Unsupervised Outlier Detection in Continuous Nonlinear Systems: Hybrid Approaches with Autoencoders and One-Class SVMs

Outlier detection in continuous nonlinear systems is essential as the presence of outliers might be indicators of faults, diseases, cyberattacks, or system malfunctions. However, the complex nature of such systems significantly increases the difficulty of developing outlier detection techniques. Due to the high complexity of such systems, accurate model based approaches are often difficult to design. While supervised outlier detection techniques yield great performance, the scarcity of labeled data motivates the necessity of unsupervised approaches. This paper introduces a data-driven approach for unsupervised outlier detection, which utilizes a hybrid combination of Autoencoders and One-Class Support Vector Machines. Experimental assessment was performed on the Tennessee Eastman Process dataset, and the performance of the proposed solution was measured using nine independent metrics, including detection delay and, true and false positive rates. Furthermore, a comparison with other recent techniques was performed, with notable results in terms of false alert rates and detection delay.