Search

CN-122017650-A - Deep learning-based battery health state and residual life collaborative prediction method

CN122017650ACN 122017650 ACN122017650 ACN 122017650ACN-122017650-A

Abstract

The invention discloses a deep learning-based battery health state and residual life collaborative prediction method, and relates to the technical field of battery health prediction. The method comprises the steps of firstly determining a current abrupt change time sequence segment, carrying out second-order differential calculation on a corresponding voltage transient response segment to generate a voltage recovery time sequence track, realizing time domain decoupling of reversible polarization attenuation and irreversible aging, then determining the irreversible aging time sequence segment through time sequence coincidence ratio, combining voltage residual variation consistency and voltage trend consistency to analyze and accurately identify a battery degradation inflection point, avoiding the problems that reversible polarization fluctuation is misjudged as irreversible aging and degradation inflection point identification offset, finally extracting voltage residual related parameters to calculate and correct SOH and RUL predicted values based on aging characteristics after decoupling, ensuring that the SOH and RUL predicted values realize collaborative prediction based on the same set of irreversible aging rules, and solving the problem of low collaborative prediction precision caused by incapacity of decoupling, misjudgment of SOH and inflection point identification error in the existing method.

Inventors

  • DING YUANYI

Assignees

  • 江苏本致网络科技有限公司

Dates

Publication Date
20260512
Application Date
20260303

Claims (10)

