AI-Driven Adaptive Robot for Micro-Crack Detection and Predictive Structural Health Intelligence

Overview

Effective structural health monitoring (SHM) of carbon steel structures is critical, but is limited by traditional methods in efficiently identifying fine cracks. Thus, this project aims to develop an AI-driven robotic system designed as a low-cost tool for the reliable indication of fine surface cracks (=230µm width) and providing predictive insights. The system uses multimodal sensor (HD visual, IR thermal, ultrasonic) integration in a mobile robotic platform controlled by a Raspberry Pi 5 and ESP32 through edge AI processing. The model comprises a Convolutional Neural Network (CNN) modified with Gabor filters and multi-scale attention layers to perform real-time crack segmentation. To improve detection confidence and reduce false positives, sensor fusion algorithms are implemented in the multimodal sensor data. Furthermore, a physics-based digital twin is developed by utilizing a Finite Element Analysis (FEA)-derived database queried by an external PC to predict potential crack growth based on detected features. A hybrid incremental learning strategy enables adaptation to new data patterns, allowing for progressive intelligence. The system has demonstrated the ability to achieve 91.7% mean Intersection-over-Union (IoU) for fine crack segmentation under baseline environmental conditions, as well as an R² = 0.93 correlation between predicted and verified crack propagation trends while consistently maintaining ~4.5 FPS and ~220ms latency during real-time inspection trials. This project shows promise of a scalable, cost-effective, and efficient integrated approach to allow for accessible and proactive SHM across industries.

Competition history

• ISEF 2025 Robotics and Intelligent Machines · Entry ROBO064

Resources

• ISEF abstract page ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.