A source positioning method based on six-component seismic data in complex environment

Abstract

The application discloses a source positioning method based on six-component seismic data in a complex environment. The method comprises the following steps: 1) obtaining training data and marking the positions of the training data to obtain training samples; constructing a direction filter based on a wave propagation model; 2) extracting a direction filter feature vector of the training samples by using the direction filter; then inputting the direction filter feature vector of the training samples and six-component seismic data in the training samples into a convolutional neural network to predict a position, and then calculating a loss value according to the predicted position and the marked position to optimize the convolutional neural network; 3) installing a six-component seismograph in a target area, collecting six-component seismic data of a source and inputting the six-component seismic data into a pre-training model obtained in step S4 to output a current position of the source; and 4) estimating a wave direction angle of a seismic wave according to the position of the source, and inverting a formation phase velocity profile. The application greatly improves the accuracy of the wave direction information in the formation structure inversion.

Description

Technical Field

[0001] This invention belongs to the field of seismic source location technology, specifically relating to a method for seismic source location using six-component seismic data in complex seismic wave propagation scenarios such as downhole drilling. It is particularly suitable for providing stable and reliable prior directional information for geophysical tasks such as stratigraphic structure inversion and medium parameter extraction. Background Technology

[0002] In various industrial scenarios, positioning technology is widely used for target monitoring and system control. Existing positioning methods mainly include those based on radio signal time of arrival (ToA), angle of arrival (AoA), and received signal strength (RSSI), which are widely used in ground navigation, indoor positioning, and moving object tracking. However, in environments with limited space, strong interference, or where multi-point equipment cannot be deployed, the positioning accuracy of these methods drops significantly, making it difficult to meet the requirements of high-reliability positioning.

[0003] In underground environments such as coal mines, accurately tracking the operating status of specific equipment (such as coal mining machines) is essential for promoting mining automation and building smart mines. It is crucial for achieving less-manned or unmanned mining operations, preventing major safety risks, and improving the inherent safety level of mines. The operation of coal mining machines continuously generates vibration signals, whose source characteristics are relatively stable and identifiable, making them a key information source for underground positioning and status monitoring. However, the underground environment limits the deployment space of multi-point sensor arrays, and wireless signal propagation is susceptible to obstruction and multipath interference, making traditional wireless positioning technologies unreliable in such scenarios.

[0004] To overcome limitations in deployment and synchronization, single-station source location technology has become an important development direction. In particular, the introduction of six-component seismographs allows for the acquisition of complete wavefield information (three translational and three rotational components) at a single point, providing a more comprehensive observational basis for estimating the direction of incoming waves. Currently, six-component data is used in areas such as stratigraphic structure inversion and medium parameter extraction, but these inversions also rely on the direction of incoming waves as a key input. For source location, accurately obtaining the direction of incoming waves is not only a core objective but also provides a priori constraints for further downhole structural imaging and condition identification.

[0005] However, most existing direction estimation methods are based on the assumption of an ideal wavefield, which makes them difficult to adapt to multi-wave interference and non-single-frequency excitation in complex environments, resulting in insufficient stability of direction estimation. Especially when applied to underground spaces, structures such as the overlying rock roof and floor, support pillars, and tunnel walls complicate the wave propagation process. Refraction and scattering phenomena caused by these structures constitute multipath effects and other complex situations, making it impossible for a single-station six-component seismograph to accurately estimate the incoming wave direction, thus limiting the effectiveness of inversion. Therefore, it is necessary to address the wave characteristics of specific sources such as underground coal mining machines, overcome the adverse effects of complex propagation environments, and develop stable and reliable single-station incoming wave direction estimation technology based on existing six-component equipment. This would provide stable and reliable prior direction information for tasks such as stratigraphic structure inversion and medium parameter extraction, fully releasing the application potential of six-component observations in underground intelligent sensing systems. Summary of the Invention

