Document Type

Conference Proceeding

Publication Date



Department of Mechanical Engineering-Engineering Mechanics


This study aims to develop an Automated Data Processing (ADP) algorithm for real-world applications of machine learning (ML). Specially, this work seeks to contribute to the state-of-the-art Automated ML domain by providing a framework by which all the processing units of any ML application could be automatically determined based on a set of prior knowledge provided to the algorithm without external assistance by an ML developer. For this purpose, the ADP algorithm was designed consisting of a Reinforcement Learning (RL) agent as the actor-critic at the top level and a classification unit at the lower level for the data processing of the given raw data. The ADP algorithm is tested on a set of raw data extracted from several lathe turning experiments for the purpose of cutting tool-wear classification. The data are collected using acceleration, acoustics, and temperature sensors mounted on a Computer Numerical Controller (CNC) lathe machine. The ADP is implemented on the collected data and the RL agent is trained for several episodes. The performance accuracy of the classification unit on the test data is the criterion to define the reward function of the RL agent upon which the agent is trained. After the training is conducted the best action policy is extracted based on higher state values and the number by which each state is chosen as the optimal action policy at the end of each learning episode. The results show that the policy chosen by the ADP algorithm exhibits better classification performance on the given data and can outperform other policies different from the optimal policy selected by the algorithm. Therefore the proposed RL framework incorporates physics-based knowledge of a turning process to systematically automate data processing for ML applications.

Publication Title

Proceedings of the ASME Design Engineering Technical Conference