Audio based wooden utility pole decay detection based on distributed acoustic sensing and machine learning
Aspects of the present disclosure describe distributed fiber optic sensing (DFOS)/distributed acoustic sensing (DAS) systems, methods, and structures that employ machine learning and provide for the automatic remote inspection and condition evaluation of wooden utility poles. Operationally, audio (acoustic) signals are obtained using DFOS/DAS when a service technician/inspector strikes the wooden utility poles with an impact tool such as a hammer. Historical audio DFOS/DAS signals that include signals resulting from hollow (decayed) utility poles and solid (good) poles are used to train one or more machine learning models and the trained machine learning models are subsequently used to evaluate real-time impact data collected from DFOS/DAS and determine utility pole condition in real-time.
1 . A distributed fiber optic sensing (DFOS)/distributed acoustic sensing (DAS) system configured for automated remote inspection of wooden utility poles, the system comprising:
a length of optical fiber sensor cable wherein at least a portion of the optical fiber sensor cable is suspended by a plurality of wooden utility poles;
a DFOS/DAS interrogator in optical communication with the length of optical fiber sensor cable;
an intelligent analyzer configured to collect and analyze DFOS/DAS sensing data received by the DFOS/DAS interrogator system;
a cloud-based machine learning model in operative communication with the intelligent analyzer, the model trained by using historical DFOS/DAS acoustic data derived from physically striking the plurality of wooden utility poles with an impact tool, the acoustic data including signals characteristic of both solid mechanical conditions and hollow mechanical conditions;
wherein during subsequent operation of the DFOS/DAS system, new acoustic data is generated by mechanically impacting a wooden utility pole with an impact tool, and the trained machine learning model determines the mechanical condition of the impacted wooden utility pole by classifying the new acoustic data based on extracted features, the extracted features comprising at least a Mel Spectrogram, Mel-Frequency Cepstral Coefficients (MFCC), and Chromagrams of the new acoustic data; and
wherein the cloud-based machine learning model comprises a Random Forests model, wherein the Random Forests model comprises a collection of decision trees, and wherein the decision trees are directed acyclic graphs (DAGs).