Few-Shot Probe Card Fault Diagnosis Using a Time-Series Foundation Model

Time-series foundation models have recently shown strong zero-shot forecasting capability across diverse domains, but their usefulness for industrial structural health monitoring (SHM) remains unclear. This paper investigates whether TimesFM, a pretrained forecasting foundation model, can support fault diagnosis of semiconductor probe cards without training a task-specific deep classifier. The dataset contains 3,750 frequency-response samples from different virtual sensors, labeled as healthy baseline, screw loosening, or crack damage. TimesFM is used as a frozen forecaster: the final part of each sensor response is predicted from the preceding signal segment, and diagnostic features are extracted from the resulting forecast residuals. These features are evaluated in few-shot three-class classification and compared with raw statistical features, PCA-based raw features, and fold-wise MiniROCKET. The results show that TimesFM residuals provide an effective few-shot diagnostic representation. With 50 labeled examples per class, TimesFM residual features achieve 0.849 balanced accuracy on the full imbalanced dataset and 0.853 on a randomly balanced dataset, outperforming fold-wise MiniROCKET by 5.0 and 6.4 percentage points, respectively. These findings suggest that forecasting foundation models may reduce the amount of labeled data required for industrial sensor diagnosis, provided that they are used as frozen feature extractors rather than direct zero-shot classifiers. The study provides an initial evaluation of foundation-model-assisted probe card SHM and identifies forecast residuals as a promising representation for label-efficient fault classification.