A frozen graph foundation model can flag multivariate time-series anomalies: Gfm4Mtsad reaches best or second-best on most metrics across three real datasets

A pretrained graph foundation model can serve unsupervised multivariate time-series anomaly detection through peripheral adaptation alone, without end-to-end retraining, emitting window-level anomaly scores plus sensor-level localization evidence. Graph detectors have generally trained task-specific networks from scratch on limited target-domain data, leaving learned topologies vulnerable to spurious edges and unable to transfer relational priors across systems. On three real datasets, MSL P-15, SMAP T-3 and SMD 1-1, with AnomalyGFM and OwlEye as backbones, comparison against nine baselines under identical data partitions, window labels and evaluation protocols yields best or second-best results on most metrics.

Adaptation proceeds in three steps: project each sensor's temporal features into the graph foundation model's native token space, constrain the window graph with connectivity priors mined from normal data, then restore high-frequency detail lost in projection at the model output. The SICA module combines a feature contract, a structure contract and an information contract, the last recovering projection-excluded components through a gated low-rank residual bypass while backbone weights stay frozen throughout. Ablations on MSL, SMAP and SMD with the OwlEye backbone, sharing one warm-start frontend and a 30-epoch supervision budget, show AUPRC and F1 drop on every entity when PCA token-bridge initialization or the IPRA residual bypass is removed.

Supervision without labels comes from local interventions on normal series: break temporal lags or neighbor collaboration on a few sensors while preserving empirical residuals, keeping only hard negatives near the normal boundary. Earlier augmentation heuristics injected random noise or shuffled global features, detached from real system spatio-temporal mechanisms, producing negatives that either overlap normal fluctuations or lie far outside the normal support. Replacing boundary ranking with uniform sampling among valid candidates lowers AUPRC consistently on MSL, SMAP and SMD, and further lowers F1 on MSL and SMD.

Window decisions no longer pool node responses naively; direct, local, relational and context evidence are each calibrated against their normal references, and only upper-tail candidates are re-ranked. Mean pooling dilutes sparse faults into missed detections while peak pooling magnifies benign local spikes into false alarms, so the normal score body needs to keep its original ranking. Comparing five scoring variants on identical frozen full checkpoints, TGEC outperforms global fusion and direct-only scoring on SMD and SMAP; on MSL-O, where no test window crosses the tail gate, TGEC retains the existing ECDF ranking while global fusion degrades markedly.

Evaluation is confined to three real series, MSL P-15, SMAP T-3 and SMD 1-1, spanning spacecraft telemetry and industrial server monitoring; training uses only anomaly-free windows, and metrics are AUROC, AUPRC and F1 without point adjustment.

The two backbones trade advantages by dataset and metric: the OwlEye instance is stronger on high-dimensional MSL, the AnomalyGFM instance steadier on SMAP and SMD, attributed to differences in native token dimensionality and scoring mechanisms. Sweeping the tail probability from 1% to 20% changes AUPRC by at most 0.0249 on SMD and SMAP, while on MSL broader tails favor AnomalyGFM and OwlEye keeps its highest AUPRC between 0.01 and 0.05. An IPRA bottleneck rank of 16 gives the highest mean native AUPRC across the six entity-backbone configurations; raising it to 32 or 64 adds parameters linearly without improving that mean.