TY  - EJOU
AU  - Ma, Tian 
AU  - An, Tong 
AU  - Wang, Jiahao 
AU  - Zhang, Ruping 

TI  - Defect Detection Method for Hoist Wire Rope Based on Flow Model Feature Embedding
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Mine hoist wire rope video inspection is hindered by low illumination, dust interference, motion blur, and vibration, which obscure small defects and destabilize frame-level responses. Given limited defect annotations, this study develops an unsupervised anomaly detection framework based on normalizing flows and trained only with defect-free videos. The framework first applies a Mine Environment Perception Enhancement Module (MEPEM) to mitigate low light degradation and motion blur before spatial anomaly modeling. It then uses a multi-scale detection network based on normalizing flows to embed enhanced features at different resolutions and compute anomaly scores for broken wires, wear, and corrosion. A Transformer temporal consistency module models anomaly score sequences over consecutive frames to suppress transient false alarms caused by illumination fluctuations and dust disturbance. Experiments on a mine hoist wire rope video dataset show that the method achieves a frame-level area under the receiver operating characteristic curve (AUC-ROC) of 93.4% with 60.1 M parameters and an inference speed of 25.1 frames per second (FPS). Compared with the evaluated models based on normalizing flows, it improves AUC-ROC by 0.5–4.5 percentage points and reduces the false alarm rate to 8.4%. These results indicate that the proposed framework is feasible for wire rope video detection under degraded underground imaging conditions.
KW  - Wire rope defect detection; normalizing flows; low light video enhancement; temporal consistency; transformer

DO  - 10.32604/cmc.2026.087060