Aplicação de machine learning para a Identificação, tratamento e antecipação de falhas em baterias de íons de lítio

Abstract

The search for electric vehicles has been growing strongly since the last decade due to governmental and climate pressures and the replacement of fossil fuels by renewable energies. Consequently, the adoption of lithium-ion batteries (LIB) for storage systems is accelerating. LIB is a dominant technology with several advantages compared to other forms of storage, such as higher energy density, higher capacity, and longer lifetime, among other characteristics. Despite this, LIB is sensitive to failure and must be operated safely and under controlled conditions. Otherwise, various abuses and faults may appear, such as external and internal short-circuit faults, overcharging, over-discharging, overheating, and thermal avalanche. This way, the Battery Management System (BMS) is used to equalize and control battery operation and prevent failures. However, the BMS is designed with inflexible programming and is only effective for new and balanced cells. Therefore, due to natural degradation and imbalance resulting from the battery construction, the BMS may fail and not operate correctly, which can cause failures and abuse. Thus, the application of Machine Learning (ML) techniques can be an excellent approach to assist in diagnosing and prognosis of failures in lithium-ion batteries. Therefore, this work presents a collection of four ML applications in batteries, which were published in high-impact international journals. The first work applied six ML models to diagnose abuse and failures in lithium-ion cells. This work used Sensitivity as a comparison metric between the models. The results indicated that the proposed models could be applied to identify faults in batteries with high accuracy. Then, the second article used the concept of anomaly detection to identify abnormalities in cells through thermal monitoring of a 3-cell pack. The work demonstrated that an anomaly is the first failure stage, and the models could anticipate overheating failures by up to 52 seconds. The third article applied the context of Random Forest and Isolation Forest to stop and handle failures in individual cells. The developed work can interrupt external short circuits, overcharging, over-discharging, and combining the three failures. As a result of the work, the applied system reduced capacity loss by up to 91%. Finally, the fourth work was developed to estimate overheating failures in a battery pack using a context composed of two steps. In the first step, the system calculates the future values of the main parameters collected from the batteries. Then, in the second stage, the system classifies failures or non-failures based on the estimated values of the first stage. The work was applied to two sets of data. Initially, the system can predict a failure about 35 to 40 seconds in advance. On the other hand, it could anticipate a failure by up to 70 seconds for the second dataset with external short circuit cases. In this way, this collection of articles can demonstrate the potential of applying ML to diagnose, treat, interrupt, and predict failures and abuses in lithium-ion cells and, in this way, help the BMS extend the battery life.