Noise reduction prior and multi-scale aggregation micro-seismic positioning method and device

Abstract

The embodiment of the invention discloses a noise reduction prior and multi-scale aggregation micro-seismic positioning method and device. According to the scheme, the method comprises the following steps: constructing an auto-encoder network model containing an encoder; a theoretical waveform data set containing noise is adopted to pre-train the auto-encoder network model, a trained encoder is generated, and the trained encoder is used for extracting theoretical waveform features without noise; splicing the trained encoder and the downstream task module, and constructing a microseism positioning network model; carrying out fine tuning on the micro-seismic positioning network model by adopting a micro-seismic data set containing a position label, and generating a trained target model; the target model is adopted to carry out seismic positioning on the microseismic data, noise reduction priori knowledge and multi-scale feature aggregation are fused to carry out microseismic source positioning, and a high-precision source positioning effect can be achieved even in the microseismic data with a low signal-to-noise ratio.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of Internet technology, and in particular to a microseismic positioning method fusing noise reduction prior and multi-scale aggregation and a device thereof. BACKGROUND

[0002] Hydraulic fracturing technology is an important means for unconventional oil and gas resource exploration and development. Through multi-stage fracturing of horizontal wells, rock deformation is caused by stress to produce fractures, forming a complex spatial fracture network to increase reservoir permeability. However, only effective fractures that are interconnected can achieve effective reservoir modification. Therefore, high-precision microseismic source positioning is the core task of microseismic monitoring technology.

[0003] With the increasing amount of seismic data collected by microseismic monitoring, traditional microseismic positioning methods cannot meet the demand for real-time positioning. At the same time, during field data acquisition, microseismic signals are easily disturbed by environmental noise, equipment noise, etc. due to their weak energy, resulting in low signal-to-noise ratio of microseismic data, which directly affects the subsequent source positioning accuracy.

[0004] Therefore, there is an urgent need for a method that can achieve high-precision source positioning even in low signal-to-noise ratio microseismic data. SUMMARY

[0005] The embodiments of the present specification provide a microseismic positioning method fusing noise reduction prior and multi-scale aggregation and a device thereof, to solve the technical problem of needing a method to achieve high-precision source positioning in low signal-to-noise ratio microseismic data.

[0006] To solve the above technical problems, one or more embodiments of the present specification are implemented as follows:

[0007] In a first aspect, the embodiments of the present specification provide a risk detection method based on small program dynamic and static analysis, comprising:

[0008] An autoencoder network model containing an encoder is constructed. A theoretical waveform dataset containing noise is used to pre-train the autoencoder network model to generate a trained encoder, wherein the trained encoder is used to extract theoretical waveform features without noise. The trained encoder and a downstream task module are spliced to construct a microseismic positioning network model. The microseismic positioning network model is fine-tuned using a microseismic dataset containing position labels to generate a trained target model. The target model is used for seismic positioning of microseismic data.

[0009] In a second aspect, one or more embodiments of the present specification provide an electronic device, comprising:

[0010] at least one processor; and

[0011] a memory in communication with the at least one processor; wherein

[0012] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of the first aspect.

[0013] In a fourth aspect, the embodiments of the present specification provide a non-volatile computer storage medium storing computer executable instructions, which, when read by a computer, cause one or more processors to perform the method of the first aspect.

[0014] The above at least one technical solution adopted by one or more embodiments of the present specification can achieve the following beneficial effects: by constructing a self-encoder network model comprising an encoder; pre-training the self-encoder network model using a theoretical waveform data set comprising noise to generate a trained encoder, wherein the trained encoder is used to extract theoretical waveform features without noise; splicing the trained encoder and a downstream task module to construct a microseismic positioning network model; fine-tuning the microseismic positioning network model using a microseismic data set comprising position labels to generate a trained target model; using the target model to perform seismic positioning on microseismic data, and fusing denoising prior knowledge and multi-scale feature aggregation to perform microseismic source positioning, even in low signal-to-noise ratio microseismic data, high-precision source positioning can be achieved. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present specification or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present specification, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0016] Figure 1 The flowchart provided by the embodiments of the present specification;

