Perceiver IO Model for Efficient Vibration Signal Compression and Anomaly Detection

Abstract

This research presents a deep learning model for vibration signal compression and anomaly detection, using Perceiver IO—an attention-based architecture—within a railway industry case study. The model is lightweight and optimized for deployment on energy- and memory-constrained sensors, making it ideal for embedded applications. Unlike traditional methods, it performs signal compression and anomaly detection simultaneously, offering an efficient and unified solution applicable across industries. To avoid the need for separate classification models or heavy architectures, contrastive learning is used to improve feature extraction. This allows the learned representations to be easily separable via a simple linear function, enabling effective anomaly detection while preserving signal reconstruction quality. The model is trained on a large vibration signal dataset using a weighted loss function combining mean squared error (mse), cosine similarity, and contrastive loss. Contrastive learning plays a key role in clustering normal and abnormal signals, achieving near-perfect classification accuracy. Compression performance is evaluated with metrics such as peak signal-to-noise ratio (PSNR), cosine similarity, and compression ratio. The model achieves a PSNR of 21.87 while maintaining a much smaller architecture. The approach shows strong potential for real-time industrial deployment, ensuring high accuracy and efficiency in resource-constrained environments.