HYBRID NEURAL NETWORKS AS A TOOL FOR THE COMPRESSOR DIAGNOSIS

An attempt to apply neural networks to the acoustic diagnosis for the reciprocating compressor is described. The proposed neural network, Hy brid Neural Network (HNN), is composed of two multi-layered neural net works, an Acoustic Feature Extraction Network (AFEN) and a Fault Discr imination Network (FDN). The AFEN has multi-layers and the number of u nits in the middle hidden layer is smaller than the others. The input patterns of the AFEN are the logarithmic power spectra. In the AFEN, t he error back propagation method is applied as the learning algorithm and the target patterns for the output layer are the same as the input patterns. After the learning, the hidden layer acquires the compresse d input information. The architecture of the AFEN appropriate for the acoustic diagnosis is examined. This includes the determination of the form of the activation function in the output layer, the number of hi dden layers and the numbers of units in the hidden layers. The FDN is composed of three layers and the learning algorithm is the same as the AFEN. The appropriate number of units in the hidden layer of the FDN is examined. The input patterns of the FDN are fed from the output of the hidden layer in the learned AFEN, The task of the HNN is to discri minate the types of faults in the compressor's two elements, the valve plate and the valve spring. The performance of the FDN are compared b etween the different inputs; the output of the hidden layer in the AFE N, the conventional cepstral coefficients and the filterbank's outputs . Furthermore, the FDN itself is compared to the conventional pattern recognition technique based on the feature vector distance, the Euclid distance measure, where the input is taken from the AFEN. The obtaine d results show that the discrimination accuracy with the HNN is better than that with the other combination of the discrimination method and its input. The output criteria of network for practical use is also d iscussed. The discrimination accuracy with this criteria is 85.4% and there is no case which mistakes the fault condition for the normal con dition. These results suggest that the proposed decision network is ef fective for the acoustic diagnosis.