| As a main mining method of thick coal seam in China,fully mechanized caving mining not only has a coal mining face that keeps the working face moving forward in front,but also has a working face that continuously releases the top coal behind,which greatly improved the efficiency of coal production in coal mining face.The top coal caving technology for fully mechanized caving mining is controlled by the expansion and retraction of the flashboard of hydraulic support.However,the key technology is still controlled by the coal workers manually operating the electro-hydraulic reversing valve at present.The coal workers control top coal caving by ear or visual observation,which can easily lead to the problem of “over caving” or “under caving”.It seriously affects the top coal mining rate in fully mechanized caving face.Besides,the health of coal workers is seriously endangered because of the poor working environment,large coal dust,poor lighting,cramped space in fully mechanized caving face.The realization of automatic top coal caving can enable the coal workers away from the working face and remotely control the top coal caving,thereby protecting the health problems of coal workers.Meanwhile,the collapsing coal and rock identification can provide theoretical basis for the control of top coal caving,and then solves the problem of “over caving” or “under caving”,increase the recovery rate and quality of top coal.Therefore,in the research task of coal science and technology basic theory identified in the “Thirteenth Five-year Plan” of the coal industry,the basic theory of coal and rock interface automatic identification is mentioned,which can provide a theoretical support for the safe and efficient development of coal.To realize the automatic identification of the collapsing coal and rock at fully mechanized caving face,the non-stationary vibration signals caused by the impact of the tail beam of hydraulic support and the collapsing coal-rock are used as the original data.The vibration signal features are extracted by using fractal theory and wavelet packet transform,and the feature dimension is reduced by using principal component analysis(PCA)and manifold learning method,and then the effective feature vectors representing the collapsing coal and rock are obtained.Based on the effective feature vectors,the classification model-BP neural network is designed and trained.Therefore,a collapsed coal-rock recognizing model using the non-stationary vibration signals at the fully mechanized caving mining field is established,and the automatic judgment and identification of the collapsed coal-rock at the fully mechanized caving face is realized,which provides a theoretical basis for the automation of the top coal caving.(1)The vibration signals in top coal caving process at fully mechanized caving mining field are accurately and completely collected.According to the significant difference between the vibration signal caused by the impact of tail beam and top coal and the vibration signal caused by the impact of tail beam and roof rock,a method to identify the collapsing coal-rock based on the vibration signal is proposed.Based on the GBC1000 acceleration sensor and YHJ(C)mine portable vibration recorder,the vibration signal acquisition test is carried out at 9201 fully mechanized caving face,and then the vibration signals of top coal caving and roof rock caving conditions(including large top coal caving and large rock caving)are completely collected at fully mechanized caving mining field.The time nodes and duration of different working conditions are also accurately recorded,which provides the rich and effective raw data for the collapsing coal and rock identification at fully mechanized caving face.The top coal caving samples and roof rock caving samples with different time lengths are used to verify the non-stationarity of the vibration signals,the variance of the two types of samples varies significantly with the length of the sample.The vibration signal exhibits non-stationarity,which determines to use the time-frequency analysis method to extract the feature of the vibration signal.(2)A method for processing vibration signals in special conditions based on EMD or based on singular value difference spectrum is proposed.In view of the vibration signals in special conditions(large top coal and large rock caving)during the process of top coal collapsing,the vibration signals are decomposed into several intrinsic mode functions(IMF)by using empirical mode decomposition(EMD)and then several appropriate IMFs are selected to reconstruct the vibration signals,the characteristics of the reconstructed signals are similar to the vibration signals in normal top coal caving and roof rock caving conditions.The Hankel matrices are constructed based on the vibration signals,and the matrices decompose into a series of component signals by using singular value decomposition(SVD),and then several large components are selected based on the singular value difference spectrum for the signals reconstruction,the characteristics of the reconstructed signals are also similar to the vibration signals in normal top coal caving and roof rock caving conditions.Both methods have achieved good results in dealing with the vibration signals in special conditions,and have enriched the original data for the identification of collapsing coal and rock in fully mechanized caving face.Furthermore,the vibration signals in top coal caving condition and roof rock caving condition(including reconstructed signals in special conditions)are continuously and isochronally cut into samples for subsequent feature extraction and collapsing coal-rock identification.