A power battery health degree online evaluation method and system based on a charging curve

Abstract

The application discloses a kind of power battery health degree online evaluation method and system based on charging curve, method includes: by real-time monitoring voltage current data in charging process, the characteristic interval of constant-current charging stage and constant-voltage charging stage is dynamically divided;Extract the time domain and frequency domain characteristic parameters in the characteristic interval;The time domain and frequency domain characteristic parameters are input into the health degree evaluation model optimized by migration learning, at least the fusion health degree index containing battery capacity attenuation coefficient and internal resistance change vector is generated;Based on the fusion health degree index, combine battery historical cycle data and operating environment parameters, output final health degree evaluation result and remaining service life prediction by adaptive weighting algorithm.Utilize the embodiment of the application, battery health degree high-precision, non-invasive online evaluation and life prediction can be realized, and the real-time, accuracy and engineering applicability of evaluation are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of batteries, and particularly relates to a power battery health degree online evaluation method and system based on a charging curve. BACKGROUND

[0002] With the rapid development of electric vehicles and large-scale energy storage, accurate evaluation of the health state of power batteries has become a key to guaranteeing safe and reliable operation of a system. In the prior art, evaluation of battery health degree is mostly dependent on complete cycle tests under laboratory conditions or empirical models based on fixed feature points. These methods are often difficult to implement in real online, non-full charging and discharging application scenarios, and have problems such as poor timeliness, dependence on historical complete data, insensitivity to early performance degradation, etc. For example, only judging by a single parameter such as voltage or internal resistance cannot comprehensively reflect the complex electrochemical state changes inside the battery, resulting in limited accuracy of the evaluation result and high prediction uncertainty. In addition, traditional models generally lack self-adaptive ability to different battery models, working conditions and environments, and are difficult to achieve accurate online life prediction in actual applications. SUMMARY

[0003] The application aims to provide a power battery health degree online evaluation method and system based on a charging curve, to solve the problems in the prior art, and to achieve high-precision, non-invasive online evaluation and life prediction of battery health degree, and to improve the real-time performance, accuracy and engineering applicability of the evaluation.

[0004] One embodiment of the application provides a power battery health degree online evaluation method based on a charging curve, which comprises the following steps:

[0005] Real-time monitoring of voltage and current data during the charging process is performed to dynamically divide the feature intervals of the constant current charging stage and the constant voltage charging stage.

[0006] Time domain and frequency domain feature parameters in the feature intervals are extracted, wherein the time domain and frequency domain feature parameters at least include a voltage rise rate curve inflection point position and a current ripple spectrum distribution feature.

[0007] The time domain and frequency domain feature parameters are input into a health degree evaluation model optimized through transfer learning to generate a fusion health degree index including at least a battery capacity attenuation coefficient and an internal resistance change vector.

[0008] Based on the fusion health degree index, in combination with battery historical cycle data and operating environment parameters, an adaptive weighting algorithm is used to output a final health degree evaluation result and a remaining service life prediction.

[0009] Optionally, the real-time monitoring of voltage and current data during the charging process to dynamically divide the feature intervals of the constant current charging stage and the constant voltage charging stage comprises the following steps:

[0010] Real-time acquisition of voltage and current raw data during charging by sensors at a fixed sampling frequency to generate raw voltage and current time series data sets;

[0011] Preprocessing of raw voltage and current time series data sets, applying a sliding window filtering algorithm to eliminate high-frequency noise and abnormal fluctuations, to generate smoothed voltage and current data sequences;

[0012] Based on the smoothed voltage and current data sequences, the moving average of the voltage and current change rates is calculated, and the critical point of the transition from the constant current charging stage to the constant voltage charging stage is automatically detected by a pre-set threshold, to generate a stage transition critical point identifier;

[0013] According to the stage transition critical point identifier, the start and end time intervals of the constant current charging stage and the constant voltage charging stage are dynamically divided, and the corresponding voltage and current data ranges of each stage are recorded, to generate a dynamic feature interval division result.

[0014] Optionally, the feature interval is extracted, and the time domain and frequency domain feature parameters at least include the voltage rise rate curve inflection point position and the current ripple spectrum distribution characteristics, which include:

[0015] From the dynamic feature interval division result, the voltage data sequence of the constant current charging stage is extracted, the first derivative curve of voltage with respect to time is calculated, and the local extreme point of the derivative curve is identified as the inflection point by using a polynomial fitting algorithm to generate the voltage rise rate curve inflection point position coordinates;

[0016] From the dynamic feature interval division result, the current data sequence of the constant voltage charging stage is extracted, and the current time series data is converted into a frequency domain signal by applying a windowed Fourier transform to generate a current frequency domain spectrum graph;

[0017] Analyze the current frequency domain spectrum graph, extract the amplitude, frequency and phase information of the main ripple frequency components, calculate the distribution entropy of the ripple energy and the harmonic distortion rate, and generate a current ripple spectrum distribution characteristic vector;

[0018] Integrate the voltage rise rate curve inflection point position coordinates and the current ripple spectrum distribution characteristic vector to construct a multi-dimensional parameter set containing time domain and frequency domain features, and generate a comprehensive time-frequency feature parameter set.

[0019] Optionally, the time domain and frequency domain feature parameters are input into a health degree evaluation model optimized by transfer learning to generate a fusion health degree index containing at least a battery capacity attenuation coefficient and an internal resistance change vector, which includes:

[0020] Load a deep neural network model pre-trained based on a large-scale battery data set, and use a small amount of samples of the current battery type to fine-tune the model using transfer learning technology to generate a health degree evaluation model optimized by transfer learning.

[0021] inputting the comprehensive time-frequency feature parameter set into a health degree evaluation model optimized by transfer learning, extracting high-dimensional feature representations through multi-layer convolution and full connection layers of the model, and generating a deep health degree feature vector;

[0022] based on the deep health degree feature vector, calculating the attenuation ratio of the battery capacity relative to the rated capacity through the output layer of the model, and estimating the change trend of the battery internal resistance under different states of charge, to generate a battery capacity attenuation coefficient and an internal resistance change vector;

[0023] weighting and fusing the battery capacity attenuation coefficient and the internal resistance change vector, combining the confidence score output by the model, and generating a standardized fusion health degree index.

[0024] Optionally, based on the fusion health degree index, combining the battery historical cycle data and the operating environment parameters, and outputting the final health degree evaluation result and the remaining service life prediction through an adaptive weighting algorithm, comprising:

[0025] extracting battery historical cycle data from the historical database of the battery management system, the cycle data including cycle number, average discharge depth and charge rate statistics, and simultaneously obtaining temperature and humidity records during operation from the environmental sensor, to generate a historical cycle and environmental parameter data set;

[0026] inputting the standardized fusion health degree index and the historical cycle and environmental parameter data set into the adaptive weighting algorithm, dynamically calculating the weight coefficient according to the sensitivity of each parameter to the health degree, and generating a weighted health degree comprehensive score;

[0027] based on the weighted health degree comprehensive score, predicting the future health degree attenuation trajectory by using a particle filtering algorithm combined with a battery degradation model, calculating the remaining cycle number of the battery reaching the failure threshold, and generating a remaining service life prediction value;

[0028] integrating the weighted health degree comprehensive score and the remaining service life prediction value, generating a final evaluation report containing health degree level and life warning, and outputting a visual result.

[0029] Another embodiment of the present application provides a power battery health degree online evaluation system based on charging curve, the system comprising:

[0030] a monitoring module for dynamically dividing the feature intervals of the constant current charging stage and the constant voltage charging stage by monitoring the voltage and current data in the charging process in real time;

[0031] an extraction module for extracting time domain and frequency domain feature parameters in the feature intervals, the time domain and frequency domain feature parameters at least including voltage rise rate curve inflection point position and current ripple spectrum distribution characteristics;

[0032] an evaluation module configured to input the time-domain and frequency-domain feature parameters into a health degree evaluation model optimized through transfer learning to generate a fusion health degree index including at least a battery capacity attenuation coefficient and an internal resistance change vector;

[0033] an output module configured to output a final health degree evaluation result and a remaining service life prediction through an adaptive weighting algorithm based on the fusion health degree index in combination with battery historical cycle data and operating environment parameters.

[0034] Another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any one of the above embodiments when running.

[0035] Another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method described in any one of the above embodiments.

