Study On Intelligent Recognition Method Of Bubble Flow State Based On Time-frequency Acoustic Texture

Bubble flow is widely distributed in nature and process industry.It is a special gas-liquid two-phase flow state,which usually exists in a state of relatively low gas velocity and relatively small gas volume.The gas phase is distributed as discrete phase in the form of bubbles in continuous liquid medium.Bubbles in the bubble flow radiate acoustic signals during their generation,rise and collapse,and the acoustic signals are closely related to bubble flow state information.Therefore,by monitoring and recording the acoustic signals of bubble flow,combined with signal processing and other means,the state information of bubble flow can be obtained,which has very important guiding significance for natural resource exploration and the actual production of chemical and other process industries.In this paper,bubble flow is taken as the research object,aiming at the passive acoustic recognition of bubble flow state.This paper designs,sets up an experimental system and conducts experimental research,performs single-domain analysis on the radiation acoustic signal,and characterizes the radiation acoustic characteristics of bubble flow based on time-frequency acoustic texture To construct the correspondence between the time-frequency acoustic texture and the state of the bubble flow,a method for identifying the state of the bubble flow combining the time-frequency acoustic texture and the convolutional neural network is proposed.The main contents of this paper include the following aspects:(1)The bubble flow simulation reactor experimental system is designed and constructed,and the experimental system can controllably generate single-column and multi-column bubble flow at different gas velocities.A single-column(1.5 mm)bubble flow experiment,a multi-column(5 × 1.5 mm)bubble flow experiment,and a multi-column(5 × 0.8 mm)bubble flow experiment were carried out.Acoustic signals radiated by the bubble flow at different air velocities and highspeed cameras were obtained.An image of the state of a bubble flow field.Single-domain analysis was performed on the collected radiation sound pressure signals to preliminarily characterize the radiation sound characteristics of single-column and multi-column bubble flow radiation.(2)Based on the time-frequency acoustic texture,three typical states(discrete,clustered,and jet)of single-column and multi-column bubble flow were characterized,and a clear correspondence between time-frequency acoustic texture and flow field structure was constructed.The characterization capabilities of three time-frequency acoustic textures(STFT,CWT,SWT)are compared.Using the SWT time-frequency acoustic texture with the highest time and frequency resolution,combined with the flow field image,the sound mechanism of single-column bubble flow at low and high speed is reasonably explained,and the mechanism of multi-column bubble flow generation is explained.(3)The convolutional neural network has great advantages in image classification,and the time-frequency acoustic texture of the bubble flow is rich in the dynamic characteristics of the state of the bubble flow field,so the recognition problem of the bubble flow state is converted to the time-frequency based on the convolution neural network.Acoustic texture image classification problem achieves high accuracy recognition of three typical states(discrete state,agglomerated state,jet state)of multi-column(5×1.5 mm)bubble flow.An implementation framework for the intelligent recognition model of the bubble flow state is constructed,and three classic models(AlexNet,VggNet,ResNet)are trained on three time-frequency acoustic texture datasets(STFT,CWT,SWT)using transfer learning methods.The optimal model is visualized and evaluated by the confusion matrix on the test set.The results show that the performance of the three models on the time-frequency acoustic texture data set is ResNet> VggNet> AlexNet,and the recognition accuracy of the models all follow the time-frequency acoustic texture.The resolution increases and improves as SWT> CWT> STFT.Through analysis and comparison,the bubble flow intelligent recognition model SWT-ResNet is finally obtained,and its recognition accuracy rate of three typical states of bubble flow is 99.67%.This paper combines the time-frequency acoustic texture and convolutional neural network to study the method of bubble flow state recognition,and builds an implementation framework for the intelligent recognition model of bubble flow state,which can realize different bubble flow states with high accuracy.The proposed method can also be applied to the fault diagnosis of various equipment in the industry,and provides technical support for the condition monitoring and health maintenance of various equipment in the industry.