Real-Time Tool Wear Prediction in Ultrasonic Machining Using Deep Learning and LSTM Networks

Abstract

Tool wear is a critical component of ensuring product quality, safety and efficiency of a process in advanced manufacturing as well as situations involving non-conventional processes like Ultrasonic Machining (USM) itself. Rapid development of Industry 4.0 highlights the necessity of predictive maintenance that is fuelled by smart models to minimise maintenance and operation costs. The recent breakthroughs in the field of deep learning have shown the powerful ability of the long short-term memory (LSTM) networks to process time-based sensor information and reliably forecast the evolution of a tool wear.

Our current paper suggests a complete, end -to-end system of the real-time prediction of tool wear in USM. The architecture provides a Python Flask server to receive and process sensor data, preprocess, and model it using a deep-learning framework, and to provide a React-based dashboard that provides interaction, uploading of data to view and manage results. High resolution sensor data, i.e. force, vibration and temperature data, undergo preprocessing and are then fed into a multilayer LSTM. The system allows them to train models, observe training dynamics, and receive predictions of the tool wear and warnings using a user-friendly interface.

Empirical tests that have been done on publicly accessible industrial datasets confirm that the suggested LSTM model beats the traditional machine-learning baselines achieving low mean absolute error rates and showing convergent stability. The web-based implementation also enables the engineers, manufacturers and researchers to interface with the predictive-maintenance solutions directly and visually thus facilitating more informed decision-making and potential integration with industrial internet of things ecosystems