[0017] Figure 2 The microseismic data preprocessing diagram provided by the embodiments of the present specification;

[0018] Figure 3 The linear model training data and data label diagram provided by the embodiments of the present specification;

[0019] Figure 4 The convolutional denoising autoencoder network model diagram provided by the embodiments of the present specification;

[0020] Figure 5a A structure diagram of an attention mechanism module provided by an embodiment of the present specification, Figure 5b A structure diagram of an attention feature extraction module provided by an embodiment of the present specification;

[0021] Figure 6 A denoising effect diagram of microseismic data provided by an embodiment of the present specification;

[0022] Figure 7 A microseismic joint positioning network model diagram provided by an embodiment of the present specification;

[0023] Figure 8 A structure diagram of using a MSAA module provided by an embodiment of the present specification;

[0024] Figure 9 A complex velocity model diagram provided by an embodiment of the present specification;

[0025] Figure 10 A training set and test set diagram generated by a complex model provided by an embodiment of the present specification;

[0026] Figure 11 A microseismic source positioning overall error diagram provided by an embodiment of the present specification;

[0027] Figure 12 A microseismic source positioning result diagram provided by an embodiment of the present specification;

[0028] Figure 13 A schematic diagram of an electronic device provided by an embodiment of the present specification. DETAILED DESCRIPTION

[0029] An embodiment of the present specification provides a microseismic positioning method and device fusing denoising priori and multi-scale aggregation.

[0030] In order for those skilled in the art to better understand the technical solutions in the present specification, the technical solutions in the present specification will be described clearly and completely in conjunction with the drawings in the embodiments of the present specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments of the present specification, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0031] This application provides a method and device for microseismic localization that integrates denoising priors and multi-scale aggregation. It achieves high-precision localization of microseismic events under low signal-to-noise ratio conditions by constructing a joint microseismic localization network model. First, a self-supervised learning pre-trained convolutional denoising autoencoder is used to denoise the microseismic data while learning important feature representations within the data. Then, the encoder portion is retained and concatenated with downstream task modules for fine-tuning training to generate the joint microseismic localization network model.

[0032] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for fusing noise reduction priors and multi-scale aggregated microseismic location, provided for one or more embodiments of this specification.

[0033] Figure 1 The process may include the following steps:

[0034] S102: Construct an autoencoder network model that includes the encoder.

[0035] For example, a denoising autoencoder based on a convolutional neural network can be built to denoise microseismic data through self-supervised learning and extract waveform features and prior knowledge from the data.

[0036] The denoising autoencoder based on convolutional neural networks employs a composite feature extraction architecture. Specifically, it includes: replacing traditional standard convolutional modules with gated residual modules, achieving dynamic feature calibration through a learnable gating mechanism; using an attention mechanism module after the gated residual modules to collaboratively optimize channel attention and spatial attention weights, achieving adaptive enhancement extraction of key features; and introducing a pyramid attention feature extraction module at the network bottleneck layer, which, through its multi-scale perception and attention weighting mechanism, effectively aggregates contextual information from different receptive fields, significantly improving the ability to represent high-level semantic features of microseismic datasets.

[0037] S104, the autoencoder network model is pre-trained using a theoretical waveform dataset containing noise to generate a trained encoder, wherein the trained encoder is used to extract the theoretical waveform features after removing noise.

[0038] A theoretical waveform dataset can be pre-output using a theoretical model, and then noise can be added to it to obtain a noisy theoretical waveform dataset. An autoencoder network model can be pre-trained using this dataset. After pre-training, the autoencoder network model will have the ability to extract theoretical waveforms from noise. At this point, it can be considered that the trained encoder can extract the features of the noise-free theoretical waveform.

[0039] S106, splice the trained encoder and the downstream task module to build a microseismic positioning network model.

