With the accelerated development of power grid infrastructure, managing safety risks in high-altitude operations has become increasingly challenging. In response to the frequent non-compliance observed in the use of self-locking differential devices during tasks such as power inspections, this paper introduces a novel joint detection framework based on multi-task learning, named CAS-MTL (Classification with Auxiliary Segmentation via Multi-Task Learning). The proposed model achieves end-to-end collaborative optimization for semantic segmentation and classification tasks through a hierarchical feature fusion architecture. Specifically, a heterogeneous dual-path encoder is constructed, where the segmentation branch leverages a multi-scale deformable convolutional network to enhance the extraction of geometric features for irregular targets, such as fall arrest lanyards, while the classification branch utilizes pyramid compression fusion to achieve cross-layer semantic abstraction. Furthermore, a Task-aware Cross-Attention (TCA) mechanism is developed to establish cross-task feature associations via global multi-head attention, with task-specific queries refining discriminative features through residual learning. Finally, a Segmentation Enhancement Attention Block (SEAB) is introduced to convert segmentation masks into spatial attention weights, thereby dynamically modulating the distribution of classification feature responses and enabling joint optimization of localization and recognition. Experimental results demonstrate that the CAS-MTL model achieves an accuracy of 94.12% in detecting non-compliant usage of differential devices, surpassing baseline models such as ViT, Swin, ConvNeXt and BiFormer by over 3.7 percentage points.