SPIRAL2 LLRF · heat-load observer sub-investigation

Does the heat-load observer see a fault coming?

Six neural-network heat-load models — dense, CNN, LSTM, LSTM+attention, and two meta-learner stacks, trained independently per cryomodule on 2022 thermal-balance data — folded into the existing EPICS precursor methodology as six new candidate channels, one per architecture. Same matched-pair distance, temporal-block classifier, VAE latent space, and LSTM-attention methods, same 31 (fault type × scope) combinations, run again with these six channels added to the original 23. Click any row for the full breakdown.

← back to the base Precursor Root-Cause Atlas (23-channel study)

Why sensitivity, not magnitude. CMA02 and CMA11 are known to run genuinely higher heat loads than the rest of the fleet — real physical differences, not a modeling artifact. This study isn't chasing an absolute-Watts calibration (separate future work); it asks whether the relative movement of each architecture's output in the hours before a fault differs from its movement before normal operation.

The existing archived channel OBSCHARGE is a different, unrelated heat-load estimate already in the base 23-channel ensemble — the two are never combined or confused here.

Modules CMA02–CMA12 only (11 of 19) Events w/ observer output 759 Folded into matching population 623 Window 4h pre-trigger, 16×15-min bins New channels 6 (one per architecture)

Headline findings

Curated virtual observer: does heat load earn a place in the top-8?

Same curation test as the base study — take the LSTM's top-8 channels by permutation importance (now chosen from all 29 candidates) and retrain a smaller model. Heat-load channels are highlighted where they make the cut.

fault type	full AUC (29 ch.)	curated AUC (8 ch.)	Δ	curated channels

Explore by fault type

All 31 scopes from the base study, re-run with the 6 heat-load channels added. no heat-load coverage marks scopes where fewer than 15 matched pairs had usable observer output for any architecture (too sparse for that specific test, not a null result).

scope	n fault	n normal	classifier AUC	matched-pair median	wilcoxon p	frac. positive

Method and honesty notes

All 6 models are per-cryomodule (trained separately for each of CMA02–12) and only exist for those 11 modules — every statistic on this page excludes CMA01 and the whole CMB side for the heat-load channels specifically; the other 23 channels retain their original full-fleet coverage.
The classical per-channel result is the strongest evidence here, and it is a single-method finding for its two fault types — VAE has no per-channel view to confirm or deny it, the joint LSTM model doesn't rank a heat-load channel in its top-10 for either, and curation surfaces heat-load channels for a different pair of fault types (quench, RF-safety-threshold) where the classical method has no data at all (too few matched pairs). Read this as real-but-single-method evidence, not the multi-method corroboration the base study found for e.g. "RF authorization missing."
Same small-n caution as the base study: classifier AUCs from slices with n<30 can swing to extreme values from block-CV instability alone — flagged with ⚠.
759 events had usable heat-load observer output; 623 overlap the base study's in-era matching population (the remainder fall just outside its 2021-06–2024-10 window).

Glossary

Heat-load observer: A neural network trained to estimate a cryomodule's real-time thermal dissipation from cryogenic sensor readings (temperatures, valve positions, pressure, level) — a virtual sensor, not a direct measurement. Six independently-trained architectures are compared here.
AUC (area under the ROC curve): How well a classifier separates fault from Normal, 0 to 1. 0.5 = coin flip, 1.0 = perfect. Treat anything near 1.0 from a small slice with suspicion (see the small-n caution), not extra confidence.
Matched-pair median: For each fault event, we found a Normal event from the same (or side-matched) cryomodule close in time, and took the fault event's anomaly-distance score minus its matched partner's. Positive = the fault event looked more unusual at the same point in time; negative = the opposite.
Wilcoxon p / Bonferroni correction: Wilcoxon tests whether matched-pair differences are consistently on one side of zero. Bonferroni tightens the bar in proportion to how many channels are tested at once (29 here), so a channel marked ** cleared that stricter bar; a single * passed the ordinary p<0.05 test but not the corrected one.
Fraction positive: Of all matched pairs, the fraction where the fault event scored higher than its Normal partner. 0.5 is a coin flip; consistently far from 0.5 is what makes a signal believable even when the raw magnitude test is noisy.
Temporal-block cross-validation: Data split into contiguous date chunks; the classifier trains on all but one and is scored on the one it never saw, rotating through every chunk once — prevents it from learning "which era" instead of the actual fault physics.
VAE (variational autoencoder) latent space: A nonlinear alternative to the classical linear (PCA) method: a small neural network compresses each 1-hour window's full channel ensemble to 8 numbers and reconstructs it back. A downstream classifier on those 8 numbers gives an independent AUC per window.
LSTM with attention: Sees the full ordered 16-step sequence across the whole 4h and learns which timesteps to weight most heavily. The chart shows fault attention minus Normal attention at each timestep.
Permutation importance: Shuffle one channel's values across all events and measure how much the trained model's AUC drops — a bigger drop means that channel mattered more to that model.
Curated model: A smaller model retrained on only the top-8 channels by permutation importance, testing whether the full ensemble was mostly carrying dead weight for that fault type.