[0036] Compared with the prior art, the power battery health degree online evaluation method based on a charging curve provided by the present application can realize high-precision, non-invasive online evaluation and life prediction of battery health degree, and improve the real-time performance, accuracy and engineering applicability of the evaluation, by monitoring voltage and current data in real time during the charging process, dynamically dividing the feature intervals of the constant current charging stage and the constant voltage charging stage, extracting time-domain and frequency-domain feature parameters in the feature intervals, inputting the time-domain and frequency-domain feature parameters into a health degree evaluation model optimized through transfer learning to generate a fusion health degree index including at least a battery capacity attenuation coefficient and an internal resistance change vector, and outputting a final health degree evaluation result and a remaining service life prediction through an adaptive weighting algorithm based on the fusion health degree index in combination with battery historical cycle data and operating environment parameters. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 FIG. 1 is a hardware structure block diagram of a computer terminal for a power battery health degree online evaluation method based on a charging curve provided by an embodiment of the present application;

[0038] Figure 2 FIG. 2 is a flowchart of a power battery health degree online evaluation method based on a charging curve provided by an embodiment of the present application;

[0039] Figure 3 FIG. 3 is a structure diagram of a power battery health degree online evaluation system based on a charging curve provided by an embodiment of the present application. DETAILED DESCRIPTION

[0040] The embodiments described below with reference to the drawings are exemplary and are used for explaining the present application only, and cannot be explained as a limitation of the present application.

[0041] The embodiment of the present application first provides a power battery health degree online evaluation method based on a charging curve, which can be applied to an electronic device, such as a computer terminal, specifically, a common computer, etc.

[0042] The following will be described in detail taking a computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal of the power battery health degree online evaluation method based on a charging curve provided by the embodiment of the present application is shown in FIG. 1. Figure 1 As shown in the figure, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0043] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can make the processor execute any kind of power battery health degree online evaluation method based on a charging curve.

[0044] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0045] The internal memory provides an environment for the running of the computer program in the non-volatile storage medium, which, when executed by the processor, can make the processor execute any kind of power battery health degree online evaluation method based on a charging curve.

[0046] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation of the computer device to which the scheme of the present application is applied. The specific computer device can include more or less components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0047] It should be appreciated that the processor can be a central processing unit (CPU), the processor can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and the like. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like.

[0048] Referring to Figure 2 Embodiments of the present application provide a power battery health degree online evaluation method based on charging curve, which can include the following steps:

[0049] S201, by monitoring the voltage and current data in the charging process in real time, the characteristic interval of the constant current charging stage and the constant voltage charging stage is dynamically divided;

[0050] Specifically, the voltage and current raw data in the charging process can be collected in real time by a sensor at a fixed sampling frequency to generate raw voltage and current time series data sets;

[0051] This step is the basic data acquisition link of health degree evaluation, and the core is to capture the dynamic changes of voltage and current in the whole charging process through high-precision sensors, to provide complete data sources for subsequent stage division and feature extraction. The specific implementation manner is as follows:

[0052] The sensor selection needs to adapt to the parameter range and precision requirement of the power battery charging scene: the voltage sensor selects a high-precision differential voltage acquisition module (model ADS1256), the measurement range is 0-5V (covering the standard charging voltage interval of 3.0-4.2V of lithium ion power battery), the resolution is 24 bits, the error is ≤±0.01V, and the subtle changes of voltage are ensured to be captured; the current sensor adopts a closed-loop Hall current sensor (model ACS724), the measurement range is-5A to 5A (adapted to the current output under 1-3C charging rate, such as 20Ah battery 3C charging current 60A, ACS758 with extended range can be selected, the measurement range is 0-100A), the linear error is ≤±0.5%, and the response time is ≤1μs, to avoid the measurement delay when the current suddenly changes.

[0053] The setting of the sampling frequency needs to balance the data accuracy and storage overhead. Combined with the characteristics of the power battery charging stage (the current changes slowly in the constant current stage, and the current decays quickly in the constant voltage stage), the fixed sampling frequency is 10 Hz (sampling period 0.1 seconds), which can not only capture the rapid decay of the current in the constant voltage stage, but also will not cause data redundancy due to high frequency. The sensor is connected to the battery management system (BMS) master unit (model STM32H743) through the SPI interface, and 1 set of voltage and current data is collected every 0.1 seconds, with a timestamp (format "YYYYMMDDHHMMSS.fff") accurate to milliseconds, for example "20251030140000.000" "20251030140000.100" "20251030140000.200" "20251030140000.300".

[0054] The generated raw voltage and current time series data set is stored in a structured format of "timestamp-voltage value (V)-current value (A)", and each data is a record, for example, a certain set of data set fragments are: "20251030140000.000,3.20,1.50; 20251030140000.100,3.21,1.49; 20251030140000.200,3.22,1.51; 20251030140000.300,3.21,1.48". The data set is stored synchronously in the BMS local flash memory (capacity ≥ 16 GB) and the cloud database. The local storage retains the last 100 charging data, and the cloud realizes long-term backup, while labeling the charging mode (such as regular charging, fast charging) and battery identification (such as battery number B20250101) at the time of collection, ensuring data traceability.

[0055] The raw voltage and current time series data set is preprocessed, and a sliding window filtering algorithm is applied to eliminate high-frequency noise and abnormal fluctuations, generating a smoothed voltage and current data sequence.

[0056] This step improves the data quality through filtering and abnormal processing, avoids distortion caused by sensor noise and electromagnetic interference, and provides reliable input for stage division. The specific implementation is as follows:

[0057] The core of the sliding window filtering algorithm is to offset random noise through weighted average of local data. The window size is set to 5 sampling points (corresponding to 0.5 seconds). If the window size is too small, it will not effectively filter out high-frequency noise, and if the window size is too large, it will cause data lag, affecting the detection accuracy of the stage transition critical point. The filtering formula is: y(i)=(x(i-2)+2x(i-1)+3x(i)+2x(i+1)+x(i+2)) / 9, where x(i) is the original value of the current sampling point, and y(i) is the filtered value. Weighted average is used instead of simple average to highlight the weight of the current sampling point and reduce data lag.

[0058] For example, the original voltage data sequence is [3.20, 3.21, 3.22, 3.21, 3.23, 3.22, 3.25] (unit V), the filter calculation for the 3rd sampling point (3.22V) is: y(3) = (3.20 + 2x3.21 + 3x3.22 + 2x3.21 + 3.23) / 9 = (3.20 + 6.42 + 9.66 + 6.42 + 3.23) / 9 = 28.93 / 9 ≈ 3.214V, retaining three decimal places ensures accuracy; for the first and last sampling points (such as the 1st and 7th), symmetric padding is used to supplement the window data (such as the 1st sampling point using its own value to fill the missing first two points), avoiding data loss.

[0059] The abnormal fluctuation processing adopts the 3σ criterion: first, calculate the mean μ and standard deviation σ of the original data sequence, for example, a certain voltage data μ = 3.22V, σ = 0.015V, set the abnormal threshold as μ ± 3σ (i.e. 3.175V to 3.265V), if a sampling point data exceeds this range (such as 3.30V), it is determined as an abnormal value. Linear interpolation is used to replace the abnormal value, the formula is x'(i) = (x(i-1) + x(i+1)) / 2, for example, the 5th sampling point is an abnormal value 3.30V, its previous sampling point 3.21V and next sampling point 3.23V, after interpolation x'(5) = (3.21 + 3.23) / 2 = 3.22V, avoiding the influence of abnormal value on the overall data trend.

[0060] The generated smoothed voltage and current data sequence needs to keep synchronization with the timestamp of the original data set, for example, the smoothed voltage sequence is [3.200, 3.208, 3.214, 3.216, 3.220, 3.222, 3.231]V, the current sequence is [1.500, 1.495, 1.502, 1.488, 1.490, 1.495, 1.492]A, the data fluctuation amplitude is reduced from the original ±0.03V, ±0.03A to ±0.01V, ±0.005A, which not only retains the real change trend of the charging process, but also eliminates high-frequency noise and abnormal interference.

[0061] Based on the smoothed voltage and current data sequence, the moving average of the voltage and current change rate is calculated, and the critical point of the transition from constant current charging to constant voltage charging is automatically detected by a pre-set threshold, and the phase transition critical point identifier is generated;

[0062] This step is the core of dynamically dividing the charging phase, which accurately locates the phase transition time by capturing the mutation of voltage and current change rate, and the specific implementation is as follows:

[0063] The calculation of voltage rate of change (dv / dt) and current rate of change (di / dt) uses the central difference method, with the formula dv / dt(i) = (v(i+1) - v(i-1)) / (2Δt), di / dt(i) = (i(i+1) - i(i-1)) / (2Δt), where Δt = 0.1 seconds is the sampling period. Compared with forward difference and backward difference, the central difference method has smaller truncation error (O(Δt 2 )) and higher calculation accuracy. For example, the third point v(3) = 3.214 V, the second point v(2) = 3.208 V, and the fourth point v(4) = 3.216 V in the smoothed voltage sequence, the calculation dv / dt(3) = (3.216 - 3.208) / (2*0.1) = 0.008 / 0.2 = 0.04 V / s; the third point i(3) = 1.502 A, the second point i(2) = 1.495 A, and the fourth point i(4) = 1.488 A in the current sequence, di / dt(3) = (1.488 - 1.495) / (2*0.1) = (-0.007) / 0.2 = -0.035 A / s.

