Team name / Company name: The Lean Boys
Team leader: Kristián Kuľka
Challenge: no. 7: Transmission system operator of the future
Problem: There is a need for checking status of vast majority of electric masts in high voltage transmission systems. The current solution of doing that is by humans which are flying on the helicopter and checking for the defects manually. This solution is costly and not very green. Challenge is to find out new and efficient ways of doing that using UAVs and automatic image recognition.
Solution: Our proposed solution lies mainly in object detection and classification of particular masts which will help drones to navigate around them and also it will help them with checking for possible defects. We've implemented PoC ML model on the provided dataset which is capable of doing that in a limited scope. We achieved a very high score for detecting multiple masts on the picture. We've also implemented base model for mast classification and proposed solutions to other subproblems.
Impact: The biggest impact of our solution is in helping societies to keep "going" on the green energy and also to lower incredible expenses which are connected to flying thousands of miles with the helicopter.
Feasibility: Our solution is based on state of the art Deep Learning frameworks and other various open source technology which are in some way greatly automatized. In our solution we also handled the problematic dataset with very few images and labels and in the end showed that it is possible to achieve reasonable results with hints for future improvements.
What you built: We've implemented Deep Learning model which is capable of predicting bounding boxes of multiple masts on the same image and also their type. The final solution can be tried by running the Jupyter Notebook (file predict.ipynb in the root of the repository). We were primarily focusing on the technical aspect of the Challenge since it was more difficult to handle because of the very small training dataset. https://github.com/Druudik/The-Lean-Boys
What you had before: On the Hackathon we were using Open Source technologies such as: tensorflow (used for training Deep Learning models), imgaug (used for Data augmentation -> it helped to increase size of the training data), Label Studio (used for labeling training dataset with bounding boxes), jupyterlab, git and basic FE template (which we did not use in the end).
What comes next: Currently we are detecting only masts and their types. However, in the future it would be needed to also predict possible defects of the masts and also improve their performance. This can be done e.g. by collecting more data (Data Lake with integrated tools for Image Labeling). With increasing size of fleet it would be also needed to optimize drone's path and power consumption. Various techniques from the AI, such as combinatorial optimization, can be applied here.