A Multifaceted Approach for Enhancing Anomaly Detection in Industrial Systems with Adaptive Synthetic Sampling and Machine Learning Evaluation

Abstract

Anomalies in industrial systems refer to deviations from expected behaviour, indicating potential malfunctions, faults, or security breaches. These anomalies can disrupt operations, compromise safety, and lead to costly downtime, making their detection and mitigation critical for industrial reliability and security. This research addresses the imperative task of enhancing anomaly detection within industrial systems by integrating a multifaceted approach encompassing various techniques and methodologies. A primary challenge in anomaly detection, the imbalanced distribution of data, is tackled by employing Adaptive Synthetic Sampling (ADASYN), which dynamically generates synthetic samples for minority classes. This effectively mitigates the impact of class imbalance and substantially enhances detection performance. Additionally, filter-based feature selection methods are utilized, incorporating statistical measures like the Chi-Square Test and ANOVA F-Value to identify pertinent features essential for anomaly detection. The study further delves into supervised anomaly detection approaches, leveraging machine learning algorithms such as Support Vector Machines (SVM) and Gradient Boosting (GB) to differentiate between normal and anomalous activities. These algorithms undergo rigorous evaluation within a Virtual Testing Environment (VTE) simulator, meticulously replicating industrial processes for comprehensive assessment. By leveraging the VTE simulator, this research ensures thorough evaluations across diverse operational scenarios, thereby fortifying anomaly detection systems and advancing the security of industrial systems. The research evaluates anomaly detection models for industrial systems using performance metrics like precision, recall, F1-score, and AUC-ROC. It compares Support Vector Machine (SVM) and Gradient Boosting (GB) models, finding GB outperforms SVM in discrimination ability. Both models show effective anomaly detection with minimal false positives and false negatives. A comparative analysis of key performance metrics aids in selecting the most suitable approach for enhancing system reliability and safety.