(3)By comparing the proposed feature extraction method fractal theory,wavelet packet transform,the feature dimensionality reduction method PCA,LTSA,the effective features characterizing collapsing coal and rock are obtained.For the non-stationarity of the collected vibration signal,the signal is described quantitatively by using fractal box dimension,the wavelet packet transform is used to decompose the signal,and then the traditional wavelet features(energy,energy entropy,energy moment,sample entropy)and wavelet packet energy flow are used to describe the detail characteristics of the signal.Finally,the effectiveness and extraction efficiency of each features are compared.The comparison shows that the fractal box dimensions of the two types of samples are different,but the difference is not significant and is likely to cause misjudgement.The wavelet packet band energy and the wavelet packet energy moment of two kinds of samples have obvious differences in multiple frequency bands,so they are effective features.The wavelet packet sample entropy only differs in frequency band 1,and it is easy to cause misjudgement as a characteristic of the collapsing coal and rock.The wavelet packet band energy entropy is similar in 16 frequency bands,so it is not an effective feature;wavelet packet energy flow has obvious differences so it is an effective feature.Therefore,a feature extraction method based on fractal box dimension and traditional wavelet packet features(energy,energy entropy,energy moment,sample entropy)is proposed,and a feature extraction method based on wavelet packet energy flow is put forward.In addition,for the problem of real-time identification of collapsing coal and rock,the time-consuming of wavelet packet feature extraction and the fractal box dimension feature extraction are innovatively compared.It is found that the time-consuming of these features extraction is less than 0.3s,which meets the requirement of real-time performance.In view of the dimensions of above feature vectors are too large,a feature reduction method based on principal vector analysis(PCA)is proposed,and a feature reduction method based on local tangent space alignment(LTSA)is proposed,and then the effectiveness of the low dimensional feature vectors and reduction efficiency are compared based on PCA and LTSA.The comparison shows that the PCA low-dimensional feature vectors of the vectors composed of fractal box dimension and the traditional wavelet packet characteristics(energy,entropy,energy moment,sample entropy)are respectively valid,not valid,valid,effective but not significant,the PCA low-dimensional feature vector of the vector formed by the wavelet packet energy flow is an effective feature,the LTSA low-dimensional embedding of the vector composed of fractal box dimension and traditional wavelet packet features(energy,energy entropy,energy moment,sample entropy)are respectively not valid feature,valid feature,not valid feature,effective feature,the LTSA low-dimensional embedding of the vector formed by wavelet packet energy flow is an effective feature.Therefore,the obtained effective features characterizing collapsing coal-rock are feature vector composed of fractal box dimension and wavelet packet energy,and its PCA low-dimensional feature,feature vector composed of fractal box dimension and wavelet packet energy moment,and its PCA low-dimensional feature,feature vector formed by wavelet packet energy flow,and its PCA low-dimensional feature,the LTSA low-dimensional embedding of the vector composed of fractal box dimension and wavelet packet energy entropy,the LTSA low-dimensional embedding of the vector composed of fractal box dimension and wavelet packet sample entropy,the LTSA low-dimensional embedding of the vector formed by wavelet packet energy flow,which not only provides the basis for the identification of collapsing coal and rock,but also provide a sufficient and effective input for the classification model for recognizing collapsing coal and rock.Meanwhile,the two dimensionality reduction methods efficiency is innovatively compared,it is found that the average time-consuming of LTSA dimension reduction is 44.199 s,which is 7.9 times that of PCA,and its efficiency is low.(4)A collapsing coal-rock recognizing model based on the PCA low-dimensional feature of the vector composed of fractal box dimension and wavelet packet energy moment and BP neural network is proposed,a recognizing model based on wavelet packet energy flow and BP neural network is also put forward.The above feature vectors are taken as the input of BP neural network respectively,and then the BP neural networks are designed and trained,the recognition accuracy and training efficiency of BP neural network with different inputs are compared.The resulting effective feature vectors characterizing collapsing coal and rock are as follows: the PCA low-dimensional feature of the vector composed of wavelet packet energy moment and fractal box dimension(recognition rate is 100%,time-consuming of feature extraction and neural network training is 8.753s),the PCA low-dimensional feature of wavelet packet energy flow(recognition rate is 100%,time-consuming of feature extraction and neural network training is 8.916s),feature of wavelet packet energy flow(recognition rate is 97.5%,time-consuming of feature extraction and neural network training is 5.797s),the PCA low-dimensional feature of the vector composed of wavelet packet energy and fractal box dimension(recognition rate is 95%,time-consuming of feature extraction and neural network training is 8.479s),the LTSA low-dimensional embedding of the vector composed of wavelet packet sample entropy and fractal box dimension(recognition rate is 95%,time-consuming of feature extraction and neural network training is 74.863s),and the LTSA low-dimensional embedding of the vector composed of wavelet packet energy entropy and fractal box dimension(recognition rate is 92.5%,time-consuming of feature extraction and neural network training is 73.863s).Further,the most effective features characterizing collapsing coal-rock are obtained,they are the PCA low-dimensional feature of the vector composed of wavelet packet energy moment and fractal box dimension,and the feature formed by wavelet packet energy flow.ACO is used to optimize the BP neural network with the input of effective feature vectors,and then the training efficiency of BP neural network and ACO-BP neural network is compared.Only with the input of the two effective LTSA low-dimensional embedding,the training time-consuming of neural network is reduced and the training efficiency is improved after optimization using ACO.Thus,for the two effective LTSA low-dimensional embedding,the effective classification model is ACO-BP neural network.While for the remaining four valid feature vectors,the effective classification model is BP neural network.Therefore,the recognizing model based on the PCA low-dimensional feature of the vector composed of fractal box dimension and wavelet packet energy moment and BP neural network and the recognizing model based on wavelet packet energy flow and BP neural network for effectively identifying the collapsing coal and rock are obtained,which provides effective theoretical support for further automation of top coal caving. |
PDF Full Text Request