| Due to the harsh underground working environment,coal and gangue stacking,acoustic vibration signal interference,and close physical and chemical properties of coal and gangue,the identification of coal and gangue in fully mechanized top coal caving is prone to misjudgment and has a high frequency of failure,which seriously restricts the intelligent mining process of top coal caving.The existing research system of coal and gangue identification in fully mechanized top coal caving is not yet perfect,and the traditional identification technology based on vibration,sound and image is very easy to be disturbed by single-factor feature mutation,which seriously affects the top coal mining rate.In this paper,based on the multimodal information characteristics of coal and gangue in the process of coal caving,the systematic research and analysis of multimodal information fusion of coal and gangue recognition in fully mechanized top coal caving were carried out;A bidirectional coupling model of complex coal collapse impact tail beam was constructed based on DEM-MFBD(Discrete Element Method-Multi Flexible Body Dynamics)technology.Taking the ZFY12000/25/42D top coal caving support and SGZ1400 rear scraper conveyor as engineering objects,based on the multi domain collaborative simulation method of machine liquid dispersion coupling,a virtual prototype test model for fully mechanized top coal caving was established,and the discrete element particle parameters were calibrated.The dynamic response law of the whole area of the tail beam impacted by the collapsed coal was analyzed,the critical areas of the tail beam surface dynamic characteristics were proposed,and the best measuring points for coal caving vibration information collection were determined;An experimental platform was constructed for top coal caving multimodal information acquisition;Comparative research on multimodal information data augmentation,time-frequency conversion,and model adaptation mechanisms were also carried out,which laid a theoretical foundation for improving the accuracy of multimodal information fusion recognition;To explore the coal and gangue multimodal information fusion strategy and method,the multi-domain,multi-platform,multi-lingual interactive coal and gangue automatic identification information feedback system was constructed.The main research contents are as follows:(1)In order to overcome the difficulties of difficult data collection underground,long period,high risk factor and difficult to achieve accurate collection for a large number of complex working conditions,based on the virtual prototype technology,a simulation test model of multi-domain collaboration and machine-liquid-dispersion coupling of fully mechanized top coal caving was constructed.And the parameters of particle size,spatial distribution and shape rules of discrete elements were studied and calibrated;the similarity theory of top coal caving system was established to determine the similarity ratio between simulated data and actual data.The results show that the multi-domain co-simulation method can ensure the reliability of the virtual prototype simulation results by setting the discrete element particle radius to 20 mm,the particle size distribution to be normal(standard deviation of 0.05),and the length-to-fine ratio A to be 1.0 when simulating fully mechanized top coal caving.Based on the structural parameters and hydraulic parameters of the top coal caving support,the multi-domain collaborative simulation model can completely simulate the multi-factor coupling behaviors of top coal caving mechanism on the opening and breaking of top coal,coal transportation by scraper conveyor,and the gradual change characteristics of hydraulic system,and then directly and accurately obtain the dynamic signals generated by coal and gangue impact in the process of top coal caving,providing a new method and means for underground real-time data acquisition and monitoring.(2)A bidirectional coupling model of complex coal seam collapse impact tail beam was established based on the DEM-MFBD method;Simulating the impact scattering behavior of coal by using discrete element aggregation method to bond particles;The mechanical transmission characteristics and vibration acceleration characteristics of each articulation point of the support and the abdomen of the tail beam were analyzed through the 3D color mapping surface method;And Clarified the signal strength and trend at different positions of the impact tail beam;Exploring the optimal detection position of coal and gangue identification and monitoring sensors under the impact of coal collapse;Through the full area collapse impact test on the surface of the tail beam,it was found that there are differences in the dynamic characteristics generated during the coal rock impact process at different hinge points of the support.The change trend of the tail beam impact position at the same hinge point is significantly different.The definition of the "low amplitude band" on the surface of the tail beam was proposed.When the impact point is located in this area,the amplitude of the hinge point has critical characteristics,and the impact energy and torque effects are small,making it easy to generate minimum points;By using polynomial curve fitting,R-square size,and absolute value of quadratic coefficient,the priority order of vibration data measurement points during the coal gangue collapse impact process was determined as follows: tail beam abdomen,left(right)piston rod hinge point,tail beam and shield beam hinge point.(3)This paper have constructed a multimodal information collection experimental platform for coal and gangue,designed coal caving experiments with different mixing rates,and created a multimodal information feature database for coal caving.FID(Frechet Inception Distance)and MMD(Maximum Mean Discrepancy)evaluation metrics were introduced to clarify the impact of different data expansion methods on network convergence,exploring the effect of the time-frequency image conversion method of vibration data on the recognition accuracy.An experimental research was carried out on matching modal features with deep learning network models using vibration time-frequency spectrum images,infrared images and RGB images during coal caving process.The experimental results showed that the FID values of WGAN-GP(Wasserstein Generative AdversarialNetwork with Gradient Penalty)network were 48.3% and 62.9% smaller than those of DCGAN(Deep Convolutional Generative AdversarialNetworks)and WGAN(Wasserstein Generative AdversarialNetwork),respectively,and the MMD index was only 56.7% of that of WGAN and also lower than that of DCGAN.The recognition rate of CWT(Continuous Wavelet Transform)time-to-frequency conversion method on ResNet-18 network was 4.8% and 1.2% higher than that of STFT(Short-time Fourier Transform)and HHT(Hilbert-Huang Transform),respectively.The average recognition rates of vibration-time spectral images,RGB images and infrared images on ResNet-18 network were 85.1%,83.9% and 87.7%,respectively,and all three modal image data performed better than other networks on ResNet-18 network,and the recognition rate of infrared images was the highest with the strongest data resolution characteristics.(4)The multimodal information fusion strategy of coal gangue was systematically studied,the early fusion and late fusion of modal features were carried out,and the degree of influence on recognition accuracy and training time under the interaction of three factors and three lightweight network models,namely,spectrum image,infrared image and RGB image during vibration,was investigated based on the orthogonal test method.It has found that the recognition accuracy of the early fusion network is slightly higher than that of the late fusion by 0.09% under the same data set and training parameters,but the accuracy of the late fusion network rises faster at the primary stage of fusion network training,it enters the steady-state interval more rapidly and can accomplish the coal gangue classification target task faster and better when dealing with small-scale data sets.The optimal combination of influencing factors with the highest accuracy and shorter training time for coal and gangue identification was obtained: the ResNet-18 network was selected for spectral images at vibration,the ResNet-50 network was selected for infrared images,and the ResNet-18 network was selected for RGB images.(5)Based on interface technology,fully utilizing Python’s massive open-source library functions and engineering packages,MATLAB/Simulink modular development environment,and interactive modeling features,a Recurdyn-AMEsim-EDEM-Py Charm-Simulink multi domain collaborative automatic recognition system for coal caving blending rate was designed,verified and analyzed.The results show that through the combination of coal and gangue multimodal information fusion and deep learning algorithm,the accurate identification task of different gangue rates can be accomplished,and the real-time transmission and sharing of coal release multimodal data,online sensing and control can be realized at the same time.This paper has 131 figures,31 tables,214 references. |
PDF Full Text Request