FusionNet Lite: Compact Federated Deep Learning for Predictive Maintenance under Label Heterogeneity and Temporal Generalization

Technical Summary: FusionNet Lite for Predictive Maintenance under Label Heterogeneity and Temporal Generalization

Problem Statement
Federated Learning (FL) offers a pathway for collaborative predictive maintenance (PdM) by training models across distributed industrial data without sharing raw observations. However, existing approaches face significant challenges: client data is often non-independent and non-identically distributed (non-IID), particularly regarding label heterogeneity; temporal dependencies and rare failure events complicate evaluation; and the need for computational efficiency on edge devices conflicts with the demand for high predictive accuracy and explainability. A critical gap exists in determining whether a compact deep learning model can maintain useful failure detection and consistent feature attribution explanations under these constrained, federated conditions.

Methodology
The study evaluates FusionNet Lite, a compact feature-axis classifier with only 1,420 trainable parameters, against four lightweight baselines: Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), Multilayer Perceptron (MLP), and CNN-LSTM.

• Architecture: FusionNet Lite processes individual observations (shape $F \times 1$) using separable convolutions, ConvMixer-style feature mixing, residual connections, and Squeeze-and-Excitation (SE) channel reweighting. It is designed to operate on a single observation rather than a temporal sequence, though chronological splits are used for evaluation.
• Datasets: Two datasets were utilized with distinct partitioning strategies:
  • AI4I 2020: A synthetic dataset with stratified random splits (65/15/20).
  • MetroPT3: A real-world compressor recording with a fixed-row chronological split (55/20/25) to simulate temporal generalization.
• Experimental Design:
  • Training Modes: Models were trained under both centralized settings and simulated synchronous Federated Averaging (FedAvg) with three clients.
  • Label Heterogeneity: Client partitions were generated using a Dirichlet distribution ($\alpha=0.5$) to create controlled label non-IID conditions.
  • Evaluation: The study employed ten matched random seeds for core comparisons. Checkpoints were selected based on validation AUROC. Performance was measured using AUROC, AUPRC, and F2 scores (with expanded grid threshold selection to address class imbalance).
  • Explainability: Class-specific Integrated Gradients (IG) were computed to compare attribution agreement between centralized and federated models.
  • Efficiency: An Energy Efficiency Index (EEI), defined as the product of parameter count and measured batch inference latency, was used as a proxy for computational cost.

Key Contributions

1. Compact Architecture: Introduction and evaluation of a 1,420-parameter architecture combining established operations (separable conv, SE, residuals) into a minimal footprint suitable for federated exchange.
2. Controlled Benchmark: A rigorous comparison across five architectures and two datasets using ten matched seeds, validation-selected checkpoints, and specific threshold policies (F1 vs. F2 optimization) to isolate architectural effects from tuning artifacts.
3. Sensitivity Analysis: Systematic examination of architectural components (ablations), feature ordering, client count (3 vs. 10), label concentration ($\alpha$), participation rates, and learning rates.
4. Attribution Agreement: A quantitative assessment of central-federated agreement using class-specific IG correlations, sign agreement, and Jaccard overlap, distinguishing between detection quality and explanatory consistency.

Results

• Predictive Performance:
  • On AI4I, FusionNet Lite achieved competitive results under FedAvg (AUROC $0.8289 \pm 0.0621$, F2 $0.3508 \pm 0.0908$), though it did not show statistically significant superiority over MLP across all metrics after multiplicity adjustment.
  • On MetroPT3, FusionNet Lite achieved high ranking discrimination (AUROC $0.9351 \pm 0.0462$ under FedAvg) but exhibited lower and more variable rare-failure detection (AUPRC $0.2197 \pm 0.1971$, F2 $0.1566 \pm 0.2044$) compared to BiLSTM and CNN-LSTM baselines.
  • Threshold Sensitivity: Switching from original-grid F1 selection to expanded-grid F2 selection significantly improved mean test F2 on MetroPT3 (from $0.0423$ to $0.1566$), though variability remained high.
• Efficiency: FusionNet Lite possessed the smallest parameter count and the lowest fixed-batch EEI (parameter-latency index) among all five architectures on both datasets, confirming its efficiency as a proxy for resource-constrained deployment.
• Attribution: Central-federated attribution agreement varied by model and dataset. On MetroPT3, FusionNet Lite showed weaker and more variable IG agreement (Spearman correlation $\approx 0.24-0.27$) compared to MLP, despite MLP's weaker predictive performance. This indicates that high explanatory agreement does not necessarily correlate with superior detection quality.
• Sensitivity: Performance rankings were sensitive to feature ordering, client allocation, and learning rates. Removing specific components (e.g., SE attention or stem residuals) yielded mixed results, sometimes improving performance on one dataset while degrading it on another.

Significance and Claims
The paper claims to establish a conditional performance–complexity trade-off. It demonstrates that a substantially smaller model (FusionNet Lite) can provide useful predictive performance and the lowest computational cost proxy (EEI) under controlled federated conditions. However, the authors explicitly state that:

• Compactness alone does not guarantee superior failure detection or consistent explanations.
• High AUROC does not equate to reliable rare-failure detection in imbalanced, temporal settings.
• Explanatory agreement (IG) is not a direct surrogate for predictive utility.
• The findings are bounded by the specific benchmark conditions (synthetic and single-record real data, synchronous FedAvg, fixed seeds).

The study concludes that while FusionNet Lite offers a viable, efficient option for federated PdM, its deployment requires careful consideration of operating points (thresholds), data heterogeneity, and the distinction between ranking discrimination and actual failure detection. The authors do not claim the model is universally superior or ready for immediate industrial deployment without further independent validation on diverse equipment and failure episodes.