Battery Health Estimation Based on Multidomain Transfer Learning

Machine learning methods are expected to play a significant role in battery state of charge (SOH) estimation, leveraging their strengths in self-learning and nonlinear fitting. One of the key challenges in SOH estimation is the concept drift issue, which refers to changes in the data distribution between the training and test datasets. General machine learning methods assume that the training data shares similar characteristics with the test data. However, in SOH estimation tasks, differences in the environment and the characteristics of the battery itself can cause concept drift, which then impacts the model's effectiveness. As a result, many data-driven models that perform well in laboratory conditions struggle to be applied to other target batteries. This is a common and significant battery diagnosis technology issue, yet it remains unresolved. This article proposes a multidomain transfer Gaussian process regression (MTR-GPR) SOH estimation approach to address this issue. In this model, training data do not directly participate in the model's learning process. Instead, the MTR-GPR model extracts information from different datasets based on the distribution similarity. This method can fully use multisource battery ageing data while reducing the negative impact of distribution differences. Experimental results prove that MTR-GPR can make reliable SOH estimates with only 20% of target battery data. On the other hand, this method can provide the posterior probability distribution of the prediction results.

I. Introduction

Due to energy shortages and environmental concerns, the electric vehicle (EV) industry has rapidly developed in the last ten years. The energy storage system is critical to EV's competitiveness. Lithium-ion batteries have several advantages over other electrochemical energy storage technologies, including high energy density, high power density, low self-discharge rate, long battery life, and excellent transient response characteristics. As a result, lithium battery has been placed in high hopes in the automotive industry. The power battery pack of electric vehicles is composed of a large number of battery cells; therefore, monitoring the health status of the battery is crucial for devising battery balancing and charging/discharging strategies, which ultimately impacts the efficiency and safety of the battery pack. Battery state of health (SOH) is a measurement that reflects the batteries' health status. SOH is frequently defined as the ratio of the available capacity to the rated capacity [1]. Ampere-hour (Ah) capacity testing a widely used SOH estimation approach. But this approach is both time-consuming and energy-consuming [2]. Recently, a considerable amount of literature has emerged on SOH estimation. These methods can be broadly categorized into three main categories: the electrochemical model [3], [4], [5], the equivalent circuit model (ECM) [6], [7], [8], [9], [10], and the data-driven approach. In recent years, data-driven technologies, represented by machine learning, have advanced rapidly. In contrast to electrochemical methods and ECM models, machine learning methods can self-learn during deployment. They possess the excellent nonlinear fitting ability, accommodate multiple features, and dramatically lessen the manual burden of modeling. Moreover, the early identification and exploration of numerous nondestructive testing features [health indexes (HIs)], including the incremental capacity (IC) curve [11], [12], differential voltage [13], ageing cycles [14], sample entropy [15], the interval of equal discharging voltage difference [16], and many others, offer favorable conditions for data-driven modeling.