Pick Wear State Recognition Based On Multi-source Heterogeneous Information And Deep Learning

With the further development of mining machinery to automation and intelligence,the intelligent diagnosis technology of mechanical fault is very important.The pick is one of the most important parts in the process of mining,whose wear state has directly influence on the mining efficiency and production cost.More effective methods to monitor the status of pick have become the research focus for a large number of scholars,and the research on pick wear state recognition technology has great potential and application value.Therefore,the pick wear state is taken as the research object,the deep learning network is used as the recognition tool,and focuses on recognition model based on signal denoising,signal characteristics,cross dimensional multi-source heterogeneous signal fusion and algorithm improved.There are a lot of AE information changes in the process of mining.The control system of pick cutting test bench,SAEU3 S signal acquisition and analysis system are constructed.A certain proportion of coal-rock specimens are poured,and according to the quality and size difference,the wear state of pick is defined,which can be divided into four different wear degrees: new teeth,initial wear,severe wear and failure.The recognition model of signal acquisition,signal analysis and BP neural network based on SAEU3 S acoustic emission system is constructed to provide a large number of AE signal samples and comparative data for multi-source heterogeneous signal fusion and deep learning recognition model.In order to filter and denoise the AE signal better,Daubechies 12 is used to transform the AE signal via wavelet packet.Three-layer wavelet packet decomposition is uesd to analyze the denoising effect of acoustic emission to obtained the reconstructed signal with high resolution and the AE signal feature database is constructed.A pick wear state recognition model based on SOM neural network is constructed,whose recognition accuracy is about 91%,which provides data support and comparative analysis of recognition effect for deep learning recognition model.In order to obtain the effective information in images,FPV image transmission integrated camera image acquisition system are constructed,and the surface of pick with different wear degree is observed via Keyence VHX-5000 ultra depth of field three-dimensional microscope in20 x magnification.the modulus maxima denoising method,wavelet coefficient correlation denoising method and threshold denoising method are used to process the images to study the the advantages and disadvantages of each of them via comparing the peak signal-to-noise ratio(PSNR),the mean square error(MSE)and the normalized correlation(NC).The wavelet packet decomposition denoising method based on Mallat algorithm is used to decompose the pick image into single-layer and two-layer.And the image effect is enhanced and the signal-to-noise ratio(SNR)is increased after two-layer decomposition through analyzing gray histogram of the original image and reconstructed image.Aiming at the problems of weak light and low contrast between object and background,the method of combining frequency domain filter enhancement and density segmentation technology is adopted to eliminate the image degradation caused by insufficient lighting,which can better protect the details and improve the contrast of features.The contour of the pick is extracted reasonably by Hough transform,and the image is smoothed properly,which provides a reasonable edge extraction method for image recognition.The motion blur degradation model of pick image and the image restoration model based on Lucy Richardson algorithm is established.The filtering effect are evaluated by SNR,PSNR and improving signal-to-noise ratio(ISNR).The results shows that when the number of iterations is 15,the restoration effect is the best.And the method provides image information with obvious features for two-source signal fusion and pick state recognition.By analyzing the data characteristics of AE signal after wavelet packet transform,a pick wear monitoring model based on long-short term memory network(LSTM)is established.After analyzing the convergence effect,the training times of LSTM recognition model is 3000,and the learning rate is 0.01.The recognition rate of this model is 93%,which is 11.9% higher than that without wavelet packet processing.It proves that the wavelet packet signal processing is reasonable,and the deep learning recognition model has higher accuracy.A multi-source heterogeneous signal fusion model based on convolution neural network(CNN)is established.one-dimensional acoustic emission signal and two-dimensional image are processed by 1D CNN and 2D CNN multiple convolution layer and pooling layer to fuse heterogeneous information.The LSTM depth recognition model based on Adam algorithm is established,and the objective function error of LSTM is minimized by self-adaptive algorithm of Adam learning rate.The results show that the average recognition rate based on CNN-Adam-LSTM recognition model is 97.5%,and the recognition time is 3.1 seconds,which greatly speeds up the training speed,and the generalization ability of the model is strong.It can be seen that the recognition accuracy of multi-source signal is 3.85% higher than that of single signal.Similarly,the recognition accuracy of LSTM is 8% higher than that of BP and SOM,which verifying the rationality of the pick wear recognition model established in this paper.This paper has 105 figures,18 tables and 159 references.