Project Overview

Concept and objectives


Modern society relies on the availability and smooth operation of complex engineering systems. Examples include electric power systems, water distributions networks, manufacturing processes, transportation systems, robotic systems, intelligent buildings, etc. The emergence of networked embedded systems and sensor/actuator networks has made possible the development of several sophisticated monitoring and control applications where a large amount of real-time data about the monitored environment is collected and processed to activate the appropriate actuators and achieve the desired control objectives. Depending on the application, such data may have different characteristics: multidimensional, multi-scale and spatially distributed. Moreover, the data values may be influenced by controlled variables, as well as by external environmental factors. However, in many cases the collected data may not make much sense! For example in an intelligent building application, the temperature sensor may be recording rapidly increasing temperatures, possibly indicating a fire, while the smoke detector sees nothing. What does the system do? For humans, the decision may be easy because we have redundant sensory information which we are able to process in real-time and correctly assess the situation. Furthermore, we have very good confidence in the state of our sensory organs.


For machines however, the decision is not as easy. Information redundancy is limited due to cost and, in addition, the state of the sensors and actuators in not always known and several things can go wrong. For example, some measurements may be missing, sensor performance may be deteriorating due to aging or environmental conditions, sensors may be drifting, etc. In some cases, data coming from different sensors (or actuators), on the same unit (e.g., with different resolution) or residing in a cluster, may become inconsistent. At the same time, the environment can be subject to nonstationarity phenomena and the electronics, e.g., the signal conditioning stage, is prone to drifts and soft/hard faults. Such problems, which will generate “nonsense data,” are generally the result of some faults in the sensor/actuator system itself or an abnormality in the monitored environment, which may be either permanent or temporary, developing abruptly or incipiently. These problems become more pronounced as sensing/actuation systems get older.

 

 

This project will focus on innovative cognitive fault diagnosis approaches that can learn characteristics or system dynamics of the monitored environment and adapt their behaviour and predict missing or inconsistent data to achieve fault tolerant monitoring and control. The main motivation is to exploit spatial and temporal correlations between measured variables and develop the tools and design methodologies that will prevent situations where a relatively “small” fault event in one or more components (e.g., sensor, actuator, communication link) may cause an overall system failure.

Most design approaches for autonomous fault diagnosis rely on the concept of analytical redundancy, which refers to the case where the redundancy required for fault detection and isolation is obtained based on an available mathematical model of the healthy system. If such a mathematical model is available, then fault diagnosis is achieved by comparing actual observations with the prediction of the model. In practice, however, such mathematical models may not be accurate or not even available at all. Therefore, there is a need to develop fault detection methods, where the time history of the observed data and the inter-relations between spatially distributed sensing are exploited.

This project will develop an innovative cognitive fault diagnosis framework with learning algorithms for approximating, during operation, key correlations between measured variables. This will allow us to handle highly unstructured environments, which goes well beyond the current state-of-the-art in the area of fault diagnosis. In the case of fault isolation, the work that has been developed so far provides a basis for differentiating between classes of known faults. However, in open-ended and unstructured environments, there may be unanticipated fault scenarios that haven’t been encountered before. How are such cases identified and handled? How is the set of fault isolation estimators updated during operation? The cognitive fault diagnosis approach that will be developed in this project provides a framework for handling unknown fault scenarios within a fault isolation and identification formulation. The unknown fault will be incorporated into the fault isolation knowledge base using a just-in-time adaptive classifier scheme.


Learning techniques for cognitive fault diagnosis are based on the concept of characterising the correlations of data gathered from multiple sensors, some of which may be faulty or they may be providing inconsistent information following a fault event in the underlying environment. As the number of sensors/actuators and the complexity of the system increases, it is often better to develop learning schemes that utilise multiple learning machines working concurrently and cooperatively. In this project we propose to take advantage of the regularized negative correlation learning algorithm for establishing the functional correlations between different sensors. This will provide a novel approach for autonomous fault diagnosis by taking into consideration the perspective of the same event as viewed from different sensing devices.

 

Photo credit: Kts Design

While fault detection and identification determine the presence and nature of a fault, ultimately it is of fundamental importance to be able to use this knowledge for handling various fault scenarios and avoiding fault propagation. Fault-tolerant control refers to feedback control systems that are designed to withstand a certain measure of fault functions by sufficiently safeguarding the key operation of the system. In this project we will investigate both passive and active fault tolerant control architectures and develop a fault tolerant model predictive control scheme, which allows the incorporation of constraints on the state and the controls.


Recent technological advances in computing hardware, sensors/actuators, communications and real-time software have facilitated the development of networked intelligent agent systems with the capability to generate huge amount of data in real time. This data is generated continuously every day in a wide range of application environments. As the volume of data generated is rapidly increasing, one of the most important research challenges in the years ahead is the development of information processing methodologies that can be used to extract meaning and knowledge out of the ever-increasing electronic information that becomes available, especially in cases where the data generated is faulty, missing or inconsistent.


Developing a systematic approach for fault diagnosis and fault tolerant control in uncertain distributed environments will have a significant impact on the availability, reliability and dependability of engineering systems, and will provide a new paradigm for fault tolerance in incompletely specified environments. To achieve its objectives, this project integrates expertise in systems and controls, fault diagnosis, machine learning, evolutionary computation, fault modelling, neural networks, sensor hardware design and manufacturing within a common framework.

 

Scientific and technological objectives


The main objective of this project is to make sense of nonsense by developing intelligent data processing methods for analyzing and interpreting the data such that faults are detected (and where possible anticipated), isolated and identified as soon as possible, and accommodated for in future decisions or actuator actions. The problem becomes more challenging when these sensing/actuation systems are deployed in a wide range of open-ended environments which are not known a priori and, as a result, it is unrealistic to assume the existence of an accurate model for the behaviour of various components in the monitored environment. Therefore, this project will focus on cognitive system approaches that can learn characteristics or system dynamics of the monitored environment, adapt their behaviour and predict missing or inconsistent data to achieve fault tolerant monitoring and control.

 

 

This project will develop the iSense Platform, which will consist of a set of intelligent agents, integrated with the sensors, actuators and feedback control system for making the overall monitoring and control system more robust, adaptive and fault-tolerant to sensor/actuator faults and system faults or abnormalities in the environment. The developed prototype of the iSense Platform will be validated for the intelligent building application domain, however, the formulation, tools and methodology developed will be transferable to other application domains such as water distribution networks, power transmission and distributions grids and more.