Fault detection in air pump assembly based on sound classification using one-dimensional convolutional neural networks

This study presents a comprehensive analysis of the operational audio signals produced by a AC rectified reciprocating air pump under various fault and healthy conditions. A total of 1200 audio data samples, each with a duration of 0.1 seconds, were collected and standardized from ten machines of the same model. These samples encompass six distinct states, including a normal state and five different assembly faults. Utilizing this data, the study investigates the performance of various one-dimensional convolutional neural network (CNN) architectures for audio signal recognition, exploring configurations with different numbers of layers (one to three) and filter sizes (10 to 50). The findings indicate that single-layer CNN architectures are ineffective in accurately identifying the different fault states. Conversely, two-layer CNN architectures, particularly those with filter sizes ranging from 20 to 30, exhibit a recognition accuracy of up to 90% and demonstrate exceptional stability. In contrast, three-layer architectures show poor performance, with a notable occurrence of depth degradation phenomena. These results underscore the importance of selecting an appropriate network depth and filter size to enhance the accuracy of audio signal recognition for air pumps.