Technical Summary: MAFL-PdM

Problem Statement

Industrial Internet of Things (IIoT) environments increasingly rely on Predictive Maintenance (PdM) to optimize operational efficiency and estimate Remaining Useful Life (RUL). However, cross-site collaboration is hindered by data sovereignty, proprietary constraints, and regulatory requirements (e.g., GDPR), which prevent the centralization of raw sensor data. While Federated Learning (FL) offers a solution by allowing distributed training without data sharing, applying standard FL protocols to heterogeneous industrial sites presents two critical challenges:

1. Data Heterogeneity (Non-IID): Industrial sites differ in machine types, sensing configurations, and degradation patterns. Standard aggregation strategies (e.g., FedAvg) often suffer from unstable convergence and reduced model quality when local updates move in conflicting directions.
2. Privacy Adaptation: Existing FL frameworks often use static, uniform privacy settings. In PdM, where rare fault signatures are critical, indiscriminate noise injection can disproportionately degrade failure-sensitive patterns. Furthermore, current approaches often treat factories as passive clients rather than autonomous agents with distinct operational constraints.

Methodology: MAFL-PdM

The paper proposes MAFL-PdM, a Multi-Agent Federated Learning framework designed for personalized PdM across heterogeneous sites. The system architecture consists of four autonomous Factory Agents and a Federated Orchestrator Agent.

Core Components

1. Local Model Architecture: Each Factory Agent trains a local model using a two-layer Long Short-Term Memory (LSTM) network with temporal attention. The model predicts RUL using a weighted Huber loss combined with FedProx regularisation to constrain local optimization.
2. Cluster-Driven Aggregation (CMFL-PdM): The Orchestrator performs K-means clustering (K=2) on the cosine similarity of gradient updates to group factories with similar operational behaviors. Aggregation is performed with an 80/20 weighting between the primary and minority clusters to prevent the exclusion of minority sites.
3. Post-Federation Personalisation: After 30 federated rounds, each Factory Agent fine-tunes the global consensus model on its own local data for eight additional epochs. This step is crucial as it adapts the shared model to local data distributions without further gradient sharing.
4. Privacy Accounting: The framework implements per-factory differential privacy (DP) accounting using a Rényi accountant (Opacus). However, in the reported configuration, gradient noise injection is disabled. The system tracks privacy budgets ($\epsilon = [3.0, 5.0, 8.0, 7.0]$) but does not enforce formal DP guarantees in the current experiments.
5. Drift-Aware Selection: Factories monitor local data drift using Kolmogorov–Smirnov tests, allowing the orchestrator to prioritize clients with significant distribution shifts.

Datasets

The framework is evaluated on two heterogeneous real-world industrial datasets:

• Elevator Door Monitoring: Three factories (Factories 0–2) using door ball-bearing revolutions, humidity, and vibration data.
• Metalworking Fluid Monitoring: One factory (Factory 3) using pH, conductivity, concentration, and temperature data.
• RUL Labels: Synthetic stochastic signals generated via an exponentially decaying curve with additive Gaussian noise, normalized to [0, 1].

Key Results

Experiments were conducted over five random seeds (42–46) comparing MAFL-PdM against centralized training, local-only training, vanilla FedAvg, FedProx, and noise-perturbed baselines.

• Performance: MAFL-PdM achieved an RMSE of 0.0498 ± 0.0012 and an F1 score of 0.819 ± 0.016.
  • It attained a lower RMSE than centralized training in all five seeds.
  • It matched centralized training on F1 score while preserving data locality.
  • It significantly outperformed all federated baselines (e.g., Vanilla FedAvg RMSE: 0.2444).
• Ablation Study: A component ablation revealed that post-federation personalisation accounts for essentially the entire performance gain.
  • Removing personalisation caused F1 to drop by 0.457, rendering the framework indistinguishable from vanilla FedAvg.
  • Removing cluster-aware aggregation and failure-sensitive loss weighting resulted in negligible changes (F1 drops of -0.0019 and -0.0096, respectively), falling within seed-to-seed variation.
• Clustering Analysis: The K-means clustering mechanism failed to consistently recover the intended dataset partition (Elevator vs. Metalworking). Instead, it grouped factories based on the difficulty of fitting the data (e.g., grouping the cross-domain Factory 3 with the hardest-to-fit Elevator Factory 2). Consequently, the clustering mechanism contributed no measurable benefit at this small federation size (N=4).

Significance and Claims

The paper makes the following specific claims regarding its contribution and limitations:

1. Primary Mechanism: The study identifies post-federation personalisation as the critical mechanism enabling federated PdM to work across heterogeneous sites, rather than the aggregation strategy or loss weighting. It demonstrates that a shared initialization followed by local fine-tuning can outperform centralized pooling in regression error (RMSE) while matching it in classification performance (F1).
2. Negative Result on Clustering: The paper explicitly reports that cluster-aware aggregation provided no measurable gain in this specific experimental setting (4 clients, K=2). It attributes this to the small federation size and the instability of the clustering partition, suggesting that the mechanism requires a larger scale to meaningfully redistribute aggregation weights.
3. Privacy Stance: The authors are explicit that no formal differential privacy guarantee is claimed for the reported results. While the framework includes DP accounting and budget tracking, gradient noise injection was disabled. The results represent a system that preserves data locality but does not yet satisfy the mathematical requirements of DP-SGD (due to issues with per-sample clipping and subsampling assumptions).
4. Reproducibility: The paper provides a fully reproducible pipeline, including code, datasets, and scripts to regenerate all tables and figures, highlighting the specific conditions under which the results hold.

Limitations

The authors acknowledge several constraints:

• Synthetic Labels: RUL labels are synthetic, limiting the applicability of results to real-world deployment without validation on ground-truth degradation datasets.
• Privacy Gap: The lack of noise injection means the system is not formally differentially private.
• Scale: The evaluation is limited to four factories; the benefits of clustering are unproven at this scale.
• Feature Misalignment: Factory 3 (metalworking) had to have its feature space truncated to match the other factories, resulting in semantic misalignment.
• Statistical Significance: With only five seeds, the comparison against centralized training reaches a two-sided p-value of 0.063, just missing the conventional 0.05 threshold.