Research On The Identification Of Gas-liquid Two-phase Flow Pattern Of The Centrifugal Pump Under Unbalanced Sample And Variable Working Conditions

With the compact structure and smooth operation,centrifugal pumps are often used to convey and pressurize overflow media.As an important power transmission device,in some industrial processes,the internal flow is often in a gas-liquid two-phase state.It has been shown that the performance of centrifugal pumps is greatly affected when mixed gas-liquid phase transport is carried out,and the transformation of its internal flow pattern is the main cause of performance degradation and increased vibration.Therefore,it is important to establish an accurate and reliable flow pattern identification model to ensure the stable operation and fault assessment of the pump which has important significance and engineering value.The current research on flow pattern identification is mainly divided into signal acquisition,feature extraction and model building in three steps.Among them,the feature extraction process relies on a priori knowledge and extraction methods,which the model has high requirements for the importance of input features.In addition,the established flow pattern identification model is difficult to adapt to the complex and changing operating conditions of centrifugal pumps.To address the above problems,this paper proposed a method for identifying gas-liquid two-phase flow pattern in end to end centrifugal pumps based on deep learning on the overlapping sampled pressure increment signal dataset.The main work of the paper is as follows:(1)For the problem that the signal feature values need to be extracted manually in the traditional flow pattern recognition,a deep learning model based on a one-dimensional pressure increment signal was proposed.The model combined the spatial feature extraction ability of convolutional neural networks for signals and the time-resolving ability of recurrent neural networks for automatic extraction of pressure increment signal features.At the same time,a Bayesian optimization algorithm was used to automatically tune the large number of hyperparameters included in the deep learning model.The results showed that the BO-1DCNN-BiGRU model proposed in this paper had excellent performance on a balanced flow pattern dataset and was far ahead of traditional machine learning methods in the accuracy of flow pattern identification.(2)For the problem that the experimentally obtained dataset was unbalanced,a data enhancement method based on generative adversarial networks(GAN)was proposed.Wasserstein distance was used instead of KL or JS scatter commonly used in traditional GAN as the loss function,as well as conditional information was introduced to constrain the data generated by WGAN in order to avoid the generated data from being too random.The results showed that the accuracy of the classification model was significantly affected when it was trained and predicted directly on datasets with unbalanced proportions,while its ability to identify the few classes of flow patterns was significantly improved after adopting the data enhancement method proposed in this paper.Compared with traditional oversampling methods and other types of GANs,the proposed model had relatively good results on all three unbalanced datasets.(3)For the problem that traditional machine learning is difficult to adapt to pattern recognition under variable working conditions,a flow pattern identification method based on generative adversarial networks and transfer learning was proposed.To address the data imbalance on the source domain,the data was enhanced based on the GAN model,while the Multiple Kernel Maximum Mean Discrepancy(MK-MMD)was also added to the loss function when training on the target domain in order to consider the domain adaptation between different domains.The results showed that the transfer learning considering the domain adaptive loss MK-MMD was more accurate than the method using a single kernel MMD loss,fine-tuning the model and retraining the model in the process of flow pattern identification under variable rotational speeds,and the proposed method had better robustness and generalization even for centrifugal pumps under variable operating conditions.