ALGORITHM FOR CLASSIFICATION OF FIRE HAZARDOUS SITUATIONS BASED ON KOLMOGOROV-ARNOLD NETWORK

Abstract

The problem of timely and accurate detection of fire hazardous situations is critical to ensure the safety of people and property. Traditional monitoring methods based on simple threshold values for smoke and temperature sensors are often insufficiently effective, as they can lead to false alarms or miss real fire hazardous situations. Modern methods using neural networks can significantly improve the accuracy of classifying an emergency situation by analyzing complex patterns in sensor data, which are complex nonlinear functions with dynamically changing parameters. The development of such models requires attention to the collection, labeling and processing of data, to the choice of neural network architecture for a specific task, because high-quality data labeling and the choice of the desired neural network architecture directly affect the selection of the desired patterns, as well as the detection of hidden patterns that are impossible or difficult to determine by traditional methods. The article examines an algorithm for classify ing fire hazardous situations based on the Kolmogorov-Arnold network (KAN). This algorithm is used to process data from a complex of interconnected fire sensors and is designed to detect and classify various types of fire hazardous situations. The key element of the development is the use of the Kolmogorov- Arnold network, which, due to its architecture, is capable of modeling complex functional dependencies between input data. Readings from a complex of interconnected fire sensors, such as temperature and smoke sensors, are used as input data. To improve the accuracy of classification, data is labeled using expert knowledge. The Python programming language was used to implement the algorithm, together with the Pytorch, pykan, and scikit-learn libraries. The article presents the results of testing the model on real data and discusses possible directions for further improvement of the algorithm. During the experiments, it was shown that the proposed model demonstrates high accuracy in classifying fire hazardous situations, which is not inferior to traditional methods of data classification.