This PhD Thesis is dedicated to the development of a fault detection and isolation system, intended to be widely applied to any industrial system from where we do not have any expert knowledge available but recorded data, i.e., we are provided with the history of the industrial process coming from the sensors and actuators placed all along the industrial system facilities. Based on the available history of the process, data mining and soft computing techniques are used to extract knowledge of the process. Thus, a set of potential running models are identified, which once trained offline are the departure point to provide a set of residual generators for the process to monitor. By this set of residual generators, the research develops techniques to provide fault detection for the process, and after it also fault isolation. The methodology developed, the methods explored, and the algorithms created were tested on two real-world scenarios, so results for them are provided. The research was done under the assumptions that neither expert knowledge nor general information of the process is available, so all the experimentation was done under the umbrella of a black-box approach.
The PhD Thesis is organized as follows:
Chapter 1 introduces the problem and the restrictions we are coping with, whereas Chapter 2 describes the state-of-the-art of the problem, showing techniques which already exist, pointing out why some are not suitable for our problem due to our restrictions.
Chapter 3 provides an overall overview of the methodology developed. Its main strength is to establish the framework for the whole research, as it explains how to preprocess the data, how to identify the residual generators using data mining and soft computing techniques, how to train the models offline and how to test them online (by using artificial faults) to check their fault detection and fault isolation capabilities.
Chapter 4 and Chapter 5 focus on the different model architectures explored. The former explains three model architectures which are applied directly as residual generators, the later explains several extensions which are to be combined with the former model architectures, thus creating new model architectures as result of the combinations.
Chapter 6 develops fault detection. It explains the basis about the residual space, introduces two measures of uncertainty applied to the residuals during the research and introduces the technique used to provide fault detection based on the residual space.
Chapter 7 develops fault isolation. Its core introduces new tools to evaluate the fault isolation capabilities of a method, and develops a fault isolation technique in two parts: analyze the partial derivatives of the residual generators and make a decision about the variable responsible for the fault indicator. It concludes analyzing the complexity of the fault isolation problem.
Chapter 8 provides the results of the research. The chapter is divided in the results for fault detection and the result for fault isolation.
Both are provided with a statistical analysis of the quantitative results, in order to establish preferences among the model architectures tested.
Chapter 9 and Chapter 10 conclude the PhD Thesis by adding conclusions and eliciting pending issues/future lines of research.