Team name / Company name: GreenSentics
Team leader: Vaclav Maixner
Challenge: no. 9: Grid Control by drones – picture recognition
Problem: Transmission towers require regular controls for possible defects and faults. Due to their remoteness, helicopters are used to collect data - flying over long distances to collect vast amounts of footage by hand. This is both taxing to the environment and expensive in manhours.
Solution: The proposed solution is using drones in order to decrease the carbon footprint and lower the number of qualified personnel needed for the task. But in order for this to work, there has to be a smart decision system in place, to know what and where to record correctly, so that the video feed can be used for quality inspection of the transmission line's vital parts. The target here is to automate this system end-to-end. Try it out at https://greensentics.github.io/
Impact: With this solution, the whole process can turn into an automated, green process - cutting down drastically on CO2 emissions by sidetracking helicopter use for the thousands of masts spread around the country. It also cuts down on the needed manpower, from pilot to the quality assurance personnel. The solution scales much more easily and enables further analysis on the collected data.
Feasibility: The feasibility of the first step of detecting masts in video has been tested by us in the field and can be easily expanded on with our experience. The trick is the holistic solution, which has the largest business value in automating quality control. ČEPS is working on having drones ready, and PylonGuard is ready to become a production ready solution on our DataSentics backbone. We are looking forward to validating our estimates of the savings in further talks with ČEPS.
What you built: We collected data for the CV algorithms that we have chosen after understanding the initial dataset. We annotated it using Diffgram tool for YOLO neural network training for all of the use-cases - detection and classification of masts + the same for their insulators. We have then created a simple web app using tensorflowJS example implementation, later we have hosted the solution on github. We also tested the solution in real life to prove real-life feasibility. Everything is open-source.
What you had before: We have used the annotation studio Diffgram on our previous projects. We have extensive experience with using computer vision algorithms, but we didn't bring anything prepared for the task. We have set up a compute in cloud for training and hosting, with which we also have previous experience.
What comes next: For us this has been a proof that the use-case can be solved using what we know from other areas. We are always looking for innovative approaches across a wide range of industries and we feel like this is a perfect and frankly very enjoyable problem that we want to help ČEPS solve. We want to talk with ČEPS about how we can turn this into a valuable asset for them.