A GAN-Based Multi-Sensor Data Augmentation Technique for CNC Machine Tool Wear Prediction

The prediction of CNC machine tool wear is limited due to the scarcity of data in industry. A sufficient amount of experimental and real-live factory data is missing to train accurate supervised machine learning prediction models. Considering that collecting a large amount of data in real industrial environments can be a great challenge, this paper investigates the applicability of data augmentation techniques for generating synthetic data. Our paper focuses on a multi-sensor approach for the classification of tool wear states in the milling process. Multi-sensor data fusion is performed in the frequency domain. Extracted features are then used for tool wear classification. In order to increase the amount of training data, a generative adversarial network (GAN) is designed for data augmentation purposes. An early stopping strategy is designed to improve the effectiveness of the proposed GAN. The experimental results show that the proposed GAN helps to significantly improve the performance of tool wear classification models. With the inclusion of GAN, the number of required real-environment industrial data can be reduced. The research shows promising results to supplement experimental data by GAN-based synthetic data for predicting tool wear.