Application of machine learning models for early detection of emergency situations based on streaming big data

Abstract

The article addresses the pressing issue of early detection of emergency situations, such as natural disasters, industrial accidents, or epidemics, through the application of machine learning and deep learning models. The focus is on analyzing streaming big data sourced from Internet of Things (IoT) sensor networks and social media platforms, which provide re-al-time information. Traditional early warning systems, relying on methods like threshold de-tectors or statistical models, often lack the speed and precision needed to process large vol-umes of heterogeneous data in real time. To overcome these limitations, the study explores the effectiveness of advanced machine learning techniques, with an emphasis on deep learning. A hybrid LSTM-CNN model is proposed, integrating convolutional neural networks (CNN) for extracting local features from time series data with long short-term memory (LSTM) networks for capturing long-term dependencies. This model was evaluated using synthetic data from the Yahoo Webscope S5 dataset and real-world streams, such as Twitter posts during an earthquake in the Ternopil region. Experimental results revealed that the hybrid LSTM-CNN model achieves high accuracy (F1=0.90) and the shortest average detection latency (≈5.5 s), significantly outperforming traditional methods like threshold detectors (F1=0.67, latency 15 s) and support vector machines (SVM, F1=0.80, latency 8.5 s). Comparative analysis also included LSTM and LSTM-autoencoder models, which yielded slightly lower performance (F1=0.88 and F1=0.77, respectively) but still surpassed classical approaches. The advan-tages of deep learning are evident not only in higher accuracy but also in the models’ ability to detect subtle anomaly signals at early stages, which is critical for timely emergency re-sponse. Additionally, the potential of social media as a data source was demonstrated: Twit-ter analysis enabled earthquake detection within 1-5 minutes of its onset, outpacing official reports. These findings highlight the promise of the proposed models for emergency monitor-ing systems, such as industrial safety platforms or crisis management centers, where speed and reliability are paramount. The study also outlines limitations, including the need for broader testing on real-world datasets and addressing false positives in social media data.