In-situ detection of Anomalies in the Acoustic Signal using Neural Network Approach

Neural network with two hidden layer is used in the process of detecting the anomalies in the input audio signal. Three types of input signals (Diesel engine sound, musical instruments, speech signal through microphone) are considered in the studies. Five different features (Wavelet, Spectral centroid, Pitch, MFCC, Phase shifted band) are extracted in the form of spectrogram images from the 1D subsampled speech signal. Later, the processing of spectrogram images by the neural network is carried out. It resulted in a prediction rate of 98.4%, 95.5% and 97.4%, for an input audio signal from diesel engines, musical instruments and speech signal, respectively. Real-time implementation of anomaly detection in audio signal is carried out using a raspberry pi board with 4 GB RAM. Average training time reported on a raspberry pi board is of 4 minutes and 30 seconds.