Search

CN-121978546-A - Battery BMS intelligent detection recognition system based on reinforcement learning

CN121978546ACN 121978546 ACN121978546 ACN 121978546ACN-121978546-A

Abstract

The invention discloses a battery BMS intelligent detection and identification system based on reinforcement learning, which belongs to the technical field of battery detection, wherein the detection limitation based on traditional surface parameters can be broken through by introducing characteristics of SEI film thickness, active lithium loss rate and the like for reflecting battery aging essence, the optimal detection performance of a model in a new state, a middle-aged state and an aged state can be kept through SOH dynamic adjustment of rewarding weights, multi-objective balance optimization is realized by comprehensively considering detection accuracy, missing report rate and false report rate, the accuracy and reliability of model detection and identification are effectively improved through introducing an aging state self-adaptive cutting threshold, and a complete concept drift compensation closed loop is formed through monitoring battery aging degree, triggering compensation mechanism, enhancing detection weight of micro-short circuit characteristics, dynamically adjusting abnormal detection threshold, verifying effect and closed loop optimization, and the detection performance and the utilization of computing resources in a deep aging stage are effectively improved only through triggering adjustment in a deep aging stage.

Inventors

  • ZHAO CHUN
  • LIANG TIANYUN

Assignees

  • 山东锋火动力通信科技有限公司

Dates

Publication Date
20260505
Application Date
20260409

Claims (10)

  1. 1. Battery BMS intellectual detection system identification system based on reinforcement study, its characterized in that includes: The self-adaptive detection model construction processing module takes the real-time cycle times, temperature sequences, SOH values, SEI film thickness and active lithium loss rate of the collected and processed battery as a core state space, and takes the detection accuracy rate, the missing report rate and the false report rate as a multi-target rewarding function, trains an intelligent body through an improved near-end strategy optimization algorithm, and enables model parameters to be dynamically and iteratively updated along with the aging process of the battery; the concept drift real-time compensation analysis processing module is used for acquiring a real-time health state and a health state sliding average value based on real-time monitoring data of the current battery, carrying out data analysis on the health state sliding average value, dynamically triggering a concept drift compensation mechanism, and realizing real-time compensation of the concept drift by dynamically adjusting micro-short circuit associated characteristic weights, updating abnormal detection thresholds, verifying real-time compensation effects and optimizing a closed loop.
  2. 2. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 1, wherein the state of health SOH is calculated in real time by a capacity fade method when performing the electrochemical feature on-line estimation; Based on a second-order RC equivalent circuit model, estimating polarization resistance in real time through extended Kalman filtering, and converting SEI film thickness through an empirical formula; The active lithium loss rate was calculated by coulombic efficiency method.
  3. 3. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 2, wherein a state space of an agent is defined as a high-dimensional electrochemical feature vector as an input state of the agent based on a result of the calculation, and a motion space is defined, a set of classification motions is fixed, a dynamic transition equation of each state element is constructed, and transition logic of each state element is packaged as a state transition function.
  4. 4. The reinforcement learning based battery BMS intelligent detection and identification system of claim 3, wherein the determination is based on a threshold of BMS real-time data, generating a true fault signature: Wherein, the method comprises the steps of, Is a true fault label; A terminal voltage measured value at the moment k; and the measured value of the charge and discharge current at the moment k.
  5. 5. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 4, wherein a multi-objective rewarding function R is defined, and three core indexes of detection accuracy, missing report rate and false report rate are integrated: Wherein, the method comprises the steps of, To detect the accuracy rewards, the value range is [0,1]; for the missing report punishment, the value range is [ -1,0]; For false alarm punishment, the value range is [ -1,0]; all are weight coefficients and are self-adaptively adjusted based on the health state SOH.
  6. 6. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 5, wherein the slave BMS system acquires real-time monitoring data of the current battery, calculates and acquires a real-time health state based on the acquired monitoring data, and dynamically triggers a concept drift compensation mechanism by using SOH sliding average values of 3 consecutive cycles as a judgment basis.
  7. 7. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 6, wherein the voltage spike frequency and the internal resistance mutation rate are extracted from the real-time monitoring data when the micro short circuit associated characteristic weight is dynamically adjusted; filling the voltage spike frequency weight and the internal resistance mutation rate weight into an original characteristic weight vector, up-regulating the voltage spike frequency weight and the internal resistance mutation rate weight, carrying out normalization processing on the regulated weight vector, and deploying the normalized weight vector into an online anomaly detection model to replace the original characteristic weight.
  8. 8. The battery BMS intelligent detection and recognition system based on reinforcement learning according to claim 7 is characterized in that when an abnormality detection threshold is updated, a sliding window with a window size W=100 cycles is adopted, micro-short circuit associated characteristic data of the previous 100 cycles at the current moment are obtained in real time, and the mean value and standard deviation of each micro-short circuit characteristic in the window are calculated; if the real-time characteristic value exceeds the threshold range, the real-time characteristic value is judged to be abnormal.
  9. 9. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 8, wherein when the real-time compensation effect verification and the closed loop optimization are performed, the abnormal detection is performed on the monitoring data of the last several cycles by adopting the adjusted characteristic weight and the abnormal detection threshold value, and the detection accuracy is calculated; If the detection accuracy is greater than or equal to the accuracy threshold, the current adjustment parameters are kept, otherwise, the characteristic distribution of the sliding window is recalculated, and the abnormal detection threshold is updated.
  10. 10. The reinforcement learning-based battery BMS intelligent detection and recognition system according to claim 1, wherein the network structure of the intelligent body adopts an Actor-Critic dual-network structure, and when training the intelligent body based on the improved PPO algorithm, the objective function of the PPO algorithm is improved, and the aging state self-adaptive cutting threshold is introduced : Wherein, the method comprises the steps of, Is a clipping strategy objective function; Is a desired operator; taking the minimum value of two items as a minimum value operator; to crop operators, input values are limited to ; As a dominance function; Is a policy network parameter; The threshold value is adaptively tailored for SOH, , The battery state of health at time k.

