J. Cent. South Univ. (2020) 27: 3078-3089

DOI: https://doi.org/10.1007/s11771-020-4530-8

Discrimination of mining microseismic events and blasts using convolutional neural networks and original waveform

DONG Long-jun(董陇军)¹, TANG Zheng(唐正)¹, LI Xi-bing(李夕兵)¹,CHEN Yong-chao(陈永超)¹, XUE Jin-chun(薛锦春)²

1. School of Resources and Safety Engineering, Central South University, Changsha 410083, China;

2. School of Energy and Mechanical Engineering, Jiangxi University of Science and Technology,Nanchang 330013, China

Central South University Press and Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract: Microseismic monitoring system is one of the effective methods for deep mining geo-stress monitoring. The principle of microseismic monitoring system is to analyze the mechanical parameters contained in microseismic events for providing accurate information of rockmass. The accurate identification of microseismic events and blasts determines the timeliness and accuracy of early warning of microseismic monitoring technology. An image identification model based on Convolutional Neural Network (CNN) is established in this paper for the seismic waveforms of microseismic events and blasts. Firstly, the training set, test set, and validation set are collected, which are composed of 5250, 1500, and 750 seismic waveforms of microseismic events and blasts, respectively. The classified data sets are preprocessed and input into the constructed CNN in CPU mode for training. Results show that the accuracies of microseismic events and blasts are 99.46% and 99.33% in the test set, respectively. The accuracies of microseismic events and blasts are 100% and 98.13% in the validation set, respectively. The proposed method gives superior performance when compared with existed methods. The accuracies of models using logistic regression and artificial neural network (ANN) based on the same data set are 54.43% and 67.9% in the test set, respectively. Then, the ROC curves of the three models are obtained and compared, which show that the CNN gives an absolute advantage in this classification model when the original seismic waveform are used in training the model. It not only decreases the influence of individual differences in experience, but also removes the errors induced by source and waveform parameters. It is proved that the established discriminant method improves the efficiency and accuracy of microseismic data processing for monitoring rock instability and seismicity.

Key words: microseismic monitoring; waveform classification; microseismic events; blasts; convolutional neural network

Cite this article as: DONG Long-jun, TANG Zheng, LI Xi-bing, CHEN Yong-chao, XUE Jin-chun. Discrimination of mining microseismic events and blasts using convolutional neural networks and original waveform [J]. Journal of Central South University, 2020, 27(10): 3078-3089. DOI: https://doi.org/10.1007/s11771-020-4530-8.

1 Introduction

Microseismic monitoring system as one of the effective methods of deep mine geo-stress monitoring, can monitor rock micro fracture and collect source parameters. The information provides reliable data support in monitoring the stability change of rock damage area [1-3], solving rock burst and other rock disaster problems [4-7], inverting of focal mechanism [8-11], and evaluating the geo-stress activity in the deep rock mining process [12-14]. Compared with traditional monitoring methods only focusing on points, the microseismic monitoring not only directly measures some mechanical parameters of the rockmass, but also determines the location and source parameters of fracture area [15-17]. Because source location and disaster control are based on accurate classification of monitoring signals, other interference signals need to be removed for obtaining more accurate and pure microseismic signals induced by geo-stress. Although the microseismic monitoring system widely apply in rockmass stability observation and analysis, it is difficult to provide accurate microseismic information in real-time under the environment mixed with various kinds of noises and inevitable blasts. Therefore, it is very important to remove the noise. The traditional discrimination between microseismic events and blasts relies on manual experience, which leads to errors in the classification and the analysis of microseismic events. Existing studies reported some classification methods, which can be mainly divided into the spectrum analysis, statistical method analysis, and machine learning.

