Proc. of JCIS'98, The Fourth Joint Conference on Information Sciences. Research Triangle Park, North Carolina, USA, Vol. 1. pp. 207-210. 1998. NEUROFUZZY COMPUTING AIDED MACHINE FAULT DIAGNOSIS C. Emmanouilidis 1 , Dr. J. MacIntyre 1 , Prof. C. Cox 2 University of Sunderland, UK 1 Centre for Adaptive Systems, School of Computing & Information Systems, PO Box 299, Sunderland SR6 0YN, UK, cs0cem@isis.sunderland.ac.uk, John. MacIntyre@sunderland.ac.uk. 2 School of Engineering and Advanced Technology, Chester Road, Sunderland, SR1 3SD, UK. chris.cox@sunderland.ac.uk ABSTRACT: Machinery malfunction problems are often sources of increased maintenance costs and disturbances in production activity across industry. Reliable diagnostic methodologies are needed to enable cost effective condition based maintenance. Among the methods that can be employed for fault diagnosis, neurofuzzy techniques offer an appealing problem solving framework. This is due to their capability for simultaneously handling numerical and linguistic information. This paper demonstrates the use of neurofuzzy techniques for machinery fault diagnosis, based on measuring vibration. Data driven diagnostic model building is facilitated by the availability of rich vibration data sources, including simulation and real data gathered from purpose-built test rigs and machinery equipment at industrial sites. This work is carried out as part of the VISION project, a European research project, which is developing a generic machinery fault diagnostic system, based on a fusion of computational intelligence techniques. 1. INTRODUCTION Rotating machinery fault diagnosis based on vibration measurements calls for reliable data analysis and decision making in the presence of uncertainty. This uncertainty can be attributed to different factors such as: • Measurement noise. • Lack of a clear deterministic relationship between the measured quantities and the machine state. • Different characteristics of the transmission path between the error point and the measurement location. • Variations in the vibration generated by different machine configurations. Many approaches can been taken for automatic fault diagnosis, including statistical (Bayesian, nearest neighbours), polynomial, neural network, fuzzy and neurofuzzy classification methods (Leonhardt and Ayoubi 1997). Vibration data is often or can easily become available from rotating machinery, but in many cases there is little provision for taking full advantage of its potential for improving reliability in machine condition diagnostics. There is also ample domain knowledge about machinery malfunction, but fault diagnosis is usually far too complex to be reliably provided by simple expert rules. There is a clear need for methods capable of simultaneously handling numerical data and human expert knowledge. Fuzzy systems methodologies are well suited for such problems, as they can naturally process both numerical data and linguistic information. Integrated hybrid fuzzy-neural representations or inherently fuzzy logic models equipped with neural network-like learning capabilities are powerful adaptive modelling tools, which combine the individual merits of both fuzzy logic systems and neural networks. Fuzzy methodologies can enter into the diagnostic problem in many different forms including: • Fuzzy definition of the severity of important fault features (Loskiewicz-Buczak and Uhrig 1993). • Sensitivity of vibration spectral features to faults, expressed with linguistic terms (Liu et al 1996). • Fuzzy predominant frequencies (Siu et al 1997). • Fuzzy or hybrid neural-fuzzy fault type recognition (Huang and Wang 1996, Tse and Wang 1996). Although significant success has been demonstrated in employing neurofuzzy methodologies for rotating machinery fault diagnosis, most applications are still rather specialised and fragmented and considerable effort is still needed in order to move towards more generic and reliable diagnostic solutions. This paper describes some steps taken towards this goal under the framework of a European research project (VISION, Brite/EuRam BE95-1313). Work currently carried out is focused on developing a generic machinery off-line intelligent diagnostic system based on vibration data. Here, neurofuzzy techniques are shown to be applicable for diagnosing common faults such as unbalance, misalignment and various types of bearing faults, with both single and multiple fault scenarios and in the presence of varying levels of noise. These methods are currently integrated with other statistical and neural network based techniques (Adgar et al 1998) in order to achieve increased diagnosis reliability and robustness. 2. NEUROFUZZY COMPUTING Among the different approaches to intelligent computation, fuzzy logic provides a strong framework for achieving robust and yet simple solutions. Fuzzy logic can be further strengthened by the introduction of learning capabilities, such as those of artificial neural networks. A large number of adaptive fuzzy or neurofuzzy models have been reported in the literature, which aim at amalgamating the benefits of both computational approaches, namely the learning capabilities of neural networks and the representation power and transparency of fuzzy logic systems. When building neurofuzzy systems for non-trivial problems, there is usually a trade-off between model interpretability and performance. This is reflected into