[0006] To address the limitations of single-station six-component seismic data extraction and stratigraphic structure inversion in complex propagation scenarios such as underground mines, where dispersion curves are extracted and stratigraphic structures are inverted using specific seismic sources like coal mining machines, the main challenge lies in the inaccurate estimation of the main incoming direction of seismic waves using single-station six-component seismographs. This leads to significant errors in the accuracy of the extracted dispersion curves and inverted underground velocity profiles. This invention provides a novel method for source location using single-station six-component data. Based on the use of a single-station six-component seismograph as a data acquisition instrument to collect environmental noise data, this new source location method solves the problem of insufficient accuracy in estimating the incoming direction of seismic waves in complex scenarios using traditional single-station six-component seismographs. It also addresses the error in estimating the incoming direction caused by multipath effects and other phenomena in complex propagation environments. By applying the constraint relationship features between the components of the seismic wavefield data to process the seismic wave time-domain sequence, this scheme can extract the relationships between the non-explicit components in the time-domain sequence and construct a mapping relationship between the time-domain sequence and the incoming direction. This allows for more accurate estimation of the back azimuth angle for various complex application scenarios.

[0007] At the hardware level, the wave direction estimation method of this invention still relies on a single-station six-component seismograph. When using a single-station six-component seismograph for stratigraphic inversion, the back azimuth angle of the incoming wave can be estimated simultaneously based on the acquired data. Compared to traditional wave direction estimation methods, this invention does not require additional signal receiving equipment or other sensors, thus reducing hardware deployment costs. Furthermore, unlike ultra-wideband (UWB) positioning technology and other solutions, this invention does not require additional signal generating devices. For example, in underground applications, before acquiring vibration signals generated by coal mining machinery using a single-station six-component seismograph and performing stratigraphic inversion, the back azimuth angle of the coal mining machinery vibration signal can be directly estimated using the acquired six-component data. Compared to traditional methods, this invention effectively solves the problem of insufficient accuracy in wave direction estimation faced by single-station six-component seismographs when performing stratigraphic inversion without increasing hardware complexity or deployment costs.

[0008] The technical solution adopted by this invention to solve its technical problem is:

[0009] This invention, relying on the hardware of a single-station six-component seismic structure inversion method, proposes a source location technique for complex seismic wave propagation scenarios such as downhole drilling. It aims to improve the accuracy of incoming wave direction information in seismic structure inversion, and is particularly suitable for source-controllable scenarios in downhole environments. Traditional single-station six-component location methods typically rely on waveform similarity between translational and rotational signals to estimate the azimuth angle after the incoming wave, such as... Figure 1 As shown. However, in underground coal mining operations, the wave field is complex and there are many interference signals. Traditional similarity calculation methods are sensitive to noise and waveform mismatch, leading to a decrease in the accuracy of incoming wave direction estimation, which in turn affects the accuracy of subsequent stratigraphic structure inversion. To overcome these shortcomings, this invention adds an incoming wave azimuth estimation module based on the constraint relationship characteristics between components to the original single-station six-component data acquisition. This module utilizes the coupling characteristics between translational and rotational components to construct the constraint relationship of the incoming wave azimuth, extract the correlation features between each component, and achieve source direction identification in a more robust way. Some parameter estimation steps can be completed offline in advance, the computational load in the inference stage is small, and it can be embedded into the existing six-component data processing flow to realize online processing and direction estimation of measured data, which is suitable for fields such as underground structural exploration and smart mine safety monitoring.

[0010] The flowchart for source location using six-component seismic data in complex seismic wave propagation scenarios such as downhole drilling proposed in this invention is as follows: Figure 2 As shown, assuming the instrument is correctly installed, in addition to adding a step for estimating the azimuth angle after the arrival wave in the data processing steps, a training data acquisition step also needs to be added before the formal acquisition of experimental data. This specifically includes the following steps:

[0011] S1: Install the six-component seismograph, align the instrument components with the actual direction according to the measurement requirements, and ensure that the seismograph is tightly coupled with the seismic wave propagation medium. Usually, after removing the surface soil, the instrument is tightly coupled with the ground. During the experiment, the ambient noise is used as the seismic source. When applied to scenarios such as underground mines, due to the limitations of the application environment, it is necessary to install it on a rock wall. In this case, the instrument is tightly coupled with the rock wall. During the experiment, the operating signal of the underground equipment is used as the seismic source. During the experiment, the seismic source needs to be continuously excited to ensure that the six-component seismograph can collect continuous and effective translational and rotational wave field data, providing a complete wave field observation basis for subsequent model training and positioning analysis.

