NEURAL NETWORK SOFT-SENSOR FOR COMPUTER-BASED MOISTURE DIAGNOSTICS

Abstract

The paper addresses the problem of computer-based diagnostics of the massecuite drying process in a drum dryer used in sugar production. One of the main challenges in automating this process is the lack of reliable real-time measurement of the outlet moisture content of massecuite, which complicates the maintenance of optimal drying conditions and adversely affects product quality. A neural network–based soft sensor is proposed for indirect estimation of the hidden moisture content of massecuite using measured technological process parameters. The distinctive feature of the proposed approach is the integration of moisture regression estimation with classification of technological drying states within a single neural network model. This enables not only quantitative assessment of product quality but also diagnostic identification of operating modes of the drying unit. The soft sensor is implemented using a multilayer perceptron that performs nonlinear approximation of the relationship between the vector of measured technological parameters and the hidden states of the drying process. To account for process inertia, an extended feature vector is formed using a sliding diagnostic window. The classification output of the model is implemented using Softmax normalization, providing probabilistic interpretation of technological drying states. The concept of integrating the neural network soft sensor into an automatic control system with a PID controller is presented through adaptation of the drying temperature setpoint. The proposed approach creates prerequisites for improving process stability, reducing energy consumption, and minimizing the influence of the human factor in industrial massecuite drying operations.