In this second instalment of the series on “AI in Action“, we will look at three techniques that we often use when developing Artificial Intelligence solutions for our customer.
- Asset Decomposition and Labelling Strategies
- Anomaly Priorities – Using of Computer Vision, Measurement & Business Rules
- Networked DNNs
We are going to introduce these techniques as a part of a case study performed in mid-2019, using AI to detect anomalies to prevent overhead powerline failure.
The challenge: identify spacer-damper defect on High Voltage overhead transmission lines
The AI challenge in this case study was to identify broken spacer-damper on high voltage lines for one of the largest Middle East energy companies.
Across the board, spacer dampers are widely used on high-voltage overhead transmission lines (OHTL) to maintain the intended spacing between conductors, preventing fatigue failures caused by wind-induced oscillations and vibrations – phenomena that could cause major damage to overhead lines.
The solution: Combine different techniques and tools
To identify this defect, you would need to collect a lot of images of the object, label them in good/bad categories and build and train the AI DNN.
This traditional approach will certainly work.
In our opinion, the following three techniques will offer a better solution for achieving the short-term objective and longer-term goal of delivering ever more intelligent and accurate AI.
Technique 1: Asset Decomposition
The images of the spacer-damper have been labelled by a third party by labelling the whole component and classifying it in good or bad condition categories.
The general labelling approach, however, does not consider:
- The engineering aspects of the component
- The characteristics of the anomaly
- How the DNN build process works, especially for small objects
Our approach is different. We understand the engineering of the object, labelling each element of it, rather than treating it as one object. We also spend the time understanding the nature of the anomaly; it is the clamp that breaks and not the main plate nor the axis joints.
Although labelling each element takes more time, we have other tools that can accelerate the process. Being engineering-driven means your labelled images are future-proofed and you have more options for tuning and experimenting.
Our approach is also motivated by the way DNN build process works. The more complex the object, the more training images you need. Labelling a spacer-damper as a whole gives the AI more data and presents a more complex object than just one element of it.
In addition, by labelling the clamps individually, you will automatically get four times as many objects. That is one of the reasons why we do our labelling in-house.
Taking an asset- and engineering-based approach has other advantages because identifying an anomaly is only the first step toward more automation and increasingly intelligent AI.
Examples of the AI DNN identify spacer-dampers but also broken clamps
Technique 2: Identify Anomaly Priorities
Developing an AI that can identify a component and assess its condition will lead to significant savings in time and costs for predictive maintenance.
Once you have identified 1,200 spacer-dampers that need maintenance, the question then becomes which ones are the most important to fix first. We then use object counting and apply business rules to assess the priority. In this case, the priority is straightforward, the spacer-damper that has the greatest number of broken clamps.
You can only do this prioritisation if you have taken the first step of labelling the parts of the spacer-damper. This shows the importance of having a labelling strategy that can future-proof your AI.
The following images of High Voltage Porcelain Insulators are another slightly more complex example. While all are contaminated, the degree to which they are contaminated varies significantly, as does their priority to fix.
Applying the same engineering approach, our AI can identify individual insulator plates. Using both object counting and pixel counting, we can assess which of these three is the most urgently contaminated.
Technique 3: Networked DNN
This is a technique we introduced in some of our earlier Proof of Concepts for customers. This has become more common in our software when developing customer AI solutions that deal with smaller image sets or real-world uncertainty.
In this case, we wanted to find a way to deal with images of objects in situations where it is not easy to tell if the object is in good condition or if it has a defect.
This is not a problem in factory or warehouse conditions, where you have complete control of the environment. While capturing images in the real world, you are subject to different weather conditions, different humans capturing the images, obstacles that prevent you from getting the right angle, poor lighting conditions, different sensors, and a variety of asset types and structures.
Example of Obscured Component
Are you able to tell the condition of the spacer-damper based on the lighting conditions or simply because the asset component is partially obscured?
There is not enough information in the image to form an opinion on the condition of these spacer-dampers, so neither an engineer nor an AI can make a judgement call. Putting the image into a good or bad classification is not the right approach. These types of images should be categorised as either Unknown or Unsure.
It is possible to do that by setting a threshold on a DNN, but then you are using one DNN with only one parameter to tune and a DNN that was not designed to detect the ‘uncertainties’ of this world.
By designing a networked DNN solution, you have more options to cross-validate images: use other Computer Vision software and adjust parameters for your particular scenario.
Simply put, our approach makes it possible to be ‘very sure’ about what is right and ‘very sure’ about what is wrong and handle the unknown correctly. As well as providing you with many more options to tune and introduce an important new classification of an asset component, networked DNNs also help improve the accuracy of the AI.
As a result of adding this new category, the accuracy of spacer-damper identification increased by 5%, to 87.5%. No additional imagery or labelling was needed, nor did DNN need to be rebuilt.
AI has enabled us to solve problems that we could not have solved with traditional methods. However, it does not solve every problem and has certain limitations. We are committed to leveraging our expertise to simplify the complexity of AI, to readily deliver real world value. Therefore, we have developed such techniques and solutions not only to overcome some of the limitations associated with AI, but also to broaden the range of problems it can solve.