One Foundation Model for Monitoring, Diagnosis, and Virtual Metrology

Anomaly detection, fault diagnosis, and virtual metrology on the shop floor are typically three separate models, each retrained from scratch whenever equipment or processes change. The team built an industrial time-series foundation model: self-supervised pre-training on unlabeled signals first teaches a shared backbone the general dynamics of time series; the backbone is then frozen and each downstream task attaches only a lightweight head, fine-tuned with minimal labeled data. Validated on real stamping and injection-molding lines, the single backbone delivers an anomaly-detection F1 of 97.0% with zero false alarms, a fault-diagnosis Macro-F1 of 94.6%, and a virtual-metrology RMSE of 0.0259 mm — all outperforming individually trained baselines.

1. 1

   Teach the Model to Read Signals Before Talking Tasks

   Pre-training mixes public time-series corpora together with manufacturing-process signals. The objective is to predict the next signal patch from the preceding context — no labels required. Industrial sites are starved of labels, yet this step turns massive unlabeled archives into training fuel.

2. 2

   Freeze the Backbone, Attach One Small Head per Task

   During fine-tuning the backbone weights stay fixed; only a task-specific lightweight module is trained. The trainable parameter count is orders of magnitude below full per-task fine-tuning, curbing overfitting on small samples and preserving the knowledge learned during pre-training.

3. 3

   Three Tasks, Three Readers, One Shared Representation

   The same backbone output is read differently by three heads. Anomaly detection reconstructs waveforms through a decoder and flags cycles whose reconstruction error exceeds a threshold set on normal training samples. Fault diagnosis uses attention pooling to weight and aggregate patches before classification. Virtual metrology aggregates likewise, then feeds a regression head to predict hard-to-measure quality indicators.

4. 4

   Attention Lands on the Process Stages That Actually Matter

   The model's attention weights map directly onto stamping process stages: slug-induced defects concentrate on the clamping and elastic-deformation stages, misfeed activates only the clamping stage, and mispunch falls on the tearing stage. Occlusion experiments confirm the alignment — masking the highest-attention patches sharply drops diagnostic accuracy, while masking random or lowest-attention patches leaves it almost unchanged.