FRANTTI et al [18] compressed the seismic and blasting signals by hundreds of times in order to convert the earthquake frequency into an acceptable range. According to the uneven distribution caused by the source and sensor location, the difference of propagation length, and the different magnitude, two thirds of seismic signals can be correctly identified. BOOKER et al [19] proposed a method for the classification between microseismic events and blasts. The research object was the category composed of collected microseismic events and blasts. The parameters were composed of the energy ratios contained in predetermined "velocity windows" on the seismograms The accuracy of the method was 85%. TAYLOR [20] established a Pg/Lg discriminant with high frequency between 0.5 and 10 Hz. With the increase of frequency, the discriminant displayed clearer difference between earthquakes and explosions. The proposed multivariate discriminant analysis was carried out using maximum likelihood Gaussian classifier and back propagation (B-P) neural network. 95% of the events can be correctly identified. MALOVICHKO [21] used multivariate maximum likelihood Gaussian classifier to classify normal events and blasts in mines. The time of occurrence, similarity of the seismic signals with the neighbouring waveforms, ratio of high-frequency and low-frequency radiation, and radiation pattern were selected as discriminating factors. The large-scale application of the discriminator indicated that about 20% of seismic events were reclassified as blasts, which provided a new way to remove blasts from normal events. JIANG et al [22] used fast Fourier transform (FFT) spectrum analysis to classify rock fracture signals and blasting signals in mine. They analyzed the collected signals and compared the difference of energy distribution in the spectrum. Results showed that the energy released by blasts was distributed in the low-frequency region of 0- 30 Hz, while the distribution of the energy released by microseismic events was mostly in the region of 30-50 Hz. Due to the high requirement on the professional knowledge in spectrum analysis, it is difficult to use this method in practical engineering.

Since the microcracks in the rock release energy in the form of seismic waves, the different source mechanism of blasts and microseismic events means that the source parameters contained in two types of signals are different. DOWLA [23] adopted the neural network to solve a series of problems including the classification of microseismic events. BENBRAHIM et al [24] proposed a method called “modified Mexican hat wavelet” for the representation stage. They proposed an algorithm based on the random projection and the principal component analysis. ZHAO et al [25] used the slope value of the starting-up trend line obtained by the linear regression to replace the angle. Two slope values related to the coordinates of the first peak and the maximum peak were extracted as the characteristic parameters. MULLER et al [26] proposed a classification method based on seismometer network for earthquakes, mining blasts, and rock bursts. This method adopted the multilayer neural network for fusion and conducted fuzzy codes for the input characteristics combined with the signal characteristics. The results of classifier could be stable and reach to greater than 90%. ORLIC et al [27] presented a new classification method for distinguishing natural earthquake events and artificial events. This method used a specially constructed genetic algorithm to search for the near optimal seismic features automatically. The accuracy of this classification method was 85%. DONG et al [28] used random forests, support vector machine, and Na1ve Bayes classifier to classify microseismic events and blasts. The results showed that the random forest algorithm had the highest classification accuracy. The models not only automatically classified microseismic events with high accuracy, but also classified the discriminant indexes according to the calculated value of weight. VALLEJOS et al [29] used logistic regression and neural network to classify microseismic events and blasts and calibrated the models with 13 seismic parameters and positioning errors provided by ESG full waveform system. The logistic regression and neural network performed well and their accuracy were both more than 95% under their own optimal decision threshold. But the generalization effect of logical regression was better than that of neural network when the third category of labels was introduced. DONG et al [30] compared the probability density of different parameters according to the different characteristics of microseismic and blasting waveform. Five typical parameters as well as time probability density functions, and probability density functions of origin time difference for neighbouring blasts were extracted as discriminating indexes. Finally, the Fisher classifier, Na1ve Bayesian classifier and logistic regression were used to classify the two kinds of waveforms. SHANG et al [31] used principal component analysis (PCA) and artificial neural network (ANN) to analyze 22 parameters of 1600 microseismic events, and compared them with logistic regression, Bayes and Fisher classifiers. The results showed that the ANN based on PCA works best. Although it is an effective method to classify the microseismic events and blasts according to source parameters, the analysis of source parameters is based on the extraction of parameters. In practical application, the problem of classification depending on individual experience has been solved, but the parameter extraction and model selection still need researchers’ subjective judgment, which determine whether the classification model is accurate or not. Secondly, the correlation between parameters is barely considered in the selection, which may cause poor recognition effect. As each parameter needs to be analyzed including their fitness with the specific model and correlation between other parameters before classification model is established, the calculation time will increase accordingly.

