Case study: Mitigating wildfire risk with computer vision-aided asset management and machine learning

A West Coast power player

Our California-based client is one of the largest combined electric energy and natural gas providers in the United States, serving more than 16 million people. Their service area encompasses over 100,000 circuit miles of electric distribution lines that deliver power to more than 5 million electricity customer accounts.

Bridging the data gap

The utility had historically relied on in-person, on-site inspections to identify possible fire risks due to equipment malfunctions. Not only was this approach costly and time consuming, but results also tended to vary from inspector to inspector. To address these challenges, our client had established a remote inspection platform using high-resolution drone imagery. They needed an efficient way to analyze a huge volume of photographs in identifying asset anomalies that require investigation—like equipment that’s worn, corroded, contaminated, or corona-damaged.

Finding needles in a massive haystack

Logic20/20’s data scientists and engineers worked with the client on using machine learning models—specifically, computer vision—to augment the asset evaluation and risk identification process. Our approach began with joining the client’s SAFe Agile environment to take ownership of the product vision, roadmap, development, analysis, and relationship with stakeholders.

Creating and enhancing AI models

The initial assessment enabled us to understand the client’s existing data architecture and create recommendations for development of the platform and the product vision and roadmap. We focused on AI models to detect insulators and their attachment hardware, which are the most common cause of the most catastrophic failures. We then enhanced these models to classify various attributes of the detected equipment that contribute to better inspection and maintenance decisions as well as decrease inspection time. Our teams improved UI/UX, developed APIs, and built front- and back-end integrations with some of the client’s existing platforms.

Delivering data-driven predictions

The AI models we created help match images with structure IDs and access paths to ensure that our client’s structure data is correct. We also developed a tabular report to deliver predictions—with confidence percentages—on equipment presence and damage, as well as other helpful data about the utility’s structures.

Enabling continuous improvement

We built a dashboard for tracking the metadata of each AI model to uncover opportunities for improvement. Our data scientists are lending their expertise to further enhance existing models and enable faster development of new ones.

Powering future goals

Another long-term goal for our client is to automate the inspection process to drive further productivity and cost-management benefits. Our team will be developing an automatic feedback loop enabling inspection results to be used in re-training and continuously improving AI models. We have also been asked to re-inspect several structures to help the client use AI-based predictions in prioritizing future re-inspections.

Building a brighter—and safer—future

As a result of this initiative, our client can quickly identify possible fire risks among the thousands of drone images streaming in on a daily basis. By using AI-driven predictive models at an early stage, they can make the necessary repairs before equipment defects can turn into fire hazards.

With Logic20/20’s help, this utility has implemented best practices for product management, software engineering, and data science to guide future enhancements, enabling continuous improvement for their wildfire mitigation system. Fire hazards can be identified promptly and accurately at a lower cost, enabling our client to play a proactive role in preventing future fires and protecting California’s communities.