Intelligent Cleaning Nozzle with AI

In this project, a data set was created, and a neural network was trained to determine the degree of contamination and the water level in a sewer pipe from the live image of a camera at a flushing nozzle.

Factsheet

Schools involved School of Engineering and Computer Science
Institute(s) Institute for Human Centered Engineering (HUCE)
Research unit(s) HUCE / Laboratory for Computer Perception and Virtual Reality
Funding organisation Innosuisse
Duration (planned) 24.09.2025 - 23.03.2026
Head of project Prof. Marcus Hudritsch
Partner WinCan AG
Keywords Sewer cleaning, image-based AI, infrastructure, maintenance, energy saving

Situation

Flushing nozzles clean sewage pipes with high-pressure water. The operator has no visibility during this process, resulting in high energy and water consumption. Therefore, efforts are underway to develop new flushing nozzles equipped with cameras, enabling a simultaneous inventory of the entire pipeline during the cleaning process.

Course of action

The following development steps were taken: - A dataset with ground truth values for pollution level and water level was created. - Google's CNN MobileNetV2 was adapted for the two regression values and trained on the dataset. - The trained model was reconverted for TensorFlow Lite so that it could run on the ARM-based camera.

Result

The model achieved an inference time of less than 50 ms on the camera and can therefore determine the two values 20 times per second. The quality in the test set is over 90%.