Research On Automatic Recognition Of The Coal-rock Interface In Top Coal Caving

Fully mechanized top coal caving is a high-yielding and efficient way of thick seam mining. In our country’s coal reserves, about44%is thick seams. So it has a particularly important significance to research the top coal caving for the development of China’s coal industry. One of the problems encountered is how to control the caving time according to the degree of coal-drop. Traditionally, the degree of the drop all rely on artificial judgment. Potential security problem posed due to the harsh environments. The coal quality will decrease due to the over caving, and the resource will waste due to the owing caving. The key technology to achieve accurate control of the degree of drop is the coal-rock interface recognition(CIR) technology. Supported by the National Natural Science Foundation of China (Grant No.51174126), the features and the recognition of the coal-rock interface are researched.In fact, the CIR is a pattern recognition problem, namely recognizing coal and waste rock via the information in the work environment. The CIR contains three parts of data acquisition, feature extraction and pattern recognition. The research work in this paper is focus on the above three key points:The model of tail beam is theoretically founded and analyzed, and the statistical law of tail beam vibration was proposed to investigate the coal and waste rock during the top coal caving. The conclusion makes sense for the subsequent feature extraction and pattern recognition. Based on the transducer selection principle, a field data acquisition system of CIR is designed. And the best installation position of transducers are founded according to the analysis results of abundant of spot data and contrasting the situations that the transducer be installed in the chute and the tail beam of the hydraulic support. Subsequently, the hydraulic support is reconstructed, thus the transducer can be installed in the best position. Then the field signal of vibration and acoustics are collected. An abundant comprehensive data is obtained for offline analysis.The vibration signals of the tail beam presents non-stationary characteristics. The traditional Fourier transform is a powerful tool to analyze stationary signals while it has not significance for non-stationary signals. The features of the coal-rock interface are extracted based on empirical model decomposition. Firstly, the vibration signals are decomposed into intrinsic mode function components (IMF). Then the further analysis is carried out. The features are proposed based on the energy characteristics of the IMF components, the kurtosis characteristics of the IMF component, and the crest factor characteristics of IMF component. The other three features are obtained in the frequency domain via the Hilbert transform. They are Hilbert spectral energy characteristics, Hilbert marginal spectrum energy characteristics and Hilbert marginal spectrum energy characteristics of the IMF.Time series analysis methods are employed to process the acoustic signals. A novel recognition method based on ARMA model parameters is proposed. Firstly, the data is pre-processed, and then the time series model is determined according to the shapes of the autocorrelation function and the partial autocorrelation function of the signal. The ARMA model is estimated using the two types of signals. The order of the model, autoregressive and the smoothing parameters has been obtained. The residuals are obtained and the model is verified by the residuals. The bispectrum is estimated using the parameters. The maximum value number of the bispectrum diagnoal energy curve is different between the coal-drop and the waste rock-drop. So it can be as the feature. Another feature based on the variance of the residuals is extracted according to the idea of system recognition. The signals are recognized using EWMA chart. The results show that the overall recognition rate is90%.The BP neural network is designed according to the features of the vibration signal and acoustic signal aim to automatically recognize the coal-rock interface. Five improved training function are compared, including the momentum BP algorithm, the adaptive learning rate algorithm, the quasi-newton algorithm, the flexible BP algorithm, and the levenberg-marquardt algorithm. Then the optimal network is designed. The weights and thresholds of the neural network layers are determined by training the network using the sample data. The number of hidden layer neurons is calculated in case of different features. Thirdly, the vibration signals are recognized based on three different features, including the energy, the kurtosis and the crest factor of the IMF. The results indicate that the three vectors all can be as the features to recognize the coal-rock interface. The recognition rate based on the IMF energy is higher than the other two methods. Meanwhile, the acoustic signals are recognized by the residual variance with the neural network. The recognition results show that the neural network method has higher recognition rate than the statistical method. Finally, the information fusion of the vibration signal and acoustic signal is executed with the neural network. The recognition results present that the information fusion method has higher recognition rate than the separate recognition method.