An unsupervised microseismic denoising method and system based on multi-channel Kalman filter network

Abstract

The present invention discloses an unsupervised microseismic denoising method and system based on a multi-channel Kalman filter network, which belongs to the intersection of earth science technology and artificial intelligence, including: obtaining multi-channel microseismic data and preprocessing; constructing a multi-channel Kalman seismic denoising network based on a recurrent neural network; constructing a loss function, and unsupervised training the multi-channel Kalman seismic denoising network based on the preprocessed multi-channel microseismic data and the loss function; denoising the real-time microseismic data through the trained multi-channel Kalman seismic denoising network. The present invention establishes a multi-channel Kalman filter state prediction and update equation, designs an RNN operation operator, realizes automatic parameter optimization, and improves the accuracy and efficiency of denoising. The method makes full use of the correlation of microseismic data, improves the filtering effect, is suitable for time-varying signals, and provides support for subsequent data processing.

Description

Technical Field

[0001] The present invention belongs to the intersection of earth science technology and artificial intelligence, and in particular relates to an unsupervised microseismic denoising method and system based on a multi-channel Kalman filter network. Background Art

[0002] Microseismic monitoring technology is a geophysical technology that identifies the properties and characteristics of underground reservoirs by monitoring microseismic events caused by small deformations of rocks or pressure release due to fracturing of unconventional reservoirs. It plays an important role in oil and gas exploration. Microseismic events last for a short time, have a high frequency of sound waves, and the energy of the source is relatively weak. In addition, because the signal is absorbed by the formation during propagation, it is seriously interfered by environmental noise and even the microseismic events are buried in the noise. Therefore, the collected microseismic data usually have the characteristics of few microseismic events, low signal-to-noise ratio, and weak effective signal energy. In order to improve the accuracy of subsequent picking and the reliability of inversion results, it is necessary to suppress the random noise in the microseismic data and improve the signal-to-noise ratio of the microseismic data.

[0003] According to the mechanism of effective signal and noise separation method, microseismic signal denoising methods can be divided into model-driven methods and data-driven methods. Model-driven methods include static model denoising and dynamic model denoising; the most representative data-driven method is deep learning denoising. The principle of static model denoising is to assume that the data conforms to a fixed distribution or simple statistical law to deal with noise. The most commonly used denoising methods in this type of method are conventional filtering, transform domain and other denoising methods. Static model denoising methods rely on the statistical characteristics of finite signal distribution and cannot adapt to signals that change dynamically over time. The denoising principle of dynamic models is to estimate the system model and measurement data according to the characteristics of signal changes over time to achieve the purpose of denoising. The most typical method of this type is Kalman filtering. The basic principle is to analyze the characteristics of microseismic signals, construct a state space model based on prior knowledge, use optimal estimation theory to model the relationship between effective signals and noise and other multivariables, and adaptively adjust the system state estimation and error covariance matrix to predict effective signals. However, the traditional Kalman filter ignores the correlation of microseismic multi-channel data in the denoising process. More importantly, the initialization of model parameters depends on prior knowledge, which limits the application effect of denoising methods.

[0004] In recent years, deep learning has performed well in seismic data denoising and inversion, and scholars at home and abroad have gradually applied deep learning to the study of microseismic data denoising. Domestic and foreign scholars have applied U-Net neural network to the denoising of microseismic data. The denoising performance of such supervised network methods depends largely on a large number of effective samples. However, the low signal-to-noise ratio of microseismic data and the high acquisition and processing costs lead to a small amount of microseismic data, making it difficult to obtain enough effective samples to train the network. Therefore, unsupervised denoising network models have received attention. Scholars have applied unsupervised network models such as variational autoencoders to the denoising of microseismic data. However, the current unsupervised denoising network is based on the pure data-driven principle, which increases the difficulty of microseismic data network learning, slows down the convergence of the model, and is prone to falling into local optimal solutions.

[0005] In view of the above analysis, the challenges faced by current microseismic denoising methods include:

[0006] (1) Model-driven methods usually rely on the rich experience of domain experts to select appropriate parameters. However, this reliance on prior knowledge not only requires a lot of professional background support, but also requires repeated parameter adjustments in different situations, resulting in inefficient algorithms and limited application scope.

[0007] (2) Although supervised learning methods can automatically learn features from data and reduce human intervention, their denoising performance depends largely on a large number of valid samples, which limits their generalization ability.

