NEURAL NETWORKS IN DATA PROCESSING SYSTEMS

Аннотация

This paper explores methodological aspects of applying neural network technologies in data processing systems. The research highlights the role of neural networks in handling large data systems (LDS) and their relevance to the Internet of Things (IoT). Key applications such as anomaly detection, pre-semantic data processing, and quality evaluation are discussed, demonstrating the efficiency and adaptability of neural networks in addressing challenges in dynamic technical environments. Keywords: Neural networks, large data systems, anomaly detection, Internet of Things (IoT), artificial intelligence, deep learning, data mining, edge computing. The rapid development of the global economy, accompanied by increasing energy consumption and the complexity of technological processes, has given rise to numerous challenges. One of these challenges is the creation of large data systems (LDS), driven by the emergence of the Internet of Things (IoT) [1]. This necessitates the integration of intelligent information technologies to enhance system performance and adaptability. An analysis of LDS capabilities and development trends reveals similarities with "living" intelligent systems. Key elements defining LDS include system, structure, goal, and technology [2]. LDS technology encompasses material, energy, and informational processes, which stem from its structure, while its objectives partially align with human goals. Humans operate on three levels of goal-setting—genetic, unconscious, and conscious [3]—shaping their functional and social technologies and influencing their movement within a "freedom space." Among the most significant technological tools in LDS is the artificial neural network (ANN). ANNs are designed to recognize patterns and perform tasks reminiscent of evolutionary processes. They operate within a flexible "freedom space" and exhibit "genetic" goal orientation embedded in their structure. During training, ANNs acquire "unconscious" goal orientation, akin to conditional reflexes in living organisms. Despite their narrow specialization, ANNs have been applied across various domains of human activity, often outperforming humans in specific tasks [4]. 301 International scientific journal “Interpretation and researches” Volume 1 issue 1 (47) | ISSN: 2181-4163 | Impact Factor: 8.2 Strategic applications of neural networks include information perception tasks, such as search, detection, recognition, and scene analysis. These tasks overlap significantly with Big Data and Data Mining technologies [5]. Neural networks are particularly effective in pre-semantic information processing within complex technical systems, addressing challenges that traditional methods often fail to resolve. However, the application of ANNs in technical systems is often constrained by legal and practical limitations. Neural networks cannot always assume responsibility in critical decision-making scenarios, and their pattern libraries may be insufficient for complex situations. Training ANNs can be accelerated using advanced methods, such as the Kullback–Leibler divergence. For example, a system based on the quality assessment method proposed in [6], enhanced with a neural network trained using this measure, efficiently detects anomalous patterns while minimizing energy costs and optimizing performance. The combinatorial capabilities of neural networks are vast. While projects like Blue Brain Project and Human Brain Project aim to simulate the human brain’s structure and dynamics, current neural networks still lack the unique combinatorial and adaptive properties of the human brain, which contains approximately 10^11 neurons and 10^15 synapses. Nevertheless, ANNs have the potential to evolve into self-replicating systems. For instance, Google's AutoML Vision demonstrates a system capable of generating optimized neural networks that interact through technologies such as Service-Oriented Architecture (SOA). As neural networks continue to develop, their integration with advanced technologies like edge computing and decentralized networks will significantly expand their potential applications. These include fields such as healthcare, education, and autonomous systems, enabling smarter, more efficient systems capable of adapting to dynamic environments. Experimental Study: Anomaly Detection with Neural Networks To illustrate the practical application of neural networks in LDS, we conducted an experiment on anomaly detection in sensor-based IoT systems. The objective was to train an artificial neural network to identify anomalous readings in a dataset collected from industrial sensors. Methodology: We used a dataset containing sensor readings over an extended period, labeled as either normal or anomalous. A deep learning model based on a convolutional neural network (CNN) and a recurrent neural network (RNN) was implemented to detect anomalies. The model was trained on historical data, using the Kullback–Leibler divergence method to optimize