Malleable neural networks in fault detection of complex systems

Industrial machining centres are composed of complex integrated subsystems with independent critical issues. Neural networks (NN) is capable of monitoring unset of faults, however, complexity of the many possible failure modes and various levels of intensity may deteriorate the accuracy of NN. This paper presents malleable neural networks architecture for condition monitoring and fault diagnosis of a subsystem of a machining centre. A central NN is trained with faulty status of operation at the core stage which is then able to discern between healthy and all possible faulty states. NNs are then modulated to learn each failure mode with their different intensity levels. Diagnostic is initially made by the central module, then, the network is reconfigured by an interprocess call to adapt to an appropriate topology and knowledgebase to detect the severity level of the fault. The monitoring system uses steady state values of sensitive parameters of current and pressure transducers. If the parameters are out of a predefined healthy range, a nondestructive test will be initiated, which produces a transient response as input pattern to the NNs. Testing the NN based monitoring system with 395 failure modes showed that in 99.2% of cases the network was successful to accurately identify the cause and severity of the failures.