Anomaly detection in industrial processes using curriculum-based deep learning

Anomaly detection is increasingly drawing attention in research as anomalies encountered in the industry become commonplace and more complex. In industrial processes, high dimensional multivariate time series (MVTS) data is generated from sensors and actuators. To guarantee the safety of the equipment and high quality products, anomalies recorded by these components have to be detected. In this thesis, the use of deep learning for industrial anomaly detection is studied. Particularly, curriculum learning (CL), a training paradigm for machine learning models is comprehensively analyzed. CL follows the learning progression of humans and animals, in which the model is first trained with "easy" data and more "dıffıcult" data is gradually introduced as training advances. This thesis proposes data-based CL for MVTS anomaly detection and introduces the CL concept to the learner (model). These frameworks are applied to two different case studies, first using a benchmark secure water treatment dataset and then a real world textile dataset. In the data-based method, extensive experiments are conducted to determine the most important MVTS features for improved model performance. In the model-based framework, a simple Long Short-Term Memory Encoder-Decoder model is trained first and then a complex Multi-Scale Convolutional Recurrent Encoder-Decoder model in utilized in the final training round. These techniques are evaluated based on the F1 score, with a data-based approach producing the best results among the proposed designs.