With the development of artificial intelligence, image feature extraction and other methods can also be used to classify different types of images. From various machine learning algorithms, the convolutional neural network (CNN) can conduct supervised deep network training without unsupervised pre-learning. The CNN is proposed from the structure of visual system, especially the structural model of cat visual cortex [32]. FUKUSHIMA [33] established the first neural network based on local connection and hierarchical structure between neurons. The CNN is a model specially used for training two-dimensional data. It reduces the number of training parameters by excavating the spatial correlation of image data and using weight sharing. The closely connection between layers and spatial information in CNN can automatically extract image features, which makes it have a good performance in image classification. The CNN has been widely used in flotation process [34], face recognition [35, 36], document classification [37, 38], voice monitoring [39], traffic identification [40], and other fields. We usually argue that the that caused by blasts is simple than P wave caused by the rock microfracture. The energy released by blasts mainly concentrated in the first half of the timeline, sometimes it is obvious in the wave window of several same waveform in the blasts. Compared with blasting waveform, microseismic waveform has obvious S wave and the energy released by microseismic events distributes in all time periods. The features of the microseismic events and blasts are shown in Figure 1.

The high professional knowledge requirement of spectrum analysis methods is not applicable to the practical engineering, and the selection of parameters and models of statistical methods still rely on individual experience. More importantly, the source parameters used in the above methods are extracted from the original waveforms according to criteria or experience, which may cause the partial loss of the information or mixed subjective factors affecting the real expression. The waveform can provide the most complete source information as the original information. The CNN classification model for discriminating microseismic events and blasts can not only retain all parameter information based on the original waveform signal, but also avoid the influence of subjective experience on the classification results. Therefore, this paper establishes a wave classification model based on CNN for the classification of microseismic events and blasts.

Figure 1 Waveform characteristic of blasts and microseismic events

2 Classification model based on convolutional neural networks

2.1 Convolutional networks

A classification model based on CNN is proposed to extract and learn the local features of waveforms. The CNN uses receptive field and weight sharing to sample images. The CNN not only extracts each pixel, but also synthesizes the local information at a deeper level to get the global information during the training process. The model includes input layer, convolutional layer, pooling layer, full connection layer and output layer. It is used to classify microseismic events and blasts.

Compared with traditional methods, the CNN can directly learn the original waveform, which avoids the information omission caused by the feature extraction. The input layer connects the waveform data set and the hidden layers. The waveforms are imported into the convolution layer to extract features.

The process on the convolutional layer is to output the new feature map through the interaction of the convolutional kernel and activation function. The convolutional kernel is a matrix or vector, which can be regarded as a kind of filter. The networks multiply and sum the input features in the receptive field by matrix elements, and stack the deviation [41]:

 (1)

where D^l, D^l+1, b, S_l+1, D(i, j), k, f, s₀ and p indicate the input, output, bias, size of D^l+1, pixel point of the feature map, convolutional kernel size, convolutional stride, and padding parameter, respectively.

As the CNN may unintentionally lose some information during the convolutional process, image features should be extracted as much as possible during the up-sampling process. The pooling layer is added after the convolutional layer, which intends to sample the feature map and output the new one with a lower data amount. The form of the pooling layer [42] is expressed as:

         (2)

When p is equal to 1, the pooling layer works in the way of average pooling. When p tends to infinity, the working pattern changes to max pooling. The length and width of the pool window are h and w, respectively. The output feature map will be reduced to 1/(h×w) of the convolutional surface after the pooling layer.

In the fully connected network, the sampled two-dimensional feature map is transformed into one-dimensional feature values and the feature values are input into the fully connected layer. Each neuron in the full connected layer is connected to all the neurons in the upper layer. The output value is the inner product of the output vector of the previous layer as well as the weight vector of the full connection layer and plus bias.

2.2 Source classification model architecture

The ideological framework of the classification model based on the waveform is shown in Figure 2. The “Database” area contains the waveform data sets in which the microseismic events and blasts are represented by 0 and 1, respectively. The “Performance” area summarizes the structure of the classification model. In the “Detection” area, the feature maps extracted from waveform are used to train the parameter. Test set and validation set are tested in the “Prediction” area to evaluate the accuracy of the model.

