Reactor Fault Diagnosis Based on CNN-LSTM Hybrid Neural Network

The acoustic signals generated by electrical insulation defects and mechanical failures during the operation of dry reactor contain a large amount of equipment state information, which can be used as an important feature parameter for the diagnosis of defects and failures. In view of the insufficiency of a single neural network to extract the temporal sequence features, this paper proposes a reactor fault diagnosis algorithm based on CNN-LSTM network, using CNN to mine the local spatial feature information of the spectrogram, and LSTM to mine the temporal sequence information of the spectrogram. Physical experimental platforms for electrical insulation defects and mechanical faults in reactors were established. Three fault samples were created for each fault type, and acoustic wave data and acoustic signals were collected during the evolution of insulation defects and mechanical faults in dry-type reactors. Experiments show that the method in this paper significantly improves the feature mining ability of the spectrogram, and effectively improves the fault diagnosis accuracy.