An Online Outlier Identification and Removal Scheme for Improving Fault Detection Performance

Abstract

Measured data or states for a nonlinear dynamic system is usually contaminated by outliers. Identifying and removing outliers will make the data (or system states) more trustworthy and reliable since outliers in the measured data (or states) can cause missed or false alarms during fault diagnosis. In addition, faults can make the system states nonstationary needing a novel analytical model-based fault detection (FD) framework. In this paper, an online outlier identification and removal (OIR) scheme is proposed for a nonlinear dynamic system. Since the dynamics of the system can experience unknown changes due to faults, traditional observer-based techniques cannot be used to remove the outliers. The OIR scheme uses a neural network (NN) to estimate the actual system states from measured system states involving outliers. With this method, the outlier detection is performed online at each time instant by finding the difference between the estimated and the measured states and comparing its median with its standard deviation over a moving time window. The NN weight update law in OIR is designed such that the detected outliers will have no effect on the state estimation, which is subsequently used for model-based fault diagnosis. In addition, since the OIR estimator cannot distinguish between the faulty or healthy operating conditions, a separate model-based observer is designed for fault diagnosis, which uses the OIR scheme as a preprocessing unit to improve the FD performance. The stability analysis of both OIR and fault diagnosis schemes are introduced. Finally, a three-tank benchmarking system and a simple linear system are used to verify the proposed scheme in simulations, and then the scheme is applied on an axial piston pump testbed. The scheme can be applied to nonlinear systems whose dynamics and underlying distribution of states are subjected to change due to both unknown faults and operating conditions.

An Integrated Learning and Filtering Approach for Fault Diagnosis of a Class of Nonlinear Dynamical Systems

“…Theorem 1: Consider the nonlinear system (1) and (2) with the adaptive approximation structure described in (7), (10), (11), (16), and (17), and the fault detection scheme described in (8), (18), and (23).…”

“…When a fault is detected at time T d , we use the same logic that was used to learn the modeling uncertainty, with the only difference that this time the task is to learn the combined effect of the modeling uncertainty coupled with the potential process fault function. Essentially, we use the same estimation model as before indicated by (10), (11), (18), and resume learning. For clarity, we rewrite the equations as follows:…”

“…Several survey papers [1]- [3] and books [4]- [7] detail the main methods that have been proposed over the years. The research conducted for fault diagnosis in linear systems is well established (the interested reader may refer to the aforementioned survey papers and books), whereas for nonlinear systems, the work is more limited and focused on specific classes of nonlinear systems by using various methods, such as observer methods [8]- [13], differential-geometric approaches [14], change-detection approaches [15], and learning methods [16]- [18]. These approaches were based on a centralized framework, whereas lately, due to advances in communications and distributed sensing, several fault diagnosis techniques for hierarchical, decentralized, and distributed systems have emerged [19]- [30].…”

Detecção Móvel E Online De Anomalia Em Múltiplos Fluxos De Dados: Uma Abordagem Baseada Em Processamento De Eventos Complexos Para Detecção De Comportamento De Condução

“…Argumentamos que a análise do comportamento de condução, utilizando tais fluxos de dados, pode ser mapeada para o problema de detecção de anomalia em fluxos de dados. Este problema refere-se à encontrar padrões nos dados que não estão em conformidade com (ou desviam suficientemente de) comportamentos esperados [14], [26], por exemplo, mudanças repentinas de faixas e frenagens bruscas. Embora a detecção de anomalia tenha sido amplamente estudada, não têm havido muitos trabalhos de pesquisa na detecção de anomalia em fluxos de dados [15], [20], [21].…”

“…Em todas estas aplicações, a maioria dos dados possuem um modelo "normal", e as anomalias são consideradas como desvios deste modelo [22]. Do grande número de pesquisas sobre detecção de anomalia, a maioria dos métodos existentes opera em todo o dataset para detectá-las [14], [30], foi projetado para lidar com a análise offline [26] de um grande volume de dados ou são ineficientes quando aplicados em fluxo de dados [21]. É desafiador adaptar as técnicas/algoritmos existentes de detecção de anomalias para realizarem o processamento dos múltiplos fluxos de dados nos próprios dispositivos móveis, uma vez que originalmente foram projetados e desenvolvidos para ambientes cujos recursos de memória e processamento são abundantes [19].…”