Método para reconhecimento automatizado de falhas construtivas na execução de fachadas com uso de drones e aprendizado de máquina

Abstract

The facade is one of the central systems of a building, and the appearance of pathologies in this system can compromise its performance. Therefore, periodic inspections and maintenance activities are necessary, especially during execution, when several anomalies arise. However, manual and visual inspections often have some isolated ones, being considered time-consuming, expensive, and unsafe. Therefore, the use of digital technologies can minimize such limitations. In this context, the main objective of this work is to propose a controlled recognition method to identify pathological manifestations in concrete wall facades during execution, using drones and Machine Learning (ML) algorithms, aiming to improve the quality of the work. To this end, the research strategy of Design Science Research (DSR) was adopted, involving three case studies along the following steps: a) Awareness through the investigation of the theoretical problem through a Systematic Literature Review-RSL, identifying the research gaps and through investigation of the practical problem in the field; b) Suggestion of the process by carrying out an exploratory case study, in which drones were used for image acquisition and AM algorithms for digital image processing; c) Development of the experiment with the realization of two case studies implementing the method of recognition of pathological manifestations through image acquisition and processing protocols, in addition to the incorporation of information arising from these activities in the quality management system of the work; d) Evaluation of the proposed method through the constructs Transparency, Efficiency and Usefulness, through measurement and sources of evidence; and finally, e) Conclusion with the formalization of the proposed method and final considerations about the study. Two standardized protocols allowed incredible speed and precision in image acquisition and processing stages. In the studies, eleven pathological manifestations were raised, of which four were visualized with a drone and followed by machine learning algorithms of the new models created in the AM software called Custom Vision, the one that presented the best food performance at 58.99% of F1 Score and 65% of mAP. Furthermore, some constructive faults were recognized correctly, with up to 97.4% probability during testing. From a managerial point of view, the information acquired through the inspections was incorporated into the work execution process through a report, images collected with a drone, and action plans based on the PDCA cycle. After evaluating the method, it was found that the information through the standard method supported managers in decision-making about quality control during the construction process of cast-in-place concrete wall facades.