Fault Diagnosis of Complex Dynamic Processes by Use of Additive Modular Knowledge Base

In this paper a fault diagnosis method for dynamic processes is proposed. It uses a special modular knowledge base, which consists of separate modules for each faulty condition and the normal condition of the process. Each module stores the most representative features of the training data for a certain faulty state in a special compact form. For such purpose, a representative set of neurons (RSN) is used that is trained by the unsupervised neural-gas learning algorithm. The introduced algorithm for fault diagnosis utilizes the concept of the average minimal distance between a set of newly collected process data and the trained RSN for each faulty condition. The fault diagnosis decision is defined as the most similar (the closest) fault to the new operation data. Real experiments on a laboratory three-buffer-tank-system are used in the paper to prove the correctness and applicability of the proposed fault diagnosis method