The number of training parameters increases with the rise of network layers. As a result, the model becomes more complex. The training time will increase exponentially. When the training parameters reach a certain quantity, over-fitting and other phenomena will even occur. To accurately and quickly classify microseismic events and blasts, we proposed a full waveform classification model with four convolutional layers. The convolutional layer, with the kernel size of 2×2, works in the way of valid padding. The extraction process of characteristic values in the convolutional layer is shown in Figure 3. Two pooling layers are selected as the maximum pooling and the size of the pooling window is 2×2. We add a fully connected layer at the end of the model. If training samples fail to meet the requirements of training parameters, it may lead to over-fitting phenomenon. “Dropout” is a simple and effective regularization technique. Adding “Dropout” into the CNN is to randomly stop some nodes from working. The basic idea is to improve the generalization ability by preventing the joint action of the feature detector. “Dropout” allows the CNN to accept the output results of different nodes at the same probability, which avoids the situation that an individual feature is effective only under some features. The function of “Dropout” in the neural network is shown in Figure 4. To improve the effect of model training and reduce the risk of over-fitting, “Dropout” is set to 0.25. The structure of the classification model is shown in Figure 5.

Figure 2 Ideological framework of classification model based on waveform

3 Monitoring data and applications

3.1 Monitoring data set

Rock deformation in the mining process mainly includes elastic deformation and inelastic deformation. The stored elastic potential energy released in the form of vibration waves during inelastic deformation. Since microseismic events are associated with rock deformation or crack growth, the microseismic events contain a lot of useful information about rock failure or evolution process of geological hazard. By analyzing the signal characteristics of the microseismic events, the state and mechanical behavior of rockmass during excavation can be inferred. We can also predict the potential failure of rock and take effective prevention and control measures to avoid the dangerous accidents [43]. The signals collected by the microseismic monitoring system mainly consist of microseismic events, blasts, and machinery vibration. Microseismic events occur during rock deformation and crack propagation, which are related to the mechanical behavior of rocks. Blasts are generated when the rockmass is directly broken by explosion.

Figure 3 Extraction process of characteristic values in convolutional layer

Figure 4 Function of “Dropout” in neural network:(neurons marked with “x” temporarily lose output)

Figure 5 CNN structure of classification model for microseismic events and blasts

The classification of microseismic events and blasts is a key problem in data processing. The CNN is an important method in image identification. The waveform images can be directly imported into the model. Compared with the traditional image identification algorithm, the CNN processes every pixel area of the image, instead of dealing with every pixel point, which enhances the continuity of image information and the speed of image identification without feature selection.

3.2 Results analysis and discussion

The training set S1 consists of 5250 blasts and 5250 microseismic events. The test set S2 composed of 1500 blasts and 1500 microseismic events is selected to test the trained model. The detailed information of the data sets is listed in Table 1. The images used are shown in Figure 6. The optimization function adopts the random stochastic gradient descent (SGD). The parameters of the CNN model include learning rate, base lr, momentum, weight decay, dropout, and maximum number of iterations, whose values are 0.01, 0.01, 0.9, 1×10^-6, 0.25 and 200, respectively.

The training results are relatively stable after 30 iterations and each iteration takes 175 s. Parameters of the input layer, hidden layer and output layer in the CNN model are shown in  Table 2. The test set S2, blasting test set S3 and microseismic test set S4 are imported into the model for testing. The accuracy is finally stabilized at more than 98%. In addition, validation set S5 and S6 are imported into the model for verification. The classification results and the accuracy-loss curve are shown in Figure 7 (the label of the microseismic event is 0 and the label of the blasts is 1).

Classification results of the CNN model on each data set is shown in Table 3. Under the same training set and test set, the final accuracy of ANN reaches 73.37% and 67.90%, respectively. The accuracy of logistic regression reaches 65.99% and 54.43% under the same training set and test set, respectively. As each neuron in the neural network is fully connected, the number and parameters will increase as the number of the samples increases. The over-fitting phenomenon will occur when the number of samples is not enough to meet the training requirements. The ROC curve and the classification results of CNN, ANN and logistic regression models are shown in Figure 8. Compared with the other two machine learning algorithms, the CNN has the characteristic of weight sharing, and the established model has a low demand for samples, which is easy to avoid over-fitting. Since most microseismic monitoring systems in mines still adopt manual classification, the proposed classification model has an absolute advantage in the discrimination of mining microseismic events and blasts.