[0012] S2: Before formally acquiring six-component seismic signals, training data acquisition is carried out first. Two types of data need to be acquired: one is the vibration signal generated during the operation of the equipment, which is acquired by a single-station six-component seismograph as training data; the other is the equipment location corresponding to the time period when the six-component seismograph acquires the equipment's working signal, which is recorded in the equipment's built-in work log as the label data of the training data. The pre-acquired data is used to optimize the model and improve the model's performance.

[0013] S3: Preprocess the collected six-component data, including removing typical outliers in the time domain, aligning data clocks, removing data trends, etc. Unify the sampling rate of the three-component acceleration data and the three-component rotation data respectively, use time-frequency analysis to identify the main frequency bands of the signal distribution, extract key signals and filter out irrelevant noise interference, and uniformly normalize the data after the above steps to construct a standardized dataset.

[0014] S4: Slice the continuously recorded seismic data into segments according to a set window, extract training samples, and generate corresponding label data for each training sample based on the device location record to ensure that each data slice has clear spatial location coordinates. Use the pre-divided training data and the matched corresponding label data as training samples, construct a directional filter based on the wave propagation model, and train a convolutional neural network model to realize the mapping relationship between the time domain data and the azimuth angle of the incoming wave. Set different training / validation ratios, use a part of the data to build the model and determine the model parameters, and use another part of the data to evaluate the model performance and verify the model generalization. The completed model is used for seismic wave azimuth angle prediction when subsequent formal data is collected.

[0015] S5: During the formal experiment, ensure that the six-component seismograph is correctly installed, collect the six-component seismic signals generated during the operation of the downhole equipment, preprocess the collected data by removing typical abnormal data in the time domain, aligning the data clock, removing data trends, etc., and resample the six-component data according to the given sampling rate and then normalize them uniformly to construct a standardized dataset. Input the dataset into the pre-trained model obtained in step S4 and output the corresponding back azimuth angle estimation parameters of the seismic wave.

[0016] S6: Based on the estimated azimuth angle of the incoming seismic wave corresponding to the six-component data, the two translational components in the horizontal direction of the six-component data are combined into a single horizontal translational component, and the two rotational components in the horizontal direction of the six-component data are combined into a single horizontal rotational component. According to the specific relationship between the six-component seismic data, the phase velocity dispersion curve at the corresponding measuring point is estimated. Then, the stratigraphic structure information at the measuring point is obtained by using the constraint relationship of the dispersion curve, and the stratigraphic phase velocity profile is inverted.

[0017] Preferably, when step S1 is applied to scenarios such as underground mining, the operating signal of underground equipment is used as the seismic source during the experiment. Underground equipment generally refers to main equipment such as tunnel boring machines and cutting machines. When there are multiple devices underground, the main equipment with high rated power, obvious signal characteristics, mobility, and the ability to provide short-term position information is preferentially selected as the seismic source. When multiple large devices are working at the same time, the signal characteristics of different devices can be identified by time-frequency analysis, polarization analysis, and other methods, and the signals of specific devices can be extracted separately for subsequent analysis.

[0018] Preferably, in step S2, to ensure the diversity of training data, it is usually necessary to collect six-component seismic signals when the equipment is working at different locations. For example, in the experimental conditions of underground coal mines, the training sample set should be the signals collected when the coal mining machine digs a complete round trip from one end to the other along the fully mechanized mining face.

