Identification of temporal anomalies of spectrograms of vibration measurements of a turbine generator rotor using a recurrent neural network autoencoder

Abstract

A method is proposed for recognizing pre-emergency conditions of rotary installations based on the use of the Hamming window and advanced Deep Learning techniques in retrospective analysis of the results of accounting for the factors of operation of a turbine generator, diagnostics and control under critical impacts. A program of experimental studies on the model of a turbine plant with simulation of faults and receiving vibration signals has been developed. An experiment based on the homostatic method of checking the signal with Hamming windows, in the frequency, time and modulation domains and common initial data, allows one to determine the most promising signal characteristics for identification. A method has been developed for monitoring the state of turbine generators in an automatic mode for timely notification of the CHPP personnel about the appearance of signs of pre-emergency situations, as well as about the nature of faults by the method of predicting the state of a pre-emergency situation using convolutional neural networks implemented in the form of a recurrent autoencoder. Clustering is applied and clusters are identified that correspond to the spectrograms of pre-emergency situations. The effectiveness of the use of the homostatic method in combination with correlation analysis is based on the decision-making model described in more detail in other works.

Keywords

• neural networks
• predictive analytics
• hamming windows
• fault prediction
• vibration diagnostics
• spectrogram analysis
• vibration stand
• turbine generator
• recurrent autoencoder