Research On Health Monitoring Of Vertical Shaft Rigid Canal Channel Based On Convolutional Neural Network

The hoisting system is an interactive tool between the mine and the ground,and its stable operation is the top priority of safe production in the mine.As the guiding device in the vertical hoisting system,the health of the rigid tank is an important guarantee for the safe operation of the hoisting system.Therefore,this article takes the rigid tank channel as the object,and the research on its health status is of great significance to the safety production of the mine.First of all,according to the actual structure and working principle of the vertical shaft hoisting system,reasonable simplifications are made according to the experimental requirements and actual experimental environment,and the hoisting system experimental platform is built.Its mechanical structure includes tank beams,rigid tank channels,lifting containers,roller tank ears and lifting motors,etc.;frequency converter and relay,etc.;Its dynamic test system includes three-direction acceleration sensor,acquisition instrument and DSPA-V11 software.Use ANSYS to solve the optimal installation position of the sensor at the middle of the top of the lifting container.Secondly,collect the vibration acceleration signals of the lifting container under the excitation of the two kinds of defects of the rigid tank path dislocation and gap,and analyze the characteristics of the two kinds of defects according to the signal and the movement track of the roller tank ears.Perform STFT with window functions of three lengths for the two signals,and perform WT with four wavelet basis functions to obtain characteristic time-frequency images.After comparison,the time-frequency image under the Complex Morlet wavelet basis function has the best time resolution and frequency resolution.It is further determined that the acceleration signals of dislocation defects,gap defects and defect-free acceleration signals have the best effect after using cmor3-3 wavelet base,cmor3-1 wavelet base and cmor3-2 wavelet base in turn.Then,according to the working principle of CNN and the requirements of the diagnosis task,the first generation CNN model was designed and built in the Pytorch environment.The vibration acceleration signals of each working condition under the conditions of multi-running speed and multi-load conditions of the lifting container under three types of defects and multiple defects of the rigid tank channel are collected,and the training set and test set are established after the signals are transformed into time images.And collect the vibration acceleration signals of 3 kinds of defects under random working conditions,each 100 sets of verification sets are set.After the image input model is trained,the average diagnosis accuracy rate is 65%,and the model has an over-fitting phenomenon and poor generalization ability.Finally,the preliminary CNN model is optimized from the three aspects of model depth,algorithm and structure,extend the model depth to 5 convolutional layers,Partially increase the BN layer and the Dropout layer,use a small-size convolution kernel and a long-stride convolution layer to locally replace the pooling layer,the average diagnosis accuracy rate of 99%or more is obtained on both the training set and the test set,and the model has good generalization and suppression of overfitting,the prediction accuracy rates of misalignment,gap and defect-free on the validation set are 100%,100%and 98%,respectively.According to the comparison with the results of four machine learning diagnosis methods,the effectiveness and superiority of this method are verified.Figure[52]Table[9]Reference[87]...