Executive Summary
We partnered with a British Distribution Network Operator (DNO) to develop a computer vision pipeline that identified over 4,000 high-risk poles on a section of their distribution network. In this trial project, we demonstrated an accuracy of up to 93%, demonstrating the capability of our Asset Inventory Verification service and its potential to improve network safety.
Overview
The DNO in question was tasked with an internal review of its network, to identify poles with characteristics that could pose a catastrophic risk. This meant our customer found themselves facing a logistical nightmare: manually inspecting all of the poles along their network would be an enormous task.
To address this challenge, the DNO turned to our Asset Inventory Verification service, using computer vision algorithms to swiftly analyse network assets. This process quickly identified high-risk poles, enabling proactive maintenance decisions.
Outdated asset inventory records cause problems
The poles at the centre of this safety review are typically configured as tension poles, with porcelain insulators.
The DNO’s network is home to a diverse array of pole types, each with distinct characteristics. Some feature porcelain or glass insulators, while others are equipped with multiple crossarms and other equipment.
We leveraged our Asset Verification foundation to develop two cutting-edge computer vision models. These models can accurately identify the configuration and type of insulators on a pole. We also created a third model that filters out low-quality images.
We conducted 33 model iterations, leading to our final pipeline, which processed over 71,000 images across 8,379 assets, in just 1 day.
Our pipeline begins by filtering out images without poles. The remaining images are then fed into two parallel processes.
First, the configuration classifier outputs one of two possibilities: “pin” or “tension”. Simultaneously, the insulator detector processes the image, identifying any insulators detected and tagging their type as either “polymer” or “porcelain”.
The predictions for each image are then rolled up into just 3 for each asset, telling us its predicted configuration and whether is has porcelain or polymer insulators installed.
Our configuration classifier achieved an average precision and recall of 90% and 91% respectively. Once we aggregate these image-level results to the asset level, we achieved an overall accuracy of 93%.
For our insulator model we ended up using a transformer based architecture, DETA, which gave favourable results when compared to the venerable FasterRCNN. On an image-level we achieved a mAP@50 of 84.5%. Once rolled up into our asset aggregation pipeline, achieved an accuracy of 86%.
With this project being a feasibility piece where the scope and available training data were small, these results are extremely promising. Notably, these results parallel our last asset verification project with transmission operator Transgrid, where we achieved image and asset level scores in the upper nineties.
The performance of the models was also influenced by the more complex nature of distribution poles, which are more likely to be equipped with a mixture of insulators.
Distribution surveys are also of a vastly different nature. Typically conducted on foot, image capture is not standardised, resulting in images with:
- Different dimensions
- Different exposure, contrast, focus, etc
- Different angles and/or areas of interest
Since these surveys weren’t conducted with this project’s needs in mind, many of the images focused on parts of the pole that didn’t show the crossarm and were therefore not usable. This is where our pre-processor came into play, weeding out images that didn’t display the necessary features. This was important because one of our criteria for aggregating up to the asset level is a frequency threshold, which states that a prediction must occur in at least n% of images.
After processing each survey for the 8,379 assets, we identified that 55% of the successfully processed assets could be classified as high risk i.e. having porcelain insulators on a tension configuration.
Validating this number of assets manually is a painstaking task, especially without the right tools. Our team faced this challenge first hand when we were preparing test data, manually going through hundreds of assets one by one.
Maintaining an accurate and up-to-date inventory of assets, including components and equipment, is crucial for running a network efficiently and safely. However, as hardware is replaced or upgraded over time, discrepancies can arise between what’s actually installed at a given location and what’s recorded in a company’s asset management system. This mismatch can lead to increased risk and unplanned maintenance expenses.
AI-assisted asset verification bootstraps the process and can work at any scale, minimising the time you go from identifying a risk to dealing with it.
See what our Asset Inventory Verification service can do for you.
