Lightweight bolt state detection based on a dynamic weighted loss function

Purpose Bolts are essential fastening components in transmission lines, and their absence can pose serious electrical safety risks. This paper aims to propose a lightweight bolt state detection method based on a dynamic weighted loss function to address the challenges of high missed detection rates and positional drift. Design/methodology/approach First, a bottleneck residual module is introduced to replace the VGGNet16 backbone in the SSD framework, thereby constructing a lightweight detection network; Second, a multilevel confidence feature extraction module is designed to capture classification features related target confidence, providing data support for the computation of the composite loss; Third, a center-focused loss is incorporated as a constraint to form a composite loss function, enhancing the overall detection performance; Finally, the model’s capability to distinguish hard samples is further improved by constructing a dynamic weight function through hard-sample mining and dynamic weight adjustment based on confidence factors. Findings Experimental results demonstrate that the proposed method increases the F1 score for bolt state detection has increased from 74.73% using the original SSD to 85.77%. Compared with other lightweight models, it also achieves better performance, providing strong technical support for intelligent inspection and maintenance of transmission lines. Originality/value The proposed method introduces a lightweight bolt state detection approach by integrating a bottleneck residual module into the SSD framework and designing a multi-level confidence feature module. A novel composite loss function and a dynamic weighting strategy for hard samples further enhance detection accuracy. This method significantly improves F1 score and reduces missed detections, offering a valuable solution for intelligent inspection of transmission lines.