An Integrated IoT-Based Predictive Maintenance Framework for Industrial Electrical Assets: A Hybrid Deep Learning Approach

Technical Summary: An Integrated IoT-Based Predictive Maintenance Framework for Industrial Electrical Assets

Problem Statement
Industrial production continuity and energy efficiency rely heavily on the reliable operation of electrical assets, particularly induction motors and distribution transformers. Despite advancements in Predictive Maintenance (PdM), existing approaches often suffer from fragmented decision-making processes and a reliance on isolated sensing modalities. Conventional methods frequently depend on vibration and thermal monitoring, which require intrusive sensor installations and face scalability challenges. Furthermore, many current studies treat fault detection, anomaly identification, and maintenance forecasting as separate tasks, failing to integrate early anomaly detection with actionable maintenance recommendations and energy performance optimization within a single framework. There is also a gap in utilizing high-resolution electrical and harmonic multivariate signatures as primary sensing modalities for detecting both electrical and electromechanical incipient faults.

Methodology
The paper proposes an integrated IoT-based PdM assistant designed to bridge the gap between electrical signature analysis and actionable maintenance strategies. The methodology is structured around four core components:

1. Hardware and Sensing Architecture: The system utilizes an analyzer-grade IoT sensing platform based on the ADE9000 integrated circuit and an ESP32 microcontroller. This setup enables high-resolution electrical monitoring with an 8 kHz sampling rate, targeting IEC 61000-4-30 Class S measurement accuracy. The hardware captures voltage, current, power, and harmonic metrics non-intrusively from existing panels. Data is transmitted via a resilient, low-latency MQTT-based telemetry stack to a cloud infrastructure.
2. Feature Engineering: A multilayer feature engineering framework expands approximately 67 primary electrical measurements (including RMS values, power components, and harmonic magnitudes) into a 328-dimensional derived feature space. This expansion captures harmonic behavior, phase imbalance, efficiency indicators, and temporal dynamics to enhance diagnostic separability for complex fault patterns.
3. Hybrid Anomaly Detection Pipeline: The core predictive engine employs a two-stage hybrid time-series modeling approach:
   • Prophet: Used as the first stage to decompose time-series data into trend, seasonality, and event components. This establishes a "normal behavior" baseline, isolating macroscopic patterns such as production shifts and load cycles.
   • LSTM (Long Short-Term Memory): The second stage utilizes an LSTM network to analyze the residual sequences generated by Prophet. The LSTM learns short- to medium-term temporal dependencies in these residuals to quantify anomaly intensity.
   • Fusion and Decision: The system combines the sequential error-based score from the LSTM with multivariate deviation scores (e.g., Mahalanobis distance) to generate a unified anomaly metric. This metric is mapped to prioritized maintenance actions (Immediate, Planned, or Monitor) and linked to specific power quality degradations.
4. Explainability: The framework includes a decision layer that translates anomaly indicators into explainable reports, identifying which features (e.g., specific harmonic bands or imbalance indicators) contributed most to the detected deviations.

Key Contributions
The study makes four primary contributions to the field of industrial PdM:

• High-Resolution Sensing: Deployment of an IoT platform capable of 8 kHz sampling and Class S accuracy, enabling the detection of subtle electrical and electromechanical anomalies.
• Advanced Feature Space: Creation of a 328-dimensional feature space derived from 67 primary measurements, specifically designed to capture harmonic and efficiency dynamics.
• Hybrid Modeling: Introduction of a hybrid Prophet-LSTM pipeline that combines the interpretability of decompositional forecasting with the predictive power of deep learning for residual analysis.
• Integrated Decision Support: Development of an explainable layer that not only predicts faults but also quantifies energy inefficiencies and carbon emissions, linking technical anomalies to economic and environmental impacts.

Experimental Results
The framework was validated through a four-month deployment on critical cooling-group motors in an industrial facility. Key findings include:

• Fault Detection: The system successfully correlated elevated trends in the 5th and 11th harmonics (negative sequence harmonics) with mechanical bearing degradation and journal sliding. It also identified loose connections and winding resistance issues through phase imbalance analysis, distinguishing them from grid-induced issues.
• Performance Metrics: The hybrid Prophet-LSTM model significantly outperformed traditional rule-based alarms. In terms of anomaly detection, the hybrid model achieved an accuracy of 0.94, precision of 0.92, recall of 0.90, and an F1-score of 0.91, compared to 0.78, 0.69, 0.65, and 0.67 respectively for the rule-based system.
• Energy and Environmental Impact: The system quantified energy losses, estimating that high harmonic distortion in one motor resulted in annual CO2 emissions equivalent to 27–28 tons. Another motor with phase imbalances contributed approximately 12 tons of CO2 annually. The analysis suggested that applying passive filters or revising connections could recover over 5% and ~3.5% efficiency, respectively.
• Ablation Study: Comparisons confirmed that the hybrid approach outperformed standalone Prophet and LSTM models, validating the synergy of using Prophet for detrending and LSTM for residual anomaly detection.

Significance and Claims
The paper claims that the proposed framework offers a comprehensive, scalable, and explainable solution for industrial PdM. By leveraging high-resolution electrical signatures, the system reduces the reliance on supplementary vibration sensors while simultaneously detecting both electrical and mechanical faults. The integration of energy efficiency analysis directly into the maintenance decision loop addresses a critical gap in existing literature, where energy optimization is often treated as a secondary objective. The authors assert that this approach provides actionable insights for maintenance optimization and contributes to industrial decarbonization by quantifying the carbon footprint of electrical inefficiencies. The study concludes that the hybrid deep learning approach effectively bridges the gap between complex electrical data and practical maintenance strategies, offering a robust alternative to static, rule-based expert systems.