Machine Learning Based Search for Access Points in Anomaly Detection Model

Abstract

Abstract Heterogeneous systems are characterized by a high degree of complexity that poses challenges to conventional anomaly detection methods. One of the challenges is the production of large amounts of data where Deep Learning based Anomaly Detection Models (DLAD) have been demonstrated to outperform conventional techniques in detecting anomalies. However, DLAD models require the manual selection of Access Points, which are specific points in a system from which the training data is recorded. The selection of Access Points is a critical task that can significantly affect the performance of the anomaly detection models. It requires domain knowledge and expertise in the intricacies of the system, which is difficult to acquire and prone to human errors. As the size of the system grows, selecting Access Points becomes increasingly challenging. In this paper, we propose a new machine learning-based algorithm called the Access Point Search Algorithm (APSA). The aim of APSA is to automate the task of finding the optimal set of Access Points that can aid DLAD models in detecting anomalies in a particular system. The algorithm utilizes a special error detector that dynamically takes in multiple signals and forecasts the signal values. The objective of finding the optimal set of Access Points is formulated as a feature selection problem, which is supervised using a binary variant of the Grasshopper Optimization Algorithm (GOA). We demonstrate the feasibility and effectiveness of the proposed algorithm by deploying it in a Simulink model. We illustrate the reliability of the proposed algorithm by feeding the signals from all the Access Points in the set provided by APSA into the DLAD model one by one. The reliability is further discussed by carrying out a fault injection experiment on the Simulink model. The proposed algorithm was able to reduce 80% of the Access Points. The Access Points selected by APSA showed a high probability of detecting anomalies over the Access Points that were not selected. The results suggest that the proposed algorithm can efficiently and effectively select the optimal set of Access Points for DLAD models to detect anomalies in component signals. It offers a promising solution to automate the tedious and error-prone task of selecting Access Points, thereby reducing the domain knowledge and expertise required.