4 Conclusions

To correctly analyze the microseismic monitoring data, we proposed a classification model between microseismic and blasts events using the CNN. After loading the trained weight, the accuracy in test set S2, blasting test set S3, and microseismic test set S4 are 99.40%, 99.33% and 99.47%, respectively. In addition, 1500 previously unused images are selected as the unlabeled validation sets. It shows that 14 waveforms out of 750 blasts are misidentified and all microseismic waveforms are correctly classified. By comparing CNN with ANN and logistic regression, we found that there is almost no over-fitting phenomenon under the same sample conditions. The accuracy of the CNN is the highest among the three models. The established CNN model, which directly accept the full waveform images as input, can be applied in mining engineering conveniently. In conclusion, the proposed classification method can minimize interference caused by the source parameters and spectrum analysis, and improve the speed and accuracy in the microseismic monitoring data processing.

Table 1 Number and category of samples for training set, test set, and verification set

Figure 6 Data of microseismic events and blasts:

Table 2 Parameters of input layer, hidden layer, and output layer in the CNN model

Figure 7 Accuracy/loss curves of classification results

Table 3 Parameters of input layer, hidden layer, and output layer in CNN model

Contributors

The overarching research goals were developed by DONG Long-jun, TANG Zheng, and XUE Jin-chun. TANG Zheng, LI Xi-bing, and CHEN Yong-chao provided the microseismic monitoring data, and analyzed the measured data. DONG Long-jun and TANG Zheng established the models and classified the microseismic events and blasts. DONG Long-jun and TANG Zheng analyzed the calculated results. The initial draft of the manuscript was written by DONG Long-jun and TANG Zheng. All authors replied to reviewers’comments and revised the final version.

Figure 8 ROC curves and classified results of CNN, ANN and Logistic Regression models (“validation” data set consists of S5 and S6):(data set “Validation” contains data set S5 and data set S6)

Conflict of interest

DONG Long-jun, TANG Zheng, LI Xi-bing, CHEN Yong-chao, and XUE Jin-chun declare that they have no conflict of interest.

References

[1] ZHANG Chu-xuan, LI Xi-bing, DONG Long-jun, MA Ju, HUANG Lin-qi. Analysis of microseismic activity parameters pre- and post roof caving and early warning [J]. Chinese Journal of Rock Mechanics & Engineering, 2016, 35: 3214-3221. DOI: 10.13722/j.cnki.jrme.2015.0928.

[2] XU Nu-wen, LIANG Zheng-zhao, TANG Chun-an, DAI Feng, ZHOU Zhong, SHA Chun. Three-dimensional feedback analysis of rock slope stability based on microseismic monitoring [J]. Chinese Journal of Rock Mechanics and Engineering, 2014, 33: 3093-3014. DOI: 10.13722/j.cnki.jrme.2014.s1.071.

[3] MONDAL D, ROY P N S. Fractal and seismic b-value study during dynamic roof displacements (roof fall and surface blasting) for enhancing safety in the longwall coal mines [J]. Engineering Geology, 2019, 253: 184-204. DOI: 10.1016/ j.enggeo.2019.03.018.

[4] GONG Feng-qiang, SI Xue-feng, LI Xi-bing, WANG Shan-yong. Experimental investigation of strain rockburst in circular caverns under deep three-dimensional high-stress conditions [J]. Rock Mechanics and Rock Engineering, 2019, 52(4): 1459-1474. DOI: 10.1007/s00603-018-1660-5.

[5] MA Tian-hui, TANG Chun-an, TANG Shi-bin, KUANG Liang, YU Qun, KONG De-qing, ZHU Xu. Rockburst mechanism and prediction based on microseismic monitoring [J]. International Journal of Rock Mechanics and Mining Sciences, 2018, 110: 177-188. DOI: 10.1016/j.ijrmms.2018. 07.016.

[6] MA Chun-chi, LI Tian-bin, ZHANG Hang, JIANG Yu-peng, SONG Tao. A method for numerical simulation based on microseismic information and the interpretation of hard rock fracture [J]. Journal of Applied Geophysics, 2019, 164: 214-224. DOI: 10.1016/j.jappgeo.2019.03.014.