[0019] Preferably, the data normalization process described in step S3 includes standardizing each channel data according to its mean and standard deviation in the training set, so that all channel data follow a normal distribution with a mean of 0 and a variance of 1, thereby improving the stability and convergence speed of network training.

[0020] Preferably, the wave propagation model construction of the constraint relationship of the wave azimuth angle in step S4 refers to modeling using the characteristics of the wave during propagation. The distance and azimuth angle of the wave during propagation are reflected in the received signal and are functions of the amplitude and phase of the received signal. Therefore, the directional features between the signal components can be extracted by constructing a directional filter and training a convolutional neural network model, thereby establishing the mapping relationship between the time-domain signal sequence and the wave azimuth angle.

[0021] Preferably, the tag data mentioned in step S4 is the device location corresponding to the time period of the six-component seismograph acquiring the device's working signal, recorded in the device's built-in work log. Typically, the device location recorded in the device's built-in work log is the distance between the device and the reference point. Combined with the downhole geometry, it can be further converted into target parameters, such as the back azimuth angle of the wave that is of interest in this invention. To reduce the error introduced by the conversion, the tag data used before the final output result is the device location information, i.e., the distance. After obtaining the estimated value of the device location in the actual measurement stage, it is further converted into the back azimuth angle by combining the sensor location and the trigonometric function relationship of the device's working surface.

[0022] Preferably, the convolutional neural network model described in step S4 preferentially adopts convolutional neural networks and their improved models to ensure the feature extraction capability of time series data. A multi-channel network model is used to process the seismic time series data, and a convolutional neural network model for estimating the azimuth angle after the arrival of seismic waves is constructed. The input of the convolutional neural network model is six-channel time series data plus multi-channel directional filtering feature vectors, where each channel corresponds to a component of time series data or a directional filtering feature vector. The output is the device location corresponding to the input multi-channel data time period. The root mean square error is used as the loss function, and the network hyperparameters are determined by the Bayesian optimization method to minimize the error function value. The model is trained until convergence to extract the non-explicit constraint features between the components of the seismic wave.

[0023] Preferably, the slicing according to the set window in step S4 refers to dividing the collected data into different time slices by applying a time window of preset width. The width of the time window is determined by comprehensively considering the main frequency bands of the signal distribution and the sampling rate of the label data obtained from the time-frequency analysis results in step S3. If the sampling rate of the label data is significantly lower than that of the seismic signal data, interpolation or label extension can be used to ensure that each data slice has a clear spatial location coordinate. The pre-divided training data and the matched corresponding label data are used as training samples and input into the convolutional neural network for training.

[0024] Preferably, the output of the corresponding back azimuth angle estimation parameter of the seismic wave in step S5 refers to the output of the estimated equipment location information, combined with the geometric relationship between the underground equipment, the fully mechanized mining face and the single-station six-component seismograph, and further converted into back azimuth angle using trigonometric functions.

[0025] The features of this invention are:

[0026] This invention provides a source location method based on six-component seismic data for complex seismic wave propagation scenarios such as underground mines. It combines joint observation data of triaxial translational acceleration and triaxial rotational angular velocity to comprehensively acquire wavefield information. In data processing, a unified sampling rate, bandpass filtering, and standardized normalization preprocessing are employed to effectively improve the signal-to-noise ratio and data comparability. This invention utilizes the frequency and time domain characteristics of underground equipment noise and the geometric features of the fully mechanized mining environment, combining directional filters and convolutional neural networks as the core regression model. It possesses deep feature extraction capabilities and good convergence performance, enabling the extraction of high-order features related to source location information from weak vibration signals in complex underground environments. The overall method requires no multiple stations or additional signal sources; it only requires a single-station six-component seismograph to perform vibration signals and hardware equipment involved in stratigraphic structure inversion to achieve high-precision seismic wave direction estimation. It features simple deployment, high-dimensional data, and strong model generalization ability.

