Application of Deep Neural Networks for Leakage Airflow Rate Estimation From Three‐Dimensional Thermal Patterns

Abstract

The employment of deep convolutional neural networks (CNNs) signifies a substantial progression in the domain of image analysis. The application of this method is particularly suitable when the image set represents a spatial structure and predictive analysis can only be performed using Gaussian processes, which are computationally complex. The uncontrolled airflow of air into buildings, known as infiltration, poses a significant challenge in terms of characterisation and quantification. The irregular contours of gaps and cracks through the enclosure create a virtually endless variety of cases, making a generalizable scientific interpretation that can be applied to existing buildings very difficult. This circumstance is always clearly manifested by an irregular, three-dimensional incoming airflow. This study presents an innovative methodology for estimating airflow rates based on three-dimensional thermal patterns captured through infrared thermography. The experimental setup employs a 3D-printed matrix of spheres, facilitating the characterisation of the spatial temperature distribution within the airflow. The resulting thermal images are processed using a CNNs, which integrates the spatial information contained in the thermograms with a scalar input representing the inlet air temperature. The model′s performance was assessed under a range of conditions, including reduced image resolutions, varying experimental configurations (involving different flow apertures) and a comparison between full thermographic inputs and thermal difference-based features. The results indicate that the model can accurately infer airflow rates within the same aperture (medium absolute error [MAE] < 2%). While generalisation to new apertures presents a greater challenge, the experiments demonstrate that a sufficiently diverse training dataset can enhance the model′s predictive capacity for configurations not included in the training phase. These findings underscore the potential of deep learning as a nonintrusive and efficient tool for estimating airflow in systems where conventional measurement techniques are either difficult to apply or impractical.