A flexible RUL prediction method based on poly-cell LSTM with applications to lithium battery data
Introduction
LIBs are the primary choice in electrical storage and conversion due to their advantages of high energy density, low pollution, and low self-consumption ([1]). Battery degradation is a major concern in long-term and reliable applications, where the failure of the battery system can cause severe accidents and thus requires a long cycle life under different operation conditions. That is a reliable and accurate prediction of RUL for LIBs is necessary to take effective maintenance measures and avoid catastrophic consequences. The LIBs have complex degradation paths leading to various aging patterns and failure modes ([2]). Anseán et al. [3] discussed these degradation patterns and the functional degradation processes of batteries. However, it is very difficult to model these degradation patterns and degradation processes simultaneously.
The current RUL prediction methods for LIBs are involved with three aspects, which are physical model-based, data-driven, and hybrid approaches. Model-based prognostics focus on the underlying physical failure mechanisms. For example, Downey et al. [4] presented a physics-based (or mechanistic) approach to LIB prognostics, which enables online prediction of RUL with consideration of multiple concurrent degradation mechanisms.. The physical-based methods usually attain pretty accurate estimation results if the degradation processes are well designed ([5]).
Note that lithium batteries have complex failure mechanisms because of temporary capacity regeneration or noise interference problems. Compared with the physical model-based prediction method, the data-driven method does not need any prior knowledge about the degradation mechanisms. For instance, particle filter ([6]), Kalman filter ([7]), and other filtering algorithms are widely used to track the recession information of the cell to obtain the optimal parameters. The stochastic processes, i.e., Wiener process ([8], [9], [10]), and Gaussians process ([11], [12], [13]) are also used to capture the degradation mechanism. Hong et al. [14] proposed an iterative model of the generalized Cauchy process with the long-range dependence characteristics for the RUL prediction of LIBs. Zhang et al. [15] proposed a parameter estimation method of Wiener degradation model by minimizing a probability divergence. The stochastic process has a better characterization to assess the lithium-ion battery health degradation. However, it is still subject to challenge to get accurate results considering the influence of random current, time-varying temperatures, and self-discharge characteristics ([16]), also see ([17])
Recently, deep learning and machine learning techniques have attracted much attention in various fields, and numerous results indicate the successful application of RUL prediction. Kim and Liu [18] proposed a new Bayesian deep learning framework that provides interval estimates of RUL. Vatani et al. [19] combined the ICA method and the support vector regression to estimate the RUL for lithium-ion batteries. Zhang et al. [20] proposed a novel online synthesis method based on the fusion of partial incremental capacity and ANN to predict RUL under constant current discharge. Furthermore, Huang et al. [21] proposed a novel deep convolutional neural network-bootstrap-based integrated prognostic approach for RUL prediction. Remadna et al. [22] employed a deep convolutional variational autoencoder framework with an attention mechanism to predict the RUL. Tang and Yuan [23] proposed an improved Res2Net-bidirectional gated recurrent unit-fully connected network to predict the remaining service life of lithium batteries.
Note that the degradation paths of LIBs have time dependence and local fluctuations (see [24], [25], for example). This leads to sequence models with time memory, i.e., RNN, LSTM, and TCN, are frequently used to learn long-term trends and time correlation information. Particularly, the LSTM has become a popular tool for RUL prediction due to its flexibility and applicability ([26]). Ren et al. [27] discussed the RUL prediction method based on improved DCNN and LSTM, and their experiments on a lithium-ion battery dataset demonstrated the effectiveness of this method. Cui et al. [28] used the TPA-LSTM model to improve lithium battery prediction by adding an attention mechanism to the feature dimension. Zhou et al. [29] employed a TCN and health characteristics for SOH estimation. Wang et al. [30] proposed a bidirectional long short-term memory with attention mechanism model to predict online RUL by continuously updating the model parameters.
Motivated by the above works and inspired by the multicellular LSTM ([31]), this paper proposes the PCLSTM to predict the RUL. Two pre-training based methods are proposed to improve the prediction performance. Specifically, the main contributions of this paper can be summarized as follows:
- •
The PCLSTM uses degradation information more effectively: This is because the PCLSTM contains a hierarchical division unit and the poly-cell unit, the information of the degradation process can be used more efficiently by integrating the global degradation trend and local degradation trend simultaneously. This leads to the PCLSTM can improve RUL prediction performance and provide a more effective reference for BMSs.
- •
The PCLSTM eases the computational burden: Compared with the multicellular LSTM, the PCLSTM only retains long-term trend, medium-term trend, and short-term trend. This reduces the model complexity and eases the computational burden.
- •
The proposed PCLSTM is also equipped with flexibility and robustness: Compared with the classical LSTM, the data importance level is considered in PCLSTM. This yields the flexibility of PCLSTM. The pre-trained procedure is employed, which leads to the robustness of RUL prediction.
- •
Finally, the effectiveness of the proposed method is illustrated by analyzing the lithium battery data. The uncertainty measurement of the RUL prediction is discussed by using the dropout technique in the neural network.
Fig. 1 depicts the structure of this paper. First, we select five valid features from the original data, use these features to pre-train the model and select the set of models that performed well, and then employ a tiny portion of the data from the test set to fine-tune some of the model’s parameters, and generate predictions. We validate the feature extraction capabilities of the PCLSTM in the experimental analysis by analyzing the intermediate output of the PLSTM cell. Finally, the model’s predictive ability under uncertainty is tested using a Monte Carlo sampling method based on the Dropout mechanism.
