Automatic fault diagnosis and health monitoring plays an important role in the timely and economic maintenance of machineries in the industry. Of late, there has been greater emphasis on machine learning based approaches to solve the fault diagnosis problem. The availability of relevant sensors and required processing capabilities has further enabled the research efforts in this area. In this work, we have developed an acoustic signal based fault diagnosis and classification system for air compressors. The acoustic signals are acquired using a microphone and National Instruments' data acquisition system. The acquired data is analyzed in time, frequency and wavelet domain. We have used a set of robust features using the Shannon's information theory. In this work, we have made two significant contributions: (i) used an information theoretic framework to determine the suitability of an acoustic data instance for classifier training. (ii) proposed a novel Bag-of-feature based approach for fault diagnosis using an acoustic signal. We evaluated the performance of these features using three well known classifiers like kNN, SVM and Maximum-Likelihood and obtained very good accuracy for the real data set.
Download Full PDF Version (Non-Commercial Use)