Classification of Anomalies in Industrial Machines Utilizing Machine Sounds and Deep Learning

Efficient maintenance of industrial machinery is crucial for ensuring seamless operations and extending the lifespan of machines. This paper introduces a novel approach to classifying anomalies in industrial machine sounds by integrating signal processing with deep learning. The proposed method utilizes continuous wavelet transform (CWT) to extract robust features by transforming acoustic signals into frequency and time representations. These representations are subsequently processed by a novel convolutional neural network (CNN), designed to identify patterns and classify machine anomalies efficiently. We experimented with the proposed approach on the Malfunctioning Industrial Machine Investigation and Inspection (MIMII) dataset, which includes eight categories based on machine type and fault type. Additionally, we conduct a comprehensive evaluation by comparing our method with various benchmark CNN architectures, including DenseNet, EfficientNet, InceptionNet, MobileNet, NASNet, VGGNet, and Xception, in terms of accuracy, inference time, and model size. The proposed model demonstrates superior performance, achieving the highest test accuracy of 99.53%, with a model size of 2.1 MB and an inference time of 32 milliseconds.