基于CNN-Transformer的离心泵空化状态识别方法OA

To address the challenges of complex network structures,large parameter size,low compu-tational efficiency,and insufficient recognition accuracy in cavitation state identification of centrifugal pumps,a novel method based on CNN-Transformer(convolutional neural networks-Transformer)was proposed.Firstly,a CNN-Transformer model was constructed,in which convolutional kernels of three different sizes were employed to extract multi-channel features from pressure pulsation signals under va-rious cavitation states,followed by channel-wise feature concatenation.Subsequently,the self-attention mechanism of the Transformer encoder was utilized to extract deep global features.The channel-concat-enated features and the global features extracted by the encoder were then fused at a ratio of 1∶1 to form a more comprehensive feature representation.Finally,classification and identification of cavitation states were performed using a fully connected layer and a Softmax classifier.The proposed model was compared with Transformer,CNN-LSTM,CNN,and LSTMmodels.The research results demonstrate that the CNN-Transformer model can accurately identify four cavitation states,namely no cavitation,incipient cavitation,developed cavitation,and critical cavitation,with an average recognition accuracy of 99.14%.This represents improvements of 2.11%,2.79%,5.09%,and 7.21%over the other four models,respectively.The findings of this study provide a new perspective for cavitation state identifica-tion in centrifugal pumps.