Applicability Boundaries of Time-Series Foundation Models: Data-Layer Mismatch in Industrial Motor Fault Diagnosis

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Electric motors are key power components in industrial applications and motor-driven household appliances and power tools. Rapid health diagnosis based on short-term operational signals before delivery is critical for quality control and maintenance cost reduction. In practical manufacturing environments, however, this task remains challenging due to fault sample scarcity, class imbalance, and non-stationary operating conditions, which hinder reliable modeling and generalization. This paper presents a large-scale empirical study on the applicability of time-series foundation models (TSFMs), including Chronos and MOMENT, for industrial motor fault diagnosis using real-world factory end-of-line inspection data collected from 46 production motors. A comprehensive experimental framework is constructed, covering multiple adaptation strategies, patch sizes, aggregation methods, and downstream classifiers. The results consistently show that, under small-sample and non-stationary conditions, traditional approaches based on physically motivated features achieve more stable and often superior performance compared to TSFM-based representations in terms of Accuracy, F1-score, and AUROC. To interpret this observation, we introduce the notion of Data-Layer Mismatch, which describes the discrepancy between the statistical characteristics of pretraining data used by TSFMs and those of raw industrial sensor signals. While TSFMs are typically pretrained on relatively smooth and structured time-series data, industrial vibration and acoustic signals are characterized by strong noise, transient impulses, and pronounced non-stationarity. We use this concept as an empirical interpretation of the observed transfer difficulty, rather than as a universal theoretical claim. Further analysis indicates that diagnostic performance strongly depends on the alignment between model inductive biases and task-specific physical properties, including temporal scale, feature aggregation, and representation stability. These observations remain consistent across vibration signals, acoustic signals, and their feature-level fusion. The results are obtained under a specific industrial motor diagnosis setting and should be interpreted as context-dependent empirical evidence, highlighting the need for further investigation into domain adaptation strategies for foundation models in industrial applications.
Funding Information
  • Key Research and Development Program of Ningxia (2024Z294)

This publication has 36 references indexed in Scilit: