How to Diagnose IGBT Aging with Just 0.7 MB of Memory

Global, Source:

The reliability of three-phase voltage-source inverters is the backbone of modern energy systems. These components power wind turbines, solar farms, and aerospace vehicles. Their health depends entirely on six insulated-gate bipolar transistors or IGBTs. Traditional monitoring often treats the inverter as a single block. This approach hides which specific switch is degrading. That blind spot can lead to unplanned outages. Such failures cost millions in lost generation. They also compromise safety in critical propulsion systems.

A new solution introduces CWT-MTNet. This lightweight multi-task convolutional neural network diagnoses the aging state of all six switches at once. It uses only a single image for input. The process begins by transforming raw voltage signals into Continuous Wavelet Transform scalograms. These create 224x224 RGB images. Each channel represents a symmetrical component of the electrical signal. This visual encoding captures subtle time-frequency anomalies. It detects bond-wire lift-off and solder-joint cracking long before hard failure occurs.

The architecture is remarkably efficient. It contains only 172,000 parameters. It occupies a mere 0.7 MB of memory. This compact footprint allows it to run directly on embedded microcontrollers within industrial drives. There is no need for external edge servers. Controller upgrades become unnecessary. The model was trained from scratch on over 15,000 simulated scenarios. These scenarios covered every possible combination of switch states.

CWT-MTNet achieves an accuracy of 98.15 percent. Its Mild recall rate stands at 95.81 percent. It performs at a level statistically indistinguishable from larger models like MobileNet-v2. Yet it uses just one-twenty-first of the parameters. The study highlights that pre-training on natural images harms performance. Domain mismatch reduces effectiveness. Scratch training proves superior for this specialized technical data.

A key innovation involves Depthwise Separable Convolutions or DS-Conv. This technique prevents gradient conflict. Gradient conflict is a common issue in multi-task learning. Competing tasks can destabilize the model. Deeper networks like ResNet-18 suffer significant performance drops when handling six simultaneous predictions. CWT-MTNet remains stable under these conditions. The research identifies a critical constraint though. The system is vulnerable to independent sensor noise. Robustness in real-world applications requires coherent averaging. The acquisition front-end must suppress broadband noise. This ensures the aging signature remains clear. This advancement marks a significant step toward predictive maintenance. It is both highly accurate and computationally feasible for widespread industrial deployment.