Lightweight Multi-Task CNN for Simultaneous IGBT Switch Aging Diagnosis Using CWT-Based RGB Images

The per-switch condition monitoring of insulated-gate bipolar transistors (IGBTs) underpins condition-based inverter maintenance, yet existing approaches collapse the six-switch state space into one system-level label, leaving individual devices unresolved. This paper presents a Continuous Wavelet Transform Multi-Task Network (CWT-MTNet), a lightweight convolutional network that predicts the four-class aging state (Healthy, Mild, Moderate, or Severe) of all six switches from one CWT-based RGB image whose channels carry the scalograms of three symmetrical-component deviation signals. Five Depthwise Separable Convolution (DS-Conv) blocks and six classification heads give 172 K parameters in a 0.7 MB footprint—a 21-fold reduction over MobileNet-v2—within the on-chip memory of an embedded inverter controller, a footprint comparison rather than a demonstrated implementation. Trained from scratch on 15,625 samples spanning all 56 state combinations, it attains 98.15 ± 0.33% accuracy and 95.81 ± 0.82% Mild recall over five seeds: within 0.84 percentage points of MobileNet-v2 in accuracy, statistically indistinguishable in Mild recall, at one twenty-first of the parameters. Scratch training outperforms ImageNet pretraining, indicating a domain mismatch with CWT scalograms. Multi-task gradient conflict costs ResNet-18 42.1 points of Mild recall under six-head operation; DS-Conv architectures change by at most 3.3. Grad-CAM attributes this to distributed time–frequency attention; Mild-to-Healthy confusion is the dominant safety risk. A representation ablation shows the CWT recovers 2.4 of the 4.0 points lost by discarding phase and makes per-switch accuracy twelve times more uniform, without being necessary here. Independent sensor noise at 0.5% of the phase RMS exceeds the aging signature sixfold and defeats every encoding examined, so the pipeline requires coherent averaging at the acquisition front end.