1. Introduction
Owing to their high specific energy and high operational voltage, lithium-ion batteries (LiB) have received great attention and are used in many applications [
1]. Unfortunately, LiB have a limited operational area mainly bound by two important parameters: voltage and temperature. As such, careful monitoring of a battery’s working temperature and voltage is necessary for its optimal and safe operation [
2]. If the battery’s voltage exceeds its limit, the battery may develop dendrites over time, which increases the battery’s internal resistance, resulting in a lower output voltage. Moreover, if the working temperature is substantially increased, the battery may release toxic gases or burst into flames [
3].
Another challenge presented by LiB technology is the accurate estimation of its available power or state of charge (SOC). SOC describes the amount of charge available in the battery at any given time during usage. SOC is often represented as a percentage value of available power vs. maximum power, or the available capacity vs. maximum capacity of the battery [
3]. The main problem in determining the SOC is the absence of instrumentation that can accurately measure SOC during the battery’s operation. This results in an estimation problem where the SOC must be estimated using indirect measurements such as the battery’s terminal voltage and current [
4].
Different techniques to estimate the LiB’s SOC have been proposed in the literature. Some techniques such as neural networks (NN) have been used with great success [
5]; however, NN make use of extensive data that must be collected beforehand and are computationally expensive compared to other solutions [
5]. Other techniques make use of electrochemical impedance spectroscopy (EIS) data, which requires special instrumentation to be installed in the system [
6].
One popular SOC estimation solution is the ampere-hour counting method. The ampere-hour counting method determines SOC based on current measurements and the remaining capacity of the battery [
3]. This method’s popularity relies on its simplistic approach. If the initial SOC is known, the previous SOC value can be subtracted or added based on the current profile. However, this method comes with many drawbacks. Its accuracy is highly dependent on the initial SOC value, correct current measurements, and accurate battery capacity readings [
3]. To ensure proper estimates of the SOC, this method must be frequently calibrated; some calibration techniques include voltage-based corrections using lookup tables [
3]. Another method was presented in [
7], where the authors were able to jointly estimate the SOC and temperature at the same time. The ability to track the temperature in conjunction with the SOC provides useful insights in terms of battery life management and operational safety.
Furthermore, Kalman filters (KFs) present other estimation techniques that, when combined with the ampere-hour counting method, have proven to be accurate at estimating SOC. KFs provide an accurate and computationally inexpensive solution, but require an accurate battery model for their successful implementation [
3]. A linear KF provides an optimal solution to the linear discrete estimation problem. However, due to the battery’s nonlinear nature, only modified versions of the KF have been used for SOC estimation. Some KF variations include the extended Kalman filter (EKF) and the unscented Kalman filter (UKF), among others [
8,
9]. Between these two strategies, the EKF is known to introduce instability in the estimation process due to the linearization process embedded in the algorithm [
10]. On the other hand, the UKF has proven to be a more robust strategy [
11,
12]. Another robust strategy, known as robust fixed-lag smoothing, attempts to overcome model uncertainties or mismatch by utilizing the least favorable model over a finite time horizon [
13]. This method is characterized by a dynamic game between two players: one player selects the least favorable model in a prescribed ambiguity set, while the other player selects the fixed-lag smoother, minimizing the smoothing error with respect to the least favorable model. Efficient implementation of the robust fixed-lag smoother may reduce computational burdens and avoid numerical instabilities, which may be helpful for battery applications.
Electrochemical and equivalent circuit models (ECMs) are among the most popular models for batteries. Electrochemical models are based on the underlying physics of the battery using 10–14 partial differential equations, resulting in highly complex and computationally demanding models, but providing high-accuracy information about the battery’s state. These types of models are often used for laboratory and battery development research [
14,
15,
16,
17].
On the other hand, ECMs represent the battery as an electric circuit using voltage sources, resistors, and capacitors. These types of models require low computational power and have low complexity, but are less accurate and yield little information about the battery [
18]. Nevertheless, these traits allow for their implementation online.
