Battery Life Prediction Using LSTM-Based Model for Tire Pressure Monitoring System

Takato Goto, Koichi Ito

Abstract: The miniaturization of sensors and batteries, along with the development of cloud systems, have made Internet-of-Things (IoT) technologies more accessible. Simultaneously, recent developments in machine learning have brought about a variety of predictive technologies that benefit our lives. Several life-prediction models for rechargeable batteries have recently been proposed and studied to replace coin-cell batteries, which are often replaced to avoid data loss, as primary energy supplies. This paper presents a battery life prediction method for a tire pressure monitoring system (TPMS) that uses coin batteries in severe environments, such as those characterized by high temperatures and pressures, where data can only be measured at fixed points. Although the measured data were irregular, noisy, and sparse, the sample size exceeded several millions. We approached limited data as time-series data, extracted features from the data, and predicted battery life from the features using a recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU). Predictive performance was evaluated using accuracy, recall, F1-value, and a confusion matrix to experimentally demonstrate the effectiveness of our method.