Xiumei Guan

Prognostics of remaining useful life for lithium-ion batteries based on a feature vector selection and relevance vector machine approach

Although the capacity is often used as a criterion to evaluate the state of health (SOH) of a lithium-ion battery, it cannot be measured on-line. Besides, degradation modeling only depending on historical capacity data will cause a large prediction error in the long term. Actually, some parameters can be monitored, such as the duration of equal discharging voltage difference, the interval of equal charging voltage difference at different experiment cycles, also exhibit a degradation trend. In order to make more accurate SOH and remaining useful life (RUL) estimations of an on-line operating battery, in this paper, the relevance vector machine (RVM) is applied to quantify the relationship between those monitoring parameters and capacity data. Based on the deduced model, the capacity could be extrapolated with the corresponding monitoring parameters. Moreover, feature vector selection (FVS) is used to remove redundant points in the input data. It improves the sparsity of relevance vectors (RVs) and decreases the memory-consuming. In the end, Battery degradation datasets from NASA demonstrated the approach has good RUL prediction accuracy, higher sparsity compared to RVM.

An improved fault diagnosis approach based on support vector machine

Fault diagnosis is extremely important for guaranteeing safe and reliable operation of modern industrial process. As an active branch of fault diagnosis, data-driven methods attract more and more attention in recent years, because they solely depend on information collected in historical databases. The support vector machine (SVM), aims at minimizing the structural risk, exhibits superior generalization ability, and succeeds in the nonlinear classification problem. This paper proposes an improved SVM based fault diagnosis framework, which consists of two primary components: (1) feature extraction; (2) classification. More specifically, multi-scale principal component analysis (MSPCA) is performed to achieve multi-scale analysis and dimension reduction. Classification combines SVM classifier with parameters optimization method, which further encompasses grid search (GS) and particle swarm optimization (PSO). To demonstrate the accuracy and efficiency of our approach, we perform experiments on the classical Tennessee Eastman (TE) process.

A Comparsion Of State Estimation Algorithms For Hybrid Systems.

The dynamic nature of hybrid systems involves discrete switching behavior between several operating modes and continuous plant dynamics governed by continuous models in each mode. State estimation is a important class of approaches for online monitoring and diagnosis of hybrid systems, which relies on the estimation of unknown variables using a filtering approach. Focused hybrid estimation methods concentrate on most likely system evolution trajectories based on probabilistic and best-first enumeration. On the other hand, switched Dynamic Bayesian Networks-based particle filter methods track the continuous behavior in individual modes of operation, and switch modes when transition conditions are met. In this paper, we study and compare these two algorithms. The theoretical analysis and experimental results show the advantages and disadvantages of both approaches.

A new hybrid hierarchy model description method

As requirements in diagnosis for hybrid systems increase, more and more researchers concentrate on hybrid models. However, common visual modeling methods such as GME (General Modeling Environment) lacks of flexibility. There is no appropriate modeling method for hybrid systems in cases containing plenty of complex components. This paper proposes a new hybrid hierarchy model description method, LLSM (Language for Large-Scale Modeling), based on concurrent probabilistic hybrid automata (cPHA) to make the process expediently. LLSM describes systems in the form of text. It settles the problem in three aspects: granularity, hierarchy and reusability. Component-oriented modeling of LLSM helps control granularity easily allowing users to create models in different scales. A special mark, which is employed to represent hierarchical relationship makes the system clearer and guides the accuracy of diagnosis. Reusability is achieved by C-style grammar which indicates component libraries for large-scale applications. In complex applications, LLSM creates models efficiently by existing libraries in the form of collaboration. Test on a switch demonstrates how it works.