Coal seam hydraulic fracturing micro-seismic signal identification method and system based on power spectrum density

Abstract

The invention discloses a coal seam hydraulic fracturing micro-seismic signal identification method and system based on power spectral density. The method comprises the following steps: preprocessing an acquired signal to obtain a micro-seismic signal; the micro-seismic signal is divided into a plurality of sub-bands, and each sub-band is used as an independent branch; extracting multi-scale time-frequency features in each independent branch by adopting continuous wavelet convolution; constructing a space-spectrum-time feature block, inputting the multi-scale time-frequency features into the feature block, and introducing a channel attention mechanism to perform adaptive weighting on the key information to obtain the multi-scale time-frequency features; introducing the power spectral density difference into a micro-seismic signal identification process, and fusing a power spectral density statistical feature and a multi-scale time-frequency feature to form a joint feature vector; and finally, analyzing the combined feature vector by using an intelligent identification model of a double-branch parallel hybrid architecture to realize automatic identification and classification of the micro-seismic event and the noise signal. According to the method, the accuracy and the stability of micro-seismic event identification are remarkably improved under the conditions of strong noise and non-stability.

Description

Technical Field

[0001] This invention belongs to the field of microseismic monitoring and signal processing technology, specifically relating to a method and system for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density. Background Technology

[0002] Hydraulic fracturing of coal seams is an important engineering technique for improving coal seam permeability and enhancing gas extraction efficiency. Microseismic monitoring technology, due to its ability to reflect the initiation and propagation of fractures in real time, has become one of the key methods for evaluating the effectiveness of hydraulic fracturing. Accurate identification of microseismic events is the foundation for microseismic monitoring data analysis and source location, and its identification quality directly affects the reliability of fracture network inversion results.

[0003] Existing microseismic signal identification methods mainly rely on time-domain features, time-frequency analysis, or manual empirical thresholds for discrimination. However, in coal mine underground or surface well fracturing monitoring, the monitoring environment is complex, with diverse background noise types, including mechanical vibration, electromagnetic interference, and environmental noise. Traditional methods have limited ability to characterize the essential spectral features of microseismic signals under strong noise conditions, and are prone to misjudgment or omission.

[0004] Power spectral density, as an important tool for describing the frequency domain distribution characteristics of signal energy, can reflect the stability and concentration of the overall spectral structure of a signal. Existing studies mostly use power spectral density for station background noise analysis or signal quality evaluation, while research on systematically incorporating power spectral density differences into the microseismic signal identification process is still relatively limited. There is a lack of an engineering-applicable microseismic signal identification method for the complex noise environment of coal seam hydraulic fracturing.

[0005] Therefore, a method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density is proposed to improve the accuracy and stability of microseismic event identification under complex noise backgrounds, which has important engineering application value. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method and system for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density, which can effectively solve the problems existing in the prior art.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density, comprising the following steps: Step 1: Preprocess the acquired continuous microseismic monitoring waveforms to obtain the processed microseismic signals; Step 2: Obtain the effective frequency range based on the processed microseismic signal, and divide it into multiple sub-bands. Each sub-band serves as an independent branch input to enhance frequency resolution. In each independent branch, continuous wavelet convolution is used to extract multi-scale time-frequency features. Step 3: Construct a spatial-spectral-temporal feature block. Input the multi-scale time-frequency features extracted in Step 2 into the feature block, extract the joint spatial and temporal features through convolution operation, and introduce a channel attention mechanism to adaptively weight the key information to obtain low-dimensional global features and high-dimensional local features respectively. Step 4: Calculate the power spectral density of the microseismic signal after processing in Step 1, then extract the statistical features of the power spectral density, and then fuse the statistical features of the power spectral density with the low-dimensional global features and high-dimensional local features obtained in Step 3 to form a joint feature vector. Step 5: Construct an intelligent recognition model with a dual-branch parallel hybrid architecture, and input the joint feature vector obtained in Step 4 into the recognition model. The model outputs the recognition results, realizing the automatic recognition and classification of microseismic events and noise signals.