[0008] (3) Although unsupervised methods can alleviate the problem of data dependence to a certain extent, they are based on pure data-driven principles and lack effective guidance, which limits the denoising performance of the model. Summary of the invention

[0009] In order to solve the above technical problems, the present invention proposes an unsupervised microseismic denoising method and system based on a multi-channel Kalman filter network to solve the problems existing in the above prior art.

[0010] To achieve the above object, the present invention provides an unsupervised microseismic denoising method based on a multi-channel Kalman filter network, comprising:

[0011] Acquire multi-channel microseismic data and perform preprocessing;

[0012] Construct a multi-channel Kalman seismic denoising network based on recurrent neural network;

[0013] Constructing a loss function, and performing unsupervised training on a multi-channel Kalman seismic denoising network based on the preprocessed multi-channel microseismic data and the loss function;

[0014] Real-time microseismic data are denoised using a trained multi-channel Kalman seismic denoising network.

[0015] Optionally, the multi-channel microseismic data includes theoretical synthetic data and actual data; the process of obtaining the multi-channel microseismic data includes: selecting a monitoring well, arranging a number of survey lines around the monitoring well, arranging detectors equidistantly on the survey lines, and obtaining actual microseismic data; generating theoretical synthetic data with added Gaussian random noise through forward modeling.

[0016] Optionally, the process of constructing a multi-channel Kalman seismic denoising network includes:

[0017] Design multiple Kalman operation operators and embed them into the basic architecture; among them, select the recurrent neural network as the basic architecture.

[0018] Optionally, the multi-channel Kalman operation operator includes a Kalman filter state prediction equation, a Kalman filter update equation, and a diagonal mapping.

[0019] Optionally, the Kalman filter state prediction equation is:

[0020] ;

[0021] in, is the prediction matrix , is the state transfer matrix, is the process noise covariance matrix, With the identity matrix Row splicing, denoted as , output Including the current moment prior estimate state vector , state covariance matrix .

[0022] Optionally, the Kalman filter update equation is:

[0023] ;

[0024] in, To update the matrix , is the Kalman gain, is the observation matrix, is the Kalman filter state prediction equation, is the seismic trace observation data at the corresponding time, Estimated from the optimal state and the error covariance matrix Row splicing composition, = , In hidden state.

[0025] Optionally, the process of unsupervised training of a multi-channel Kalman seismic denoising network includes;

[0026] The trainable parameters of the multi-channel Kalman seismic denoising network are set as square matrices and initialized as random numbers. The theoretical synthetic data and the actual microvibration data are divided into blocks through fixed-size window sliding, and then input into the multi-channel Kalman seismic denoising network. Back propagation operation is performed based on the loss function to obtain the error value. The parameters of the multi-channel Kalman seismic denoising network are optimized based on the error value and the Adam optimization algorithm until the convergence condition is met and the training is terminated.

[0027] Optionally, the process of denoising the real-time microseismic data includes:

[0028] The real-time microseismic data is input into the trained multi-channel Kalman seismic denoising network, and the initial hidden state is obtained through the Kalman filter state prediction equation; the final hidden state of the real-time microseismic data at the corresponding moment is obtained based on the real-time microseismic data, the initial hidden state and the Kalman filter update equation, and the final hidden state includes an optimal state estimate and an error covariance matrix; the optimal state estimate is mapped by diagonal mapping to obtain a denoising result.

[0029] The present invention also provides an unsupervised microseismic denoising system based on a multi-channel Kalman filter network, comprising:

[0030] A data acquisition module is used to acquire multi-channel microseismic data and perform preprocessing;

[0031] Model building module, used to build a multi-channel Kalman seismic denoising network based on recurrent neural network;

[0032] A model training module, used for constructing a loss function, and performing unsupervised training on a multi-channel Kalman seismic denoising network based on the preprocessed multi-channel microseismic data and the loss function;

[0033] The denoising module is used to denoise the real-time microseismic data through the trained multi-channel Kalman seismic denoising network.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] The present invention provides an unsupervised microseismic denoising method and system of a multi-channel Kalman filter neural network, which combines Kalman filtering with a recurrent neural network. Firstly, a Kalman filter state prediction and update equation for multi-channel microseismic data is established, and multi-channel correlation is fully utilized to improve the characterization ability of Kalman filter parameters; secondly, a RNN operation operator for multi-channel Kalman filter state prediction and update is designed, and the parameters of the Kalman filter are optimized by automatically obtaining a chain gradient, and a multi-channel Kalman network denoising based on a recurrent neural network mode is constructed; thirdly, an unsupervised microseismic denoising training method is combined to realize automatic optimization of Kalman parameters and avoid excessive reliance on valid data labels; finally, experimental results of theoretical forward modeling and actual microseismic data show that the present invention is superior to similar methods such as traditional Kalman filtering and autoencoders in terms of microseismic denoising accuracy and efficiency. Compared with the prior art, the technical solution of the present invention combines Kalman filtering with recurrent neural networks, and proposes multi-channel Kalman filtering to achieve synchronous processing of multi-channel microseismic data, thereby improving computational efficiency; fully utilizing the intra-channel and inter-channel correlations of microseismic data to improve the characterization capability and filtering effect of Kalman filtering; and having a good effect on denoising time-varying microseismic signals, thereby providing reliable data support for further processing and analysis of subsequent microseismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of a multi-channel Kalman filter according to an embodiment of the present invention;

