Detection of faulty power line insulators using convolutional neural networks

Publish date:

A convolutional neural networks‐based model to detect dirt and cracks on insulators to initiate a new paradigm for overhead (OH) line inspection.

Inspection of overhead (OH) power lines and their subsequent maintenance is one of the major activities of electric utilities. Patrolling OH lines, which includes both distribution and transmission lines, is still an old-fashioned job and is treated as a tedious work. The traditional visual inspection of OH assets is highly error prone and costly. All the different types of insulators on OH lines may appear perfectly ok to the naked eye, but the presence of cracks and dirt can lead to flashover and subsequent tripping of the OH circuits. Proper detection of cracks and the amount of dust and dirt on insulators is still a challenging task. All these challenges ultimately affect the overall reliability indices and customer satisfaction.

The limitations in traditional line inspection can be overcome by using deep learning algorithms, such as convolutional neural networks (CNNs). CNN is a type of deep neural networks, commonly used to analyze images. Its high level of learning ability is the key to detecting dirt and cracks in insulators. Keeping this unique learning capability in mind, we are proposing a CNN‐based model to detect dirt and cracks on insulators to initiate a new paradigm for OH line inspection.

In developing a CNN‐based model, the reference image dataset volume or training data must be large enough to make the model perfect. The process of developing a CNN classifier involves three broad steps:

  1. Crack and dirt image database creation
  2. CNN training
  3. Validation of the trained CNN classifier.

In the first step, the raw HD images are cropped to make smaller images. Then, the cropped small images are manually classified into different categories (with crack, without crack, with dirt, without dirt, etc.). Once this is done, 80% of the images are kept for training the CNN classifier and the remaining 20% for validation purposes. Subsequently, the images are imported into CNN classifier for training and validation. The output of the whole process is a trained CNN classifier that can pick out images with cracks and dirt from a set of images.

In order to improve the overall efficiency of the inspection process, unmanned aerial vehicles (UAVs) or drones can be used in conjunction with a CNN classifier. The use of remote-controlled HD camera drones is the most efficient way to inspect OH lines in place of traditional OH line patrolling. Thus, we can design an inspection process with three broad steps:

  1. Capturing images of OH line insulators by remote controlled HD camera drones
  2. The image files are processed to create image database
  3. Processed images are imported into the CNN classifier to identify the cracks and dirt on insulators fixed on the inspected line segment.

Currently, we are planning to develop a CNN classifier in our global Energy Utility CoE jointly with Emerging Technology and Architecture Practice CoE. Our proposed solution mainly aims in proving the concept with a live environment that can be customized to meet the requirement of the geographies.

You can connect with me to find out more about this topic.

Related Posts

AI

Beyond the health crisis – Embracing the idea of enterprise as a platform

Lanny Cohen
Date icon September 23, 2020

In this third article to a series of blogs, Lanny Cohen shares insights on the concept of...

AI

Implementing an AI strategy and framework

Lee Beardmore
Date icon September 17, 2020

The journey to an AI-driven future operating model requires continued executive support, a...

AI

Beyond the health crisis – Positioned for success?

Lanny Cohen
Date icon September 16, 2020

In this second article to a series of blogs, Lanny Cohen focuses on importance of Position –...