[0008] Furthermore, the preprocessing in step one includes sequentially performing detrending processing, bandpass filtering, and amplitude normalization to eliminate background drift and low-frequency interference while retaining the main energy distribution frequency bands.

[0009] Furthermore, the continuous wavelet convolution in step two is implemented using the Morlet wavelet function.

[0010] Furthermore, the spatial-spectral-temporal feature block in step three consists of two convolutional layers and a channel attention mechanism (SEBlock), specifically: I. First convolutional layer: A two-dimensional convolutional kernel is used to operate on the time-frequency image, thereby simultaneously extracting spatial and temporal features: in, The input time-frequency feature tensor is obtained by continuous wavelet convolution and contains information in the time dimension, frequency dimension, and channel dimension. This represents the weight parameters of the first-layer two-dimensional convolution kernel; Represents a two-dimensional convolution operation; Indicates the bias parameters of the convolutional layer; (⋅) represents a nonlinear activation function; This represents the output feature map of the first convolutional layer; the convolutional kernel weights ,in: 10 represents the size of the convolution kernel in the frequency dimension; 10 represents the size of the convolution kernel in the time dimension; this makes the convolution kernel match the number of channels in the input feature map in the input channel dimension, with 32 output channels.

[0011] II. Channel Attention Mechanism: Channel weights are obtained through global average pooling. The importance coefficient is obtained through nonlinear mapping, and the calculation method is as follows: in For global average pooling, For ReLU, It is Sigmoid; This represents the weight parameters of the second-layer two-dimensional convolutional kernel; the final enhanced feature is: III. Second Convolutional Layer: Further extracting local temporal features based on the enhanced channel features: The output channel count is 64; then, by using a channel attention mechanism to enhance key information, we obtain: in, It is a low-dimensional global feature (global branch input). It is a high-dimensional local feature (local branch input).

[0012] Furthermore, the power spectral density statistical characteristics in step four include power spectral Shannon entropy, dominant frequency, average frequency, energy, and duration.

[0013] Furthermore, in step four, the power spectral density is calculated using the Welch method or a multi-window spectral estimation method.

[0014] Furthermore, the dual-branch parallel hybrid architecture of the intelligent recognition model in step five includes a branch for deep time-frequency feature learning and a branch for power spectral density statistical feature analysis.

[0015] The system formed by the above identification method includes: The data acquisition module is used to acquire continuous microseismic monitoring waveform data generated during the hydraulic fracturing of coal seams; The signal preprocessing module is used to preprocess the continuous microseismic monitoring waveform data to eliminate the effects of background drift and environmental noise. The frequency division and time-frequency feature extraction module is used to perform multi-branch frequency division on the preprocessed microseismic signal and extract multi-scale time-frequency features through continuous wavelet convolution. The spatial-spectral-temporal feature extraction module is used to perform joint spatial, spectral and temporal modeling of multi-scale time-frequency features, and enhances discriminative feature information through an attention mechanism to obtain low-dimensional global features and high-dimensional local features. The power spectral density feature extraction module is used to calculate the power spectral density of the signal and extract the statistical features of the power spectral density. The feature fusion module is used to fuse low-dimensional global features, high-dimensional local features, and power spectral density statistical features to construct a joint feature vector; The signal recognition and classification module takes the joint feature vector as input and uses a dual-branch parallel hybrid architecture to automatically recognize and classify microseismic events and noise signals.

[0016] Compared with the prior art, the present invention has the following advantages: 1. This invention constructs an innovative spatial-spectral-temporal feature block, extracts joint spatial and temporal features through convolution operations, and introduces a channel attention mechanism to adaptively weight key information, which is used to jointly extract multi-scale time-frequency depth features of microseismic signals in terms of spatial distribution, spectral structure and temporal changes, providing the required feature data for subsequent fusion.

[0017] 2. This invention introduces the power spectral density difference into the microseismic signal identification process and extracts power spectral density statistical features to reflect the essential differences between microseismic events and noise signals in frequency domain energy distribution and temporal characteristics. Then, the power spectral density statistical features are fused with multi-scale time-frequency depth features to form a joint feature vector. Since this vector contains two different features of microseismic events and noise signals, it provides accurate data support for subsequent accurate differentiation between microseismic events and noise signals, and significantly improves the accuracy and stability of microseismic event identification under strong noise and non-stationary conditions.

[0018] 3. Through the above-mentioned processing, the present invention has good noise resistance and is suitable for complex noise environments in coal seam hydraulic fracturing, and has wide applicability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method in this invention.

[0020] Figure 2 This is a schematic diagram of the system in this invention. Detailed Implementation

[0021] The present invention will be further described below.

[0022] like Figure 1 As shown, a method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density includes the following steps: Step 1: Since the original signal generally contains trend drift, low-frequency interference and environmental noise, the acquired continuous microseismic monitoring waveform is preprocessed to obtain the processed microseismic signal. The preprocessing includes detrending processing, bandpass filtering and amplitude normalization in sequence to eliminate background drift and low-frequency interference and retain the main energy distribution frequency band.

[0023] Let the preprocessed input waveform be: Sampling rate The preprocessed signal is then represented as: in For bandpass filtering operators, This represents the trend.

[0024] Before entering the network, the input data is reshaped as follows: To adapt to the subsequent wavelet convolution processing method.

[0025] Step 2: Based on the processed microseismic signal, the effective frequency range is obtained and divided into five sub-frequency bands, each serving as an independent branch input. In each independent branch, the Morlet wavelet function is used for continuous wavelet convolution to extract multi-scale time-frequency features, specifically: The effective frequency range is divided into five sub-bands: Each subband is used independently as a branch input to enhance frequency resolution; subsequently, continuous wavelet convolution (CWConv) is used to extract time-frequency features; the Morlet wavelet function is: The convolution operation is as follows: in The center frequency is Learnable wavelet kernel. The output shape of CWConv is: By applying continuous wavelet convolution to the signals in each of the above branches to perform time-frequency mapping, multi-scale time-frequency features are generated to characterize the evolution of microseismic signals at different time and frequency scales.

[0026] Step 3: Construct a spatial-spectral-temporal feature block. Input the multi-scale time-frequency features extracted in Step 2 into the feature block, extract joint spatial and temporal features through convolution operations, and introduce a channel attention mechanism to adaptively weight key information. This is used to jointly extract key information about the spatial distribution, spectral structure, and temporal variations of the microseismic signal, obtaining low-dimensional global features and high-dimensional local features respectively. Specifically, the spatial-spectral-temporal feature block consists of two layers of convolution and a channel attention mechanism (SEBlock). I. First convolutional layer: A two-dimensional convolutional kernel is used to operate on the time-frequency image, thereby simultaneously extracting spatial and temporal features: in, The input time-frequency feature tensor is obtained by continuous wavelet convolution and contains information in the time dimension, frequency dimension, and channel dimension. This represents the weight parameters of the first-layer two-dimensional convolution kernel; Represents a two-dimensional convolution operation; Indicates the bias parameters of the convolutional layer; (⋅) represents a nonlinear activation function; This represents the output feature map of the first convolutional layer; the convolutional kernel weights ,in: 10 represents the size of the convolution kernel in the frequency dimension; 10 represents the size of the convolution kernel in the time dimension; this makes the convolution kernel match the number of channels in the input feature map in the input channel dimension, with 32 output channels.

[0027] II. Channel Attention Mechanism: Channel weights are obtained through global average pooling. The importance coefficient is obtained through nonlinear mapping, and the calculation method is as follows: in For global average pooling, For ReLU, It is Sigmoid; This represents the weight parameters of the second-layer two-dimensional convolutional kernel; the final enhanced feature is: III. Second Convolutional Layer: Further extracting local temporal features based on the enhanced channel features: The output channel count is 64; then, by using a channel attention mechanism to enhance key information, we obtain: in, It is a low-dimensional global feature (global branch input). These are high-dimensional local features (local branch inputs); used for subsequent feature fusion.

