Automated non-invasive wooden pole health assessment with natural language post-hoc model explanation

Power distribution networks in rural areas comprise large fleets of wooden pole assets that form overhead line circuits. The service life of these poles depends critically on their structural integrity, which is often compromised by decay, typically beginning from the inside of the asset. Effective condition assessment for asset management purposes tends to focus on destructive methods which can end the service life of the pole or non-destructive methods which are inherently subjective. Supervised machine learning classifiers have been used to map health assessment results to simple non-destructive tests such as hammer testing. However, two key challenges obstruct this operationally: first, it is difficult to obtain reliable and subjectivity-free ground truth labels without extensive destructive testing, which limits the training of accurate classifiers. Second, the resulting models must be transparent and interpretable to support trusted decision-making in the field. This paper presents an end-to-end, hammer-based solution that addresses these challenges. Ground truth labels are derived from established drilling test results to provide binary health classifications. Machine learning based condition classifiers are trained on features extracted from hammer response waveform signals, and the resulting decision process is interpreted using SHapley Additive exPlanation to explain feature importance. To further enhance field usability, we apply a retrieval-augmented generation system with large language models that generate human-readable explanations tailored to field operatives needs. The full solution is evaluated using real-world deployments on pole assets located in the far north and south of mainland Britain, highlighting its practicality and adaptability across diverse environments.