Deep Learning for Intelligent CNC Machine Condition Monitoring and Prognostics

Abstract

The study of Condition Monitoring (CM) and Prognostics and Health Management (PHM) of manufacturing systems have been prevalent interests in research of intelligent manufacturing assets, enabling machines to predict impending failures, characterise normal or abnormal operating conditions, and estimate the Remaining Useful Life (RUL) of the system and component. Nevertheless, supervised deep learning (DL) models, praised for their ability to learn features from raw data, typically leverage extensive amounts of labelled data and computational resources to be trained for their intended tasks. Engineered systems experience degradation phenomena whose progression abide by certain physical characteristics, necessitating different methods of modelling the supervised DL predictive problem. Furthermore, uncertainty from ambient environmental conditions causes operational errors and may impact the ongoing machining process.Tool wear is a significant cost burden in machining, increasing production costs and lead times through resource-inefficient time-based tool replacement processes. Towards developing intelligent predictive capabilities for CNC machine tool wear and RUL prediction from visual inspection data and machining process parameters (MPPs), a novel DL methodology was developed integrating Deep Transfer Learning (DTL) and multimodal sequence learning. The experimental data used for this research pertains to the side milling of AlCrSiN. On one hand, DTL overcomes high labelled data volume limitations which are typically costly to collect in a manufacturing setting. In this work, a variety of intermediate training tasks are proposed based on weakly supervised learning, to improve downstream task performance. On the other hand, additional modalities (e.g. MPPs) which encode further process knowledge could further supplement models with limited data in one modality. The proposed methodology attains test performance of 0.052 MAE (mm), 0.0074 RMSE (mm) and 0.848% R2 on 5 test sequences (of 25 total experimental trials) for tool wear prediction. Meanwhile, for tool RUL prediction, test performance of 18.83 MAE (min), 21.70 RMSE (min) and 0.984% R2 was achieved on the same sequences using all the features as inputs to the model. These results and furtheranalyses based on MPP ablation settings and prognostics analytics indicate the viability of the proposed methodology for intelligent prognostics based on deep feature learningTowards intelligent CNC machine thermal error prediction and compensation, a novel deep semi-supervised learning approach was developed based on temperature sensor data modelling. By first learning a sparse representation of sensor data which enables the data to be reconstructed from a lower-dimension encoding, a deep unsupervised learning model based on sparse Autoencoders (SAEs) was developed. This model was used to encode the high-dimensional sensor data space into a lower dimension, thus mitigating against potentially collinear observations which may induce further uncertainty into the predictions. Subsequently, a sequence learning framework based on recurrent models was used to predict the thermal error of the CNC machine in an idle air cutting setting, using the reconstructed sensor data and lagged response as inputs. The proposed methodology attains 0.00292 MAE (mm), 0.0478 RMSE (mm), and 96% R2, offering competitive results in terms of prediction accuracy and computational resource usage compared to several approaches from the literature.The developed approaches in this work can be readily integrated into intelligent self-predictive and PHM software tools for CNC machines, which can further enhance CNC machine performance, resource efficiency, and robustness to ambient environmental conditions.

Date of Award	Dec 2021
Original language	English
Awarding Institution	Coventry University
Sponsors	Unipart Group
Supervisor	Weidong Li (Supervisor), Nazaraf Shah (Supervisor) & Kuo-Ming Chao (Supervisor)