APPLICATION OF NEURAL NETWORKS FOR INTELLIGENT PROCESSING OF SENSOR SIGNALS IN THE CONTROL OF TECHNOLOGICAL PROCESS PARAMETERS

This scientific article investigates the problem of analyzing technological process parameters in the fields of chemistry, energy, and metallurgy based on sensor data and applying intelligent signal processing methods. The main objective is to evaluate the effectiveness of artificial intelligence and deep learning models for intelligent analysis, forecasting, and anomaly detection of data obtained from sensors. Time-series data collected from industrial sensors were analyzed using LSTM (Long Short-Term Memory) and Autoencoder neural networks, as well as the Kalman filter. At the first stage of the study, sensor signals were denoised and their true state was estimated using the Kalman filter. According to the results, the filter achieved a high level of accuracy (RMSE = 0.001), allowing precise reconstruction of the overall signal trend. In the next stage, time-series forecasting was performed using the LSTM network. The LSTM model effectively learned long-term dependencies and was able to predict dynamic variations in industrial sensor signals. The obtained results (MAE = 8.079, RMSE = 10.626, MAPE = 9.194%) demonstrate that the model achieved a high level of accuracy. The Autoencoder architecture was applied for unsupervised anomaly detection. The model achieved a Precision of 0.934 and a PR-AUC value of 0.929, indicating high accuracy in identifying anomalous conditions. However, due to a relatively low Recall (0.249), some abnormal cases were not detected. Therefore, it is recommended to improve the Recall–Precision balance in future studies through threshold calibration and ensemble-based approaches.