Automation of the thermographic inspectiondevelopment of thermo-geometric algorithms for detection of pathologies supported by machine learning strategies

Zusammenfassung

Inspections of buildings, heritage sites and civil infrastructures are common tasks for both professional technicians and researchers. The importance of maintaining buildings, heritage sites and civil infrastructures in optimal conditions stays on their usefulness to the general population. Thermal comfort and the reduction of energy demand are also essential tasks in the building sector. Therefore, the use of the most advanced inspection techniques is recommended for the application of the prevention and conservation tasks required in every instant and in each structure, as well as for the maintenance in healthy conditions and the reduction of the energy demand in buildings. InfraRed Thermography (IRT) is one of the most successful inspection techniques to identify and thermally characterize a pathology in a building, heritage site or civil infrastructure, if it is located both superficially or subsuperficially. Its advantages over other inspection techniques keep this technology updated over time, increasingly improving its performance. Proof of this continuous improvement is the recent publication of works in which IRT data collected in a test campaign are interpreted in an automated way. Automation offers higher advantages over manual interpretation, and is becoming more and more successful, especially thanks to the continuous emergence of new and better Machine Learning (ML) models. This Doctoral Thesis proposes different workflows to advance in the automatic IRT interpretation within the fields of building, heritage site and civil infrastructure. Specifically, the advances are focused on lines that had not yet being contemplated in the literature at the beginning of the Doctoral Thesis, from the analysis of largescale structures and heritage elements to severe pathologies not previously contemplated in the automatic IRT interpretation. The automatic thermal characterization of the identified pathologies, and the improvement of the learning processes of ML models in a more efficient way than increasing the resolution of thermal images and the number of elements of the input dataset, are advances also conducted in this Doctoral Thesis. Two reviews of IRT methodologies in inspections of buildings, heritage sites and civil infrastructures, one focusing on the acquisition stage and the other focusing on the post-acquisition stage, the development of four automatic IRT data processing algorithms and the proposal of a new approach to improve the ML model learning process were the results of this Doctoral Thesis. Each developed methodology has started from the same hypothesis based on thermal fundamentals. The first methodology automatically detects pathologies and the rest also delimit the contour of each pathology area, with an increasing improvement of the performance with each new methodology proposed. In addition, two thermophysical parameters are obtained, with and without the help of other Non-Destructive Techniques (NDTs), which also provided new information beyond the scope of the IRT. Residential building facades, heritage building facades, a wall of a cemetery, a mosaic and interior walls of a building were the structures analyzed, being thermal bridges and moisture the pathologies of interest. Regarding the new proposed approach, the combination of automatic IRT data processing algorithms, as a preprocessing step, with ML models, specifically with Deep Learning (DL) models, has proved to be better than the separate application of both data processors. In this case, the new approach proposed is applied to marqueteries with resin pocket effect, missing tesserae, voids and detachments as pathologies of interest. The results obtained from each workflow were promising, in such a way that each workflow has been presented in a scientific article published in international journals with high impact indices and in international conferences. So, this Doctoral Thesis is structured as a compendium of seven scientific publications, specifically six were published in international journals indexed on the Journal Citation Report (JCR), and one is a double-blind peer reviewed conference proceeding. The main contribution of the Doctoral Thesis to society is the optimization of the decision making in prevention and conservation actions in the fields of building, heritage site and civil infrastructure, as well as the benefit to inspectors to maintain a building in salubrious conditions and to rehabilitate it energetically.