Sensor and actuator selection for the identification of partial observation models of discrete-event systems with concurrent behavior for fault detection

Abstract

In the classical methods of fault diagnosis of Discrete-Event Systems (DES), the complete knowledge of the system behavior is required to obtain a white-box monolithic system model, which can be a difficult or even impossible task to be carried out for complex systems. In these cases, system identification techniques can be applied using sequences of vectors formed of the input (sensor) and output (actuator) signals of the controller, called I/O vectors. However, when the system is composed of several subsystems with concurrent behavior, it may be necessary to observe a very long length sequence of I/O vectors generated by the closed-loop system to obtain an accurate monolithic model. This motivates the identification of partial observation models, where each model is identified using part of the binary signals exchanged between plant and controller, which are selected such that the identified model can be obtained using less I/O vector observations. After the identification of each partial observation model, a fault detection strategy that plays the partial observation models in parallel, and verifies the occurrence of non-expected behaviors, can be defined. We propose, in this paper, a method for selecting the inputs and outputs of the controller for identification of the partial observation models such that each model has predominantly sequential behavior, which leads to models that can be obtained using less I/O vectors than the monolithic model. In addition, the causal relationship between sensors and actuators is represented in the partial observation models, which allows the correct representation of the system dynamics. The proposed method is based only on the observation of the system I/O vectors, without using any knowledge of the system behavior. A virtual plant, simulated using a 3D simulation software, and controlled by a programmable logic controller, is used to illustrate the proposed method. In addition, we have applied the proposed method to the mechatronic plant, installed at the Laboratory of Control and Automation of the Federal University of Rio de Janeiro, to obtain the partial observation models.