The use of recurrent neural networks and direct propagation networks to determine piezoaccelerometer parameters in the information and measurement system of vibration monitoring of a turbine unit in the presence of interference

The study presents work aimed at improving the efficiency of the information, measurement and control system of vibration monitoring of turbine units. A key aspect of the study is the analysis of the effect of vibrations with frequencies not multiple of the reverse frequency of the turbine units on the accuracy of determining the parameters of piezo accelerometers. Piezoaccelerometers act as primary measuring devices in the vibration monitoring system of turbine units, and their correct operation is critically important for reliable monitoring of equipment condition. The authors improved the previously developed method for determining piezoaccelerometer parameters, moving from using a neural network with one hidden layer to using deep neural networks. As part of the study, two modifications of the method were studied: using a recurrent neural network and a direct propagation network. The article describes in detail the architecture and key parameters of these networks, as well as the training parameters and the composition of the training dataset. A series of experiments was conducted, during which the results of training and testing neural networks were analyzed. The maximum allowable amplitude of vibration components with frequencies not multiple of the reverse frequency of the turbine unit has been established, at which the neural network provides the required accuracy in determining the parameters of free oscillations of the piezoelectric accelerometer sensor element and its conversion coefficient. The results obtained using deep neural networks are compared with the data obtained using prototype methods and an artificial neural network with one hidden layer. The study showed that deep neural networks provide higher accuracy in determining parameters. Among the studied types of networks, the direct distribution network has demonstrated the best results, which makes it the most preferable for practical use. In addition, it is noted that the use of deep neural networks reduces the number of neurons in hidden layers, which leads to a reduction in computational operations and resource consumption. The results obtained indicate the possibility of using the proposed method in equipment operating conditions that are more stringent than those in which prototype methods are effective.