A hybrid supervised/unsupervised neural network architecture for health and usage monitoring

The role of a health and usage monitoring system (HUMS) is to provide continuous online fault detection, isolation, and (ideally) prognostics. The hybrid supervised/unsupervised neural network HUMS architecture outlined is designed to achieve these goals. Indeed, with prevalent life extension programs and ever increasing maintenance costs such a system is essential to the successful operations of today´s complex systems. The primary role of the supervised networks is fault isolation. Rather than training a single diagnostic network with n outputs, one for each fault or fault precursor in the database, multiple neural networks operating in parallel provide superior performance while simultaneously facilitating the neural network training process. More importantly, one can limit the number of inputs to each network to those which are most significant to the isolation of the fault or fault precursor to which that network is dedicated, thereby reducing the number of neurons in the network, while reducing both the run time and training time required for the network. Finally, by using separate networks for each fault or fault precursor, one can easily add additional networks in-service as new faults and fault precursors are added to the diagnostic database, without retraining the existing networks. Any of the standard feedforward neural networks with a supervised training algorithm can be used for the diagnostic networks