An Incremental Voltage Difference Based Technique for Online State of Health Estimation of Li-ion Batteries.

Accurate state of health (SOH) estimation of rechargeable batteries is important for the safe and reliable operation of electric vehicles (EVs), smart phones, and other battery operated systems. We propose a novel method for accurate SOH estimation which does not necessarily need full charging data. Using only partial charging data during normal usage, 10 derived voltage values ([Formula: see text]) are collected. The initial [Formula: see text] point is fixed and then for every 1.5% increase in the Coulomb counting, other points are selected. The difference between the [Formula: see text] values ([Formula: see text]) and the average temperature during the charging form the feature vector at different SOH levels. The training data set is prepared by extrapolating the charging voltage curves for the complete SOH range using initial 400 cycles of data. The trained artificial neural network (ANN) based on the feature vector and SOH values can be used in any battery management system (BMS) with a time complexity of only [Formula: see text]. Less than 1% mean absolute error (MAE) for the test cases has been achieved. The proposed method has a moderate training data requirement and does not need any knowledge of previous SOH, state of charge (SOC) vs. OCV relationship, and absolute SOC value.

Internal short circuit detection in Li-ion batteries using supervised machine learning.

With the proliferation of Li-ion batteries in smart phones, safety is the main concern and an on-line detection of battery faults is much wanting. Internal short circuit is a very critical issue that is often ascribed to be a cause of many accidents involving Li-ion batteries. A novel method that can detect the Internal short circuit in real time based on an advanced machine leaning approach, is proposed. Based on an equivalent electric circuit model, a set of features encompassing the physics of Li-ion cell with short circuit fault are identified and extracted from each charge-discharge cycle. The training feature set is generated with and without an external short-circuit resistance across the battery terminals. To emulate a real user scenario, internal short is induced by mechanical abuse. The testing feature set is generated from the battery charge-discharge data before and after the abuse. A random forest classifier is trained with the training feature set. The fault detection accuracy for the testing dataset is found to be more than 97%. The proposed algorithm does not interfere with the normal usage of the device, and the trained model can be implemented in any device for online fault detection.