[0064] To further smooth the rate curve and avoid misjudgment, the calculated dv / dt and di / dt sequences are subjected to moving average processing, with a moving window size of 10 sampling points (corresponding to 1 second). For example, a certain dv / dt sequence is [0.03, 0.04, 0.05, 0.04, 0.03, 0.02, 0.01, 0.005, -0.002, -0.001] V / s, and after moving average processing, it becomes [0.032, 0.035, 0.034, 0.028, 0.021, 0.015, 0.008, 0.003, -0.001] V / s, highlighting the overall trend of the rate of change.

[0065] The determination of the preset threshold is based on the charging characteristics of the power battery: in the constant current stage, the voltage slowly rises (dv / dt ≈ 0.02-0.05 V / s) and the current is basically stable (di / dt ≈ -0.01-0.01 A / s); when switching to the constant voltage stage, the voltage tends to be stable (dv / dt drops to ≤0.01 V / s) and the current rapidly decays (di / dt absolute value increases to ≥0.1 A / s). Therefore, the threshold values are set as dv / dt threshold Thv = 0.01 V / s (lower than this value to determine that the voltage tends to be stable) and di / dt threshold Thi = -0.1 A / s (less than this value to determine that the current rapidly decays). When the moving average dv / dt ≤ Thv and di / dt ≤ Thi, and this continues for 3 sampling points (0.3 seconds), it is determined that it is the critical point of the constant current to constant voltage conversion.

[0066] For example, at a certain moment, dv / dt = 0.008 V / s (≤ 0.01 V / s), di / dt = -0.12 A / s (≤ -0.1 A / s), and the subsequent two sampling points maintain this state, corresponding to the timestamp 20251030140500.000, a phase transition critical point identifier is generated: "Critical point ID: CP-202510301405, timestamp: 20251030140500.000, dv / dt: 0.008 V / s, di / dt: -0.12 A / s, determination basis: dv / dt ≤ 0.01 V / s and di / dt ≤ -0.1 A / s for 3 sampling points", and the identifier is stored synchronously to the dataset, providing clear basis for phase division.

[0067] According to the phase transition critical point identifier, the start and end time intervals of the constant current charging phase and the constant voltage charging phase are dynamically divided, and the corresponding voltage and current data ranges of each phase are recorded to generate the dynamic feature interval division result.

[0068] This step clearly defines the boundaries of the two charging phases through the critical point identifier, and accurately defines the range for subsequent feature extraction. The specific implementation is as follows:

[0069] The start and end time intervals of the constant current charging phase: the start time is the charging start time (i.e. the timestamp of the first record of the original dataset, such as 20251030140000.000), and the end time is the timestamp of the phase transition critical point (20251030140500.000). The core feature of this phase is that the current is basically constant and the voltage continues to rise. The start and end time intervals of the constant voltage charging phase: the start time is the critical point timestamp, and the end time is the charging termination time (i.e. the timestamp when the current drops to the charging cutoff current, such as 0.05 A, 20251030141500.000). The core feature of this phase is that the voltage is constant and the current continues to decay.

[0070] When recording the corresponding voltage and current data range of each phase, the maximum, minimum, average and standard deviation of the interval need to be calculated to fully reflect the phase data characteristics:

[0071] Constant current phase (20251030140000.000-20251030140500.000): voltage range 3.20V-3.80V (average 3.50V, standard deviation 0.15V), current range 1.45A-1.50A (average 1.48A, standard deviation 0.02A), consistent with the characteristics of constant current phase with stable current and linearly rising voltage;

[0072] Constant voltage phase (20251030140500.000-20251030141500.000): voltage range 3.80V-3.82V (average 3.81V, standard deviation 0.005V), current range 0.05A-0.80A (average 0.40A, standard deviation 0.20A), consistent with the characteristics of constant voltage phase voltage stability and current exponential decay.

[0073] The generated dynamic feature interval division result is structured data, containing fields such as "phase name, start and end timestamp, voltage data statistics, current data statistics, critical point association ID", for example, the complete division result is: "dynamic feature interval division result ID: ZR-2025103014, constant current charging phase: start time 20251030140000.000, end time 20251030140500.000, voltage statistics (3.20V-3.80V, mean 3.50V, standard deviation 0.15V), current statistics (1.45A-1.50A, mean 1.48A, standard deviation 0.02A); constant voltage charging phase: start time 20251030140500.000, end time 20251030141500.000, voltage statistics (3.80V-3.82V, mean 3.81V, standard deviation 0.005V), current statistics (0.05A-0.80A, mean 0.40A, standard deviation 0.20A); associated critical point ID: CP-202510301405". This result dynamically adapts to the differences in charging curves of different batteries, even if the charging curve of the same battery changes at different cycle times and different temperatures, it can also be accurately divided through critical point detection.

[0074] S202, extracting time domain and frequency domain feature parameters in the feature interval, the time domain and frequency domain feature parameters at least including voltage rise rate curve inflection point position and current ripple spectrum distribution characteristics;

[0075] Specifically, the voltage data sequence of the constant current charging phase can be extracted from the dynamic feature interval division result, the first derivative curve of voltage with respect to time is calculated, the local extreme points of the derivative curve are identified as inflection points using a polynomial fitting algorithm, and the voltage rise rate curve inflection point position coordinates are generated;

[0076] This step focuses on the core features of the constant current phase voltage change, captures the mutation points of the voltage rise rate through derivative calculation and fitting analysis, reflects the phased changes of the battery internal chemical reaction, and the specific implementation is as follows:

[0077] First, the complete voltage data sequence of the constant current charging stage is extracted from the dynamic characteristic interval division result. The time interval of this stage is 20251030140000.000 to 20251030140500.000 (a total of 5 minutes, 3000 sampling points), the voltage data range is 3.20V-3.80V, and the smoothed voltage data sequence is as follows: 3.20V at t=0.0s, 3.208V at t=0.1s, 3.214V at t=0.2s,..., 3.80V at t=300.0s, the data is arranged in ascending order of timestamp to ensure time continuity.

[0078] The first derivative curve of voltage with respect to time, i.e. the voltage rise rate curve, is calculated using the central difference method, with the formula dv / dt(i)=(v(i+1)-v(i-1)) / (2Δt), where Δt=0.1s is the sampling period, and i is the sampling point index (from 2 to 2999 to avoid missing the first and last data). For example, the voltage at the 100th sampling point (t=9.9s) is v(100)=3.30V, the voltage at the 99th sampling point is v(99)=3.298V, and the voltage at the 101st sampling point is v(101)=3.303V. Substituting these values into the formula gives dv / dt(100)=(3.303-3.298) / (2×0.1)=0.005 / 0.2=0.025V / s, which reflects the speed of voltage rise at this moment. By traversing all valid sampling points, a voltage rise rate curve containing 2998 data points is generated. The initial stage of the curve is about 0.04V / s (battery active material reacts quickly), the middle stage gradually decreases to 0.02V / s (reaction rate slows down), and the later stage fluctuates slightly.

[0079] The polynomial fitting algorithm is used to smooth the derivative curve and identify local extreme points. A cubic polynomial is selected (a cubic polynomial can accurately fit the rate curve of nonlinear changes, and a higher order is prone to overfitting, and a lower order cannot capture the inflection point), and the fitting formula is f(t)=a3t 3 +a2t 2 +a1t+a0, where t is the relative time of the constant current stage (0 to 300s), and a3, a2, a1, a0 are the fitting coefficients. The coefficients are solved by the least squares method, for example, the fitting gives a3=-1.2×10 -7 , a2=8.5×10 -5 , a1=0.038, a0=0.002, the mean square error (MSE) of the fitted curve and the original derivative curve is 1.5×10 -6 V 2 / s 2 , the fitting degree (R 2 ) reaches 0.98, ensuring the consistency of the curve trend.

[0080] The identification of local extreme points is achieved by solving the first derivative of the fitting function for zero, i.e. f'(t) = 3a3t 2 + 2a2t + a1 = 0, substituting the coefficients into the calculation gives two solutions t1 = 80s, t2 = 220s, corresponding to the maximum point and the minimum point of the rate curve, i.e. the inflection point of the voltage rising rate curve. The generated inflection point position coordinates are expressed in the format of "relative time (s) - voltage rising rate (V / s) - corresponding voltage (V)", for example: "inflection point 1: (80.0s, 0.032V / s, 3.45V), inflection point 2: (220.0s, 0.018V / s, 3.68V)". Inflection point 1 corresponds to the reaction acceleration stage after the preliminary formation of the SEI film on the surface of the positive electrode material of the battery, and inflection point 2 corresponds to the reaction deceleration stage caused by the decrease in active material utilization rate, and the positions and rate values of the two inflection points directly reflect the battery health status (the rate of inflection point 1 of a healthy battery is usually ≥0.03V / s, and a decaying battery may decrease to ≤0.02V / s).