[0039] Figure 3 A multi-channel Kalman operation operator model of an embodiment of the present invention;

[0040] Figure 4 A flow chart of forward propagation and reverse propagation of a multi-channel Kalman network according to an embodiment of the present invention;

[0041] Figure 5 A multi-channel Kalman network model diagram of an embodiment of the present invention;

[0042] Figure 6 Schematic diagram comparing the denoising effects of different algorithms of the embodiments of the present invention on synthetic microseismic data, wherein (a) is microseismic synthetic data, (b) is microseismic synthetic noisy data, (c) is the model effect of the present invention, (d) is the traditional Kalman model effect, (e) is the wavelet transform model effect, and (f) is the VAE model effect;

[0043] Figure 7 This is a curve diagram of SNR value changes for 19 verifications of an embodiment of the present invention;

[0044] Figure 8 Schematic diagram of the comparison of denoising effects of different algorithms on actual microseismic data according to an embodiment of the present invention, wherein (a) is the original microseismic data, (b) is the ideal microseismic data after being processed by traditional denoising technology, (c) is the model effect of the present invention, (d) is the traditional Kalman model effect, (e) is the wavelet transform model effect, and (f) is the VAE model effect. DETAILED DESCRIPTION

[0045] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0047] Embodiment 1

[0048] like Figure 1 As shown, this embodiment provides an unsupervised microseismic denoising method based on a multi-channel Kalman filter network, comprising:

[0049] Step 1: Dataset preparation:

[0050] The theoretical synthetic data uses multi-channel microseismic data generated by 25 Hz Ricker wavelet forward modeling. In order to evaluate the denoising ability of the model, Gaussian random noise is added in the experiment to simulate the noise in actual microseismic monitoring.

[0051] The actual data comes from the actual fracturing monitoring data of the Z-5 well in the Songliao Basin. The horizontal section of the fracturing well is 2025~2895m deep and the vertical depth is 1789.7~1794.2m. The Z-5 well is a ground monitoring method, in which the observation system is a star array, with six survey lines arranged around the monitoring well. The designed survey network radius is 770m, and 11 detectors are arranged equidistantly on each survey line. The microseismic data in segy format is converted into a more easily operated mat file, and further data segmentation and other operations are performed to achieve efficient reading and processing of multi-channel microseismic data.

[0052] The samples are organized in a data block manner, and the microseismic data is divided into blocks by sliding a fixed-size window, and a new sample data is generated each time it moves. The sliding window size of the theoretical synthetic data is set to 128×128, and the sliding step is 32; the sliding window size of the actual data is set to 64×64, and the sliding step is 3.

[0053] Step 2: Multi-channel Kalman network model G structure setting:

[0054] The network model G is an unsupervised microseismic denoising network that combines Kalman filtering and recurrent neural networks. First, the Kalman filter state prediction and update equations for multi-channel microseismic data are established, and the multi-channel correlation is fully utilized to improve the characterization ability of the Kalman filter parameters. Secondly, the RNN operator for multi-channel Kalman filter state prediction and update is designed, and the parameters of the Kalman filter are optimized by automatically obtaining chain gradients, and a multi-channel Kalman network denoising based on the recurrent neural network model is constructed. Thirdly, combined with the unsupervised microseismic denoising training method, the Kalman parameters are automatically optimized to avoid excessive reliance on valid data labels.

[0055] The network model is set as:

[0056] (1) Multi-channel Kalman filter modeling

[0057] Multi-channel Kalman filter modeling makes full use of the correlation of multi-channel microseismic data to improve the characterization ability of the Kalman filter model. The principle of multi-channel Kalman filter modeling is: yes Multi-channel microseismic data at all times, hidden state It consists of a state vector and a state covariance matrix. Multi-channel Kalman filtering uses Time input and Hidden state at the moment Filter and output the hidden state at the current moment , as The partial input at each moment forms a continuous state update process on this basis. A Kalman filter method for parallel processing of multi-channel microseismic data is proposed, which considers the temporal and spatial connection of multi-channel microseismic data and performs state prediction and update of Kalman filter synchronously.