The organization of this paper can be briefly listed as follows. In Section 2, we propose the hierarchical division unit and poly-cell unit with different update rules. Based on the division unit and poly-cell unit, in Section 3 we construct the PCLSTM and discuss its application in RUL prediction. As applications of the proposed method, PCLSTM is employed to predict the RUL of lithium battery in Section 4 to demonstrate the effectiveness of the proposed methods. The approximation probability model of the RUL prediction is discussed to measure the model’s uncertainty. Finally, some concluding remarks are provided in Section 5.
Section snippets
Hierarchical division unit and update rules
As mentioned in Section 1, the hierarchical division unit and the poly-cell unit are included to improve the accuracy and robustness of the PCLSTM model. We first focus on these two units in this section before discussing the PCLSTM.
The poly-cell structure
The PCLSTM is constructed using a hierarchical division unit and poly-cell unit. The PCLSTM can be regarded as a new variant of LSTM, and its hidden layer state is shown in Fig. 2. From Fig. 2 and the update Eqs. (2.3)–(2.6), the iterative Eq. of PCLSTM can be given as follows where denotes the
The lithium battery data analysis
This section focuses on the illustration of proposed RUL prediction method based on the lithium battery data. The dataset can be found at https://data.matr.io/1/, which contains battery cycle aging test data. The battery has a nominal capacity of 1.1 Ah and a nominal voltage of 3.3 V. The data were obtained by allowing the battery to continuously charge–discharge cycle until the end of life (EOL) of the battery arrived. The EOL of the battery is defined as 80% of the nominal capacity, that
Conclusion remarks
This paper proposes an improved LSTM (PCLSTM) and a deep learning framework based on PCLSTM to predict the RUL of products’ degradation. The model adds a hierarchical division unit and poly-cell unit based on the classical LSTM. It can update different modes according to different levels of data, to integrate the global degradation trend and the local degradation trend. In applications, the proposed method is used to predict the RUL of the lithium battery capacity. The uncertainty of the
CRediT authorship contribution statement
Jiaolong Wang: Writing – original draft, Software, Methodology. Fode Zhang: Writing – review & editing, Writing – original draft, Supervision, Funding acquisition. Jianchuan Zhang: Writing – original draft, Visualization, Validation. Wen Liu: Methodology, Investigation, Formal analysis. Kuang Zhou: Validation, Software, Data curation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work is supported by the National Natural Science Foundation of China (No. 12071372, 11528102, 11571282), the Fundamental Research Funds for the Central Universities (No. JBK1901053, JBK1806002, and JBK140507) of China.
References (36)
- et al.
Machine learning for reliability engineering and safety applications: Review of current status and future opportunities
Reliab Eng Syst Saf
(2021) - et al.
Physics-based prognostics of lithium-ion battery using non-linear least squares with dynamic bounds
Reliab Eng Syst Saf
(2019) - et al.
Remaining useful life estimation of lithium-ion battery based on interacting multiple model particle filter and support vector regression
Reliab Eng Syst Saf
(2021) - et al.
Degradation data analysis and remaining useful life estimation: A review on Wiener-process-based methods
European J Oper Res
(2018) - et al.
Mis-specifcation analysis of Wiener degradation models by using f-divergence with outliers
Reliab Eng Syst Saf
(2020) - et al.
An iterative model of the generalized Cauchy process for predicting the remaining useful life of lithium-ion batteries
Measurement
(2022) - et al.
Remaining useful life prediction and predictive maintenance strategies for multi-state manufacturing systems considering functional dependence
Reliab Eng Syst Saf
(2021) - et al.
Synchronous estimation of state of health and remaining useful lifetime for lithium-ion battery using the incremental capacity and artificial neural networks
J Energy Storage
(2019) - et al.
A novel deep convolutional neural network-bootstrap integrated method for RUL prediction of rolling bearing
J Manuf Syst
(2021) - et al.
A hybrid approach based on decomposition algorithm and neural network for remaining useful life prediction of lithium-ion battery
Reliab Eng Syst Saf
(2022)
Accelerated degradation tests with inspection effects
European J Oper Res
Modeling and estimation of system reliability under dynamic operating environments and lifetime ordering constraints
Reliab Eng Syst Saf
Online remaining useful life prediction of lithium-ion batteries using bidirectional long short-term memory with attention mechanism
Energy
Multicellular LSTM-based deep learning model for aero-engine remaining useful life prediction
Reliab Eng Syst Saf
Aircraft engine remaining useful life estimation via a double attention-based data-driven architecture
Reliab Eng Syst Saf
Electric vehicle charging scheme for a park-and-charge system considering battery degradation costs
IEEE Trans Intell Veh
Lithium-ion battery degradation indicators via incremental capacity analysis
IEEE Trans Ind Appl
Cited by (11)
Remaining useful life estimation of bearing via temporal convolutional networks enhanced by a gated convolutional unit
2024, Engineering Applications of Artificial IntelligenceJoint electric vehicle routing and battery health management integrating an explicit state of charge model
2024, Computers and Industrial EngineeringA feature disentanglement and unsupervised domain adaptation of remaining useful life prediction for sensor-equipped machines
2024, Reliability Engineering and System SafetyA light-weight feature extractor for lithium-ion battery health prognosis
2023, Reliability Engineering and System SafetyPrognosis for stochastic degrading systems with massive data: A data-model interactive perspective
2023, Reliability Engineering and System SafetyMultifeature-based online remaining useful life prediction of lithium-ion batteries in stages using cascaded data-driven algorithm
2024, Quality and Reliability Engineering International