[0081] The current data sequence of the constant voltage charging stage is extracted from the dynamic characteristic interval division result, and the windowed Fourier transform is applied to convert the current time series data into a frequency domain signal to generate a current frequency domain spectrum graph.

[0082] This step captures the frequency characteristics of the current ripple through frequency domain analysis. The current ripple is determined by the switching characteristics of the charging power supply and the equivalent impedance of the battery, and its spectral distribution changes with the battery health degree (when the internal resistance increases, the ripple amplitude at a specific frequency will change significantly). The specific implementation is as follows:

[0083] The current data sequence of the constant voltage charging stage is extracted, and the time interval of this stage is 20251030140500.000 to 20251030141500.000 (a total of 10 minutes, 6000 sampling points), and the current decays from 0.80A to 0.05A. An example of the smoothed current data sequence is: 0.80A at t = 300.0s, 0.795A at t = 300.1s, 0.792A at t = 300.2s,..., 0.05A at t = 900.0s, the data presents an exponential decay trend, with superimposed small high-frequency ripples.

[0084] The time-domain current data is converted into frequency-domain signals by applying the windowed Fourier transform (STFT). The core parameter settings are as follows: the window function is Hanning Window, the window length is set to 256 sampling points (corresponding to 25.6s, taking into account the frequency resolution and time resolution), the window overlap rate is 50% (to avoid spectral leakage), the Fourier transform point number (FFTSize) is 256 points, the frequency resolution Δf = 1 / 25.6 ≈ 0.039Hz, and the main ripple frequency range of 0-5Hz is covered (the switching frequency of the charging power supply is usually 1-10kHz, but after filtering, the battery end current ripple is concentrated in the low frequency band).

[0085] During windowing, 6000 current data points are divided into 46 data segments with a window length of 256 and an overlap of 128. Each data segment is multiplied by a Hanning window function and then subjected to FFT transformation. For example, the current data of the first data segment (t = 300.0s-325.6s) is windowed to eliminate the abrupt changes at the beginning and end of the data segment. After FFT transformation, 256 frequency point amplitudes are obtained, where the frequency f = k × Δf (k = 0 to 127, the Nyquist frequency is 5Hz, and frequencies with k > 127 are mirror frequencies, which are discarded).

[0086] The generated current frequency spectrum graph has frequency (Hz) as the horizontal axis (0-5Hz) and current amplitude (A) as the vertical axis. In the spectrum graph, several obvious peaks corresponding to different frequency ripple components can be seen, such as amplitude 0.02A at f = 0.156Hz, amplitude 0.015A at f = 0.312Hz, and amplitude 0.008A at f = 0.625Hz. These peak frequencies are related to the PWM modulation frequency harmonics of the charging power supply and the resonance frequency of the battery equivalent circuit. The amplitude distribution and peak position of the spectrum graph provide intuitive evidence for subsequent feature extraction.

[0087] The current frequency spectrum graph is analyzed to extract the amplitude, frequency, and phase information of the main ripple frequency components, calculate the distribution entropy of the ripple energy and the harmonic distortion rate, and generate a current ripple spectrum distribution feature vector.

[0088] This step quantitatively extracts key features from the frequency spectrum, reflects the distribution pattern and distortion degree of the current ripple, and indirectly represents the changes in battery resistance, capacitance, and other parameters. The specific implementation is as follows:

[0089] First, the main ripple frequency components are extracted, sorted by amplitude from large to small, and the top 5 peak frequencies (cumulative energy accounting for more than 85% of the total ripple energy) are selected. Three core information items are extracted for each frequency component: amplitude (A), frequency (Hz), and phase (°). The amplitude is the absolute value of the complex amplitude after FFT transformation, the frequency is the horizontal axis coordinate of the corresponding peak, and the phase is the argument of the complex amplitude (range 0-360°). For example, the extracted main ripple frequency components are:

[0090] Frequency f1 = 0.156 Hz, amplitude A1 = 0.02 A, phase φ1 = 30°;

[0091] Frequency f2 = 0.312 Hz, amplitude A2 = 0.015 A, phase φ2 = 60°;

[0092] Frequency f3 = 0.625 Hz, amplitude A3 = 0.008 A, phase φ3 = 90°;

[0093] Frequency f4 = 1.25 Hz, amplitude A4 = 0.005 A, phase φ4 = 120°;

[0094] Frequency f5 = 2.5 Hz, amplitude A5 = 0.003 A, phase φ5 = 150°.

[0095] The distribution entropy of the ripple energy is used to measure the uniformity of the ripple energy in the frequency domain. The greater the distribution entropy, the more dispersed the energy distribution, and the more complex the frequency domain characteristics of the equivalent impedance of the battery (usually the health decreases). The calculation steps are: 1. Calculate the energy of each main frequency component = 0.5 × 2 (unit A 2 ), for example, E1 = 0.5 × 0.02 2 = 2 × 10 -4 A 2 , E2 = 0.5 × 0.015 2 = 1.125 × 10 -4 A 2 , the total energy E_total = E1 + E2 + E3 + E4 + E5 = 4.3 × 10 -4 A 2 ; 2. Calculate the energy proportion of each frequency component = / E_total, for example, p1 = 2 × 10 -4 / 4.3 × 10 -4 ≈ 0.465, p2 = 1.125 × 10 -4 / 4.3 × 10 -4 ≈ 0.262; 3. Substitute the distribution entropy formula H = -Σ( × ln ), and the calculation is H = -(0.465 × ln 0.465 + 0.262 × ln 0.262 + 0.074 × ln 0.074 + 0.029 × ln 0.029 + 0.011 × ln 0.011) ≈ 1.23 (dimensionless).

[0096] The total harmonic distortion (THD) is used to measure the distortion degree of the current ripple, which is defined as the ratio of the total effective value of all harmonic components to the effective value of the fundamental component, and the fundamental component is selected as the frequency component with the largest amplitude (f1=0.156 Hz). The calculation steps are as follows:

[0097] 1. Fundamental effective value: ;

[0098] 2. Total harmonic effective value:

[0099] ;

[0100] 3. Harmonic distortion rate .

[0101] The generated current ripple spectrum distribution feature vector is a 17-dimensional parameter vector, arranged in the order of "amplitude of 5 frequency components-frequency of 5 frequency components-phase of 5 frequency components-distribution entropy-harmonic distortion rate", and an example is: [0.02, 0.015, 0.008, 0.005, 0.003, 0.156, 0.312, 0.625, 1.25, 2.5, 30, 60, 90, 120, 150, 1.23, 8.87]. Each dimension parameter retains three decimal places (angle retains integer), ensuring the standardization of the feature vector.

[0102] Integrate the coordinates of the inflection point position of the voltage rise rate curve and the current ripple spectrum distribution feature vector to construct a multi-dimensional parameter set containing time domain and frequency domain features, and generate a comprehensive time-frequency feature parameter set.

[0103] This step organically integrates time domain and frequency domain features to form a multi-dimensional parameter set that comprehensively reflects the battery charging characteristics, providing rich and effective input for the subsequent health degree evaluation model. The specific implementation is as follows:

[0104] First, parameterize and expand the coordinates of the inflection point position of the voltage rise rate curve. Each inflection point contains three core parameters: relative time (s), voltage rise rate (V / s), and corresponding voltage (V). Two inflection points have six time domain parameters, and an example is: [80.0, 0.032, 3.45, 220.0, 0.018, 3.68]. To enhance the representativeness of the time domain features, the rate difference (0.032-0.018=0.014 V / s) and the time interval (220.0-80.0=140.0 s) of the two inflection points are calculated, and two derived time domain parameters are added, resulting in a total of eight time domain feature parameters, and an example is: [80.0, 0.032, 3.45, 220.0, 0.018, 3.68, 0.014, 140.0].

[0105] The current ripple spectrum feature vector is 17-dimensional frequency domain parameters, which is directly included in the multi-dimensional parameter set without additional processing. When integrating, ensure the dimensional consistency of time domain and frequency domain parameters, and normalize all parameters (map to the 0-1 interval). The normalization formula is x_norm=(x-x_min) / (x_max-x_min), where x_min and x_max are the physical limit range of each parameter: relative time 0-300s, voltage rise rate 0-0.05V / s, corresponding voltage 3.0-4.2V, rate difference 0-0.03V / s, time interval 0-200s, ripple amplitude 0-0.1A, frequency 0-5Hz, phase 0-360°, distribution entropy 0-3, harmonic distortion 0-20%. For example, the inflection point 1 relative time 80.0s is normalized to (80.0-0) / (300-0)=0.267, the ripple amplitude 0.02A is normalized to (0.02-0) / (0.1-0)=0.2, and the phase 30° is normalized to 30 / 360≈0.083.