[0027] Compared with the prior art, the positive effects of the present invention are as follows:

[0028] This invention proposes a source location method based on six-component seismic data for complex seismic wave propagation scenarios such as underground mines. Compared with traditional methods relying on source inversion, such as travel-time methods, three-component location methods, or multi-station methods based on array measurements, this invention only requires a single six-component seismograph to accurately estimate the incoming wave direction, greatly simplifying sensor deployment and reducing system costs and operation and maintenance difficulties. Compared with traditional single-point incoming wave direction estimation methods, this invention utilizes the frequency and time domain characteristics of underground equipment noise and the geometric features of the fully mechanized mining environment, combined with directional filters and convolutional neural network frameworks, to mine the nonlinear mapping relationship between complex wave patterns and target parameters in noisy environments, extract non-explicit constraint features between components, and significantly improve the stability and accuracy of the estimation results. It solves the problem of insufficient accuracy in seismic wave incoming wave direction estimation due to wavefront consistency and propagation path complexity in near-field location in underground applications, and provides necessary back azimuth information for single-station six-component seismographs to acquire six-component seismic data for stratigraphic structure inversion underground. Experimental results show that the positioning accuracy of the proposed method is better than that of the traditional single-station six-component wave localization method. It is especially suitable for geological exploration and positioning tasks in complex near-field environments such as coal mines and tunnels, and has good engineering application prospects. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the traditional single-station six-component seismic source location method.

[0030] Figure 2 This is a flowchart of source localization based on six-component seismic data in complex environments, as proposed in this invention.

[0031] Figure 3 and Figure 4 This is a diagram illustrating the specific implementation effect of the invention applied to measured data in a domestic coal mine. Detailed Implementation

[0032] The flowchart for estimating the direction of arrival of seismic waves using six-component seismic wave data in complex environments is as follows: Figure 2 As shown, assuming the instrument is correctly installed, this embodiment provides a specific method for source location based on a six-component single-station seismograph. It utilizes the natural vibration signals generated during the operation of downhole equipment, combined with a directional filter and a convolutional neural network model, to estimate the source direction and achieve formation phase velocity profile inversion. The specific steps include:

[0033] S1: Install the six-component seismograph. Align the instrument components with the actual direction according to the measurement requirements. Point the positive direction of the translational and rotational x-components to due north, the positive direction of the translational and rotational y-components to due east, and the positive direction of the translational and rotational z-components to the opposite direction to the gravitational acceleration. Ensure that the seismograph is tightly coupled to the seismic wave propagation medium. Usually, after removing the surface soil, the instrument is tightly coupled to the ground. During the experiment, use environmental noise as the seismic source. When applied in scenarios such as underground mines, due to the limitations of the application environment, it is necessary to install it on a rock wall. In this case, the instrument is tightly coupled to the rock wall. During the experiment, use the underground equipment operation signal as the seismic source. During the experiment, the seismic source needs to be continuously excited to ensure that the six-component seismograph can collect continuous and effective translational and rotational wavefield data, providing a complete wavefield observation basis for subsequent model training and location analysis.

[0034] S2: Before formally acquiring the six-component seismic signal, training data acquisition is required. This involves acquiring two types of data: one is the vibration signal generated during equipment operation, denoted as a. x (t), a y (t),a z (t),r x (t), r y (t),r z (t), where a(t) is the translational component of the seismic data, r(t) is the rotational component of the seismic data, and the subscript indicates the direction along which the variable is along the subscript (e.g., a). x (t) represents the x-direction translational component of the seismic data. The data is collected by a single-station six-component seismograph as training data. Alternatively, the device location l(t) corresponding to the time period of the six-component seismograph's working signal is recorded in the device's built-in work log and used as label data for training the convolutional neural network model.

