Design of an in-pipe inspection robotic system (IPIRS) with YOLOv8–LSTM integration for real-time in-pipe navigation

TL;DR

This paper introduces a robotic system for inspecting pipes that uses advanced AI to navigate and detect objects in real time, improving pipeline inspection accuracy and safety.

Contribution

The novel integration of YOLOv8 and LSTM for real-time in-pipe navigation and object detection in varying-diameter pipes is presented.

Findings

The system achieves 97.9% mean average accuracy (mAP (0.5)) and an F1 score of 0.95 for object detection.

The LSTM model shows high accuracy in motor voltage prediction with an MSE of 0.00037 and MAE of 0.00581.

The modular robot design successfully navigates pipes with diameters from 100 to 150 mm using ROS and simulation tools.

Abstract

Conventional methods are still labor-intensive, hazardous, and have limited coverage, even though reliable pipeline inspection is crucial for oil, gas, and water distribution networks, emphasize key challenges of pipeline inspection and automated navigation. Although robotic systems have advanced with mechanisms like helical motion and multi-joint telescopic designs, the problem of obtaining dependable adaptability in curved or varying-diameter pipes has not been solved. Additionally, there are still gaps in precise navigation and predictive decision-making under complicated pipeline conditions because of the limited integration of deep learning and computer vision into real-time autonomous platforms. To provide real-time navigation and predictive analysis, this work introduces an improved In-Pipe Inspection Robotic System (IPIRs) that integrates robotic simulation and deep learning. The system’s strong performance in navigation and visual object detection using YOLOv8 is demonstrated by its mean average accuracy (mAP (0.5)) of 97.9% and F1 score of 0.95. In addition to spatial scanning, the long short-term memory (LSTM) model analyzes temporal and group-action IMU data. The mean square error (MSE) of 0.00037 and mean absolute error (MAE) of 0.00581 test show that the model ensures accurate motor voltage prediction for smooth navigation and object detection in curved pipelines. The modular robot design, developed using the Robot Operating System (ROS) and evaluated using Gazebo and Rviz simulations, demonstrated its ability to automatically navigate and recognize objects in pipes with diameters ranging from 100 to 150 mm. The results demonstrate how the YOLOv8–LSTM architecture enhances inspection accuracy and enables predictive maintenance.

Figures