Machine Learning for Pattern Analysis Using Continuous Secondary Data and Sparse Primary Data

  • Kelechi Okegbe

Student thesis: Doctoral Thesis


The harsh environment around manufacturing processes such as machining makes it difficult to extract primary measurement data that correlates directly with the machining conditions and also the quality of manufactured parts. The alternative to getting data directly from the cutting process is to get sparse data from post process inspection of components using a coordinate measuring machine or infrequent tests such as probing (carried out during machine down time) which ultimately causes unproductivity. However, there are opportunities to embed sensors into the machine to measure secondary parameters like vibration, motor power, and temperature in-process without disrupting machine operation. It must be noted that these secondary parameters have loose correlation due to process variability, high levels of noise and system non-linearity. A thermal imaging camera was used to decide the optimal spots on the CNC machine to place the temperature sensors. After the sensors have been placed on the CNC machine, Pearson’s correlation coefficients were used to decide on the sensors to be used to create the neural network models. This was a trade-off between having a high amount of data and having just the data from temperature sensors that are most correlated to the machining process. An FEA model was used to simulate the thermal properties of the CNC machine including the environmental effects of ambient temperature on the CNC. The data from CNC machine testing and FEA are used to create a hybrid neural network thermal error prediction system. The robustness of the system is tested by data from different feed rates and environmental conditions and compared to different machine tool axis. A hybrid neural network model is developed in this research. On testing, the hybrid neural network model achieved an average residual value between the actual thermal error and the predicted thermal error of 0.025um. The hybrid neural network model also showed a good fit 4to long term machining data with an RMSE of 1.36 compared to a transfer function model, and an ANFIS model that had RMSE of 2.08 and 1.60 respectively.
Date of Award19 Feb 2024
Original languageEnglish
SupervisorSimon Fletcher (Main Supervisor) & Andrew Longstaff (Co-Supervisor)

Cite this