AI Enhances Fault Detection in Induction Furnaces

At the heart of steel foundries, induction furnaces melt metal with remarkable efficiency. Yet when these systems fail, production halts come at staggering costs. A breakthrough in explainable artificial intelligence (XAI) now offers a solution for preemptive fault detection and diagnosis in these critical industrial systems.

Induction furnaces (IFs) have become indispensable for industrial heating, melting, welding, and metal hardening due to their efficiency, speed, controllability, and clean operation. These complex systems consist of four primary components:

• Rectifiers that convert AC to DC using silicon-controlled rectifiers (SCRs)
• DC links with current-limiting reactors and filtering capacitors
• Inverters that regulate power by adjusting switching frequency
• Resonant circuits comprising furnace coils and AC capacitors

The systems require precise cooling to dissipate heat from power semiconductors, busbars, capacitors, and induction coils. Any failure in these components can cascade into catastrophic production interruptions.

A groundbreaking research initiative has developed a deep learning framework for IF fault diagnosis, enhanced by post-hoc XAI modules that interpret complex model decisions. This dual approach delivers both accurate predictions and understandable explanations, building operator trust in AI recommendations.

The system provides six key operational benefits:

• Prevents secondary damage to critical components like power semiconductors
• Reduces repair time and overall downtime through precise fault localization
• Lowers maintenance costs by preventing cascading failures
• Enhances overall system health and performance
• Increases productivity by minimizing unplanned stoppages
• Mitigates safety risks through early hazard detection

The research team collected comprehensive electrical parameter data from operational 15-ton, 5MW induction furnaces, including:

• Voltage, current, and power measurements
• Harmonic distortion metrics up to the 22nd order
• Power quality indicators like THD, OHD, and EHD
• System imbalance ratios and K-factors

After preprocessing 218 parameters collected from power quality analyzers, researchers employed a Local Outlier Factor algorithm for semi-supervised anomaly detection. The labeled data then trained a streamlined deep neural network (DNN) architecture optimized for real-time performance.

The system's true innovation lies in its integration of LIME and SHAP algorithms to explain DNN predictions. When the model detects potential faults like:

• Capacitor degradation
• Control circuit failures
• Switching contact corrosion
• Ground leakage incidents
• Semiconductor faults

the XAI module identifies the most influential parameters contributing to each diagnosis. For example, in ground fault scenarios, the system consistently highlighted the significance of 13th harmonic voltage in Phase III (V3H13), along with total power factor (CosPhiT) and specific current harmonics.

Comparative testing demonstrated the DNN's superiority over traditional machine learning approaches:

• Average F-measure of 0.9187, outperforming gradient boosting (0.8998) and random forest methods
• 15.22% higher accuracy than naive Bayes classifiers
• Consistent performance across all evaluation metrics (precision, recall, accuracy)

The research confirms that odd-order harmonics (particularly 3rd, 11th, 13th, and 17th) serve as reliable indicators for various fault conditions in induction furnace systems. This finding aligns with electrical engineering principles regarding rectifier-induced harmonics in industrial power systems.

While the current implementation shows remarkable promise, researchers note limitations including class imbalance in training data and the technical complexity of interpreting some XAI outputs. Future work will focus on refining these aspects to further enhance system reliability and usability in high-stakes industrial environments.