[0106] The constructed multi-dimensional parameter set has a total of 25 dimensions (8 time domain + 17 frequency domain), and an example of the normalized comprehensive time-frequency feature parameter set is: [0.267, 0.640, 0.375, 0.733, 0.360, 0.567, 0.467, 0.700, 0.200, 0.150, 0.080, 0.050, 0.030, 0.031, 0.062, 0.125, 0.250, 0.500, 0.083, 0.167, 0.250, 0.333, 0.417, 0.410, 0.443]. This set contains the time domain dynamic characteristics of voltage change and the frequency domain distribution characteristics of current ripple during battery charging, which not only reflects the stage change of battery chemical reaction, but also embodies the frequency domain characteristics of battery equivalent impedance, providing comprehensive and accurate input basis for the health degree evaluation model. At the same time, normalization processing ensures the convergence speed and accuracy of model training and reasoning.

[0107] S203, input the time domain and frequency domain feature parameters into the health degree evaluation model optimized by transfer learning to generate a fusion health degree index containing at least a battery capacity attenuation coefficient and an internal resistance change vector;

[0108] Specifically, a deep neural network model pre-trained based on a large-scale battery data set can be loaded, and a transfer learning technique is used to fine-tune the model with a small number of samples of the current battery type to generate a health degree evaluation model optimized by transfer learning;

[0109] This step solves the model training problem in the small sample scenario through transfer learning, uses the general feature extraction ability of the pre-trained model, and fine-tunes the current battery type-specific data to ensure that the model has both generalization and adaptation to specific battery characteristics. The specific implementation is as follows:

[0110] The pre-trained deep neural network model selects a hybrid architecture of "Convolutional Neural Network (CNN) + Fully Connected Network (FCN)" designed specifically for battery charging curve features. The input layer of the model adapts 25-dimensional integrated time-frequency feature parameters, the hidden layer contains 3 convolutional layers (extracting local features) and 3 fully connected layers (fusing global features), and the output layer is divided into two branches (corresponding to capacity attenuation coefficient and internal resistance change vector respectively). The pre-training dataset covers 100,000 groups of charging curve data of different types of power batteries (lithium-ion ternary batteries, lithium iron phosphate batteries), including full life cycle data from new batteries to failed batteries (cycle times 0-2000 times). Each group of data is labeled with the true capacity attenuation coefficient and internal resistance measurement value, ensuring that the pre-trained model learns the general mapping relationship between battery health and charging features.

[0111] The core of transfer learning is "freeze general feature layer + fine-tune specific adaptation layer": first, freeze the first two convolutional layers of the model (responsible for extracting voltage rising rate inflection point, ripple spectrum, etc. general features), and only unfreeze the last convolutional layer and three fully connected layers; Then collect a small number of labeled samples (50 groups, including charging curve features and corresponding health indicators of different cycle times) of the current battery type (such as 20Ah ternary lithium battery) to build a fine-tuning dataset. The fine-tuning process uses the Adam optimizer with a learning rate of 0.0001 (much lower than the pre-training phase of 0.001 to avoid damaging the pre-training features), and the loss function is mean square error (MSE). Iterative training for 100 rounds, with a batch size of 10. During training, the early stopping mechanism is introduced, and when the validation set loss does not decrease for 10 consecutive rounds, training is stopped. The final fine-tuned model's validation set mean square error decreases from 0.05 to 0.015, the capacity attenuation coefficient prediction error is ≤3%, and the internal resistance change vector prediction error is ≤5mΩ. The health evaluation model optimized by transfer learning retains the generalization ability of large-scale data training and precisely adapts to the characteristics of the current battery.

[0112] Input the integrated time-frequency feature parameter set into the health evaluation model optimized by transfer learning, extract high-dimensional feature representations through the multi-layer convolution and fully connected layers of the model, and generate a deep health feature vector;

[0113] This step converts low-dimensional time-frequency features into high-dimensional vectors that can accurately represent the battery health state through the feature extraction ability of the model, providing a core basis for subsequent health index calculation. The specific implementation is as follows:

[0114] The integrated time-frequency feature parameter set is a 25-dimensional normalized vector (such as [0.267, 0.640, 0.375, …, 0.443]), which is input into the model and processed according to the process of “convolution to extract local features → pooling to reduce dimension → full connection to fuse global features”. The parameters of each layer of the model are designed to match the structural characteristics of the time-frequency features:

[0115] Convolution layer: a total of 3 layers, using 1D convolution kernel (adapted to one-dimensional feature vector), the first layer has 32 convolution kernels, kernel size 3×1 (sliding along the feature dimension), step 1, activation function ReLU, used to extract the local correlation between adjacent features (such as the correlation between voltage inflection point time and ripple frequency), output feature dimension 25-3+1=23; the second layer has 64 convolution kernels, kernel size 3×1, step 1, ReLU activation, output 21; the third layer has 128 convolution kernels, kernel size 2×1, step 1, ReLU activation, output 20. After each convolution, a batch normalization (BN) layer is added to speed up training convergence and suppress gradient vanishing.

[0116] Pooling layer: average pooling layer is inserted after the second and third convolution layers, with a pooling kernel size of 2×1 and a step of 2, used to reduce dimension and preserve key features. The output of the second layer after pooling is 10-dimensional, and the output of the third layer after pooling is 10-dimensional (since the output of the third layer is 20-dimensional, it is halved after pooling).

[0117] Full connection layer: a total of 3 layers, the first layer inputs 10-dimensional, outputs 64-dimensional, ReLU activation, and introduces a dropout layer (dropout rate 0.2, to prevent overfitting); the second layer inputs 64-dimensional, outputs 32-dimensional, ReLU activation; the third layer inputs 32-dimensional, outputs 128-dimensional, without activation function, directly as a deep health feature vector.

[0118] For example, after inputting a 25-dimensional integrated time-frequency feature vector, the first layer extracts the “correlation feature between voltage inflection point rate and low frequency spectrum amplitude”, the second layer captures the “coupling feature between high frequency ripple distribution and inflection point time interval”, and the full connection layer fuses it into a 128-dimensional high-dimensional vector, such as [0.123, 0.345, 0.217, 0.468, …, 0.592] (128-dimensional). This vector is no longer an isolated time-frequency parameter, but an abstract representation of deep health information such as battery internal chemical reaction rate, electrode aging degree, and equivalent impedance change, and each dimension corresponds to a composite feature related to health status, providing rich and accurate input for subsequent index calculation.

[0119] Based on the deep health feature vector, the model output layer calculates the attenuation ratio of the battery capacity relative to the rated capacity, and estimates the change trend of the battery internal resistance at different states of charge, generating a battery capacity attenuation coefficient and an internal resistance change vector;

[0120] This step is the core prediction link of the model, which maps high-dimensional features to specific health metrics through a double-branch output layer, directly reflecting the core health status of the battery. The specific implementation is as follows:

[0121] The model output layer adopts a double-branch structure, corresponding to two core health indicators, and the output layer parameters are precisely adapted to the deep feature vectors:

[0122] Battery capacity attenuation coefficient calculation branch: the output layer has 1 neuron, and the activation function is Sigmoid (mapped to the 0-1 interval), used to output the capacity attenuation ratio. The capacity attenuation coefficient is defined as "(rated capacity-current actual capacity) / rated capacity", where the rated capacity is the factory calibration capacity of the battery (such as 20Ah), and the current actual capacity is predicted by the model based on deep features. For example, after the deep health feature vector is input into this branch, the output value is 0.12, which means the capacity attenuation coefficient is 0.12, corresponding to the current actual capacity = 20Ah x (1-0.12) = 17.6Ah. The training goal of this branch is to minimize the mean square error between the predicted attenuation coefficient and the true attenuation coefficient. During training, the network weights are continuously optimized through backpropagation to ensure that the prediction error is ≤3%.

