Alternating curvature-driven evolutionary genetic algorithm for cable insulation thickness measurement

Abstract

Accurate insulation thickness measurement in industrial cables presents a critical yet challenging optimization problem, characterized by an extremely large solution space and stringent requirements for sub-millimeter precision under real-time constraints. Traditional computer vision approaches struggle with computational complexity when processing irregular cross-sections, while conventional evolutionary algorithms exhibit poor convergence characteristics due to the non-convex search landscape. To address these dual challenges, we propose an Alternating Curvature-Driven Genetic Algorithm (ACD-GA) that innovatively integrates geometric prior knowledge with evolutionary computing. Our key advancement lies in establishing a multi-scale curvature gene information database that synergizes differential geometry with adaptive genetic operators. This hybrid architecture features three core innovations: (1) Adaptive generation of initial populations based on multi-scale curvature characteristics to ensure gene quality and diversity; (2) Alternating crossover-mutation operations guided by curvature information to accelerate solution convergence; (3) A population renewal mechanism combining crossover-mutation with original populations and post-selection updates to preserve high-quality genes. Experiments on both regular and irregular IEC standard specimens demonstrate the superiority of ACD-GA in terms of accuracy, repeatability, and convergence efficiency. Compared with traditional manual inspection and fitted ray methods, ACD-GA reduces the average detection time by over 95.8%, decreases the minimum insulation thickness deviation by more than 89.3%, and significantly improves measurement repeatability. When compared with classical evolutionary algorithms (GA, PSO, ACO, SA, DE) and the latest intelligent methods (PSOCO, HPDE, TDE), ACD-GA achieves reductions of 21.1%–86.7% in minimum thickness deviation and up to 88.5% in repeatability, while maintaining comparable detection efficiency.