Fault diagnosis via local neural networks

Abstract

This article deals with the fault diagnosis problem of plants with unknown description. The problem is approached via a modeling technique which is based on the local model network (LMN) structure. Local models (LMs) perform local linearization and their structure is quite similar to the Takagi–Sugeno fuzzy models. Local neural models (LNMs) function as linear estimators, giving a satisfying estimation for the plant’s output within parts of the operating regime. Their training is performed off-line, which ensures a reliable method with false alarm avoidance. This is critical since false alarms may cause a production line to pause. Plant modeling is followed by an on-line comparison between plant and model behavior. The model is created on a healthy plant, so any mismatch leads to suspicions concerning the presence of faults. The comparison leads to the creation of residuals. Residuals are signals that trigger a decision mechanism to conclude for presence, size and cause of possible faults. Change detection algorithms are used at this point, to avoid misinterpretation of plant model mismatches not caused by faults. The method is tested via simulation on the three-tank system which is a well-known benchmark.