[0123] Internal resistance change vector calculation branch: the output layer has 5 neurons without activation function (directly outputs real numbers), used to output the internal resistance change values at different states of charge (SOC), with SOC selected uniformly at 20%, 40%, 60%, 80%, and 100% (covering the full charging range of the battery). The internal resistance change vector is defined as "actual internal resistance at current SOC-new battery standard internal resistance at the same SOC", with units of milliohms (mΩ). The standard internal resistance of a new battery is calibrated through experiments (for example, the standard internal resistance of a 20Ah ternary lithium battery: 22mΩ at SOC20%, 20mΩ at SOC40%, 18mΩ at SOC60%, 16mΩ at SOC80%, and 15mΩ at SOC100%). For example, after the deep feature vector is input into this branch, the output value is [4.2, 3.8, 3.5, 3.1, 2.7]mΩ, which means the current battery's internal resistance at each SOC is increased by 4.2mΩ, 3.8mΩ, 3.5mΩ, 3.1mΩ, and 2.7mΩ compared to the new battery, generating the internal resistance change vector [4.2, 3.8, 3.5, 3.1, 2.7]mΩ. The training goal of this branch is to minimize the average absolute error of the 5 internal resistance change values, ensuring that the single-SOC internal resistance prediction error is ≤0.5mΩ.

[0124] The outputs of the two branches are independent of each other and related to each other, the capacity attenuation coefficient reflects the overall energy storage capability degradation of the battery, and the internal resistance change vector reflects the dynamic change of the internal ion transmission resistance of the battery, which together constitute the core quantitative indicators for describing the health state of the battery and are one-to-one corresponding to the abstract features in the deep health degree feature vector (for example, the "high-frequency component amplitude of ripple" dimension in the high-dimensional feature is directly related to the internal resistance change, and the "voltage inflection point rate" dimension is directly related to the capacity attenuation).

[0125] The battery capacity attenuation coefficient and the internal resistance change vector are weighted and fused, and the confidence score output by the model is combined to generate a standardized fusion health degree index.

[0126] This step eliminates the one-sidedness of a single indicator by scientifically fusing the two core indicators, and introduces a confidence score to ensure the reliability of the results. Finally, a standardized health quantitative value is generated. The specific implementation is as follows:

[0127] The core of weighted fusion is to assign weights according to the importance of the indicators. The weights are determined based on the battery health degree influence factor analysis: the capacity attenuation coefficient directly determines the available energy storage capacity of the battery and is the core evaluation dimension of the health degree, with a weight allocation of 0.6; the internal resistance change vector reflects the charging and discharging efficiency and safety risk of the battery, with a weight allocation of 0.4. Before fusion, the internal resistance change vector needs to be normalized (since the internal resistance change value at each SOC is a multi-dimensional vector, it needs to be converted to a single-dimensional scalar). The "mean normalization" method is used: calculate the average value of the internal resistance change vector, and then divide it by the average value of the standard internal resistance of the new battery (the average value of the standard internal resistance of the new battery = (22+20+18+16+15) / 5 = 18.2 mΩ), to get the normalized value of the internal resistance change. For example, the average value of the internal resistance change vector [4.2, 3.8, 3.5, 3.1, 2.7] mΩ is 3.46 mΩ, and the normalized value is 3.46 / 18.2 ≈ 0.19.

[0128] The confidence score output by the model reflects the reliability of the prediction results, which is generated based on the prediction error statistics in the model training process, with a range of 0-1 (1 represents complete reliability, and 0 represents unreliability). Its calculation logic is "1-(prediction error / maximum allowed error)", for example, the capacity attenuation coefficient prediction error is 0.003 (the maximum allowed error is 0.03), the internal resistance change vector prediction error is 0.2 mΩ (the maximum allowed error is 0.5 mΩ), and the confidence score is (1-0.003 / 0.03) x 0.5 + (1-0.2 / 0.5) x 0.5 = 0.95 x 0.5 + 0.6 x 0.5 = 0.775.

[0129] The calculation formula of the standardized fusion health index is: standardized index = (capacity attenuation coefficient x 0.6 + internal resistance change normalized value x 0.4) x 100 x confidence score, wherein x 100 is to map the index to the intuitive score range of 0-100 (0 indicates complete failure, and 100 indicates a brand-new state). Substituting the example data, (0.12 x 0.6 + 0.19 x 0.4) x 100 x 0.775 = (0.072 + 0.076) x 77.5 = 0.148 x 77.5 ≈ 11.47 points, and the standardized fusion health index is 11.5 points after rounding.

[0130] The index not only integrates the two core health dimensions of capacity and internal resistance, but also filters unreliable prediction results through the confidence score. After standardization, it has intuitive comparability, and the health status of different batteries and different cycle times can be directly quantified and compared through the index, providing a core quantitative basis for subsequent final evaluation.

[0131] S204, based on the fusion health index, combining battery historical cycle data and operating environment parameters, outputting a final health degree evaluation result and a remaining service life prediction through an adaptive weighting algorithm.

[0132] Specifically, the battery historical cycle data can be extracted from the historical database of the battery management system, and the cycle data includes cycle times, average discharge depth and charge rate statistics. At the same time, the temperature and humidity records during operation are obtained from the environmental sensor to generate a historical cycle and environmental parameter data set.

[0133] This step integrates the historical operation data and environmental data of the whole life cycle of the battery to provide multi-dimensional reference for health degree evaluation and life prediction, making up for the limitations of single charge curve characteristics. The specific implementation is as follows:

[0134] The historical database of the battery management system (BMS) adopts structured storage, and the core fields include "battery identification, cycle times, single discharge depth, charge rate, charge start time, charge end time". The data is arranged in ascending order of cycle times, and a data is automatically recorded every time a charge and discharge cycle is completed. When extracting historical cycle data, the current evaluation time is taken as a node, and the last 100 complete cycles (covering the recent health status change trend of the battery) are traced back. The core parameter definition and extraction rules are as follows:

[0135] Cycle times: directly read the database field, for example, the current battery has completed 800 cycles, and the last 100 cycles (701-800 times) are extracted;

[0136] Average discharge depth (DOD): Discharge depth = (rated capacity - remaining capacity) / rated capacity, single discharge depth is calculated according to the formula, the average discharge depth is the average of 100 cycle discharge depths, for example, 100 cycle discharge depths are 0.8, 0.75, 0.82…, the average is 0.78 (i.e. 78%), reflecting the discharge intensity of the battery in daily use;

[0137] Charging rate statistics: charging rate = charging current / rated capacity (for example, a 20Ah battery with a charging current of 10A is 0.5C), statistics of different charging rates in 100 cycles, for example, 0.5C accounts for 60%, 1C accounts for 30%, and 2C accounts for 10%, generating charging rate distribution statistics, high proportion of high rate charging will accelerate battery degradation, which needs to be considered.

[0138] The environmental sensor is deployed in the battery cabin, using a digital temperature and humidity sensor (model SHT30), measuring range temperature -40℃ to 85℃ (accuracy ±0.2℃), humidity 0-100%RH (accuracy ±2%RH), sampling frequency 1 / minute, data real-time synchronization to BMS database, associated with charging cycle data through time stamp. When extracting environmental parameters, according to the charging time period of the last 100 cycles, match the temperature and humidity data of the corresponding period, calculate the average temperature and average humidity during each cycle charging, for example, the charging time of the 750th cycle is 202510251400-1415, the corresponding temperature record is 25℃, 25.2℃…, the average temperature is 25.1℃, the average humidity is 55%RH.

[0139] The generated historical cycle and environmental parameter data set is structured and integrated according to "cycle batch-cycle number-average discharge depth-charge rate distribution-average temperature-average humidity", for example: "batch ID: His-20251030, cycle number: 701-800 times, average discharge depth: 0.78, charge rate distribution: 0.5C (60%), 1C (30%), 2C (10%), average temperature: 24.8℃, average humidity: 52%RH". The data set is also labeled with data integrity indicators (such as "100 cycle data complete, no missing"), ensuring the reliability of subsequent algorithm input.

[0140] Input the standardized fusion health index, historical cycle and environmental parameter data set into the adaptive weighted algorithm, dynamically calculate the weight coefficient according to the sensitivity of each parameter to the health index, and generate a weighted health index comprehensive score;

[0141] This step balances the contribution of different parameters to the health evaluation through dynamic weight distribution, avoids the evaluation deviation caused by fixed weight, and makes the comprehensive score more consistent with the actual health status of the battery. The specific implementation is as follows:

[0142] The core logic of the adaptive weighting algorithm is "sensitivity analysis + dynamic adjustment". First, the set of evaluation parameters is determined: 1 standardized fusion health index (denoted as F, example value 11.5 points), 3 historical cycle parameters (cycle number C = 800 times, average discharge depth D = 0.78, high-rate charging proportion R = 10%), 2 environmental parameters (average temperature T = 24.8℃, average humidity H = 52%RH), a total of 6 evaluation parameters.

[0143] Sensitivity analysis determines the influence priority of each parameter based on the battery attenuation mechanism: the standardized fusion health index directly quantifies the current health status, with the highest sensitivity; the cycle number is the core driving factor of attenuation (the more cycles, the more serious the attenuation), with the second highest sensitivity; the average discharge depth, high-rate charging proportion, and average temperature affect attenuation through accelerated chemical reactions, with medium sensitivity; the average humidity has a relatively small effect on attenuation, with the lowest sensitivity. The algorithm calculates the feature importance of each parameter through Gradient Boosting Decision Tree (GBDT) as the basis for quantifying sensitivity. The example calculation of feature importance is: F(0.4), C(0.25), D(0.15), R(0.1), T(0.08), H(0.02), and the sum of feature importance is 1, which is directly used as the dynamic weight coefficient.