[0035] S3: For the acquired seismic signals [a x (t), a y (t),az (t),r x (t), r y (t),r z [t] is preprocessed, including outlier removal, clock synchronization, and detrending in the time domain, and the three-component acceleration data and three-component rotation data are sampled at a unified rate f. s The main frequency bands of signal distribution are identified using time-frequency analysis. Taking a coal mining machine as an example, its main signal distribution is between 0.1Hz and 5Hz, therefore a unified sampling rate f can be used. s =50Hz, extract key signals and filter out irrelevant noise interference, and perform uniform normalization on the data after the above steps to construct a standardized dataset [a′ x (t), a′ y (t),a′ z (t),r′ x (t), r′ y (t),r ′ z (t)];

[0036] S4: Slice the continuously recorded seismic data into segments with a set window width T = 15s as training data. For each training sample [a ′ x,i (t), a′ y,i (t),a′ z,i (t),r′ x,i (t), r′ y,i (t),r ′ z,i (t)], Generate corresponding tag data l(t) based on device location records. i ), ensuring that each data slice has a clear spatial location coordinate, and using the pre-divided training data and the matched corresponding label data as training samples {[a ′ x,i (t), a′ y,i (t),a′ z,i (t),r′ x,i (t), r′ y,i (t),r ′ z,i (t)], i = 1, 2, 3, ...} is used as the input to the directional filter to obtain the directional filter feature vector P(θ). m After that, the training data, along with the time series data (six-component seismic data) in the training sample, is input into the convolutional neural network. The training / validation ratio is set from 0.9 to 0.1 in increments of 0.1 to optimize the neural network parameters. The trained convolutional neural network is then used for predicting the azimuth angle of incoming seismic waves during subsequent formal data acquisition.

[0037] S5: During the formal experiment, ensure the six-component seismograph is correctly installed and acquire the six-component seismic signals generated during the operation of the downhole equipment. The collected data was preprocessed using methods such as removing typical outliers in the time domain, aligning data clocks, and removing data trends. The preprocessed six-component data was then sampled at a given sampling rate f. s =After resampling at 50Hz, the data is uniformly normalized to construct a standardized dataset. The input direction filter and the trained convolutional neural network model output the corresponding estimated azimuth angle α of the incoming seismic wave. e (t);

[0038] S6: After obtaining the estimated azimuth angle α of the seismic wave arrival corresponding to the six-component data, e Based on (t), the two translational components in the horizontal direction are combined into a single horizontal translational component, and the two rotational components in the horizontal direction are combined into a single horizontal rotational component. The specific formula is as follows:

[0039] a T (t)=a N (t)sinα e (t)-a E (t)cosα e (t)

[0040] r T (t)=r N (t)sinα e (t)-r E (t)cosα e (t)

[0041] Based on the specific relationships between the six-component seismic data, the phase velocity dispersion curve at the corresponding measuring point is estimated using the following formula:

[0042]

[0043] Then, the constraint relationship c of the dispersion curve is utilized. L (f) and c R (f) Obtain the stratigraphic structure information at the measuring point and invert the stratigraphic phase velocity profile.

[0044] Preferably, the data normalization process described in step S3 includes standardizing each channel's data according to its mean and standard deviation in the training set, specifically using the following formula:

[0045]

[0046] Where x is the original signal data, μ is the mean of the channel in the training data, and σ is the standard deviation of the channel in the training data. Normalizing all channel data makes all channel data follow a normal distribution with a mean of 0 and a variance of 1, thereby improving the stability and convergence speed of network training.

[0047] Preferably, the tag data mentioned in step S4 is the device position l(t) corresponding to the time period of the six-component seismograph acquiring the device's working signal, recorded in the device's built-in work log. Typically, the device position recorded in the device's built-in work log is the distance between the device and the reference point. Combined with downhole geometry, this can be further converted into target parameters. The specific formula is as follows:

[0048]

[0049] Where d is the length of the right-angled side of the right triangle formed by the projection of the instrument installation point onto the working surface of the equipment, including the instrument installation point; α(t) is the azimuth angle of the incoming seismic wave; to reduce the error introduced by the conversion, the label data used in the model training stage is the equipment location information, i.e., the distance l(t), and the estimated value of the equipment location l is obtained in the prediction stage. e After (t), it is further converted into the rear azimuth angle α by combining trigonometric function relationships. e (t).

