Virtual sensors for erroneous data repair in manufacturing a machine learning pipeline

Abstract

Manufacturing converts raw materials into finished products using machine tools for controlled material removal or deposition. It can be observed using sensors installed within and around machine tools. These sensors measure quantities, such as vibrations, cutting forces, temperature, currents, power consumption, and acoustic emission, to diagnose defects and enable zero-defect manufacturing as part of the Industry 4.0 vision. The continuity of high-quality sensor data streams is fundamental to predicting phenomena, such as geometric deformations, surface roughness, excessive coolant use, and imminent tool wear with adequate accuracy and appropriate timing. However, in practice, data acquired by some sensors can be of poor quality and unsuitable for prediction due to sensor faults stemming from environmental factors. In this paper, we answer if we can repair erroneous data in a faulty sensor based on data simultaneously available in redundant sensors that observe the same process. We present a machine learning pipeline to synthesize virtual sensors that can step in for faulty sensors to maintain reasonable quality and continuity in sensor data streams. We have validated the synthesized virtual sensors in four industrial case studies.