Evaluating weakly supervised deep segmentation trained on radiometric pseudo labels for thermal fault screening in electrical substations

Fixed radiometric rules often supply training labels for infrared inspection of substations, where dense expert masks are rarely available. We examined what segmentation networks learn from such a rule and whether the result supports fault screening. On 193 calibrated FLIR frames, pixels exceeding the frame’s 30th-percentile temperature by more than \(16^{\,\circ }\) C served as pseudo labels for U-Net, DeepLabV3 and a compact attention-based network under five-fold cross-validation; outputs were assessed against held-out expert labels and independent field notes. With per-frame min–max input normalisation, agreement with the rule was low (Dice 0.34–0.44), and two networks ranked expert-positive frames below expert-negative ones (area under the curve 0.26 and 0.30). Supplying the temperature difference on a fixed physical scale raised agreement to Dice 0.78–0.81 and removed this inversion, whereas architecture had no significant effect. The pseudo labels encoded the extent of warm equipment rather than faults (labelled-area AUC 0.40), and the rule’s maximum temperature difference remained the best screening signal (0.75 against expert labels, 0.91 against field notes). A local thermal reference made labelled area informative (0.60–0.69). Network inputs should preserve the physical units in which supervision is defined, and radiometric pseudo labels benefit from local rather than frame-level references.