Drone Based Asset Inspection

Customer: ECI

ECI provides vegetation management consulting, field services, and remote sensing/software consulting services to the utility industry.

https://www.eci-consulting.com/

 

Time Frame: Dec 2017 – Mar 2018

 

Use Cases:

 

Find ways to improve the efficiency of asset inspection process through the use of AI / machine learning. Following tasks were being done manually which the ECI wanted to automate:

 

  • For a given set of images of a transmission tower or pole, find the best image that include all or most of the structure
  • Identify the type of structure – lattice tower, pole, h-frame, etc…
  • Within the image, find the insulators and create a “blown-up” image of each – find cracks &/or damage
  • Within the image, find the tops of each wood pole (inspect for rot)
  • Within the image, find the conductor attachment points and create a “blown-up” image

 

Input:

 

Labelled images from ECI’s existing drone image set were shared with us. 1000 images containing towers labelled with their types (appx 150 images/class) and 500 images without a tower or confusing structures were loaded to a repository.

 

Implementation:

 

ECI wanted to test IBM Watson capabilities and explore other options in the market as well. The first two use cases were chosen to test the different technologies & see how closely the machine could classify the three classes of poles, not poles and partial poles according to visibility of power lines and towers in the images gives. Precision and recall were used to evaluate the technologies.

 

 

 

 

In the first stage, IBM’s Visual Recognition was compared with Microsoft Custom Vision and Clarifai. IBM gave precision of 53% and recall of 39% while Microsoft Custom Vision gave a precision of 61.1% and a recall of 59.7%. Clarifai returned a much better precision rate of 98.2% while the recall decreased to 43.1%.

 

In the second stage of implementation, TensorFlow was used to train our own neural network. The precision and recall results improved to above 80% for both.

 

Outcome:

 

ECI wanted to improve the time & reduce the costs for this process. The manual rate was $15 per structure and human inspectors could process approximately 50 towers a day. After incorporating AI/machine learning, the rate reduced to $2 per structure and the number of structures per day increased by many times.

Would you like to learn more or need to schedule an appointment? Please click here