Description

Battery BMS intelligent detection recognition system based on reinforcement learning Technical Field The invention relates to the technical field of battery detection, in particular to a battery BMS intelligent detection and identification system based on reinforcement learning. Background The battery BMS intelligent detection and identification refers to a technical system for carrying out high-precision estimation, early fault early warning and automatic abnormal mode classification on the battery state in a Battery Management System (BMS) by utilizing Artificial Intelligence (AI), big data analysis and advanced algorithm, and exceeds the traditional BMS which only relies on a mode of carrying out passive protection by means of a fixed threshold value such as overhigh/overlow voltage and overhigh temperature, and is in turn driven by data. In the prior art, most intelligent detection models are trained on the data of a new state or a specific aging stage of a battery, however, the battery is a strong time-varying system, the electrochemical characteristics of the battery change in a nonlinear way along with the cycle times, the temperature history and the storage conditions, namely, the concept drift, for example, the accuracy of micro short circuit identification of the existing detection models is up to 99% in the initial use period (the previous 500 cycles) of the battery, but in the middle-aged period (after 1500 cycles) of the battery, the aging failure of the detection models caused by the concept drift exists due to the change of internal side reaction mechanisms, such as SEI film thickening and active lithium loss proportion, the original characteristic fingerprints deviate, and the systematic omission or false alarm rate of the model rises suddenly. Disclosure of Invention The invention aims to provide a battery BMS intelligent detection and identification system based on reinforcement learning, which is used for solving the technical problems in the background technology. The aim of the invention can be achieved by the following technical scheme: battery BMS intellectual detection system identification system based on reinforcement study includes: The self-adaptive detection model construction processing module takes the real-time cycle times, temperature sequences, SOH values, SEI film thickness and active lithium loss rate of the collected and processed battery as a core state space, and takes the detection accuracy rate, the missing report rate and the false report rate as a multi-target rewarding function, trains an intelligent body through an improved near-end strategy optimization algorithm, and enables model parameters to be dynamically and iteratively updated along with the aging process of the battery; the concept drift real-time compensation analysis processing module is used for acquiring a real-time health state and a health state sliding average value based on real-time monitoring data of the current battery, carrying out data analysis on the health state sliding average value, dynamically triggering a concept drift compensation mechanism, and realizing real-time compensation of the concept drift by dynamically adjusting micro-short circuit associated characteristic weights, updating abnormal detection thresholds, verifying real-time compensation effects and optimizing a closed loop. Further, the cycle times, the temperature sequence, the terminal voltage and the charge-discharge current of the battery pack are collected in real time through the BMS sensor network. Further, when the electrochemical characteristics are estimated on line, the health state SOH is calculated in real time through a capacity attenuation method; Based on a second-order RC equivalent circuit model, estimating polarization resistance in real time through extended Kalman filtering, and converting SEI film thickness through an empirical formula; The active lithium loss rate was calculated by coulombic efficiency method. Further, based on the result of calculation, defining a state space of the intelligent agent as a high-dimensional electrochemical feature vector as an input state of the intelligent agent, defining an action space, fixing a two-class action set, constructing a dynamic transfer equation of each state element, and packaging transfer logic of each state element as a state transfer function. Further, based on the threshold value of the BMS real-time data, judging is carried out, and a real fault label is generated: Wherein, the method comprises the steps of, Is a true fault label; A terminal voltage measured value at the moment k; and the measured value of the charge and discharge current at the moment k. Further, defining a multi-target rewarding function R, and fusing three core indexes of detection accuracy, false alarm rate and false alarm rate: Wherein, the method comprises the steps of, To detect the accuracy rewards, the value range is [0,1]; for the missing report punishment, the value ra