Real-Time Deep Anomaly Detection Framework for Multivariate Time-Series Data in Industrial IoT

The Industrial Internet of Things (IIoT) relies on interconnected devices to monitor, collect, analyze, and provide real-time insights into industrial processes. This data, generated by millions of devices, enables quick decision-making and intelligent automation. However, unexpected disruptions, computation, storage complexities, and network issues can lead to data anomalies, posing risks, especially in safety-critical scenarios. Real-time anomaly detection is crucial for identifying irregularities and improving industrial processes through actionable insights.

IoT anomaly detection is a process that uses machine learning (ML), deep learning (DL), and advanced ML/DL to identify patterns in time-series data that differ from the system's normal behavior. Designing a real-time system for IIoT environments faces challenges like dealing with temporal dependencies, high-dimensional data, hardware resource limitations, and identifying standard behavior patterns. The data at a given point in time is intricately linked to data from previous time points, near or distant, known as short-term and long-term dependencies.

The paper proposes a hybrid end-to-end deep anomaly detection (DAD) framework for detecting anomalies in IoT streaming data, aiming to enhance the existing system by simulating normal behavior and reconstructing errors.

The framework uses two critical deep learning models, including Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, to develop an advanced anomaly detection framework. CNNs are known for their success in computer vision tasks and are used to extract meaningful features from data. LSTM is a specialized type of Recurrent Neural Network (RNN) designed to handle time-series data efficiently, making it suitable for predicting future values.

The system architecture has three layers: an industrial setup with sensors and machines, an edge layer for real-time data processing, and a cloud layer for offloading processing. The process involves collecting data from IIoT sensors, preprocessing it, training DL models in the cloud, and deploying them on edge devices for real-time inference. In this process, CNN is used to extract high-level features, and LSTM-based AEs (Autoencoder) are used for further detailed analysis. The system continuously analyzes incoming data streams to detect anomalies based on learned standard patterns. The researchers also introduce the concept of a scoring function to identify anomalies based on the model's output.

The proposed DAD model was evaluated on publicly available datasets to measure the performance with metrics like accuracy, recall, precision, and F1-score. The model outperformed other existing models with 82.49% accuracy and 84.40% precision on the rare event dataset and 98.59% accuracy and 96.30% precision on the ECG5000 dataset.

It's also important to fine-tune specific settings, such as the sliding window length for best performance, the model's weights to minimize a loss function, etc. Various model versions, such as cells and dropouts, must be trained for selection to obtain maximum accuracy with optimal settings. Overall, the goal is to create a well-tailored model with optimal settings for accurate predictions.

The proposed anomaly detection system demonstrates efficiency in training and prediction times, with strong performance on diverse datasets. It is suitable for industries such as aerospace, manufacturing, and healthcare.