Guided wave-based condition assessment of in situ timber utility poles using machine learning algorithms
- Publication Type:
- Journal Article
- Structural Health Monitoring, 2014, 13 (4), pp. 374 - 388
- Issue Date:
Files in This Item:
|Paper_SHMIJ_ANSHM Special Issue_2013_Ulrike Dackermann_Timber Poles and Machine Learning_Revision1_v5_FinalProof(UD).pdf||Accepted Manuscript Version||1.42 MB|
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
This paper presents a machine-learning-based approach for the structural health monitoring (SHM) of in-situ timber utility poles based on guided wave (GW) propagation. The proposed non-destructive testing method combines a new multi-sensor testing system with advanced statistical signal processing techniques and state-of-the-art machine learning algorithms for the condition assessment of timber utility poles. Currently used pole inspection techniques have critical limitations including the inability to assess the underground section. GW methods, on the other hand, are techniques potentially capable of evaluating non-accessible areas and of detecting internal damage. However, due to the lack of solid understanding on the GW propagation in timber poles, most methods fail to fully interpret wave patterns from field measurements. The proposed method utilises an innovative multi-sensor testing system that captures wave signals along a sensor array and it applies machine learning algorithms to evaluate the soundness of a pole. To validate the new method, it was tested on eight in-situ timber poles. After the testing, the poles were dismembered to determine their actual health states. Various state-of-the-art machine learning algorithms with advanced data pre-processing were applied to classify the poles based on the wave measurements. It was found that using a support vector machine classifier, with the GW signals transformed into autoregressive coefficients, achieved a very promising maximum classification accuracy of 95.7±3.1% using 10-fold cross validation on multiple training and testing instances. Using leave-one-out cross validation, a classification accuracy of 93.3±6.0% for bending wave and 85.7±10.8% for longitudinal wave excitation was achieved. © The Author(s) 2014.
Please use this identifier to cite or link to this item: