Elsevier

Computers & Geosciences

Volume 166, September 2022, 105164
Computers & Geosciences

Identifying microseismic events using a dual-channel CNN with wavelet packets decomposition coefficients

https://doi.org/10.1016/j.cageo.2022.105164Get rights and content

Highlights

  • The network combines time-domain information and wavelet packet decomposition coefficients of the microseismic data.

  • The more obvious feature map of the microseismic is extracted, which makes the classification performance better.

  • The wavelet packet decomposition coefficients can avoid the dimensional disaster of the input data.

Abstract

Microseismic monitoring is a widely used technique in coal mine safe production. Microseismic events classification is one of the most important steps in microseismic monitoring. CNN has been proven to be the potential tool to identify microseismic events automatically. When the noise is serious and the signal is weak, the recognition ability of CNN is limited. Some scholars use denoising methods, such as wavelet transform, to enhance the signal-to-noise ratio (SNR) of the input data. However, the effective signal will be broken if the incorrect threshold parameters are set in this wavelet transform-based denoising method. In this paper, we intend to make the data volume of the time-frequency graph obtained by WPT as the input of CNN. But it may lead to a dimensional disaster if taking the time-frequency domain signals as input directly. To utilize the effective information extraction and reduce the data dimension, we propose a dual-channel CNN model by combining time domain information and wavelet packet decomposition coefficients (T-WPD CNN). The wavelet packet decomposition coefficients highlight the characteristics of the signal and suppress the characteristics of noise. In addition, the coefficients have the same dimension as the original signal. These advantages are useful for enhancing the performance of CNN. Two field datasets are used to test the network. One from Australia and another from Jiayang coal mine, Sichuan, China. The feature map of the convolutional layer with different inputs are shown to illustrate the influence effect of raw time domain signal and WPD coefficient on the classification performance. The final results show that the T-WPD CNN is superior to the traditional CNN method in accuracy and robustness.

Introduction

Microseismic monitoring is an efficient technology for safety production in the coal mine (Ge, 2005; Ma et al., 2020).

Microseismic signal classification plays an essential role in data processing. In the early researches, experts carried out this tough job manually. However, the manual categorizing method has many distinct disadvantages on economy and time (Dong et al., 2016; Jiang et al., 2020). Researchers have proposed various automatic methods to conduct accurate identification of the microseismic event. The most commonly used classification algorithm is short- and long-time average ratio (STA/LTA) (Allen, 1978; Duan et al., 2015; Li et al., 2019). The classical one computes the energy ratio between the short- and long-time windows. Once the ratio exceeds the preset threshold, the identification is activated. Although the STA/LTA has advantages on computation and implementation, it is hard to set proper threshold and time windows length. In the situation of low signal-to-noise ratios (SNR), it cannot distinguish between noise and signal. There are also other algorithms used to classify microseismic events, such as template matching method (Li and Zhan, 2018), shearlet transform (Cheng et al., 2017). But these traditional methods are not suitable for large and streaming data volume calculations.

Generally, these traditional methods require manual parameter setting or additional feature extraction operations. For more automated methods, some scholars tend to use the artificial neural network (ANN) for microseismic events recognition since on 1990s (Dysart and Pulli, 1990; Fedorenko et al., 1999). It provides an intelligent method to process large amounts of data automatically. However, it did not well because of the lack of training data set and the low computational efficiency caused by the hardware performance limitations at that time. Along with the rapid development of technology and hardware, many different kinds of neural networks can be used for microseismic event classification and the first arrival picking, such as recurrent neural network (RNN) (Jan et al., 2021; Zheng et al., 2017), U-Net Network (Zhang and Sheng, 2020), long short-term memory (LSTM) (Zhang et al., 2018).

In recent years, many scholars used convolution neural network (CNN) to classify microseismic events and achieved excellent results. CNN can automatically extract the waveform characteristics without extra operating. It can learn the intrinsic characteristics of signals which overcome the limitations of the traditional method. Furthermore, CNN has made more achievements in studying first arrival picking in advance (Ross et al., 2018a, 2018b; Lin et al., 2019; Zheng et al., 2020). Nevertheless, the problem of low SNR from mine data has a significant influence on CNN, which leads to some denoising methods are used to remove noise before CNN.

Wavelet packet transform (WPT) is a common denoising method in seismic signal processing (Liu et al., 1998). It can highlight the characteristics of signal component and suppress characteristics of noise component, which is beneficial to signal recognition. However, the denoising method based on WPT is always compute-consuming, and the effective signal is broken because of the incorrect threshold parameters settings. At here, we intend to make the data volume of the time-frequency graph obtained by WPT as the input of CNN. But it may lead to a dimensional disaster if taking signals of the time-frequency domain as input directly. To utilize the information obtained by WPT and reduce the data dimension, we use wavelet packet decomposition (WPD) coefficients as CNN input. Coefficients represents the similarity between the wavelet basis function and the signal, which provide information for identifying microseismic events. As Daubechies functions are similar to seismic signals, the high precision db4 is used as the basis function of wavelet packet analysis (Daubechies, 1988; Yan et al., 2007). Nevertheless, the basis wavelet cannot fit completely with the original microseismic signals, which may cause errors. To solve this problem, we combine the time domain signal with the wavelet packet coefficients as CNN inputs. It improves the classification performance and enhances robustness.

In this paper, we propose a dual-channel CNN to classify microseismic events by jointing time domain information and wavelet packet decomposition coefficients (T-WPD CNN). The structure of the paper mainly consists of several parts: model and method introduction, data preparation, experimental results analysis. We applied the T-WPD CNN on two field datasets from different regions. At last, the computer code is made public.

Section snippets

Network architecture

The structure and parameters of T-WPD CNN are shown in Fig. 1. The T-WPD CNN accepts two inputs: time domain waveforms and WPD coefficients. The input layer is followed by two feature learning modules that each module consists of a convolutional layer, a dropout layer, and a max pooling layer. In feature extraction modules, the number of convolutional kernels is different between two convolutional layers. For the convolutional kernels, a relatively larger scale of receiving field can capture

T-WPD CNN learning approach

Data format needs to be reshaped before feeding into the network. We consider waveform as an image, and the T-WPD CNN receives two 4-dimensional inputs. The first dimension is the number of input data. The last dimension is the number of channels. The rest is the height and width of the image. For making full use of the training data set and rigorously evaluating the learning performance of CNN, stratified κ-fold cross validation is performed with 10 folds. Here with the same operating, steps

Data preparation

In this paper, we evaluate the performance of T-WPD CNN on two field datasets. The one is from an undisclosed mine in Australia. Another is collected in Jiayang coal mine in Sichuan, China. The following two datasets are introduced.

T-WPD performance

Now we use dual-channel CNN to classify signals by combining time domain data and WPD coefficient data. The events from dataset 1 are low SNR, which are the main targets of this paper. So we focus on this dataset for experiments and analysis. The size of training set is (880, 8, 1500, 1). After data augment, the number of training data increased to 1596. After trimming data and obtaining WPD coefficients, the training size of the time domain channel is (1596, 8, 1200, 1) and the wavelet domain

Conclusions

In this paper, we proposed a dual-channel CNN named T-WPD CNN by jointing time domain signal and WPD coefficients. The wavelet transform can enhance the difference of the signal and noise, but taking the time-frequency domain signals transformed by wavelet as the input of CNN will make a dimensional disaster. This method takes the WPD coefficients and the time domain signal as the input of CNN, which improves the characteristics of the microseismic event signals without increasing the dimension

Authorship contribution statement

Yaojun Wang: Conceptualization, Methodology, Investigation, Writing- Original draft preparation, Writing - Reviewing and Editing, Funding acquisition, Resources. Qian Qiu: Software, Data curation, Writing- Original draft preparation. Zhiqiang Lan: Software, Investigation, Writing - Reviewing and Editing. Keyu Chen: Investigation, Data curation. Jun Zhou: Data curation. Peng Gao: Data curation. Wei Zhang: Investigation.

Code availability section

T-WPD_CNN.ipynb

Contact: [email protected].

Program language: Python 3.

Software required: tensorflow>=2.2, keras, pandas, numpy.

Program size: 390 KB.

The source codes are available for downloading at the link: https://github.com/qiuqian96/T-WPD_CNN.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

Thank to Jiayang coal mine for the help in experiments. We would like to acknowledge financial support from the National Natural Science Foundation of China (Grant No. 42174151, 41804126).

References (24)

  • I. Daubechies

    Orthonormal bases of compactly supported wavelets

    Commun. Pure Appl. Math.

    (1988)
  • P.S. Dysart et al.

    Regional seismic event classification at the NORESS array: seismological measurements and the use of trained neural networks

    Bull. Seismol. Soc. Am.

    (1990)
  • Cited by (10)

    View all citing articles on Scopus
    View full text