Development of machining condition monitoring system with multi-domain feature fusion-based multi-scale CNN

• Main Author: Voon, Jamie Ji Xiang
• Other Authors: Chen Chun-Hsien
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Industrial engineering::Information systems
• Online Access: https://hdl.handle.net/10356/157858
• Institution: Nanyang Technological University

The manufacturing industry has a long history and has always incorporated new concepts and technologies to improve production efficiency or to unlock new methods for new products to be created. Machining condition monitoring has therefore always been a crucial part of the manufacturing process, to ensure quality and efficiency on the shop floor with the goal of increasing yield and reducing waste. The objective of this project is to develop a machine learning model that will improve machining condition monitoring. Specifically, the model should be able to identify the processes based on sensory input provided by the industrial Internet-of-Things technology. The proposed approach consists of two main components. The first one is the Multi-Scale Convolution Neural Network (MCNN). CNN can extract distinct features and patterns from the sensory input that will be used to train the model. With the model trained on these features, it can then identify these features in real data, and this would aide in the classification process. The multi-scale refers to the variety of filters applied to the CNN process to allow the model to extract more comprehensive features that exist in the input data and improve the performance of the model’s ability to perform condition identification. The second component is an Attention mechanism. This module will allow the model to focus on specific parts of the input data, bringing features of interest to the foreground while pushing the rest to the background. Specifically in this project, self-attention will be used to make the model focus on the important parts of the input. Self-attention is the model’s ability to use neighbouring data points and identify if any relationship between points exists. The self-attention will be implemented in the form of a multi-head attention function that will allow the model to bring different portions of the input into focus. An additional feature of the proposed solution is the use of multi-domain data. Multi-domain refers to the extraction of additional features from the original time series data set. In the proposed solution, the frequency and wavelet domains were extracted and used as input to the model in conjunction with the time domain. The proposed solution will be validated using real data collected from a Computerized Numerical Control (CNC) machining set up. The data set will consist of 5 vibration signals as a result of variation of CNC machining parameters such as spindle speed, feed rate, path space and the depth of cut. These parameters were varied randomly within a specified range. The resultant vibration signals will then be transformed into both the frequency and wavelet domain and fed into the models as inputs. Once trained, the proposed MCNN, combined with the attention module, was compared with traditional CNN models to validate the performance of the proposed solution. Two traditional methods were used as comparison. CNN, using multi-domain input and single kernel size obtained a 97.1% classification accuracy and an MCNN, using single domain would have an accuracy ranging from 90.7% to 95.5%, depending on domain used as input. The proposed solution performed better with classification accuracy of 97.8%. However, without the attention module, the MCNN function achieved an accuracy result of 98.5%. The results show that the multi-domain input and use of multi-kernel is beneficial in improving the performance of models however, the use of attention may require further development.