Artificial neural network performance based on different pre-processing techniques

Condition monitoring systems aim to detect machinery faults, preferably in their early stages. This provides important information to the maintenance team and reduces the plant operational and maintenance costs. A popular tool for automated condition monitoring systems uses artificial neural networks, whose performance is heavily dependent on the quality of the input training data, which must accurately contain the main features of the signal to be classified. This work compares the usage of two feature extraction techniques (cepstrum and cepstrum reconstruction), as pre-processing tools for automated monitoring systems. Here, a neural network is used to diagnose the condition of a model drive-line, consisting of many rotating parts. Including a pair of spur gears, bearings, and an electric motor. Firstly, the model drive-line was ran in its normal condition, and later it was run with different gear faults (simulated cracks of different sizes) introduced intentionally. The real time domain vibration signatures from the drive-line under different conditions were pre-processed using the different pre-processing techniques. The pre-processed signals are used as input to neural networks that perform fault detection. It is shown here, that the cepstrum reconstruction technique does in fact outperform the basic cepstrum technique as a feature extraction tool for automated monitoring systems. Hence although more computationally expensive, cepstral reconstruction is a better choice of preprocessing technique for automated condition monitoring systems.