[0144] Parameter normalization ensures dimensional consistency: the cycle number is mapped to the 0-1 interval (normalization formula C_norm = C / 2000, 2000 times is the common failure cycle number), example C_norm = 800 / 2000 = 0.4; average discharge depth D_norm = D (already in 0-1 interval); high-rate charging proportion R_norm = R / 100 = 0.1; average temperature T_norm = (T-25) / 20 (with 25℃ as the benchmark, temperature range -40℃ to 85℃ is mapped to 0-1), example T_norm = (24.8-25) / 20 = -0.01, taking the absolute value as 0.01; average humidity H_norm = H / 100 = 0.52; standardized fusion health index F_norm = F / 100 = 0.115 (100 points for full marks).

[0145] The calculation formula of the weighted health degree comprehensive score is: comprehensive score = Σ (parameter normalized value × corresponding weight coefficient) × 100. The calculation formula of the comprehensive score is: comprehensive score = (0.115 × 0.4 + 0.4 × 0.25 + 0.78 × 0.15 + 0.1 × 0.1 + 0.01 × 0.08 + 0.52 × 0.02) × 100 = (0.046 + 0.1 + 0.117 + 0.01 + 0.0008 + 0.0104) × 100 = 0.2842 × 100 ≈ 28.4 points, and the score is 28.4 points after rounding off. The score comprehensively reflects the current health characteristics, historical use intensity and environmental influence, and is more comprehensive than a single index in reflecting the actual state of the battery.

[0146] Based on the weighted health degree comprehensive score, a particle filtering algorithm is used to predict the future health degree attenuation trajectory in combination with a battery degradation model, to calculate the remaining cycle number of the battery reaching the failure threshold, and to generate a remaining use life prediction value.

[0147] This step captures the uncertainty of battery attenuation by a probability prediction method, accurately predicts the remaining life in combination with a degradation model, and provides a decision basis for battery replacement and maintenance. The specific implementation manner is as follows:

[0148] The battery degradation model selects an exponential degradation model, which conforms to the attenuation law of the capacity and internal resistance of the power battery. The model expression is: S(N) = S0 × e^(-k × N), wherein S(N) is the weighted health degree comprehensive score after N cycles, S0 is the current comprehensive score (28.4 points), k is the attenuation coefficient (obtained by fitting historical data), and N is the future cycle number. The calculation of the attenuation coefficient k is based on the health degree change of the last 50 cycles: assuming that the comprehensive score of the 750th cycle is 35 points and the comprehensive score of the 800th cycle is 28.4 points, the model is 28.4 = 35 × e^(-k × 50), and k is solved as -ln(28.4 / 35) / 50 ≈ -ln(0.8114) / 50 ≈ 0.0041 (times -1 ).

[0149] The particle filtering algorithm is used to process the uncertainty of the degradation model. The core steps are “particle initialization-state prediction-weight update-resampling”:

[0150] Particle initialization: 1000 particles are generated, each particle represents a possible attenuation coefficient k_i (randomly distributed around the fitting value 0.0041, ranging from 0.0035 to 0.0047), and the particle weight is initialized to 1 / 1000;

[0151] State prediction: based on the exponential degradation model, the health degree score S_i(N) of each particle under the future cycle number N is predicted as 28.4 × e^(-k_i × N);

[0152] Weight update: With the battery failure threshold (the failure threshold is usually 30% of the new battery score, assuming that the full score of the new battery is 100, and the failure threshold is 30), calculate the matching degree of each particle prediction value and the true attenuation trend, and update the weight (the higher the matching degree, the greater the weight);

[0153] Resampling: Keep the top 30% of the particles, regenerate 1000 particles, and repeat the iterative prediction until the prediction trajectory converges.

[0154] During the prediction process, set the future cycle step size to 1, and iterate until S(N)≤30, record the value of N at this time as the remaining cycle number. Example prediction result: when N=120, S(120)=28.4×e^(-0.0041×120)=28.4×e^(-0.492)≈28.4×0.611≈17.3 min≤30 min, combined with the probability distribution of particle filtering, take the median of the 95% confidence interval, and finally generate the remaining service life prediction value as 115 cycles. If the average use frequency of the battery is 3 days / cycle, the remaining service life is about 345 days (115×3), and the prediction error is ≤10%, which reflects the deterministic trend of attenuation and considers the uncertainty factors such as environmental fluctuations and changes in usage habits.

[0155] Integrate the weighted health score and the remaining service life prediction value to generate the final evaluation report containing the health level and life warning, and output the visualization results.

[0156] This step converts complex evaluation data into intuitive and easy-to-understand information through structured reports and visual displays, making it easy for users to quickly grasp the battery status and take appropriate measures. The specific implementation is as follows:

[0157] The final evaluation report adopts a three-layer structure of "core conclusion-detailed indicators-warning suggestions", the core conclusion briefly summarizes the current health status and remaining life of the battery, the detailed indicators list all evaluation parameters and calculation processes, and the warning suggestions give targeted measures according to the health level:

[0158] Core conclusion: "Battery identification B20250101, current weighted health score 28.4, health level 'warning' (30 points or less for warning, 20 points or less for failure warning); remaining service life prediction is 115 cycles (about 345 days), it is recommended to develop a replacement plan within the remaining 100 cycles (about 300 days)".

[0159] Detailed indicators: including standardized fusion health degree indicators (11.5 points), historical cycle parameters (800 cycles, 78% average discharge depth, 10% high rate ratio), environmental parameters (average temperature 24.8℃, average humidity 52%RH), weighted coefficients (F:0.4, C:0.25, D:0.15, etc.), comprehensive score calculation process, and remaining life prediction basis;

[0160] Early warning suggestions: "1. Reduce the charging rate and reduce the use of 2C high-rate charging (current ratio is 10%, recommended to reduce to less than 5%); 2. Control the discharge depth and avoid full charge and full discharge, and recommend daily discharge depth control within 70%; 3. Keep the battery compartment temperature at 20-25℃, avoid high temperature (>35℃) or low temperature (<0℃) environment charging; 4. Perform calibration charging (constant current and constant voltage to full charge and then float for 1 hour) every 30 cycles to correct capacity estimation deviation".

[0161] The visualization results are output synchronously through the BMS local display screen and the cloud monitoring platform, and the core display content includes:

[0162] Health trend chart: with cycle number as horizontal axis and health score as vertical axis, draw the score change curve of the last 100 cycles and the predicted attenuation trajectory of the next 115 cycles, mark the failure threshold line (30 points) with red color, and intuitively show the attenuation speed;

[0163] Remaining life countdown: displayed in "remaining cycle number / remaining days" dual format, such as "remaining 115 cycles (about 345 days)", and color changing according to remaining life proportion (>50% green, 30%-50% yellow, <30% red);

[0164] Parameter contribution pie chart: shows the weight proportion of each evaluation parameter, helping users understand the main reasons for the decline of health degree (such as cycle number contribution 25%, discharge depth contribution 15% in the example);

[0165] Early warning popup: when the health degree level is "warning level" or "failure warning level", trigger sound and light alarm, and display screen pops up early warning prompt and recommended measures.

[0166] The report and visualization results support export to PDF format (local storage) and cloud synchronization, making it easy to trace and maintain later, and finally forming a complete closed loop of "data collection-evaluation calculation-result display-decision support", providing accurate support for the whole life cycle management of power batteries.

[0167] It can be seen that by monitoring the voltage and current data in the charging process in real time, the characteristic intervals of the constant current charging stage and the constant voltage charging stage are dynamically divided; the time domain and frequency domain feature parameters in the characteristic intervals are extracted; the time domain and frequency domain feature parameters are input into a health degree evaluation model optimized through transfer learning to generate a fusion health degree index including at least a battery capacity attenuation coefficient and an internal resistance change vector; based on the fusion health degree index, combined with battery historical cycle data and operating environment parameters, the final health degree evaluation result and the remaining service life prediction are output through an adaptive weighting algorithm, so that high-precision, non-invasive online evaluation and life prediction of the battery health degree can be realized, and the real-time, accuracy and engineering applicability of the evaluation are improved.

[0168] Another embodiment of the present application provides a power battery health degree online evaluation system based on a charging curve, referring to Figure 3 , the system can include:

[0169] The monitoring module 301 is configured to dynamically divide the characteristic intervals of the constant current charging stage and the constant voltage charging stage by monitoring the voltage and current data in the charging process in real time.

