High-performance computing (HPC) is indispensable for meeting the computational demands of extreme industrial defect detection—specifically, it enables real-time processing (Latency \(\le\) 20 ms) of massive datasets (6000+ images) within the high-temperature ( \(1000-1200^{\circ }\) C) and \(\textrm{SiCl}_{4}\) -corrosive environment of polysilicon production equipment, where traditional hardware struggles to perform. Notably, HPC serves solely as a necessary hardware prerequisite rather than a technological innovation—this distinction is crucial for clarifying the core contribution of this research. In the photovoltaic industry, surface defects (e.g., thermal stress-induced microcracks (Cr), \(\textrm{SiCl}_{4}\) corrosion pits (Ps), “diffuse speckle” silicon deposits (SD)) on steel components of polysilicon production equipment impair device performance and production continuity. Existing detection methods face three core challenges: difficulty in distinguishing high-temperature oxidation patterns from defect grayscale features, limited capability in characterizing dynamic irregular defects such as SD/Cr, and the difficulty in detecting the macroscopic manifestations of micro-defect clusters (with microscopic origins at 1–5 \(\mu\) m) amidst high-temperature reflection noise. These issues, combined with industrial real-time inspection requirements and substantial data processing needs, underscore the importance of HPC as a foundational support technology. To address these challenges, this study first constructed the DDTE dataset (6,252 field images, 6 defect categories) replicating high-temperature ( \(1000-1200^{\circ }\) C) and \(\textrm{SiCl}_{4}\) corrosion conditions. Subsequently, we proposed the Multi-Scale Feature Fusion Network (MSEC-Net) with three specialized modules: 1) MTECAAttention employs a “multi-branch separable convolution + zero-dimensional attention” mechanism that preserves feature dimensions to mitigate the loss of subtle features from micro-defect clusters, thereby helping to resolve grayscale confusion between high-temperature oxidation textures and defects; 2) ODConv utilizes a “four-dimensional dynamic weight” mechanism spanning spatial, channel, kernel, and sample dimensions, enabling improved adaptation to dynamic morphological variations in SD and thermal stress-induced Cr defects; 3) SA-FEM implements a “gated denoising + multi-scale attention” approach that suppresses high-temperature reflection noise while enhancing the detectable signals of micro-defect clusters. In the DDTE experiment, supported by high-performance computing (NVIDIA A30 GPU, Intel Xeon Silver 4314 CPU, 251GB RAM), MSEC-Net achieved 81.4% mAP50 and 65.9 FPS (Latency 15.2 ms), essentially meeting the real-time requirements for polysilicon production line inspection. Statistical analysis indicates that these improvements exhibit relative reproducibility across different random seeds. The system maintains relatively stable detection under fluctuating factory lighting conditions (mAP50 relative decline \(\le\) 0.9% in extreme glare environments). After target channel pruning (retaining only the SA-FEM channel), the params size was reduced by 41% (from 42.47M to 24.81M), with mAP50 suffering only a 0.6% relative loss. (HPC technology effectively prevents params overflow and accelerates evaluation.) MSEC-Net also demonstrates robust performance on the public NEU-DET dataset, validating its cross-scenario generalization capability. This study proves that this network (rather than HPC technology) holds promise for transforming complex defect detection algorithms into industrial solutions for polysilicon devices, while providing practical guidance for industrial defect detection.
