Learning Multi‐Pattern Vibration Representations for In‐the‐Wild Bolt Looseness Detection on Power Transmission Towers

ABSTRACT

Bolt looseness on in‐service transmission towers poses a serious safety risk, yet routine inspection remains costly, labour intensive and difficult to perform thoroughly. Existing vibration‐based approaches have been developed almost exclusively on single, laboratory‐built specimens and therefore failed to capture the pronounced pattern variability caused by different tower materials, geometries and ambient environments. To bridge this gap, we construct, to the best of our knowledge, the first labelled multi‐pattern tower‐excitation dataset, comprising 46,696 hammer‐induced vibration segments from 346 in‐service transmission towers with diverse structures, materials and field conditions, enabling bolt‐looseness detection from truly in‐the‐wild vibration inputs. Building on this task, we propose a baseline architecture tailored to multi‐pattern vibration classification, which combines a hybrid filterbank and an STFT representation module with learnable analysis‐window coefficients and an attention‐based decoder inspired by auditory perception and structural vibration physics. Extensive experiments show that this baseline substantially outperforms conventional hand‐crafted features and standard classifiers, yielding more accurate and robust bolt‐looseness detection across heterogeneous towers and fault scenarios.