  1. 1. The battery state of health and remaining life collaborative prediction method based on deep learning is characterized by comprising the following steps: Performing time sequence first-order differential calculation and current change rate calculation on the current time sequence segment to determine a current abrupt change time sequence segment; performing time sequence second order differential calculation on a voltage transient response segment corresponding to the current mutation time sequence segment to generate a voltage recovery time sequence track; Calculating the time sequence coincidence ratio of the start point and the end point of the track according to the voltage recovery time sequence track, and determining an irreversible aging time sequence segment; performing voltage residual variation consistency and voltage trend consistency analysis on the irreversible aging time sequence segment, and determining a battery degradation inflection point; Performing time sequence feature coding processing on the irreversible aging time sequence fragments after the battery degradation inflection point, extracting the instantaneous variation of the voltage residual error and the cumulative amount of the voltage residual error, and calculating and correcting an SOH predicted value; And calculating the irreversible aging rate after the battery degradation inflection point according to the voltage residual error, determining the RUL predicted value according to the irreversible aging rate, and correcting.
  2. 2. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 1, wherein determining a current abrupt timing segment is: Calculating the difference value of the current sampling values at adjacent sampling moments in the current time sequence segment, and taking the ratio of the difference value to the interval of the corresponding sampling moments as a first-order differential value of the current to form a first-order differential sequence of the current; Traversing the first-order differential sequence of the current by a preset sliding window, averaging all the first-order differential values of the current in each sliding window to obtain the corresponding current change rate of each sliding window, and forming a current change rate sequence; if the absolute values of all the first-order differential values of the currents in a certain sliding window are larger than a non-zero threshold value, and the variances of the corresponding current change rates and the current change rates corresponding to the adjacent front sliding window and the adjacent rear sliding window are larger than a judging threshold value, judging that the sliding window is an effective abrupt change window; The continuous N effective mutation windows are marked as current mutation time sequence fragments.
  3. 3. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 1, wherein the process of generating a voltage recovery timing trajectory is: Calculating the difference value of the voltage sampling values at adjacent sampling moments in the voltage transient response fragment, and taking the ratio of the difference value to the interval of the corresponding sampling moments as a voltage first-order differential value to form a voltage first-order differential sequence; performing difference calculation on voltage first-order differential values of adjacent sampling moments in a voltage first-order differential sequence, and taking the ratio of the difference value to the interval of the corresponding sampling moments as a voltage second-order differential value to form a voltage second-order differential sequence; positioning a voltage transient jump inflection point and a polarization relaxation termination point corresponding to the voltage transient response fragment based on the voltage second-order differential sequence, and intercepting a relaxation time sequence interval from the voltage transient jump inflection point to the polarization relaxation termination point; and combining the value of the relaxation time sequence interval with the reference voltage, and performing time sequence feature reduction and baseline alignment to generate a voltage recovery time sequence track.
  4. 4. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 3, wherein the positioning of voltage transient jump inflection points and polarization relaxation termination points corresponding to voltage transient response segments based on a voltage second order differential sequence comprises: taking a sampling time aligned with a time axis of the current abrupt change time sequence segment as a time sequence reference, and traversing a voltage second-order differential sequence by adopting a preset inflection point detection sliding window; Calculating the maximum value of the absolute value of the voltage second-order differential value in each inflection point detection sliding window; When the maximum value of the absolute value exceeds a preset jump judgment threshold value and the time difference between the sampling time corresponding to the maximum value of the absolute value and the starting time of the current abrupt change time sequence segment is in a preset transient response time window, calibrating the sampling time as a voltage transient jump inflection point; traversing a voltage second-order differential sequence in a time sequence interval after a voltage transient jump inflection point by adopting a preset relaxation termination detection sliding window; When the absolute values of all the voltage second-order differential values in the continuous preset number of relaxation termination detection sliding windows are smaller than a preset relaxation termination threshold value and the variance of the voltage second-order differential values in the corresponding relaxation termination detection sliding windows is smaller than a preset stability judgment threshold value, locking the continuous preset number of relaxation termination detection sliding windows; And calibrating the initial sampling time of the first relaxation end detection sliding window of the continuous preset number of relaxation end detection sliding windows as a polarization relaxation end point.
  5. 5. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 4, wherein the performing the time series feature reduction and baseline alignment, generating a voltage recovery time series trace, comprises: Intercepting a numerical value corresponding to a relaxation time sequence interval in the voltage second-order differential sequence to obtain a relaxation time interval second-order differential sequence; Carrying out point-by-point first-order numerical integration on the second-order differential sequence of the relaxation interval according to the step length of the original sampling time sequence to obtain a first-order differential reduction sequence of the voltage of the relaxation interval; Performing point-by-point second-order numerical integration on the first-order differential reduction sequence of the voltage in the relaxation interval according to the step length of the original sampling time sequence to obtain a voltage amplitude reduction sequence in the relaxation interval; And taking the reference voltage as an alignment reference, performing baseline translation correction on the voltage amplitude reduction sequence of the relaxation interval, and generating a voltage recovery time sequence track completely synchronous with the original sampling time sequence.
  6. 6. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 1, wherein the process of calculating the time sequence coincidence of the start point and the end point of the track according to the voltage recovery time sequence track and determining the irreversible aging time sequence segment is as follows: respectively calculating the absolute values of the differences between the voltages corresponding to the start point and the end point of the track and the reference voltage to obtain a start point deviation value and an end point deviation value; Calculating the time sequence coincidence degree of the track starting point and the track end point based on the starting point deviation value and the end point deviation value: , wherein, In order to achieve the time sequence coincidence degree, As the value of the end point deviation, As the value of the deviation of the starting point, Is that And (3) with Is used to determine the absolute value of the difference of (c), Is the reference voltage; if the time sequence coincidence ratio is lower than the aging judgment threshold value, judging that the time sequence interval corresponding to the voltage recovery time sequence track is an irreversible aging candidate interval; And calibrating the time sequence range covered by the continuously distributed irreversible aging candidate intervals as an irreversible aging time sequence segment.
  7. 7. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 1, wherein the process of determining battery degradation inflection points by performing voltage residual variation consistency and voltage trend consistency analysis on irreversible aging time sequence segments is as follows: Intercepting voltage residual errors of each voltage recovery time sequence track in the irreversible aging time sequence segment, and generating an irreversible aging residual error time sequence according to time sequence; Traversing the irreversible aging residual error time sequence by adopting a fixed-step non-overlapping sliding window to obtain a plurality of continuous time sequence analysis windows, and respectively calculating the voltage residual error change consistency and the voltage trend consistency of each time sequence analysis window; respectively calculating a consistency statistics median value and a consistency statistics median value based on the voltage residual variation consistency and the voltage trend consistency of the full-scale time sequence analysis window; Screening out a time sequence analysis window, wherein the consistency of the voltage residual variation is not lower than the median value in consistency statistics, and the consistency of the voltage trend is not lower than the median value in consistency statistics, and marking the time sequence analysis window as an effective trend window; Sequentially calculating the aging change rate of each effective trend window according to the time sequence; and when the absolute value of the difference value of the aging change rates of the two adjacent effective trend windows exceeds 2 times of standard deviation of the aging change rates of the full-quantity effective trend windows, calibrating the junction moment of the two adjacent effective trend windows as a battery degradation inflection point.
  8. 8. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 1, wherein the process of calculating and correcting SOH predicted values is: Extracting voltage residual errors corresponding to all voltage recovery time sequence tracks in the irreversible aging time sequence segment after the battery degradation inflection point, performing equal time step time sequence feature coding, inputting the coded voltage residual errors into a light-weight deep learning time sequence model, and simultaneously extracting the instantaneous variation of the voltage residual errors and the accumulated quantity of the voltage residual errors after the inflection point; Calculating the ratio of the voltage residual error accumulation amount after the inflection point to the theoretical voltage residual error total increment from the degradation inflection point to the complete failure state of the battery to obtain the residual error aging duty ratio; subtracting the product of the residual aging duty ratio and the adaptive coefficient by 100% to obtain an SOH predicted value, and carrying out linear weighted summation on the SOH predicted value and the SOH model predicted value output by the light-weight deep learning time sequence model to obtain an initial SOH predicted value; And multiplying the initial SOH predicted value by a difference value between 1 and the fluctuation coefficient to obtain a final SOH predicted value.
  9. 9. The deep learning-based battery state of health and remaining life collaborative prediction method according to claim 1, wherein determining and correcting the RUL predicted value comprises: extracting voltage residual errors of corresponding time sequences of irreversible aging time sequence fragments after battery degradation inflection points, and generating voltage residual error time sequence sequences after the inflection points; Dividing the difference value of the termination time voltage residual and the starting time voltage residual in the voltage residual time sequence after inflection point by the total time sequence length corresponding to the voltage residual time sequence after inflection point to obtain an irreversible aging rate; Dividing the difference value between the battery voltage residual error failure threshold and the voltage residual error at the current moment by the irreversible aging rate to obtain an initial RUL predicted value; And multiplying the initial RUL predicted value by a difference value between 1 and the fluctuation coefficient to obtain a final RUL predicted value.
  10. 10. The deep learning based battery state of health and remaining life collaborative prediction method according to any of claims 8-9, wherein the ripple factor is a ratio of standard deviation to average of instantaneous variation of voltage residuals.

