rebound factor and delay response exponent calculated in all time windows are combined into feature pairs, and all feature pairs are organized into feature vector groups according to the perturbation segment index and time window order, which are used as perturbation behavior input features for the subsequently trained Bayesian weighted adaptive mapping network.

6. The method for analyzing and rapidly evaluating battery self-discharge characteristics according to claim 1, characterized in that, The process of fitting a trend curve to the estimated battery self-discharge rate over at least three consecutive cycles and verifying its consistency to obtain a battery self-discharge level assessment result includes: The estimated self-discharge rate of the battery for at least three consecutive cycles is constructed into an ordered estimation sequence in chronological order, and the rebound amplitude corresponding to each cycle is recorded together to form a pairing sequence of self-discharge rate and rebound amplitude. The self-discharge rate estimate between any two adjacent periods in the estimated sequence is calculated by difference, and its direction of change is determined. The number of times the direction of change reverses in the entire sequence is further counted. If the number of reversals is greater than or equal to two, the trend fluctuation indicator of the sequence is output. Based on the trend fluctuation indicator, the paired sequence is traversed, the period with the largest increase in rebound amplitude is extracted, and the change in the self-discharge rate estimate of the period relative to the previous period is calculated as the offset sensitive difference. When the trend fluctuation indicator exists and the offset sensitivity difference is greater than the preset offset expansion threshold, the judgment result that does not meet the consistency condition is output, and the aforementioned operation of obtaining the instantaneous voltage change sequence and current micro-perturbation response sequence of multiple short time segments of the target battery under micro-load disturbance is triggered; otherwise, the ordered estimation sequence is output as the basis for generating the battery self-discharge level evaluation result.

7. The method for analyzing and rapidly evaluating battery self-discharge characteristics according to claim 1, characterized in that, Based on the battery self-discharge level assessment results and the corresponding estimated battery self-discharge rate, a structured inspection report is generated, including the initial offset, rebound amplitude, and estimated confidence level. Based on the battery self-discharge level assessment results and the corresponding battery self-discharge rate estimate, the starting offset, rebound amplitude and estimated confidence of the current cycle are extracted and arranged in the order of calculation when the Bayesian weighted adaptive mapping network generates the battery self-discharge rate estimate. The three are then constructed into a list of detection fields whose structural field order corresponds to the calculation path logic. Based on the disturbance segment index of the feature vector group corresponding to the detection field list, determine the data segment to which the corresponding original instantaneous voltage change sequence and current perturbation response sequence belong. Based on the disturbance segment index, time window number and recovery segment start voltage, generate a disturbance segment encoded fingerprint. This disturbance segment encoded fingerprint is attached to the structured detection report as a unique identifier for data tracking and report tracing. Combining the battery internal resistance dynamic evolution tensor, the forward rebound velocity and reverse voltage fall velocity of each disturbance segment recovery segment are extracted, and the maximum difference between the two is calculated as the time period rebound asymmetry index. This time period rebound asymmetry index is added to the detection field list as a supplementary field to assist battery quality rating. A structured detection report is generated, which includes the initial offset, rebound amplitude, estimated confidence level, perturbation segment encoded fingerprint, and time-period rebound asymmetry index. The structured detection report is accompanied by a field-level hash check value generated based on the field content and records the current data version identifier to support the content integrity verification of the report results and the comparative analysis of historical versions.

Images