Multi-output MLP Architecture for Machining Process Parameter Classification

Remote monitoring of manufacturing processes plays a crucial role in enhancing productivity, reducing production costs, and improving product quality in machining related industries. Traditional remote monitoring methods rely on invasive sensor installation, which can be costly to install and may interfere with production. This paper uses an efficient, non-invasive method of using microphones and machine learning to classify key machining parameters: spindle speed (RPM), depth of cut (DOC), and feed rate (FR). Conventional machine learning approaches typically classify one parameter at a time, while the alternative deep learning approaches tend to be more computationally expensive, featuring a high number of parameters. In this paper, we explore a simpler neural network model – a multi-layer perceptron (MLP) with a branched architecture capable of simultaneous multi-parameter classification. A dataset of audio recordings from a microphone array on machining operations was first collected. Recorded test data was then autocorrelated, processed through a Fast Fourier Transform (FFT), and then normalized to decrease noise. Optimal model hyperparameters were determined through Bayesian hyperparameter tuning. The developed multi-output model achieved a classification accuracy of 93% (FR), 99% (DOC), and 100% (RPM). The multi-output model featured fewer parameters (1.85 M) compared to that of deeper architectures such as VGG16 (135 M parameters), which was used for the same task in other studies, highlighting the effectiveness of the MLP architecture.