Hybrid Classification Method for Anomaly Detection in Injection Moulding Process

Ng Yao Xuan,Hailin Ge,Lu Wen Feng,Doan Ngoc Chi Nam,Xiang Li
DOI: https://doi.org/10.1109/iai63275.2024.10730246
2024-01-01
Abstract:Industrial and environmental challenges faced in injection moulding processes push manufacturing companies to embrace smart manufacturing and adopt machine learning to deal with anomaly detection. Challenges arisen from imbalanced dataset gives rise to the development of our novel hybrid classification method. A novel hybrid classification method for anomaly detection in injection moulding is proposed in this paper. The proposed method involves feature selection with K-Means Clustering to remove noise data prior to Multi-Layer Perceptron. Feature selection helps to reduce the high-dimension of the dataset. K-Means Clustering helps to remove noise data that will have similar patterns with data of another class label. Multi-Layer Perceptron is used to verify the training accuracy and to build train model. Real industrial dataset of injection moulding process from a local manufacturing company is used to demonstrate the effectiveness of the method in this paper. The proposed hybrid method outperforms other classical models such as Multi-Layer Perceptron (MLP), Decision Tree (DT) and CNN (Convolutional Neural Network). It successfully increases the failure detect rate and reduces false reject rate up to 90%.
What problem does this paper attempt to address?