[0028] Step 4: Calculate the power spectral density of the microseismic signal processed in Step 1 using the Welch method or multi-window spectral estimation method. Then, extract the statistical features of the power spectral density, including the Shannon entropy, dominant frequency, average frequency, energy, and duration, to reflect the essential differences between microseismic events and noise signals in frequency domain energy distribution and temporal characteristics. The specific calculation formula is as follows: Power spectrum Shannon entropy (SpEn): Among them, frequency characteristics: main frequency average frequency It reflects the center of energy distribution; Energy characteristics: Time-domain energy.

[0029] Duration characteristic: Event duration .

[0030] Then, the power spectral density statistical features are fused with the low-dimensional global features and high-dimensional local features obtained in step three to form a joint feature vector, specifically: The joint feature vector serves as the final input to the intelligent recognition model.

[0031] Step 5: Construct a dual-branch parallel hybrid architecture intelligent recognition model. In this embodiment, it includes two branches: a first branch for deep learning of the multi-scale time-frequency features of the microseismic signal; and a second branch for feature mapping and representation of the power spectral density statistical features. A feature fusion unit is used to fuse the features output from the two branches and complete the classification. A classification unit is used to output the recognition result based on the input joint feature vector, realizing the automatic recognition and classification of microseismic events and noise signals. The first branch takes a time-frequency feature map as input and includes: at least one two-dimensional convolutional layer for extracting local time-frequency features; at least one nonlinear activation layer; an optional normalization layer or pooling layer for enhancing feature stability; and a feature compression unit for generating a fixed-dimensional deep feature vector. This branch is used to extract a joint high-dimensional representation of the microseismic signal in the time and frequency dimensions, realizing deep time-frequency feature learning.

[0032] The second branch takes the power spectral density statistical feature vector as input and includes: a feature mapping unit for performing linear or nonlinear transformations on the statistical features; and a feature enhancement unit for improving the discriminative ability of the statistical features. This branch is used to characterize the differences in the frequency domain energy distribution and statistical structure of microseismic signals, thereby realizing power spectral density statistical feature analysis.

[0033] The outputs of the two branches are fused through a feature fusion and classification unit. The fusion methods include, but are not limited to, vector concatenation, weighted fusion, and feature mapping fusion. The resulting joint feature vector is then passed through a classification unit to output the identification results of microseismic events and noise signals.

[0034] The system formed by the above identification method, such as Figure 2 As shown, it includes: The data acquisition module is used to acquire continuous microseismic monitoring waveform data generated during the hydraulic fracturing of coal seams; The signal preprocessing module is used to preprocess the continuous microseismic monitoring waveform data to eliminate the effects of background drift and environmental noise. The frequency division and time-frequency feature extraction module is used to perform multi-branch frequency division on the preprocessed microseismic signal and extract multi-scale time-frequency features through continuous wavelet convolution. The spatial-spectral-temporal feature extraction module is used to perform joint spatial, spectral and temporal modeling of multi-scale time-frequency features, and enhances discriminative feature information through an attention mechanism to obtain low-dimensional global features and high-dimensional local features. The power spectral density feature extraction module is used to calculate the power spectral density of the signal and extract the statistical features of the power spectral density. The feature fusion module is used to fuse low-dimensional global features, high-dimensional local features, and power spectral density statistical features to construct a joint feature vector; The signal recognition and classification module takes the joint feature vector as input and uses a dual-branch parallel hybrid architecture to automatically recognize and classify microseismic events and noise signals.

[0035] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density, characterized in that, Includes the following steps: Step 1: Preprocess the acquired continuous microseismic monitoring waveforms to obtain the processed microseismic signals; Step 2: Obtain the effective frequency range based on the processed microseismic signal, and divide this range into multiple sub-frequency bands, with each sub-frequency band serving as an independent branch input; in each independent branch, use continuous wavelet convolution to extract multi-scale time-frequency features; Step 3: Construct a spatial-spectral-temporal feature block. Input the multi-scale time-frequency features extracted in Step 2 into the feature block, extract the joint spatial and temporal features through convolution operation, and introduce a channel attention mechanism to adaptively weight the key information to obtain low-dimensional global features and high-dimensional local features respectively. Step 4: Calculate the power spectral density of the microseismic signal after processing in Step 1, then extract the statistical features of the power spectral density, and then fuse the statistical features of the power spectral density with the low-dimensional global features and high-dimensional local features obtained in Step 3 to form a joint feature vector. Step 5: Construct an intelligent recognition model with a dual-branch parallel hybrid architecture, and input the joint feature vector obtained in Step 4 into the recognition model. The model outputs the recognition results, realizing the automatic recognition and classification of microseismic events and noise signals.

