Sensor fusion and use of reconfigurable neural networks in condition monitoring

High-speed and precise industrial machinery demand continuous real-time monitoring to alert an tinsel of faults. This paper presents an adaptive Modular Neural Networks System (MNN) for condition monitoring and fault diagnosis. In designing the system, steady state values of sensitive parameters of current, voltage and pressure transducers are continuously monitored. Deviation of two out of three sensors causes an alarm for a possible malfunction. To detect the cause of malfunction a rapid non-destructive test, which generates a transient signal is carried. This signal is tested by the knowledge base of the MNN which is trained with ensemble of transient patterns of the known failure modes of the system tinder test. The MNN design is capable of detecting faulty/unfaulty conditions, and if faulty it detects the degree of deterioration of the system. MNN is trained with separate input (transient patterns of faults)-target (causes and levels of faults). Initially, MNN detects cause of malfunction and an interprocess communication up-loads the appropriate trained weight vector to test and detect the level of deterioration. Test result on over 350 cases indicated a fault detection accuracy of more than 99%.