Some ECMs studied include Rint model, Thevenin model, PNGV model, and Dual Polarity (DP) model [
19]. These models can be differentiated by the number of Resistor–Capacitor (RC) branches in the circuit. Adding more RC branches allows the capture of higher-order nonlinearities, resulting in a more accurate model [
19]. However, adding more RC branches increases the complexity and computational time of the algorithms.
In summary, a battery monitoring system (BMS) should be implemented to ensure safe operation of LiB. The BMS’s main function is the accurate estimation of the battery’s current SOC and operating temperature. In addition, the BMS can also track other parameters such as the battery’s state of health (SOH); SOH is a measurement of the current health of the battery and is sometimes calculated based on its available maximum capacity [
20]. As the battery is subject to aggressive current profiles, excessive cycling, or regular use, its maximum battery capacity degrades over time [
20]. Moreover, accurate estimation of the battery’s SOH can significantly increase the accuracy of the ampere-hour counting method, since it is dependent on the battery’s capacity [
20]. Lastly, accurate tracking of the battery’s SOH allows for an effective planned retirement of the battery, which ensures that the system continues to operate optimally.
A battery is referred to as due for retirement once its SOH is at 80%, or in other words, when the battery’s maximum available capacity is at 80% or less of its designed capacity [
20]. Battery retirement can be presented as a fault diagnosis problem, where a SOH value of 80% or lower signals a fault in the battery [
21]. A recent paper presented a degradation empirical model-free battery end-of-life prediction framework [
22]. This method utilized the KF and Gaussian process regression. It is important to note that the SOH should be rapidly tracked and updated for improved performance and reliability. The authors in [
23] introduced a fast capacity estimation method as well as a fast accelerated degradation fault diagnosis strategy for SOH estimation. This article offers insights into the importance of tracking micro-health parameters in batteries, which directly correspond to the overall SOH of the battery or set of batteries.
The multiple model (MM) strategy has been used to detect faults in batteries [
24]. In the MM strategy, several models representing different behaviors of the system are generated to make the algorithm resilient against uncertainty [
25]. Moreover, [
25] presented an interacting multiple model (IMM) strategy where the IMM was combined with the EKF to accurately estimate the SOC of a LiB. The IMM was given allowed two different variations of noise in the battery model to account for the different degrees of parameter shift during the estimation process. Lastly, in [
26], a multiple model adaptive estimation (MMAE) technique was used for fault diagnosis. The proposed strategy made use of EIS data and EKFs to generate residual signals that were fed into an MMAE block to detect a fault in the battery.
This paper focuses on the implementation of a MM strategy, i.e., the IMM strategy, to estimate the battery’s capacity degradation while accurately estimating the SOC of a battery under cycling conditions [
27]. This is a unique contribution to the field of battery monitoring, particularly when utilizing the relatively new sliding innovation filter (SIF). The experimental dataset was partitioned into sections representing a 100% SOH, 75% SOH, 50% SOH, 25% SOH, and 0% SOH, where each section can be identified as a mode to be used within the IMM algorithm. The motivation behind this partition is that the IMM would yield the best matching mode, thus identifying the current SOH of the battery.
The main contribution of this paper is the development of the SIF in conjunction with the IMM (the so-called SIF-IMM) for determining the SOC and SOH of a battery. The IMM algorithm is used for SOH estimation by partitioning the experimental dataset into several SOH modes. This strategy has not been presented in the literature. In addition to introducing this method, the paper compares the performances of SIF-IMM and KF-IMM in estimating SOH.
The remainder of the paper is structured as follows:
Section 2 presents the battery and parameter models.
Section 3 details the experimental data and estimation algorithms.
Section 4 covers the artificial measurements.
Section 5 describes the model parameter identification results.
Section 6 presents the experimental setup and details the results of the proposed strategy.
Section 7 presents the concluding arguments of the work.