ANN-based predictive models are becoming increasingly popular in machining technologies. Our research focused on the potential applications of AI-based predictive models in micro-milling, using a dataset from a cutting experiment designed to analyze the tool life. This dataset was previously utilized solely to develop traditional regression models, so our goal was to create an Artificial Neural Network (ANN) that could more efficiently predict tool life based on this data. Given the small sample size of the dataset, leave-one-out cross-validation (LOOCV) was employed during validation. By experimenting with various network structures—modifying the numbers of layers and neurons, and types of activation functions—we determined an appropriate ANN model to outperform the original regression models. Fully-connected feed-forward neural networks were trained using the Adam optimizer for up to 200, 500, and 1000 epochs. The model complexity was adjusted by varying the number of hidden layers from 1 to 10 in steps of one, and the number of neurons per layer from 5 to 50 in increments of five. Each model’s evaluation was based on the Mean Absolute Error (MAE) and the Coefficient of Determination (R2) and Standard Deviation of repeated training. The optimized ANN structure outperformed the second-order linear regression method in terms of both evaluation metrics and monotonicity analysis between the data points.
Science Journal 