[7] DOU Lin-ming, CAI Wu, CAO An-ye, GUO Wen-hao. Comprehensive early warning of rock burst utilizing microseismic multi-parameter indices [J]. International Journal of Mining Science and Technology, 2018, 28: 767-774. DOI: 10.1016/j.ijmst.2018.08.007.

[8] MA Ju, DONG Long-jun, ZHAO Guo-yan, LI Xi-bing. Discrimination of seismic sources in an underground mine using full waveform inversion [J]. International Journal of Rock Mechanics and Mining Sciences, 2018, 106: 213-222. DOI: 10.1016/j.ijrmms.2018.04.032.

[9] ZHANG Peng-hai, YANG Tian-hong, YU Q L, XU Tao, SHI Wen-hao, LI Shu-cai. Study of a seepage channel formation using the combination of microseismic monitoring technique and numerical method in Zhangmatun iron mine [J]. Rock Mechanics and Rock Engineering, 2016, 49: 3699-3708. DOI: 10.1007/s00603-016-1007-z.

[10] AMINZADEH F, TAFTI T A, MAITY D. An integrated methodology for sub-surface fracture characterization using microseismic data: A case study at the NW Geysers [J]. Computers & Geosciences, 2013, 54: 39-49. DOI: 10.1016/ j.cageo.2012.10.015.

[11] ZHAI Hong-yu, CHANG Xu, WANG Yi-bo, YAO Zhen-xing. Inversion for microseismic focal mechanisms in attenuated strata and its resolution [J]. Chinese Journal of Geophysics-Chinese Edition, 2016, 59: 3025-3036. DOI: 10.6038/cjg20160825. (in Chinese)

[12] YANG Cheng-xiang, LUO Zhou-quan, TANG Li-zhong. Study on rule of geostatic activity based on microseismic monitoring technique in deep mining [J]. Chinese Journal of Rock Mechanics and Engineering, 2007, 26(4): 818-824. DOI: 10.1016/S1872-2067(07)60020-5.

[13] DONG Long-jun, SUN Dao-yuan, LI Xi-bing, DU Kun. Theoretical and experimental studies of localization methodology for AE and microseismic sources without pre-measured wave velocity in mines [J]. IEEE Access, 2017, 5: 16818-16828. DOI: 10.1109/ACCESS.2017.2743115.

[14] DONG Long-jun, SHU Wei-wei, LI Xi-bing, HAN Guang-jie, ZOU Wei. Three dimensional comprehensive analytical solutions for locating sources of sensor networks in unknown velocity mining system [J]. IEEE Access, 2017, 5: 11337-11351. DOI: 10.1109/ACCESS.2017.2710142.

[15] DONG Long-jun, SHU Wei-wei, HAN Guang-jie, LI Xi-bing, WANG Jian. A multi-step source localization method with narrowing velocity interval of cyber-physical systems in buildings [J]. IEEE Access, 2017, 5: 20207- 20219. DOI: 10.1109/ACCESS.2017.2756855.

[16] DONG Long-jun, LI Xi-bing, MA Ju, TANG Li-zhong. Three-dimensional analytical comprehensive solutions for acoustic emission/microseismic sources of unknown velocity system [J]. Chin J Rock Mech Eng, 2017, 36: 186-197. DOI: 10.13722/j.cnki.jrme.2015.1765.

[17] DONG Long-jun, LI Xi-bing, TANG Li-zhong, GONG Feng-qiang. Mathematical functions and parameters for microseismic source location without pre-measuring speed [J]. Chinese Journal of Rock Mechanics and Engineering, 2011, 30: 2057-2067. DOI: 10.1016/j.cnsns.2011.01.018.

[18] FRANTTI G E, LEVEREAULT L A. Auditory discrimination of seismic signals from earthquakes and explosions [J]. Bulletin of the Seismological Society of America, 1965, 55: 1-25. DOI: 10.1016/0014-4835(75)900 80-9.

[19] BOOKER A, MITRONOVAS W. An application of statistical discrimination to classify seismic events [J]. Bulletin of the Seismological Society of America, 1964, 54: 961-971. DOI: 10.1596/1813-9450-4793.