[0050] Preferably, the directional filter described in step S4 is based on the assumption that the signal consists of multiple sparse directions or wavenumber components. It iteratively estimates the covariance matrix and spectral density function to improve spatial resolution, serving as feature enhancement information for subsequent convolutional neural networks. This is based on the six-component data [a] ′ x,i (t), a′ y,i (t),a′ z,i (t),r′ x,i (t), r′ y,i (t),r ′ z,i (t)], The initialization formula for the direction vector is as follows:

[0051] α m (θ)=[k ax (θ m ),k ay (θ m ),k az (θ m ),k rx (θ m ),k ry (θ m ),k rz (θ m )]T

[0052] in n(θ m Let P be the unit vector representing the direction of wave propagation, and initialize the pseudo-power spectrum P. (i) =1, and construct the weighted covariance matrix using the obtained direction vectors. The specific formula is as follows:

[0053]

[0054] Based on this, the power of each spectral point in the pseudo-power spectrum is iteratively updated. and the weighted covariance matrix R (i) The specific formula is as follows:

[0055]

[0056] Where x(t) is the time-domain sequence of the corresponding component data, i.e., six-component seismic data; when the pseudo-power spectrum difference between two adjacent iterations is less than the preset tolerance or the iteration reaches the maximum number of iterations, the estimated value of the back azimuth angle θ is output. m Pseudopower spectrum P(θ) m As feature enhancement information, it is input along with the temporal data into the subsequent convolutional neural network for training; H (θ m ) refers to a(θ) m The conjugate transpose of ).

[0057] Preferably, in addition to the basic model, the convolutional neural network described in step S4 can also employ improved convolutional neural network structures such as residual networks, densely connected networks, and multi-scale convolutional networks. Furthermore, typical models suitable for time-domain feature extraction, such as long short-term memory networks and time series transformers, can also be used to construct a convolutional neural network model for estimating the azimuth angle after the arrival of seismic waves. Its basic input is six-channel time-series data, where each channel corresponds to a component of the time-series data; that is, the training sample data for the six channels corresponds to six component data [a] ′ x,i (t), a′ y,i (t),a′ z,i (t),r′ x,i (t), r′ y,i (t),r ′ z,i (t)], Simultaneously, directional filtering feature vectors are used as auxiliary information to enhance data features. The output is the device location l(t) corresponding to the six-channel time-series data period of the input. The network hyperparameters are determined by the Bayesian optimization method, and the root mean square error is used as the loss function. The specific formula is as follows:

[0058]

[0059] The predicted value is The actual label value is y i The total number of samples is N, and the model is trained until it converges.

[0060] Preferably, the slicing according to the set window width mentioned in step S4 refers to dividing the collected data into different time slices using a time window of preset width T. Based on the time-frequency analysis results in step S3, the main frequency bands of the signal distribution are obtained [f]. low ,f high ] and label data sampling rate f l The time window width T is determined by a comprehensive calculation, and the specific formula is as follows:

[0061]

[0062] If the label data sampling rate f l Significantly lower than the seismic signal data sampling rate f s By using interpolation or label extension, we can ensure that each data slice has a clear spatial coordinate. The pre-divided training data and the corresponding matched label data are then used as training samples and input into the convolutional neural network for training.

[0063] Preferably, in step S5, the output of the corresponding seismic wave arrival azimuth estimation parameter refers to, after outputting the estimated equipment location information, combining the geometric relationship between the underground equipment, the fully mechanized mining face, and the single-station six-component seismograph, further converting it into the rear azimuth using trigonometric functions. The specific formula is as follows:

[0064]

[0065] Where d is the vertical distance from the instrument installation point to the coal mining face, and l e (t) represents the device location estimate predicted by the convolutional neural network model based on the input data, where α is the device location estimate. e (t) is its final estimated back azimuth angle.

