Learning machining stability diagrams from data using neural networks

Abstract

Machining instabilities are detrimental. model predicted stability charts help identify cutting parameters for stability. Since models disregard speed-varying cutting force characteristics and dynamics, charts fail to guide stable cutting in industrial praxis. This study shows how supervised neural networks can learn stability charts from data. The learning capacity of this machine learning model depends on the size of the training dataset, its train-test split, the learning rate, the activation function, the number of hidden layers, and the number of neurons in each layer. This is the first study to examine how hyperparameters influence learning machining stability diagrams. Learnings from a linear stability dataset are transferrable to nonlinear datasets, demonstrating the prediction model is physics-agnostic. Predictions accuracies of up to 97.2% were obtained. Since the data used to train the model includes all the vagaries and uncertainties of the cutting process, the results can inform self-optimizing and autonomous machining systems.