A deep learning process anomaly detection approach with representative latent features for low discriminative and insufficient abnormal data

Anomaly detection in industrial processes is vital for yield improvement and cost reduction. With the development of sensor system and information technology, industrial big data provide opportunities to detect the abnormalities of processes and raise alarms by using operating parameters. However, the slight deviations in operating parameters and the insufficient abnormal data may hinder the effectiveness of existing anomaly detection models. To cope with the above problems, a more effective process anomaly detection framework combining shallow feature fusion learning with unsupervised deep learning is constructed. Specifically, the extracted statistical features that can reflect the slight deviations of operating parameters and the original measured features are firstly concatenated to enrich the available information. Then, a combined feature selection method of SMOTE & Tomek Links and random forest is developed to further discover the abstract features closely relevant to the quality characteristics of finished products with imbalanced data. After that, an unsupervised anomaly detection method is developed, where only normal process data are available for training the stacked denoising autoencoder. The utilized autoencoder can alleviate the effect of imbalanced data as the reconstruction error would be larger when the abnormality occurs. Lastly, the anomaly discrimination criteria, which consist of the monitoring index construction and the threshold determination, are formulated to detect the state of the production process. The experimental results demonstrate that the proposed method can detect the abnormalities effectively and achieves better performance than other state-of-art anomaly detection methods in commutator spot welding of a practical motor manufacturing process.