Description

Deep learning-based battery health state and residual life collaborative prediction method Technical Field The invention relates to the technical field of battery health prediction, in particular to a battery health state and residual life collaborative prediction method based on deep learning. Background In the actual scene of the collaborative prediction of the SOH and the RUL of the battery, the battery voltage response signal is an ohmic voltage drop transient, reversible polarization relaxation and irreversible aging steady-state superposition signal, the sampling noise or the fluctuation interference of working conditions are obvious, the time domain boundary of the polarization relaxation and the aging steady-state is fuzzy, and the following defects are exposed in the prior art: firstly, when a battery voltage time sequence signal is processed by the existing collaborative prediction method, original voltage fragments are directly intercepted for feature extraction, so that time domain decoupling is not carried out on a reversible polarization relaxation process and an irreversible aging steady-state component in the voltage signal, start and stop boundaries of the polarization relaxation are not accurately positioned through differential operation, interference signals such as ohmic voltage drop transient and sampling noise cannot be stripped, and polarization attenuation and aging signals in the extracted features are mutually coupled. Then, when the battery degradation inflection point is identified and SOH and RUL are calculated in the existing scheme, trend judgment is not carried out based on the decoupled pure aging characteristics, but the aging rate and residual ratio are directly calculated based on the coupling signals, instantaneous fluctuation of reversible polarization is easily misjudged as the degradation trend of irreversible aging, SOH estimated values deviate from a real health state, and meanwhile time offset is identified due to degradation inflection points caused by polarization interference. Therefore, a method for collaborative prediction of battery health status and residual life by decoupling reversible polarization and irreversible aging signals through second-order differential operation, precisely positioning polarization relaxation boundary and battery degradation inflection point, and calculating SOH and RUL based on pure aging characteristics is needed to solve the above technical problems and improve precision and reliability of collaborative prediction of SOH and RUL. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a battery health state and residual life collaborative prediction method based on deep learning, which solves the problems that the prior collaborative prediction method cannot decouple reversible polarization attenuation and irreversible aging, is easy to misjudge SOH and causes wrong identification of a degradation inflection point, and finally causes reduction of collaborative prediction precision of SOH and RUL. The method for collaborative prediction of the battery state of health and the residual life based on deep learning comprises the following steps of performing time sequence first-order differential calculation and current change rate calculation on a current time sequence segment, and determining the current abrupt change time sequence segment. And performing time sequence second order differential calculation on the voltage transient response segment corresponding to the current abrupt change time sequence segment to generate a voltage recovery time sequence track. And calculating the time sequence coincidence degree of the start point and the end point of the track according to the voltage recovery time sequence track, and determining the irreversible aging time sequence segment. And (3) carrying out voltage residual variation consistency and voltage trend consistency analysis on the irreversible aging time sequence segment, and determining a battery degradation inflection point. And (3) carrying out time sequence feature coding processing on the irreversible aging time sequence fragments after the battery degradation inflection point, and calculating and correcting an SOH predicted value after extracting the instantaneous variation of the voltage residual and the cumulative quantity of the voltage residual. And calculating the irreversible aging rate after the battery degradation inflection point according to the voltage residual error, determining the RUL predicted value according to the irreversible aging rate, and correcting. Compared with the prior art, the method has the advantages that firstly, the current abrupt change time sequence segment is determined, the second-order differential calculation is carried out on the corresponding voltage transient response segment to generate the voltage recovery time sequence track, the time domain decoupling of reversible polarization attenuation and irrev