[0058] (2) Design of multi-channel Kalman operator

[0059] The observation data input to the multi-channel Kalman filter is the amplitude value of multiple microseismic data at the same time. ,In order to realize the Kalman filter state prediction and update process of multi-channel microseismic data, parallel prediction and update calculation equations and network implementation operators are defined.

[0060] Kalman filter prediction equation for multi-channel microseismic data Defined as:

[0061] ;

[0062] in, is the prediction matrix , is the state transfer matrix, is the process noise covariance matrix, and is defined as Represented as the concatenation of two matrices A and B , With the identity matrix Row splicing, denoted as Output The state vector is estimated a priori at the current moment , state covariance matrix constitute.

[0063] ;

[0064] The Kalman gain is calculated by the above formula and is used to combine the predicted state estimate with the observed data to generate an updated state estimate. is the observation matrix, is the transpose of the observation matrix; To simplify the calculation steps and ensure the effective processing of state update and covariance adjustment, the present invention optimizes the calculation operation of state vector and state covariance matrix, and proposes the Kalman filter update equation for multi-channel microseismic data. for:

[0065] ;

[0066] in, To update the matrix , is the observation data of all seismic channels at that moment, Estimated from the optimal state and the error covariance matrix The rows are concatenated and used as the hidden state at that moment. , participate in the prediction equation calculation at the next moment. The proposed multi-channel Kalman filter state prediction and update equation can make full use of multi-channel correlation to improve the characterization ability of Kalman filter parameters.

[0067] Optimal state estimation The main diagonal elements integrate the intra-channel and inter-channel data correlation. In order to convert the optimal state estimation from the system dimension to the observation dimension, the present invention proposes a diagonal mapping (matrix diagonal mapping, mp), that is:

[0068] ;

[0069] pass Mapping Mapping , and use it as the observation data at that moment The denoising result.

[0070] The multi-channel Kalman operator can enhance the model's ability to obtain the correlation between different seismic traces and significantly improve the processing efficiency of microseismic data.

[0071] (3) Design of automatic optimization method for multi-channel Kalman network parameters

[0072] Model G uses the Kalman parameters F, H, Q, and R as trainable parameters of the network, optimizes the Kalman filter parameters through a chain-like automatic solution method, and realizes automatic optimization of the Kalman parameters, thereby better adapting to the complex characteristics of microseismic data.

[0073] When the Kalman parameters are used as the trainable parameters of the network, during the forward propagation process, the network uses the Kalman filter parameters to calculate the state vector and state covariance matrix through multiple Kalman operators. The mean square error (MSE) is set as the loss function, and the loss function is defined as the model output With input data The difference between

[0074] ;

[0075] The loss function The prediction accuracy of the multi-channel Kalman network for the input multi-channel sequence is measured.

[0076] During the back propagation process, the chain rule is used to calculate Relative to the gradient of the Kalman parameters, and passing the gradient calculated by back propagation from the last time step T to the first time step 1, the gradient update formula of the state transfer matrix F is:

[0077] ;

[0078] The gradient calculation process of other Kalman parameters is similar. After the gradient is calculated by back propagation, the Adam optimization algorithm is used to update the parameters. The update formula is:

[0079] ;

[0080] in, represents the Kalman parameters of the network, Represents the updated value of the weight parameter, Indicates the value before updating. is the learning rate. Representative parameters The gradients are accumulated at each time step.

[0081] (4) Establishment of multi-channel Kalman network model

[0082] In order to achieve networking, model G uses RNN as the architecture, and combines the state prediction and update equations of multiple Kalman filters, and The mapping is embedded in the RNN network as an operator to optimize the time series state estimation process of the Kalman filter and improve the accuracy of state estimation. Through RNN, matrix parallel operation is realized to synchronize microseismic data and improve the efficiency of microseismic data processing.

[0083] A batch of noisy microseismic data After entering the multi-channel Kalman network, the Multi-channel microseismic data at time and Always hide status After multi-channel Kalman filter state prediction and update equation processing, the hidden state at that moment is output , go through Mapping to obtain the denoising result of microseismic data at this moment ; and Multi-channel microseismic data After multiple Kalman operation operators, we get Partial input with the next moment , and so on, the denoising result of noisy microseismic data is finally obtained. .

[0084] Step 3: Loss function setting:

[0085] The loss is defined as:

[0086] ;

[0087] Where: is the model output; Input data; Mean Squared Error (MSE) is used as the loss function to measure the prediction accuracy of the multi-channel Kalman network for the input multi-channel sequence. Model G is combined with an unsupervised microseismic denoising training method to achieve automatic optimization of the Kalman parameters, avoid excessive reliance on valid data labels, and overcome the problems of low signal-to-noise ratio of microseismic data, high acquisition and processing costs, resulting in a small amount of microseismic data, and difficulty in obtaining a large number of data samples.

[0088] Step 4: Network model training:

[0089] Due to the short duration of microseismic events, the theoretical synthetic samples were cut to 128×128 size, and the actual microseismic samples were cut to 64×64 size. During training, the data were input into the multi-channel Kalman network G in batches, and each batch contained 3 samples. The Adam method was used to optimize the network parameters. The microseismic data set was divided into two parts, 80% of the data set was selected as the training set, and the parameters of the model were continuously updated. The remaining 20% ​​was used as the validation set, which did not participate in the training process of the training set and was used to verify the performance of the network. During the model training process, 1 validation was performed after every 5 iterations of training. The overall training was performed 95 times and 19 times of validation.

[0090] Step 5: Microseismic data denoising test:

[0091] Input the sample into the multi-channel Kalman network model G with trained and adjusted parameters, and output the denoising result With clean data The SNR was calculated to test the denoising performance of the model. Microseismic data of different sizes were input into the multi-channel Kalman network model G and the traditional Kalman filter algorithm to compare the denoising calculation time of the two methods under different data sizes and test the computational efficiency of the model G. The network model G was trained 95 times and verified 19 times through overall training, and the average SNR of each verification was output to test the adaptive optimization of network parameters.

[0092] This embodiment also discloses an unsupervised microseismic denoising system based on a multi-channel Kalman filter network, comprising:

[0093] A data acquisition module is used to acquire multi-channel microseismic data and perform preprocessing;

[0094] Model building module, used to build a multi-channel Kalman seismic denoising network based on recurrent neural network;

[0095] A model training module, used for constructing a loss function, and performing unsupervised training on a multi-channel Kalman seismic denoising network based on the preprocessed multi-channel microseismic data and the loss function;

[0096] The denoising module is used to denoise the real-time microseismic data through the trained multi-channel Kalman seismic denoising network.

[0097] Reference Figure 6-8The training of the network model in the present invention mainly includes two parts: forward propagation of data and back propagation of errors. First, the trainable parameters of the multi-channel Kalman network are set as square matrices, and the Kalman parameters are initialized as random numbers; then the forward propagation process of model training is carried out, and the forward data and actual data are respectively divided into blocks by sliding the fixed-size window on the microseismic data, and the divided data is used as the input of the network model, and then the multi-channel Kalman filter state prediction and update equations are used, and After mapping and obtaining the output of the model, back propagation operation is performed to compare the output of the model with the input to obtain the error between the two. The hyperparameters of the model are adjusted through the Adam optimization algorithm until the convergence conditions are met, and the training of the network model is terminated to obtain a trained network model.

[0098] The experimental platform configuration of this embodiment is: Inter(R) Core(TM) CPU i7-10700K, 16 GB RAM, NVIDIA GeForce GTX 3080 Ti, operating system is Ubuntu20.04, CUDA11.1 acceleration is used, experimental environment is python3.9, and deep learning framework is Pytorch1.13.0 version. The specific implementation process is as follows:

[0099] Preprocessing of microseismic data training set

[0100] 1.1 Reading mat file

[0101] Read the segy file through obspy, extract the seismic data from the segy file and save it as a numpy array, save the relevant variables through the mat_dict dictionary, and finally save the .mat file.

[0102] Get all .mat files in the data_dir dataset directory through glob.glob(), define the data list to store the datasets that meet the requirements after cropping, loop the mat files in the folder through the for loop, and then call the gen_patches method to read the mat file and crop it, where the mat file is read through h5py.File().

[0103] 1.2 Sample cropping of microseismic data

[0104] After reading the mat file through h5py.File(), the microseismic data of the mat file is then cropped.

[0105] Due to the short duration of microseismic events, the synthetic data is cropped to m=128, and the actual microseismic data is small, n=64; the theoretical synthetic data is cropped to 128×128, and the actual microseismic data is cropped to 64×64. The sample patches are cut through a sliding window. If the proportion of the amplitude of 0 in the patchsh does not exceed 10%, the patches are retained. Finally, the patches that meet the conditions in the mat file are returned and saved in the data.

