Data Driven Anomaly Detection Framework for Industrial Cyber Physical Systems Using Deep Learning

Abstract

The current paper proposes an anomaly detector based on deep learning to monitor Industrial Cyber-Physical Systems (CPS), which uses both Long Short-Term Memory (LSTM) networks and autoencoders. It is assessed on the Industrial Anomaly Detection Dataset available on Kaggle, which is multivariate time-series data of sensors in the form of temperature, pressure, vibration, and current. The model proposed has a higher performance in comparison with the conventional methods, such as Support Vector Machines (SVM) and Decision Tree, with a precision (0.92), recall (0.89), F1-score (0.90), and the area under the ROC curve (AUC) of 0.94. Such outcomes indicate that the framework can be effective in identifying known and unknown anomalies in real-time to make sure that failures and performance degradation in industrial systems are detected in time. The hybrid model also utilizes LSTMs as a representation of temporal dependencies and autoencoders as a representation of normal behavior, and detects anomalies on the basis of reconstruction errors. The scaling of the model and its real-time processing capabilities ensure that it can be applicable in a wide range of industrial processes, both in small and large manufacturing factories. Further studies would help to enhance the model in terms of interpretability to allow a more effective decision-making process by the operator and to combine the framework with real-time monitoring and predictive maintenance. Finally, the suggested framework can contribute to the performance of operational efficiency, safety, and lower the expenses of maintenance in industrial CPS.