A High-accuracy Hyperdimensional Computing Pipeline for IoT Anomaly Detection Based on a High-quality Real-World Dataset

Abstract This work introduces an anomaly classification framework based on hyperdimensional computing (HDC), specifically designed for industrial monitoring scenarios. Inspired by cognitive computing principles, HDC leverages high-dimensional vector representations to achieve robustness, low computational complexity, and hardware-efficient operation. The proposed framework covers the complete processing pipeline, from data encoding and learning to anomaly classification. The HDC framework has been implemented on two platforms, a PC and a Raspberry PI 4 (RPI), which demonstrates the feasibility of its implementation on systems with limited resources. The framework is evaluated using a real-world dataset collected from 118 emergency lighting devices, originally comprising 126 monitored features recorded over multiple days of operation. An exploratory analysis and data processing stage is conducted to assess the relevance of the available variables and to remove those that do not contribute meaningful information to the anomaly detection task. This data curation process has a direct impact on the learning and classification stages, leading to improved performance. We have evaluated the HDC framework in terms of memory consumptions, training and testing times and accuracy. Experimental results show that the proposed approach achieves high classification accuracy (close to 98% in the best cases), confirming its suitability for deployment in scalable industrial systems with limited computational resources.