Deep-dive AI/ML methods for predictive maintenance (2022–2025)

The landscape of Predictive Maintenance (PdM) has undergone a seismic shift between 2022 and 2025. Driven by the need for more reliable energy grids and industrial efficiency, new architectures have emerged that push the boundaries of what's possible in forecasting equipment failure.

Transformer architectures dominate RUL prediction

The attention mechanism's ability to capture long-range dependencies in multivariate time series has made transformer-based architectures the dominant paradigm for remaining useful life (RUL) prediction. On the benchmark NASA C-MAPSS turbofan dataset (four subsets: FD001 through FD004, ranging from single to six operating conditions and one to two fault modes), arXiv state-of-the-art models achieve remarkable precision.

The TMSCNN (Transformer-based Multiscale CNN, 2024) achieves RMSE of 11.3 on FD001 and a state-of-the-art 14.25 on FD004 — the most challenging subset — with an average 8.19% RMSE reduction across all subsets compared to prior methods. The architecture combines MMD-based domain adaptation for transfer learning, addressing the critical challenge of limited fault data in real deployments. TCRSCANet (2024) reduces RMSE by over 23% and Score by over 49% on FD002 versus prior SOTA, using temporal convolution with recurrent skip components and attention.

The DSFormer (Dual-Scale Transformer, 2024) introduces a dual-attention module extracting features across both sensor and temporal dimensions simultaneously, while GAT-DAT (2024) integrates Graph Attention Networks with Deep Adaptive Transformers, capturing spatial relationships between sensors — directly applicable to multi-sensor equipment monitoring.

For battery RUL specifically, the SOH-KLSTM (2025) hybrid LSTM-KAN architecture achieves RMSE of 0.001682 on NASA B0005, a 97% improvement over standalone LSTM. The CNN-WNN-WBiLSTM-AM (2025) achieves MAPE of 0.0042 on NASA B0018 with error reductions of 37.6–78.2% versus CNN-BiLSTM baselines.

Performance Comparison

Method	Architecture	FD001 RMSE	FD002 RMSE	FD004 RMSE	Year
LSTM (baseline)	Recurrent	14.57	16.14	22.24	2022–2024
Attention-DCNN	CNN + Attention	~11.5	~19	~20	2023
TMSCNN	Transformer + CNN + DA	~11.3	14.79	14.25	2024
TCRSCANet	TC + Recurrent + Attn	~11.4	~14.5	~15	2024
DSFormer	Dual-Scale Transformer	Best FD001–3	—	—	2024
Transformer-KAN-BiLSTM	Hybrid	Low RMSE/MAPE	—	—	2024

Graph neural networks exploit electrical topology for fault prediction

GNNs represent the natural architecture for power grid applications because distribution networks are inherently graphs — buses map to nodes (with voltage, angle, and net power features) and transmission lines map to edges (with impedance, power flow, and loading features).

The 1-D Convolutional Graph Attention Network (2024) achieves ~99.5% fault detection accuracy on the IEEE 123-bus distribution system, with 4% average improvement in fault type identification and 5% in fault localization compared to baselines. Using bus voltages and branch currents sampled at 1 kHz as node/edge features, the architecture combines one graph attention layer, a 1-D convolutional layer, and three fully connected layers.

PowerGraph (NeurIPS 2024) provides the first comprehensive GNN benchmark for power grids, featuring node-level regression (power flow analysis), graph-level classification (cascading failure prediction), and ground-truth explanations for explainability — critical for utility trust. Spatio-temporal architectures combining GRU cells with GCN/GAT/GraphSAGE (RGNN pipeline, 2025) handle evolving topologies as switches operate and DERs connect/disconnect.

Explainable GNN methods (2025) address the black-box criticism through gradient-based, feature-based, and decomposition explanations — essential for regulatory acceptance in a sector where operators must understand and trust automated decisions.

Federated learning preserves data sovereignty across utility boundaries

The first study applying federated learning specifically to high-voltage substation equipment PdM (2025) evaluated FedAvg, FedProx, and FedBN strategies across circuit breakers, power transformers, and emergency generators. FedBN yielded the highest F1 score of 0.88 with a moderately low Federated Information Criterion (FIC) of 4.35 — a novel metric proposed to evaluate federated model quality. For non-IID data distributions typical of cross-utility deployments (different equipment ages, manufacturers, and operating conditions), FedBN's batch normalization adaptation proves most robust.

This is directly relevant to GridTokenX's data sovereignty requirements: PEA's asset data is sensitive operational intelligence that cannot be centralized, yet prosumer DER data from thousands of installations would benefit enormously from collaborative model training. Federated approaches with differential privacy and secure aggregation enable this collaboration without data leaving edge devices.

Physics-informed neural networks bridge the data-physics gap

PINNs embed physical laws — thermal degradation kinetics, electrochemical models, conservation equations — directly into the neural network's loss function, enabling physically consistent predictions even with limited training data. Wang et al. (Nature Communications, 2024) demonstrated PINN-based battery SOH estimation across 55 NCM batteries using the XJTU dataset plus NASA and CALCE datasets, achieving superior stability with 30–65% reduction in prediction volatility compared to pure data-driven approaches.

The LSTM-PINN framework (2025) combines LSTM temporal feature extraction with a Deep Hidden Temporal Physical Module, training on three losses simultaneously: SOH prediction loss, implicit PDE loss, and PDE gradient loss. Results achieve MAE of 0.594–0.746% and RMSE of 0.791–0.897% — approaching the practical accuracy ceiling for field deployment.

For transformer aging, PINNs integrate Arrhenius thermal models with neural networks, enforcing physically monotonic degradation trajectories that pure data-driven models frequently violate. The ability to generalize across battery chemistries and transformer designs with limited target-domain data makes PINNs particularly valuable for PEA's diverse asset fleet.

Digital twins, anomaly detection, and reinforcement learning

Digital twin platforms for microgrids have matured significantly. An AI-enhanced IoT digital twin framework (2025) using multi-agent RL for coordinating maintenance across interconnected microgrids demonstrated 28% reduction in overall maintenance costs, 15% improvement in system availability, and 12% reduction in energy losses.

Anomaly detection methods show strong performance: Transformer-GAN (2024) achieves 95.18% accuracy for power distribution anomalies; LSTM autoencoders deliver F1 scores of 0.9565 for equipment monitoring; and VAE-based approaches effectively detect battery cluster anomalies from cell-level SOC, current, and voltage data.

Reinforcement learning for maintenance scheduling achieves ≥10% operating cost reduction versus traditional methods, with system recovery time reduced to under 30 minutes after disruptions. The IDAPPO framework (2025) uses GNN-enhanced PPO to capture interactive degradation among connected machines — directly applicable to microgrid assets where one component's failure accelerates degradation in neighboring equipment. This connects naturally to GridTokenX's existing DRL work for V2G fleet dispatch.