[0106] Ideal microseismic data preparation:

[0107] The ideal microseismic data is obtained by processing the actual microseismic data with traditional denoising technology. The ideal microseismic data preparation process is as follows: when the actual microseismic data sample is judged by the sliding window to retain the condition, if the actual microseismic data meets the condition, the clean data corresponding to the actual microseismic data is obtained by cutting the corresponding ideal microseismic data index position, and the actual microseismic data and its ideal microseismic data are corresponding one by one to form a path_pair according to the organization method of the Data class. Finally, the path_pair is formed into patches and returned to the data to form a data set for training the network.

[0108] Design network model structure: The network of the present invention combines Kalman filtering with recurrent neural network, proposes multi-channel Kalman filtering, constructs multi-channel Kalman operation operators, realizes automatic optimization of multi-channel Kalman parameters, and finally constitutes a multi-channel Kalman network model.

[0109] 3.1 Multi-channel Kalman Filter

[0110] Multi-channel Kalman filtering makes full use of the correlation of multi-channel microseismic data to improve the characterization ability of the Kalman filtering model. The principle of multi-channel Kalman filtering is: yes Multi-channel microseismic data at all times, hidden state It consists of a state vector and a state covariance matrix. Multi-channel Kalman filtering uses Time input and Hidden state at the moment Filter and output the hidden state at the current moment , as The partial input at each moment forms a continuous state update process on this basis. A Kalman filter method for parallel processing of multi-channel microseismic data is proposed, which considers the temporal and spatial connection of multi-channel microseismic data and performs state prediction and update of Kalman filter synchronously.

[0111] 3.2 Multi-channel Kalman operator

[0112] The observation data input to the multi-channel Kalman filter is the amplitude value of multiple microseismic data at the same time. ,In order to realize the Kalman filter state prediction and update process of multi-channel microseismic data, parallel prediction and update calculation equations and network implementation operators are defined.

[0113] The Kalman filter prediction and update equations for multi-channel microseismic data are defined as:

[0114] ;

[0115] ;

[0116] ;

[0117] in, is the prediction matrix , To update the matrix ,definition Represented as the concatenation of two matrices A and B , Estimated from the optimal state and the error covariance matrix The rows are concatenated and used as the hidden state at that moment. , participate in the prediction equation calculation at the next moment. The proposed multi-channel Kalman filter state prediction and update equation can make full use of multi-channel correlation to improve the characterization ability of Kalman filter parameters.

[0118] Optimal state estimation The main diagonal elements integrate the intra-channel and inter-channel data correlation. In order to transform the optimal state estimation from the system dimension to the observation dimension, this paper proposes a diagonal mapping (matrix diagonal mapping, mp), that is:

[0119]

[0120] Through the above formula Mapping Mapping , and use it as the observation data at that moment The denoising result of the microseismic data is obtained by looping the microseismic data. When the microseismic data processing at all times is completed, the denoising result of the microseismic data can be output. .

[0121] 3.4 Multi-channel Kalman network model

[0122] In order to achieve networking, model G uses RNN as the architecture, and combines the state prediction and update equations of multiple Kalman filters, and The mapping is embedded in the RNN network as an operator to optimize the time series state estimation process of the Kalman filter and improve the accuracy of state estimation. Through RNN, matrix parallel operation is realized, multiple microseismic data are input into the network in the form of tensors, and microseismic data are synchronously processed to improve the efficiency of microseismic data processing.

[0123] Three noisy microseismic data After entering the network, Multi-channel microseismic data at time and Always hide status After multi-channel Kalman filter state prediction and update equation processing, the hidden state at that moment is output , go through Mapping to obtain the denoising result of microseismic data at this moment ; and Multi-channel microseismic data After multiple Kalman operation operators, we get Partial input with the next moment , and so on, the denoising result of noisy microseismic data is finally obtained. .

[0124] Design loss function:

[0125] After the network architecture of the network model is determined, the present invention defines the loss as The mean square error loss function measures the prediction accuracy of the multi-channel Kalman network for the input multi-channel sequence, and back-propagation updates the Kalman parameters to achieve adaptive optimization of the network parameters.

[0126] 5. Train and save the network model:

[0127] Using the multimodal neural network model G established in the implementation process 2, the samples are input into the multi-channel Kalman network model G, and the microseismic data denoising results are output after the network model is forward-transferred. During the model training process, after every 5 iterations of training, one validation is performed. The overall training is performed 95 times and 19 times of validation. Finally, the trained model is saved, torch.save(model, os.path.join(results_directory, 'generator.pth')), where the save address is results_directory, and the model parameters are saved with generator.pth as the file name.

