Edge-Based Real-Time IIoT Anomaly Detection Using Semi-Supervised CNN-Attention Model with Cross-Protocol Capabilities

Abstract

With the deepening of industrial digital transformation, industrial IoT network anomaly detection faces challenges such as high real-time requirements and diverse security threats, making traditional IT network methods difficult to directly apply; This paper aims to propose a real-time anomaly detection system based on edge computing, which can solve the problems of real-time, detection coverage, resource adaptation, cross protocol detection capability and adaptive learning mechanism. Methodologically, the system integrates edge computing and semi supervised learning technology, uses lightweight CNN model (three layers of convolution layer, core size 5, 3, 3, depth 32, 64, 128, ReLU activation and batch normalization) and Seq2Seq architecture to add attention mechanism for pre training and re training, combines random, systematic, and cluster sampling strategies to optimize data processing, and edge intelligent framework integrates dynamic computing unloading algorithm to optimize resource allocation; The experimental setup used NVIDIA Jetson AGX edge servers, industrial PC local devices, and Alibaba Cloud cloud servers, with software environments of Ubuntu, TensorFlow, and PyOD libraries. Validation was conducted on FastBee, Sagooiot, and Baowu Group IoT datasets. As shown in the results, the system accuracy reaches 91.6%, 93.8%, and 94.5%, respectively. In cross protocol detection, the recognition rates of abnormal traffic in Modbus, PROFINET, and OPCUA all exceeded 93% (OPCUA F1 score of 96.2%). Quantization technology reduces memory usage by 72.9%, latency by 65.5%, and accuracy by only 0.6%. The conclusion is that the system effectively improves real-time performance and accuracy, but there are limitations in dynamic load and cross vendor collaboration. In the future, load balancing algorithms and privacy protection frameworks will be optimized. The contribution lies in the innovative combination of edge computing and semi supervised learning to achieve a lightweight, high-precision and cross protocol anomaly detection solution for the industrial Internet of Things.