[20] TAYLOR S R. Analysis of high-frequency Pg/Lg ratios from NTS explosions and western US earthquakes [J] Bulletin of the Seismological Society of America, 1996, 86: 1042-1053. http://library.sciencemadness.org/lanl2_a/lib-www/la-pubs/00326269.pdf.

[21] MALOVICHKO D. Discrimination of blasts in mine seismology [C]// Deep Mining. 2012: 161-171. DOI: 10.36487/ ACG_rep/1201_11_malovichko.

[22] JIANG Wen-wu, YANG Zuo-lin, XIE Jian-min, LI Jia-fu. Application of FFT spectrum analysis to identify microseismic signals [J]. Science & Technology Review, 2015, 33: 86-90. DOI: 10.3981/j.issn.1000-7857.2015.02. 013.

[23] DOWLA F U. Neural networks in seismic discrimination [J]. Nato Asi, 1995, 303: 777-789. DOI: 10.1007/978-94-011- 0419-7_41.

[24] BENBRAHIM M, DAOUDI A, BENJELLOUN K, IBENBRAHIM A. Discrimination of seismic signals using artificial neural networks [C]// The Second World Enformatika Conference, WEC’05. Istanbul, Turkey, 2005, http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.106.144.

[25] ZHAO Guo-yan, MA Ju, DONG Long-jun, LI Xi-bing, CHEN Guang-hui, ZHANG Chu-xuan. Classification of mine blasts and microseismic events using starting-up features in seismograms [J]. Transactions of Nonferrous Metals Society of China, 2015, 25: 3410-3420. DOI: 10.1016/S1003-6326(15)63976-0.

[26] MULLER S, GARDA P, MULLER J D, CANSI Y. Seismic events discrimination by neuro-fuzzy merging of signal and catalogue features [J]. Physics and Chemistry of the Earth, Part A: Solid Earth and Geodesy, 1999, 24: 201-206. DOI: 10.1016/S1464-1895(99)00019-8.

[27] ORLIC N, LONCARIC S. Earthquake—explosion discrimination using genetic algorithm-based boosting approach [J]. Computers & Geosciences, 2010, 36: 179-185. DOI: 10.1016/j.cageo.2009.05.006.

[28] DONG Long-jun, LI Xi-bing, XIE Gong-nan. Nonlinear methodologies for identifying seismic event and nuclear explosion using random forest, support vector machine, and naive Bayes classification [J]. Abstract and Applied Analysis, 2014: 459137. DOI: 10.1155/2014/459137.

[29] VALLEJOS J A, MCKINNON S D. Logistic regression and neural network classification of seismic records [J]. International Journal of Rock Mechanics and Mining Sciences, 2013, 62: 86-95. DOI: 10.1016/j.ijrmms.2013.04. 005.

[30] DONG Long-jun, WESSELOO J, POTVIN Y, LI Xi-bing. Discrimination of mine seismic events and blasts using the fisher classifier, naive bayesian classifier and logistic regression [J]. Rock Mech. Rock Eng, 49(1): 183-211. DOI: 10.1007/s00603-015-0733-y.

[31] SHANG Xue-yi, LI Xi-bing, MORALES-ESTEBAN A, CHEN Guang-hui. Improving microseismic event and quarry blast classification using artificial neural networks based on principal component analysis [J]. Soil Dynamics and Earthquake Engineering, 2017, 99: 142-149. DOI: 10.1016/ j.soildyn.2017.05.008.

[32] HUBEL D H, WIESEL T N. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex [J]. The Journal of physiology, 1962, 160: 106-154. DOI: 10.1113/jphysiol.1962.sp006837.

[33] FUKUSHIMA K. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position [J]. Biological Cybernetics, 1980, 36: 193-202. DOI: 10.1007/BF00344251.

[34] LI Zhong-mei, GUI Wei-hua, ZHU Jian-yong. Fault detection in flotation processes based on deep learning and support vector machine [J]. Journal of Central South University, 2019, 26(9): 2504-2515. DOI: 10.1007/s11771- 019-4190-8.

[35] TIVIVE F H C, BOUZERDOUM A. A new class of convolutional neural networks (SICoNNets) and their application of face detection. [C]// Proceedings of the International Joint Conference on Neural Networks, IEEE, 2003. https://ieeexplore.ieee.org/docum ent/1223742.

