TSFM in-context learning for time-series classification of bearing-health status

TL;DR

The paper presents a framework to classify bearing health states from vibration signals using time-series foundation models (TSFMs) with in-context learning. By encoding spectral features as covariates and targets within few-shot prompts, a General Time Transformer (GTT) can forecast the distribution of health states without finetuning, using a Gaussian mixture forecast head and a sink token for targets. Experiments on a servo-press motor dataset show about 96% accuracy across four health classes, despite the data not being part of the model’s training corpus, demonstrating strong generalization and scalability. This approach moves toward AI-driven maintenance by enabling rapid, asset-agnostic health classification across varied operating conditions with minimal labeled data.

Abstract

This paper introduces a classification method using in-context learning in time-series foundation models (TSFM). We show how data, which was not part of the TSFM training data corpus, can be classified without the need of finetuning the model. Examples are represented in the form of targets (class id) and covariates (data matrix) within the prompt of the model, which enables to classify an unknown covariate data pattern alongside the forecast axis through in-context learning. We apply this method to vibration data for assessing the health state of a bearing within a servo-press motor. The method transforms frequency domain reference signals into pseudo time-series patterns, generates aligned covariate and target signals, and uses the TSFM to predict probabilities how classified data corresponds to predefined labels. Leveraging the scalability of pre-trained models this method demonstrates efficacy across varied operational conditions. This marks significant progress beyond custom narrow AI solutions towards broader, AI-driven maintenance systems.

Figures (4)

• Figure 1: A sequence of random samples from the servo-press dataset with one-hot encoded targets (orange) and corresponding covariates (blue) forms the few-shot prompting context. The model outputs a prediction of the targets w.r.t. covariates in the forecast horizon (blue block to the right of the dashed line).
• Figure 2: Preprocessing: Signal spectrum (magnitude of FFT) is transformed into a $60\times64$ matrix ($N=60$ data channels and $M=64$ frequency sub-bands per channel).
• Figure 3: Example FFTs from the servo-press dataset. Note similarity between 1_Normal and 3_SandBearing and between 2_OuterRing and 4_InnerRing.
• Figure 4: Correct classification of class 2_OuterRing (upper right), based on corresponding covariate matrix available in the forecast part of the GTT. Note, the prediction intensity stays/develops towards 0 for all other classes.