Introduction to our solution to use thermal drone images and a deep learning model to detect delamination defects in concrete walls.
(Japanese version is here)
1. Why is a method to inspect retention walls needed?
Japan is a country with many mountains. Therefore, to avoid landslides during the rainy seasons countless concrete slopes were built along side streets and in urban areas. Most of the concrete structures are now 40 or more years old and have started to exhibit cracks and delamination. Delamination is especially dangerous as the fallen pieces of cement can directly injure people or lead to traffic accidents when falling onto roads. For example, see the news article(Japanese).
The current state of the art is to check concrete walls for defects using a hammering method. The process requires an expert to climb onto the structure, hammer in intervals against it, and note areas with changed sounds that indicate delamination defects. This process is time-intensive and costly, further the number of experts that can conduct these inspections is scarce. This means tha
t concrete structures can only be sporadically inspected.
We propose a new inspection method to solve this problem utilizing drone images and deep learning algorithms.
2. Methodology and Results
To detect the delamination defects a drone took optical and thermal images of a concrete wall in Shimane prefecture. The delamination defects are detectable in the thermal images as the defect leads to a different thermal conductivity. These differences are especially visible in the morning and evening hours when the temperature gradient between the concrete and the air is the highest.
To train our deep learning algorithm we matched the thermal and the optical images to increase the amount of information fed to the algorithm. To train the algorithm we annotated the training images using the results from a hammering inspection of an expert. After training the model with a limited number of approx. 100 images, we were able to detect 7/8 delamination areas in our validation image (see Figure 1). This shows that this algorithm can find defective areas that are not recognizable by the human eye in thermal images.
Figure 1: Prediction results for the validation image, show good detectability of areas that the expert classified with hatch sound and mortar stripping.
Overall, this approach is very promising. With the current model pre-inspections of cement walls could be conducted to reduce the labor of hammering inspections. And after further training and fine-tuning of the AI algorithm, this method might be able to replace the old hammering method. Now we are looking forward to further developing this inspection method with our partner DroneCreate Inc.