[36] SHAO Jie, QIAN Yong-sheng. Three convolutional neural network models for facial expression recognition in the wild [J]. Neurocomputing, 2019, 355: 82-92. DOI: 10.1016/ j.neucom.2019.05.005.

[37] KOLSCH A, AFZAL M Z, EBBECKE M, LIWICKI M. Real-time document image classification using deep CNN and extreme learning machines [EB/OL] 2017-11-03. https://arxiv.org/pdf/1711.05862v1.pdf. DOI: 10.1109 /ICDAR.2017.217.

[38] SHARAD J, SURAJ S, NITIN K. First steps toward CNN based source classification of document images shared over messaging app [J]. Signal Processing: Image Communication, 2019, 78: 32-41. DOI: 10.1016/j.image.2019.05.020.

[39] CHEN Tong, JIANG Ying, JIAN Wang, QIU Lan-xin, LIU Huan, XIAO Zhao-lin. Maintenance personnel detection and analysis using mask-rcnn optimization on power grid monitoring video [J]. Neural Processing Letters, 2020, 51(2): 1599-1610. DOI: 10.1007/s11063-019-10159-w.

[40] SONG Shi-jin, QUE Zhi-qiang, HOU Jun-jie, DU Sen, SONG Yue-feng. An efficient convolutional neural network for small traffic sign detection [J]. Journal of Systems Architecture, 2019. DOI: 10.1016/j.sysarc.2019.01.012.

[41] GOODFELLOW I, BENGIO Y, COURVILLE A. Deep learning(Vol.1) [M]. Cambridge: MIT Press, 2016: 326-366.

[42] ESTRACH J B, SZLAM A, LECUN Y. Signal recovery from pooling representations [C]// Proceedings of the International Conference on Machine Learning (ICML). 2014: 307-315. http://www.oalib.com/paper/4115463#.X2QRMHot2Uk.

[43] DONG Long-jun, WESSELOO J, POTVIN Y, LI Xi-bing. Discriminant models of blasts and seismic events in mine seismology [J]. Int J Rock Mech Min Sci, 2016, 86: 282-291. DOI: 10.1016/j.ijrmms.2016.04.021.

(Edited by HE Yun-bin)

中文导读

基于卷积神经网络与原始波形的微震与爆破事件辨识方法

摘要：微震监测系统作为深部矿山地压监测的有效手段之一，其作用原理是分析微震事件包含的力学参数，为岩体稳定性分析提供最准确的信息。准确地辨识微震事件与爆破事件决定了微震监测技术预警的时效性与准确性。鉴于微震事件和爆破事件的地震波形具有不同的特征，本文建立了一种基于卷积神经网络的微震事件和爆破事件辨识模型。首先将采集到的微震与爆破事件波形分别组成训练集(微震爆破事件各5250个)、测试集(微震爆破事件各1500个)以及验证集(无标签的微震爆破事件各750个)，将分类得到的数据集进行预处理并在CPU模式下输入至构建好的卷积神经网络模型中进行训练。结果显示训练集中微震事件识别的准确率为99.46%，爆破事件识别的准确率为99.33%，验证集中微震事件识别的准确率为100%，爆破事件识别的准确率为98.13%。与其他机器学习方法进行对比，该方法拥有较高的辨识准确率。逻辑回归模型与人工神经网络模型在相同测试集下的准确率仅为54.43%和67.90%。通过绘制并对比三种模型的ROC曲线，可以看出由于使用原始波形训练模型，CNN在辨识微震与爆破事件中表现出了绝对的优势。这不仅减少了个体经验差异的影响，而且消除了提取震源参数和波形参数过程中产生的误差。证明了本文所建立的微震与爆破事件辨识方法提高了微震数据处理的速度和精度。

关键词：微震监测；波形辨识；微震事件；爆破事件；卷积神经网络

Foundation item: Projects(51822407, 51774327, 51664016) supported by the National Natural Science Foundation of China

Received date: 2020-07-10; Accepted date: 2020-08-31

Corresponding author: XUE Jin-chun, PhD, Professor; Tel: +86-13803515297; E-mail: 1023817019@qq.com; ORCID: https://orcid.org/ 0000-0001-8519-4534