[0066] Figure 3 The performance metrics of the convolutional neural network model were plotted based on different training set partitioning ratios during an experiment conducted in a coal mine in China. Figure 4 The relationship between the accuracy of the back azimuth estimation and the proportion of the training set was plotted. Compared with traditional methods that are limited by the complex propagation environment downhole and cannot accurately estimate the back azimuth, the method described in this invention has significant advantages in accurately estimating the back azimuth.

[0067] The present invention has been described in detail above, but it is obvious that the specific implementation of the present invention is not limited thereto. For those skilled in the art, various obvious modifications made to the method without departing from the spirit and scope of the claims are within the protection scope of the present invention.

Claims

1. A source positioning method based on six-component seismic data in a complex environment, comprising the steps of: 1) obtaining training data, by continuously collecting six-component seismic data generated by a seismic source device in each working period using a six-component seismometer, recording the device position corresponding to each working period collected by the six-component seismometer using the working log of the seismic source device as the position label data corresponding to the six-component seismic data, and obtaining the training data; 2) slicing the continuously collected six-component seismic data according to a set window; generating the position label data corresponding to each slice based on the working log to obtain a training sample; and constructing a directional filter based on the wave propagation model; 3) extracting the directional filter feature vector of the training sample using the directional filter; then inputting the directional filter feature vector of the training sample and the six-component seismic data in the training sample into a convolutional neural network to predict the position corresponding to the six-component seismic data in the training sample, and then calculating the loss value according to the predicted position and the labeled position to optimize the convolutional neural network; 4) installing a six-component seismometer in a target area, collecting six-component seismic data of the seismic source, and inputting the pre-trained model obtained in step S4 to output the current position of the seismic source; 5) estimating the seismic wave arrival direction angle according to the position of the seismic source, and inverting the formation phase velocity profile.

2. The method of claim 1, wherein, The method for inverting the formation phase velocity profile is: combining two horizontal translational components in the six-component seismic data into one horizontal translational component, combining two horizontal rotational components in the six-component seismic data into one horizontal rotational component, estimating the phase velocity dispersion curve of the target area according to the relationship between the six-component seismic data, and then obtaining the formation structure information of the target area using the constraint relationship of the phase velocity dispersion curve to invert the formation phase velocity profile.

3. The method of claim 1, wherein, The method for extracting the directional filter feature vector of the training sample is: the directional filter initializes a directional vector α m (θ) according to six-component seismic data in the training sample, and initializes a pseudo power spectrum P (i) =1; then a weighted covariance matrix R m is constructed by using the directional vector α (i) (θ); then iteration is started to update the power of each spectral point of the pseudo power spectrum and the weighted covariance matrix R (i) ; when the difference between the pseudo power spectra of two adjacent iterations is less than a preset tolerance or the iteration reaches a maximum iteration number, the pseudo power spectrum P(θ m ) about the estimated value of the back azimuth θ m is output as the directional filter feature vector of the training sample.

4. The method of claim 3, wherein, Directional vector α m (θ) = [k ax (θ m ), k ay (θ m ), k az (θ m ), k rx (θ m ), k ry (θ m ), k rz (θ m )] T ; wherein, n j (θ m ) is a unit vector of the propagation direction of the jth component in the six-component seismic data.

5. The method of claim 4, wherein, where x(t) is the six-component seismic data, a H (θ m ) is the conjugate transpose matrix of a(θ m ).

6. The method according to claim 1 or 2 or 3, characterized in that, The main frequency band [f low ,f high ] of six-component seismic data is identified by using time-frequency analysis method, and the time window width T of the set window is determined according to the main frequency band [f low ,f high ] and the sampling rate f l of the six-component seismic data.

7. The method of claim 6, wherein, Window width 8. The method according to claim 1 or 2 or 3, characterized in that, The convolutional neural network is a residual network, a densely connected network, a multi-scale convolutional network, a long short-term memory network, or a time series transformer.

9. The method according to claim 1 or 2 or 3, characterized in that, The loss value is calculated using the root mean square error as the loss function; and the convolutional neural network is optimized according to the loss value by the Bayesian optimization method.

Images