[0128] The optimization function used in this implementation is the Adam function to optimize the network parameters. The function call statement is optimizer = optim.Adam (model.parameters(), lr = opt.lr), model.parameters() obtains all trainable parameters of the model, and opt.lr sets the learning rate to 0.00001. This implementation sets the number of training iterations to 95 times, and sets epoch = 95.

[0129] 6. Test network model performance:

[0130] The network adopts a validation method after completing 5 iterations of training. The noisy microseismic data is input into the network model G to obtain the denoising results of the microseismic data and the SNR calculated by the ideal microseismic data to measure the denoising effect; the microseismic data of different sizes are input into the multi-channel Kalman network model G and the traditional Kalman filter algorithm to compare the denoising calculation time of the two methods under different data sizes and test the computational efficiency of the model G; during the model training process, the validation method is used once after completing 5 iterations of training. The overall training is carried out for 95 trainings and 19 validations to test the adaptive optimization ability of the network parameters.

[0131] Implementation effect:

[0132] Randomly select a segment of microseismic synthetic data such as Figure 6 As shown in (a), Gaussian noise with a standard deviation of 10% is added to the energy of the effective signal in the microseismic synthetic data. Figure 6 (b) shows that the signal-to-noise ratio is greatly reduced. The output SNR of the model of the present invention is 18.94dB; the output SNR of the traditional Kalman model is 15.83dB; the output SNR of the wavelet transform is 16.14dB; the output SNR of the VAE model is 15.11dB; the effects of the four methods on the test set are as follows Figure 6 (c)- Figure 6As shown in (f), it can be seen that the background noise residue in the denoising result of the traditional Kalman filter method is relatively obvious, the edge details are distorted, and at the same time, the data has obvious vertical texture. The reason is that the algorithm only uses the time correlation within the single-channel data to dynamically estimate the state of the signal, which cannot fully describe the connection between the data between channels, affecting the denoising effect, especially the large randomness of the lateral distribution of the data; the wavelet transform method assumes that the data conforms to a simple statistical distribution in the wavelet domain, establishes a denoising model and sets a threshold function. Due to the weak expression ability of the model, the results show blurred details and excessive smoothing of the high-frequency signal part; the VAE network is based on the principle of pure data drive, and the lack of effective guidance in the unsupervised denoising process will lead to blurred edges and loss of details; the multi-channel Kalman network method proposed in the present invention makes full use of the intra-channel and inter-channel correlation of microseismic data, combined with the time-varying characteristics of microseismic signals, and can achieve a higher denoising effect using the same number of sample training.

[0133] The traditional Kalman filter denoising method based on GPU acceleration is compared with the invention on microseismic data of different sizes, and the denoising calculation time of the two methods under different data sizes is compared. The results are shown in Table 1. It can be seen that under sample data of different sizes, the time consumption of the invention is about 50% less than that of the traditional Kalman filter method.

[0134] Table 1

[0135]

[0136] The overall training was performed with 95 training and 19 validation. Figure 7 The change of SNR value during the verification process is given. It can be observed from the figure that with the continuous training of the network, the average SNR value of the verification gradually increases, indicating that the Kalman parameters have been effectively updated and optimized during the network back propagation process, which can more accurately capture and characterize the distribution characteristics of microseismic data.

[0137] The original microseismic data of the first section of the Z-5 well were randomly selected as follows: Figure 8 (a) is shown. Only the actual data of the well is used to train the network model, so that the model can better learn the specific geological environment and noise characteristics. When the parameters of the trained multi-channel Kalman network are applied to time-varying data denoising, the real microseismic signals and noise can be separated more specifically, enhancing the overall reliability of the system. Figure 8(b) The ideal microseismic data after traditional denoising technology is given as the SNR benchmark for subsequent calculation. To compare the effects, the traditional Kalman filter, wavelet transform, VAE and the present invention are used to suppress the noise of the actual microseismic data. The output SNR of the model of the present invention is 26.89dB; the output SNR of the traditional Kalman model is 14.98dB; the output SNR of the wavelet transform is 19.50dB; the output SNR of the VAE model is 12.94dB; the four methods are as follows on the test set: Figure 8 (c)- Figure 8 As shown in (f), it can be seen that various methods have certain effects, but the traditional Kalman filter method fails to make full use of the inter-channel correlation, resulting in inaccurate separation of the output signal and noise, unclear event axis, and more residual noise; the wavelet transform method has excessive smoothing, resulting in loss of event edge information; the VAE network has a low signal-to-noise ratio of the input microseismic data and lacks effective data feature guidance, so the network output results show unclear event axis and distortion of edge signal details; in contrast, the present invention relatively well separates the effective signal from the noise, improves the clarity of the event axis, and is better than other algorithms in SNR value. This proves the effectiveness of the present invention on theoretical synthetic data and actual data.

