Utility Pole Classification via Contrastive Learning on DAS Signals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Distributed acoustic sensing (DAS) systems face challenges in extracting intrinsic information about utility poles from DAS signals, as these signals are intermixed with extrinsic factors like environmental noises and vibrations, making it difficult to ascertain ground truth during data collection, and existing preprocessing procedures are inadequate due to high variability and volume of data.
Innovation Solution
A contrastive learning-based approach is employed, using a data collection procedure that introduces variabilities in excitation signals and collects data on utility poles, allowing for the extraction of intrinsic properties and remote extrinsic influences without knowing the ground truth, using a model trained adaptively on modern deep learning frameworks like PyTorch, which preserves unique pole characteristics in low-dimensional feature vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional preprocessing procedures are used on DAS signals, then processing simplicity is maintained, but extraction precision of intrinsic pole information deteriorates due to intermixed extrinsic factors and high data variability
Solution Approach 1:
The patent extracts intrinsic pole information from DAS signals by separating it from extrinsic factors through contrastive learning. The system identifies and extracts only the relevant pole characteristics while filtering out environmental noises and vibration signals from distant sources, achieving precise extraction without complex preprocessing
Solution Approach 2:
The patent introduces an embedding space as an intermediary representation layer between raw DAS signals and final pole information extraction. This embedding space serves as a mediator that transforms high-dimensional signal data into compressed latent representations, enabling efficient and accurate extraction of intrinsic pole properties
2Measurement precision
If comprehensive experiments accounting for all factor combinations are conducted, then ground truth accuracy is improved, but experimental cost and time consumption increase prohibitively
Solution Approach 1:
The system performs self-supervised learning where the model learns to distinguish intrinsic pole information from extrinsic factors automatically without requiring manual annotation of ground truth for each experimental condition. The contrastive learning framework enables the system to self-correct and improve through self-supervised signals derived from the data structure itself
Solution Approach 2:
The patent changes the parameter space by transforming the problem from one requiring exhaustive experimental conditions to one solvable through learned representations. By parameterizing pole information in terms of embedding vectors rather than explicit ground truth labels, the system achieves accurate extraction without comprehensive experiments
3Reliability
If signal filters are applied to remove extrinsic factors, then signal purity is improved, but information loss about intrinsic pole properties increases due to high data variability
Solution Approach 1:
The patent employs dynamic contrastive learning where the separation between intrinsic and extrinsic factors is not fixed but adapts to the specific characteristics of each pole and environmental condition. The embedding space dynamically adjusts to preserve relevant information while filtering noise, avoiding the rigid information loss associated with static signal filters
Solution Approach 2:
The patent transitions from filtering in the original signal dimension to separation in an embedding dimension. By projecting DAS signals into a latent embedding space, the system achieves purification of intrinsic information without the information loss that occurs when applying filters directly to the raw signal, as the embedding transformation preserves relevant features while discarding noise
4Reliability
If manual data collection with controlled excitation signals is performed, then data quality for training is improved, but data collection time and operational complexity increase
Solution Approach 1:
The patent performs preliminary data collection with intentionally introduced variabilities in excitation signals, impact locations, and impact strengths. This preliminary action prepares diverse training data that enables the model to learn robust representations of intrinsic pole information that are invariant to extrinsic factors, improving training quality without requiring exhaustive controlled experiments
Solution Approach 2:
The patent applies partial control over excitation signals by introducing specific variabilities (impact strength, location, time ambiguity) rather than attempting to control all possible parameters. This partial action approach collects sufficient training data quality while maintaining practical data collection efficiency, avoiding the excessive complexity of fully controlled experiments
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This method effectively extracts utility pole intrinsic properties and remote extrinsic influences, resulting in a robust model that can classify pole integrity and detect changes, facilitating selective classification and prioritization for maintenance, with high accuracy in identifying pole profiles and change of status.
Implementation Method 1
Distributed acoustic sensing techniques measure strain changes (stretch or compression) of optical fiber cores
Implementation Method 2
DAS signals captured at a location near a utility pole may include information not only from the characteristics of the utility pole
Data Source
AI summary
Systems and methods for operating a distributed fiber optic sensing (DFOS)/distributed acoustic sensing (DAS) system include a length of optical sensing fiber suspended aerially by a plurality of utility poles and in optical communication with a DFOS interrogator/analyzer. The method includes operating the DFOS/DAS system while manually exciting more than one of the poles to obtain frequency response(s) of the excited poles; contrastively training a convolutional neural network (CNN) with the frequency responses obtained; classifying the utility poles using the contrastively trained CNN; and generating a profile map of the excited poles indicative of the classified utility poles.