2. The method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density according to claim 1, characterized in that, The preprocessing in step one includes sequentially performing detrending processing, bandpass filtering, and amplitude normalization to eliminate background drift and low-frequency interference while retaining the main energy distribution frequency bands.

3. The method for identifying microseismic signals of hydraulic fracturing in coal seams based on power spectral density according to claim 1, characterized in that, In step two, the continuous wavelet convolution is implemented using the Morlet wavelet function.

4. The method for identifying microseismic signals of coal seam hydraulic fracturing based on power spectral density according to claim 1, characterized in that, The spatial-spectral-temporal feature block in step three consists of two convolutional layers and a channel attention mechanism, specifically: I. First convolutional layer: A two-dimensional convolutional kernel is used to operate on the time-frequency image, thereby simultaneously extracting spatial and temporal features: in, Represents the input time-frequency feature tensor; This represents the weight parameters of the first-layer two-dimensional convolution kernel; ,in: 10 represents the size of the convolution kernel in the frequency dimension; 10 represents the size of the convolution kernel in the time dimension. Represents a two-dimensional convolution operation; Indicates the bias parameters of the convolutional layer; (⋅) represents a nonlinear activation function; This represents the feature map output by the first convolutional layer; II. Channel Attention Mechanism: Channel weights are obtained through global average pooling. The importance coefficient is obtained through nonlinear mapping, and the calculation method is as follows: in For global average pooling, For ReLU, It is Sigmoid; This represents the weight parameters of the second-layer two-dimensional convolutional kernel; the final enhanced feature is: III. Second Convolutional Layer: Further extracting local temporal features based on the enhanced channel features: Then, by using a channel attention mechanism to enhance key information, we obtain: in, As a low-dimensional global feature, It represents a high-dimensional local feature.

5. The method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density according to claim 1, characterized in that, The power spectral density statistical characteristics in step four include power spectral Shannon entropy, dominant frequency, average frequency, energy, and duration.

6. The method for identifying microseismic signals in coal seam hydraulic fracturing based on power spectral density according to claim 1, characterized in that, In step four, the power spectral density is calculated using the Welch method or the multi-window spectral estimation method.

7. The method for identifying microseismic signals of coal seam hydraulic fracturing based on power spectral density according to claim 1, characterized in that, The dual-branch parallel hybrid architecture of the intelligent recognition model in step five includes a branch for deep time-frequency feature learning and a branch for power spectral density statistical feature analysis.

8. A system utilizing the identification method according to any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire continuous microseismic monitoring waveform data generated during the hydraulic fracturing of coal seams; The signal preprocessing module is used to preprocess the continuous microseismic monitoring waveform data to eliminate the effects of background drift and environmental noise. The frequency division and time-frequency feature extraction module is used to perform multi-branch frequency division on the preprocessed microseismic signal and extract multi-scale time-frequency features through continuous wavelet convolution. The spatial-spectral-temporal feature extraction module is used to perform joint spatial, spectral and temporal modeling of multi-scale time-frequency features, and enhances discriminative feature information through an attention mechanism to obtain low-dimensional global features and high-dimensional local features. The power spectral density feature extraction module is used to calculate the power spectral density of the signal and extract the statistical features of the power spectral density. The feature fusion module is used to fuse low-dimensional global features, high-dimensional local features, and power spectral density statistical features to construct a joint feature vector; The signal recognition and classification module takes the joint feature vector as input and uses a dual-branch parallel hybrid architecture to automatically recognize and classify microseismic events and noise signals.

Images