Organizations: Department of Electrical and Electronics Engineering, Antalya Bilim University, Antalya, Türkiye · Department of Computer Engineering, Antalya Bilim University, Antalya, Türkiye · Graduate School of Natural and Applied Sciences, Ankara University, Ankara, Türkiye · Department of Radiology, Feinberg School of Medicine, Northwestern University, Chicago, IL, 60611, USA
Industrial predictive maintenance increasingly depends on learning from equipment spread across sites whose sensor data cannot easily be pooled. Federated averaging (FedAvg) solves this with a central aggregation server; gossip learning removes the server, but its behaviour for recurrent failure-detection models has not been measured under controlled conditions. We compare synchronous ring gossip with FedAvg, isolated local training and a centralized reference for a stacked LSTM that detects imminent failure on the NASA C-MAPSS turbofan benchmark. All methods share one open implementation, architecture, initialization, optimizer, data split and training budget, and the primary endpoint uses one terminal window per test engine to avoid the statistical dependence of overlapping windows. On FD001 (five seeds), gossip reached a terminal-window F1 of 89.6 +/- 1.3%, compared with 89.9 +/- 1.1% for FedAvg, 83.6 +/- 6.7% for local training and 93.5 +/- 2.1% for centralized training, while transmitting the same payload as FedAvg without a coordinator. Node models agreed closely but not exactly (1.8% pairwise decision disagreement versus 5.6% without communication). Across FD002-FD004, peer communication improved terminal-window F1 over local training by 13-28 points; gossip matched FedAvg on FD003 and FD004 but was 4.3 points lower on the multi-condition FD002 subset. Simulated message loss, node failure and server outage changed neither method appreciably, whereas larger rings degraded gossip faster. Ring gossip is therefore a practical serverless alternative when data heterogeneity is moderate, and faster-mixing topologies become important as heterogeneity grows.
Figures & tables
Figure 1: Training protocols and evaluation design. (a)–(c) Local-only training, FedAvg and ring gossip on N=10 nodes. (d) Each communication round consists of one local epoch per node, a synchronization barrier and one mixing step. (e) For a truncated test engine with official terminal RUL R observed up to cycle Tobs , labels follow from RUL(t)=Tobs+R−t ; the primary endpoint scores only the terminal window of each engine, whereas the secondary analysis scores all overlapping windows.
Figure 2: Main comparison on FD001. (a) Terminal-window and (b) all-window F1 for each seed (dots), with the mean (black tick) and ± 1 SD (shaded bar). (c) Paired per-seed differences between ring gossip and each comparator on the terminal (circles) and all-window (squares) populations.
Figure 3: Agreement between node models on FD001 (mean over five seeds; shaded band: range across seeds). (a) Root-mean-square distance of node parameters from the network mean (log scale). (b) Pairwise decision disagreement of node models on a fixed probe of validation windows, evaluated every five rounds.
Figure 4: FedAvg and ring gossip under additional scenarios on FD001 (three seeds). (a) All-window F1 under faults, a lifetime-sorted non-IID partition and removal of the cycle input with N=10 (dots: seeds; black tick: mean; bar: ± 1 SD). (b) All-window F1 when the same training data are split across N=5 to 40 nodes.
Figure 5: Replication across all four C-MAPSS subsets with an unchanged protocol. Bars show the mean terminal-window (top) and all-window (bottom) F1; white dots are individual seeds.