One self-supervised encoder lifts CWRU fault-classification macro-F1 from 0.350 to 0.812, but not remaining-useful-life accuracy

Multi-dataset self-supervised pretraining transfers to industrial fault classification, and the gain comes mainly from the representation rather than from extra fine-tuning optimization. PHM models are typically designed and evaluated per dataset and per task, which makes it hard to tell whether a reported gain reflects a transferable inductive bias or specialization to one controlled environment. On CWRU, withheld entirely from pretraining, the centralized encoder raises macro-F1 from 0.350 to 0.812, and on HSG18 from 0.569 to 0.951; a frozen encoder with a single linear layer already reaches 0.705 and 0.906.

The same pretraining gives no reliable benefit on remaining-useful-life estimation, and the conclusion depends on which error model is adopted. Under the protocol's validation-selection rule the best checkpoint, FedAvgM, tests at 44.04 RMSE, statistically indistinguishable from the centralized 100k-step checkpoint at 44.16, while the from-scratch reference reaches 43.67. On CMAPSS, full fine-tuning of the centralized encoder tests at 44.33 RMSE, 0.65 cycles worse than from scratch; the same single-seed comparison reverses under the benchmark's asymmetric score, which penalizes late predictions that overestimate remaining life more heavily, giving 645,389 for full fine-tuning against 754,754 from scratch.

Federated simulated pretraining retains part of the representation but does not close the gap to the centralized encoder at matched budget. The centralized 25k-step checkpoint leads in all four pairwise comparisons, by 16.2 macro-F1 points on CWRU and 15.5 on HSG18. Both arms fix architecture, objective, source data and local optimizer steps, but the centralized arm is a step-25,000 snapshot of a schedule dimensioned for 100,000 steps while the federated arm uses a constant learning rate, a mismatch that favours the centralized arm, and each regime has a single pretraining run.

The self-supervised reconstruction loss is not a valid criterion for selecting among pretrained checkpoints. The centralized encoder's validation reconstruction loss of 0.3467 is less than half the federated 0.7844, yet that gap has no counterpart on the CMAPSS downstream task. Sweeping FedAvgM server momentum moves the self-supervised validation loss within a band of 0.021 while downstream validation macro-F1 varies by 0.21 on CWRU and 0.29 on HSG18.

The results rest on 47 public PHM datasets harmonized to one tensor contract, 36 used only for pretraining and 11 withheld in full; multi-seed downstream evaluation covers three tasks, CWRU, HSG18 and CMAPSS, with an 801,808-parameter encoder, and the authors make no claim about scale.

The centralized 25k checkpoint and the federated arm differ in optimizer schedule, so the authors read the classification comparison as an upper bound on the cost of federating; each regime has one pretraining run and the three seeds vary only downstream adaptation; the from-scratch baselines use one learning rate and a single seed and are described as lower bounds; CMAPSS uses causal windows with stride 5 and every window is padded, departing from the pretraining contract; the CWRU test split lacks one of four fault classes, so macro-F1 averages three classes.