Automating Fire Protection Engineering Plan Analysis: Deep Learning for Symbol Detection

Resumo

This study presents a deep learning-based approach for the automatic detection of graphic symbols in fire protection engineering plans, aiming to automate the inspection of fire protection system designs. Fire protection plays a vital role in safeguarding lives, property, and the environment against the destructive effects of fires. In this context, engineering plans serve as critical tools for ensuring that buildings and facilities are equipped with appropriate safety measures. The manual verification of these plans is a highly visual task that demands sustained concentration, is prone to human error, and consumes a significant amount of time from public agents responsible for technical review. The integration of computer vision techniques with machine learning algorithms can reduce the analysts' workload while enhancing the efficiency and reliability of the inspection process.A dataset comprising 1,000 real images of fire protection plans submitted to the Military Fire Brigade of the State of Espírito Santo, Brazil, was used. A total of 5,291 graphic symbols were manually annotated, distributed across 12 imbalanced classes representing common elements in fire protection system layouts. A YOLOv11 convolutional neural network was trained to detect these symbols in real time. A test set was created by randomly extracting 10% of the original dataset. Model evaluation was performed using a stratified k-fold cross-validation technique, yielding a mean Average Precision at IoU 0.5 (mAP@50) of 90.6%. After training, the model was tested on test set, achieving an mAP@50 of 96.3%.To assess practical applicability, two software prototypes were developed for the automatic detection of symbols in engineering plans. Although detection time metrics were not included in this phase, the study demonstrates the feasibility of automating fire protection project verification using AI-based methods. Future work includes expanding the dataset, addressing class imbalance, and exploring alternative neural network architectures to enhance model robustness and generalization capabilities.This research contributes to the advancement of intelligent systems in fire protection engineering, offering a promising foundation for the integration of computer vision tools into regulatory workflows and technical project evaluation.