[0170] The extraction module 302 is configured to extract the time domain and frequency domain feature parameters in the characteristic intervals, and the time domain and frequency domain feature parameters include at least a voltage rise rate curve inflection point position and a current ripple spectrum distribution feature.

[0171] The evaluation module 303 is configured to input the time domain and frequency domain feature parameters into a health degree evaluation model optimized through transfer learning to generate a fusion health degree index including at least a battery capacity attenuation coefficient and an internal resistance change vector.

[0172] The output module 304 is configured to output the final health degree evaluation result and the remaining service life prediction through an adaptive weighting algorithm based on the fusion health degree index, combined with battery historical cycle data and operating environment parameters.

[0173] The embodiment of the present application also provides a storage medium, and the storage medium stores a computer program, wherein the computer program is set to execute the steps in any one of the method embodiments.

[0174] The embodiment of the present application also provides an electronic device including a memory and a processor, and the memory stores a computer program, and the processor is set to execute the computer program to execute the steps in any one of the method embodiments.

[0175] Specifically, the electronic device can further include a transmission device and an input and output device, wherein the transmission device is connected with the processor, and the input and output device is connected with the processor.

[0176] The above detailed description of the structure, features and effects of the present application is based on the embodiments shown in the drawings. The above description is only the preferred embodiments of the present application, but the present application is not limited to the embodiments shown in the drawings. Any changes or modifications made in accordance with the concept of the present application, or equivalent embodiments with equivalent changes, are still within the scope of the present application.

Claims

1. A method for online evaluation of state of health of power battery based on charging curve, characterized in that, The method comprises: by real-time monitoring of voltage and current data during charging, dynamically dividing the characteristic interval of constant current charging stage and constant voltage charging stage; extracting time domain and frequency domain feature parameters in the characteristic interval, the time domain and frequency domain feature parameters at least including voltage rising rate curve inflection point position and current ripple spectrum distribution characteristics; wherein, extracting voltage data sequence of constant current charging stage from dynamic characteristic interval division result, calculating first derivative curve of voltage with respect to time, identifying local extreme point of derivative curve as inflection point by polynomial fitting algorithm, generating voltage rising rate curve inflection point position coordinates; extracting current data sequence of constant voltage charging stage from dynamic characteristic interval division result, applying windowed Fourier transform to convert current time series data into frequency domain signal, generating current frequency domain spectrum graph; analyzing current frequency domain spectrum graph, extracting amplitude, frequency and phase information of main ripple frequency components, calculating distribution entropy of ripple energy and harmonic distortion rate, generating current ripple spectrum distribution characteristic vector; integrating voltage rising rate curve inflection point position coordinates and current ripple spectrum distribution characteristic vector, constructing multi-dimensional parameter set containing time domain and frequency domain characteristics, generating comprehensive time-frequency feature parameter set; inputting the time domain and frequency domain feature parameters into the health degree evaluation model optimized by transfer learning to generate a fusion health degree index containing at least a battery capacity attenuation coefficient and an internal resistance change vector; based on the fusion health degree index, combining battery historical cycle data and operating environment parameters, outputting the final health degree evaluation result and the remaining service life prediction through an adaptive weighting algorithm.

2. The method of claim 1, wherein, The method comprises: real-time acquisition of voltage and current raw data during charging by sensors at a fixed sampling frequency to generate raw voltage and current time series data set; preprocessing the raw voltage and current time series data set, applying sliding window filtering algorithm to eliminate high-frequency noise and abnormal fluctuations, generating smoothed voltage and current data sequence; based on the smoothed voltage and current data sequence, calculating the moving average of voltage and current change rate, automatically detecting the critical point of constant current charging stage to constant voltage charging stage conversion through a preset threshold, generating stage conversion critical point identifier; according to the stage conversion critical point identifier, dynamically dividing the start and end time interval of constant current charging stage and constant voltage charging stage, and recording the corresponding voltage and current data range of each stage, generating dynamic characteristic interval division result.

3. The method of claim 2, wherein, The method comprises: loading a deep neural network model pre-trained based on a large-scale battery data set, fine-tuning the model using a small amount of samples of the current battery type using transfer learning technology to generate a health degree evaluation model optimized by transfer learning; inputting the comprehensive time-frequency feature parameter set into the health degree evaluation model optimized by transfer learning, extracting high-dimensional feature representation through the multi-layer convolution and fully connected layer of the model to generate a deep health degree feature vector; Based on the deep health degree feature vector, the attenuation ratio of the battery capacity relative to the rated capacity is calculated through the model output layer, and the change trend of the battery internal resistance under different states of charge is estimated to generate the battery capacity attenuation coefficient and the internal resistance change vector; The battery capacity attenuation coefficient and the internal resistance change vector are weighted and fused to generate a standardized fusion health degree index combined with the confidence score output by the model.

4. The method of claim 3, wherein, Based on the fusion health degree index, the historical cycle data of the battery and the operating environment parameters are combined to output the final health degree evaluation result and the remaining service life prediction through an adaptive weighting algorithm, including: Extract the battery historical cycle data from the historical database of the battery management system, including the cycle number, average discharge depth and charge rate statistics, and obtain the temperature and humidity records during operation from the environmental sensors to generate a historical cycle and environmental parameter dataset; Input the standardized fusion health degree index and the historical cycle and environmental parameter dataset into the adaptive weighting algorithm to dynamically calculate the weight coefficient according to the sensitivity of each parameter to the health degree, and generate a weighted health degree comprehensive score; Based on the weighted health degree comprehensive score, a particle filtering algorithm is used to predict the future health degree attenuation trajectory combined with the battery degradation model, calculate the remaining cycle number of the battery reaching the failure threshold, and generate the remaining service life prediction value; Integrate the weighted health degree comprehensive score and the remaining service life prediction value to generate a final evaluation report containing the health degree level and life warning, and output the visualization result.

5. A charging curve based power battery state of health online evaluation system, characterized in that, The system comprises: A monitoring module for dynamically dividing the feature intervals of the constant current charging stage and the constant voltage charging stage by monitoring the voltage and current data in real time during the charging process; An extraction module for extracting time domain and frequency domain feature parameters in the feature intervals, including at least the voltage rise rate curve inflection point position and the current ripple spectrum distribution characteristics; wherein the voltage data sequence of the constant current charging stage is extracted from the dynamic feature interval division result, the first derivative curve of voltage with respect to time is calculated, the local extreme point of the derivative curve is identified as the inflection point by using the polynomial fitting algorithm to generate the voltage rise rate curve inflection point position coordinates; the current data sequence of the constant voltage charging stage is extracted from the dynamic feature interval division result, the current time series data is converted into a frequency domain signal by applying a windowed Fourier transform to generate a current frequency domain spectrum graph; the amplitude, frequency and phase information of the main ripple frequency components are extracted by analyzing the current frequency domain spectrum graph, the distribution entropy of the ripple energy and the harmonic distortion rate are calculated to generate a current ripple spectrum distribution characteristic vector; the voltage rise rate curve inflection point position coordinates and the current ripple spectrum distribution characteristic vector are integrated to construct a multi-dimensional parameter set containing time domain and frequency domain features to generate a comprehensive time-frequency feature parameter set; An evaluation module for inputting the time domain and frequency domain feature parameters into a health degree evaluation model optimized by transfer learning to generate a fusion health degree index containing at least a battery capacity attenuation coefficient and an internal resistance change vector; An output module is configured to output a final health degree evaluation result and a remaining service life prediction based on the fusion health degree index, in combination with battery historical cycle data and operating environment parameters, and through an adaptive weighting algorithm.

6. The system of claim 5, wherein, The monitoring module is specifically configured to: acquire voltage and current raw data in a charging process in real time through a sensor at a fixed sampling frequency, to generate raw voltage and current time series data sets; perform preprocessing on the raw voltage and current time series data sets, apply a sliding window filtering algorithm to eliminate high-frequency noise and abnormal fluctuations, and generate smoothed voltage and current data sequences; based on the smoothed voltage and current data sequences, calculate moving average values of voltage and current change rates, automatically detect a critical point of a transition from a constant-current charging phase to a constant-voltage charging phase through a preset threshold, and generate a phase transition critical point identifier; according to the phase transition critical point identifier, dynamically divide start and end time intervals of the constant-current charging phase and the constant-voltage charging phase, and record voltage and current data ranges corresponding to each phase, to generate a dynamic feature interval division result.

7. A storage medium, characterized by The storage medium has stored therein a computer program, wherein the computer program is configured to execute the method of any one of claims 1-4 when running.

8. An electronic device comprising a memory and a processor, characterized in that The memory has stored therein a computer program, and the processor is configured to execute the computer program to execute the method of any one of claims 1-4.

Images