In this chapter, the use of neural networks in diagnostics is investigated. Diagnostics is critical to any continuous process set up, to detect the occurrence of failures with accuracy and speed, particularly when failures can be catastrophic. Previous approaches to diagnostics problems have been rule-based. They suffered from slow response time and failure under missing information. Neural networks are fast because they use parallel processing concepts. They are tolerant of missing information because information is not stored locally but distributed over several nodes. In this research, a two-layer feedforward neural network is proposed as a modeling paradigm for continuous process diagnostics. Specific diagnostic problems in the domain of nuclear reactor control are studied. The diagnostic knowledge gathered by experts is represented in binary patterns. Different input and output representations are derived and the performance under each is analyzed. The results indicate that input representation is particularly important in determining the neural network’s recognition capability and the amount of tolerable noise in parameter readings.
[1]
Dana H. Ballard,et al.
Cortical connections and parallel processing: Structure and function
,
1986,
Behavioral and Brain Sciences.
[2]
James L. McClelland,et al.
Parallel distributed processing: explorations in the microstructure of cognition, vol. 1: foundations
,
1986
.
[3]
Steve Nann.
A Decision Support System for Control and Automation of Dynamical Processes
,
1990
.
[4]
Robert Milne,et al.
Strategies for Diagnosis
,
1987,
IEEE Transactions on Systems, Man, and Cybernetics.
[5]
Jerome A. Feldman,et al.
Connectionist models and parallelism in high level vision
,
1985,
Comput. Vis. Graph. Image Process..
[6]
D Zipser,et al.
Learning the hidden structure of speech.
,
1988,
The Journal of the Acoustical Society of America.