Fault Diagnosis with 
Neuro-Fuzzy Methods

Publications

The area of supervision and fault detection/diagnosis of technical processes has gained more and more attention. Several approaches have been developed, comprising traditional signal analysis as well as model-based techniques.

If one considers the supervision of technical processes as a two-stage process consisting of a feature or symptom generation followed by a diagnostic system, the two basic problems of a supervision concept become clear:

• How can significant symptoms be constructed that contain as much information about the fault as possible?
• How can a good diagnostic system be designed, which reflects the fault symptom relationship and performs a reliable decision?

My research deals with the second question.

Fig. 1: Self-Learning Diagnosis: From measured symptoms ...

The diagnostic system serves several purposes: In the existence of noise and model inaccuracies, it identifies the optimal decision boundaries between the different fault situations. As such, it works pretty much like any other classification system. A diagnosis can additionally provide insight into the relevance and power of the symptoms that are used. But most importantly, the system should make the fault decision transparent. This increases the acceptance of the system and allows changes to be easily applied. 

Common approaches are fault symptom trees and fuzzy rule bases that are constructed manually. Systems based on measurement data use multilayer perceptron networks in most cases. Occasionally, self-organizing maps, ART-networks and simple clustering techniques are implemented. Radial basis function networks, regression trees and neuro-fuzzy approaches have also gained more attention. 

A particular problem of methods that use an explicit knowledge base (like fault symptom trees or fuzzy rule bases) is the lack of sufficient information. Sometimes, only parts of such a knowledge base exist, or they are tedious to derive.

Most of the data-driven techniques, on the other hand, carry the disadvantage that they rely completely on measured data. While this is perfectly valid as long as the data set is a good representation of the data present in real life, it neglects the capacity of human interaction and knowledge that can improve the design. Even if sufficient data is present, the diagnosis system can be improved when considering aspects like transparency, simplicity or robustness.

Fig. 2: ... to a transparent fault diagnosis system.

Common approaches to incorporate a-priori knowledge into diagnostic systems integrate only some kind of basic knowledge (like fuzzy rules). In contrast to that, I developed methods to incorporate knowledge, which naturally arises from the knowledge of the process, supporting the design of a diagnosis system. The aim is to develop diagnosis system design tools that support the engineer as much as possible. This leads to approaches which combine the strength of data-driven approaches with human interaction in form of fuzzy logic. 

Application areas include: Electrical motor diagnosis, process components, sensors, actuators diagnosis, pump diagnosis

Keywords: Neural Networks, Fuzzy Systems, Neuro-Fuzzy, SELECT, Fault Detection and Diagnosis, Fault Trees

Publications

Related links: IFAC SAFEPROCESS Technical Committee

Contains links to other fault detection and diagnosis related material and people.

Previous Research Topics

At the State University at Buffalo I worked in the area of Multidisciplinary Optimization (MDO). There, Optimization techniques are investigated that are used for global optimization purposes covering several engineering areas. A simple example would be to find a minimum cost design of an airplane part considering both the mechanics, aerodynamics and also control aspects of the part. This requires a controls analysis as well as - for instance -some kind of finite element analysis which run under one global optimization objective. My work tried to balance the accuracies of a control and a finite element analysis against each other to yield a minimum computational cost.
If you are interested in more detail, please hook up to the Multidisciplinary Design and Optimization Lab (MODEL) at the SUNY in Buffalo.