Inspection² has been using Artificial Intelligence since 2015 to perform condition assessments and analyse industrial assets for defects and anomalies across the Oil & Gas, Telecommunication and Transmission and Distribution sectors.

We have developed our proprietary AI methods, supporting software applications and AI accelerators/plugins. The aim of these was to speed up the development of AI, overcome some of the inherent problems associated with AI and address real-world scenarios that require a higher level of sophistication.

The case studies in this series “AI in Action” cover our AI approach and solution that enable us to overcome some of the limitations of AI on challenging real-world problems and complex industrial assets.

AI challenge: Data scarcity

Artificial Intelligence is powerful and has brought considerable benefits to businesses and economies by contributing to productivity growth and innovation. It is not without drawbacks.

The deployment of an AI system begins with defining the problem to be solved, collecting the data, labelling the images and creating the AI DNNs (Deep Neural Networks, namely, Convolutional Neural Networks), training the system, and then deploying it.

The dataset used to train the AI system is a core element, crucial to delivering the full value of AI. The better the training data is, the better the AI DNN performs.

So how many images do you need to train the AI?

The short answer is that it can range from several thousand to millions of images, depending upon the complexity of the problem sets.

As an example, it is easier to get decent accuracy from a DNN to identify everyday objects (e.g. people, cars, birds, lamp posts, etc.) than a specific disease from a CT scan since there are fewer data available on the latter.

When it comes to industrial assets, quality images are scarce and difficult to gather and collate. The more complex or sophisticated the object, the more images you will need. To identify anomalies and defects, which are even rarer in any image set, you must examine a greater number of images. Moreover, industrial assets, which are often mission-critical, require a higher level of AI accuracy, thus requiring even more images.

These have been a common problem with global customers that we have worked with in the last 5 years. 

As a solution to this problem, we have developed deep learning systems that are more efficient and work with fewer data.

Real-world application: Detection of anomalies on Medium Voltage infrastructure

During the last few weeks, we have been working on an AI solution with a major power utility customer to identify and map high-risk structures, thus increasing productivity and improving grid management.

In this custom AI development project, Inspection2 AI software is used to analyse data captured during medium and high voltage structures inspection to detect over 30 different types of anomalies.

Use Case 1: Cotter Pin Detection

The customer wants to automatically assess the condition of lightning conductors on electrical pylons. The inspection included the condition of the lightning conductor, the presence of nuts and bolts, and the cotter pin that holds the nuts and bolts together.

A cotter pin is a small object in an image. It will look different depending on the angle at which it was captured. It will also look different depending on installation and level of corrosion or damage.

It is standard practice to compile several thousand images of cotter pins and label them, then build and tune the AI DNN.

This approach is valid, as several companies have gotten pretty good results using this approach. In this example, 52,331 images have been fed into the Cotter Pin AI model. In terms of accuracy, they levelled out around 3,000 images.

When you already have this volume of data, the resources to label images, and the money to build one AI DNN, you will get the results.

Example of AI Cotter Pin Detection Solution

Example of one of the solutions on the market

We know from experience that our customers do not always have that scale of images to work with. These constant real-world pressures have led us to develop unique techniques and AI software to expedite this process.

In our recent custom AI development for electrical pylons, we were provided with less than 200 images to develop an AI DNN to detect the cotter pins. We achieved over 85% accuracy.

The detection of the cotter pin was just one of the problem sets. Thus, we had to build a set of AI DNNs to form a checklist to assess the condition of the lightning conductor.

Inspection2 AI detection of cotter pin nuts and bolts on lightning conductor medium voltage

Example of Inspection² AI detection results to identify lightning conductor components such as cotter pin, nuts and bolts on medium voltage infrastructure

Use Case 2: Wood Cavity Detection

Cotter pins are a manufactured component that has consistent characteristics and form.

Do our techniques work with organic materials with more natural anomalies, which have a more random and chaotic nature?

Using the same methods, we build an AI model to detect ‘wood cavities’ in Medium Voltage wooden cross beams

Inspection2 AI detection of wood cavity and drill holes on medium voltage wooden cross beam

Example of Inspection² AI solution that detects wood cavity and drill holes on medium voltage wooden cross beam

We were still able to achieve 71% accuracy, again with less than 200 images.

This also gave us a base AI DNN to speed up the automatic labelling of images when more data was collected.

The above examples illustrate what can be accomplished with the right tools and techniques, such as image augmentation, image quality standards, component isolation, threshold adjustment. We believe this is the way forward for industrial assets and anomalies, rather than relying on thousands and thousands of images to train an AI DNN.

The goal of this project was not to implement the solution straight into production but to demonstrate what AI can achieve in the short term (8 weeks) and provide the customer with a clear perspective of a scalable solution. Considering the small dataset, inconsistent image quality, and few labelling instructions, these were very positive results.

Pin It on Pinterest

Share This