Research On Lithium-ion Battery Remaining Useful Life Estimation With Relevance Vector Machine

Due to its excellent performance, lithium-ion battery has been widely used in various fields. However, the management has to be focused on lithium-ion batteries for safety and reliability in applications. In last years, as a key technology of Battery Management System (BMS), the Remaining Useful Life (RUL) estimation for the lithium-ion battery becomes one of the hotspots in electronic systems RUL estimation research.As a kind of complex electrochemical system, the performance of lithium-ion battery will degrade under continuous charging and discharging. In practical applications, it is difficult to measure Lithium-ion battery capacity and internal resistance while is power system is working. Thus, it is impossible to illustrate the degradation without available parameters. At the same time, environment conditions and dynamic load conditions could cause unstability of battery degradation. To solve those problems, more and more researches are focusing on lithium-ion battery RUL prediction methods. However, fewer new methods could be used in practical applications due to high complexity and low efficiency of algorithms. In view of this, the Relevance Vector Machine (RVM) is adopted as the basic approach to achieve the RUL prediction due to its uncertainty expression capability. The aim of our research is to solve the problems on prediction accuracy, efficiency and adaptability of RVM algorithm, and find a new method for degradation recognition of lithium-ion battery. Furthermore, to provide a novel technical solution to online RUL prediction methods, reconfigurable computing with the FPGA platform is used to realize the embedded computing of machine learning algorithm. The main contributions of this dissertation can be summarized as follows.(1) To overcome the low precision problem in long-term prediction with RVM algorithm for lithium-ion battery, this work proposes a dynamic grey RVM (DGM-RVM) algorithm. The DGM-RVM algorithm utilizes the predicted trend as the input data for the RVM forecasting, and dynamically updates the RVM model according to the prediction result. With this process, the long-term prediction precision can be improved. Comparied with the basic RVM and non-DGM-RVM model, the experimental results indicate that the proposed approach can obtain high RUL prediction precision, as well as the uncertainty repsentation for the RUL results.(2) To extend the RVM algorithm to the on-line forecasting application of lithium-ion battery RUL, this work proposes a RUL predciton method based on the incremental optimized RVM. To solve the low efficiency problem of traditional on-line incremental algorithm, we make use of high sparse relevance vectors of RVM and new training data samples as the on-line training data which reduces the size of the on-line training samples and increases theon-line prediction efficiency. Comparied to the basic RVM and retraining RVM, experimental results show that this method can increase the computing efficiency while keep a satisfying prediction accuracy. Moreover, it also proves a new approach for the embedded computing in actual application.(3) On the basis of above research, to further manage the difficulty in on-line measurement of capacity or internal resistance for RUL estimation, this work proposes a framework for health indicator (HI) construction based on the discharge voltage difference in equal time interval and the Box-Cox transformation. With Box-Cox transformation, the nonlinear transformation of the extracted discharge voltage difference in equal time interval is realized. Then the performance degradation of lithium-ion battery can be expressed with discharge voltage parameter which can be measured on-line..With this constructed HI, we can realize indirect RUL estimation. Experimental results verified the effectiveness and reasonableness of the proposed indirect RUL estimation with novel HI compared to the HI with battery capacity.(4) To overcome the limitation of computing resource and capability in actual embedded battery monitoring and management system, this work devleps a RUL predciton system for lithium-ion battery based on dynamic configurable RVM. The system takes the reconfigurable FPGA device as the hardware platform. Then the partition of RVM computing tasks is realized by the multi-objective optimization method. And at last the calculating efficiency problem under the limited resource of embedded system is solved by the usage of parallel computing, pipe-line and time-multiplexing of computational resources. Experimental results show that the proposed approach can achieve high resource utilization and high computing efficiency compared with the general processors with similar computing precision.