This paper proposes a novel foundation model, ECHO, designed for modeling general machine signals with arbitrary sampling rates. This model integrates an advanced band-segmentation architecture with frequency-local embedding to capture spectral localization, and uses sliding patches to support variable-length inputs without padding or truncation. Experimental results on the DCASE task and industrial signal datasets demonstrate state-of-the-art performance in machine signal anomaly detection and fault classification, demonstrating the model's effectiveness and generalization ability.