[0138] The main contributions of the present invention are: (1) using the statistical characteristics of microseismic data, the Kalman filter prediction and update equations for multi-channel microseismic data are proposed, and the multi-channel correlation is fully utilized to improve the characterization ability of the Kalman filter parameters and improve the computational efficiency; (2) the RNN operator for state prediction and update of the multi-channel Kalman filter is designed, the Kalman filter parameters are optimized by the chain gradient automatic calculation method, and the multi-channel Kalman network denoising based on the RNN mode is constructed; (3) the network training proposed in this paper is combined with the unsupervised microseismic denoising training method to realize the automatic optimization of the Kalman parameters and avoid the dependence on the effective data labels. Finally, the effectiveness and practicality of the invention are verified by testing the multi-channel Kalman network with theoretical forward modeling and actual microseismic data.

[0139] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An unsupervised microseismic denoising method based on a multi-channel Kalman filter network, characterized in that: The following steps are involved: Acquire multi-channel microseismic data and perform preprocessing; The multi-channel microseismic data includes theoretical synthetic data and actual data; The process of obtaining the multi-channel microseismic data includes: selecting a monitoring well, arranging a number of survey lines around the monitoring well, arranging detectors at equal intervals on the survey lines, and obtaining actual microseismic data; generating theoretical synthetic data with Gaussian random noise added by forward modeling; Construct a multi-channel Kalman seismic denoising network based on recurrent neural network; The process of constructing a multi-channel Kalman seismic denoising network includes: Design multiple Kalman operators and embed them into the basic architecture; among them, a recurrent neural network is selected as the basic architecture; The multi-channel Kalman operation operator includes a Kalman filter state prediction equation, a Kalman filter update equation, and a diagonal mapping; The Kalman filter state prediction equation is: ； in, is the prediction matrix , is the state transfer matrix, is the process noise covariance matrix, With the identity matrix Row splicing, denoted as , output Including the current moment prior estimate state vector , state covariance matrix ; The Kalman filter update equation is: ； in, To update the matrix , is the Kalman gain, is the observation matrix, is the Kalman filter state prediction equation, is the seismic trace observation data at the corresponding time, Estimated from the optimal state and the error covariance matrix Row splicing composition, = , is a hidden state; Constructing a loss function, and performing unsupervised training on a multi-channel Kalman seismic denoising network based on the preprocessed multi-channel microseismic data and the loss function; Real-time microseismic data are denoised using a trained multi-channel Kalman seismic denoising network.

2. The unsupervised microseismic denoising method based on a multi-channel Kalman filter network according to claim 1 is characterized in that: The process of unsupervised training of a multi-channel Kalman seismic denoising network includes; The trainable parameters of the multi-channel Kalman seismic denoising network are set as square matrices and initialized as random numbers. The theoretical synthetic data and the actual microvibration data are divided into blocks through fixed-size window sliding, and then input into the multi-channel Kalman seismic denoising network. Back propagation operation is performed based on the loss function to obtain the error value. The parameters of the multi-channel Kalman seismic denoising network are optimized based on the error value and the Adam optimization algorithm until the convergence condition is met and the training is terminated.

3. The unsupervised microseismic denoising method based on a multi-channel Kalman filter network according to claim 1 is characterized in that: The process of denoising real-time microseismic data includes: The real-time microseismic data is input into the trained multi-channel Kalman seismic denoising network, and the initial hidden state is obtained through the Kalman filter state prediction equation; the final hidden state of the real-time microseismic data at the corresponding moment is obtained based on the real-time microseismic data, the initial hidden state and the Kalman filter update equation, and the final hidden state includes an optimal state estimate and an error covariance matrix; the optimal state estimate is mapped by diagonal mapping to obtain a denoising result.

4. An unsupervised microseismic denoising system based on a multi-channel Kalman filter network for implementing the method according to any one of claims 1 to 3, characterized in that: include: A data acquisition module is used to acquire multi-channel microseismic data and perform preprocessing; Model building module, used to build a multi-channel Kalman seismic denoising network based on recurrent neural network; A model training module, used for constructing a loss function, and performing unsupervised training on a multi-channel Kalman seismic denoising network based on the preprocessed multi-channel microseismic data and the loss function; The denoising module is used to denoise the real-time microseismic data through the trained multi-channel Kalman seismic denoising network.

Images