[0040] Specifically, an end-to-end microseismic joint positioning network can be built, and the encoder part of the pre-trained denoising autoencoder is used as the feature extraction backbone network. Meanwhile, the encoder-decoder bridging structure of the traditional U-Net is improved, and a multi-scale attention aggregation module (MSAA) is used to aggregate multi-scale features output by the encoder at different levels. In each level, the output of the trained encoder is used as the corresponding input of the downstream task module. In the downstream task module, a multi-scale attention aggregation module is used to aggregate multi-scale features output by the encoder at different levels. A downstream task module is connected to the output layer of the downstream task module, which is used to perform secondary enhancement on the output preliminary positioning Gaussian heat map to suppress noise interference and sharpen the spatial position features of the seismic source.

[0041] S108, fine-tune the microseismic positioning network model using a microseismic dataset containing position labels to generate a trained target model.

[0042] The microseismic dataset containing position labels can be generated in the following way: using the finite difference forward method to generate the required dataset within the velocity model range, and then using the two-dimensional Gaussian probability density formula to generate the corresponding position labels.

[0043] The joint positioning network is trained using a fine-tuning strategy, which optimizes the model parameters by minimizing the error between the predicted Gaussian heat map containing the seismic source position and the true label. The model parameters here include the parameters in the encoder and the downstream task module. Specifically, it includes:

[0044] The following loss function is used The trained encoder and downstream task module are fine-tuned simultaneously, where N is the total number of training samples, l i is the i-th microseismic data source position information predicted by the network, represents the true data label of the i-th microseismic data source.

[0045] The difference between the predicted heat map and the true Gaussian heat map is compared pixel by pixel, so that the probability distribution image output by the network approximates the heat map distribution of the true seismic source position in overall form. The optimization mechanism of this loss function has a double effect: on the one hand, it makes the peak value of the network prediction heat map correspond to the corresponding true seismic source peak position, and on the other hand, it promotes the network to learn the Gaussian distribution characteristics that conform to the physical law, including reasonable attenuation gradient and spatial distribution characteristics. This design not only enables the model to be more accurate in positioning, but also reflects the spatial confidence of the seismic source position estimation, providing a more reliable basis for subsequent analysis.

[0046] S110, performing seismic positioning on the microseismic data by using the target model.

[0047] By constructing a self-encoder network model containing an encoder; pre-training the self-encoder network model using a theoretical waveform data set containing noise to generate a trained encoder, wherein the trained encoder is used to extract noise-removed theoretical waveform features; splicing the trained encoder and a downstream task module to construct a microseismic positioning network model; fine-tuning the microseismic positioning network model using a microseismic data set containing position labels to generate a trained target model; and performing seismic positioning on microseismic data using the target model, fusing denoising prior knowledge and multi-scale feature aggregation for microseismic source positioning, even in low SNR microseismic data, high-precision source positioning can be achieved.

[0048] To make the scheme of the present application more specific and easy to understand, a specific experimental example is given below.

[0049] A linear velocity model is used, and the linear velocity model has a size of 630m x 630m; the grid spacing is 10m x 10m, and a surface receiving mode is used, with 64 detectors uniformly arranged on the surface, and the trace spacing is 10m. A 50Hz main frequency Ricker wavelet is selected as the source, and a microseismic data set containing position labels is generated, which can specifically include single-source, double-source and three-source event microseismic data sets, 1260 of each type of source event, a total of 3780 synthetic data sets, and the generated synthetic data is divided into a training set and a validation set in a ratio of 4:1.

[0050] At the same time, in order to improve the quality of network input data and reduce the computational complexity, data preprocessing operations are performed, including data segmentation, first arrival picking, and downsampling, as shown in Figure 2 The corresponding position label is established using a two-dimensional Gaussian probability density formula, and the data set and position label of the linear model are as shown in Figure 3 .

[0051] Figure 4 For the constructed self-encoder network model, the microseismic data is denoised by self-supervised learning, and the waveform features and prior knowledge in the data are extracted. In the network architecture, a gated residual module is innovatively used instead of a traditional convolutional module to dynamically adjust the data features, and an attention module is used after the gated residual module to dynamically adjust the importance of the feature map through the dual mechanisms of channel attention and spatial attention, thereby enhancing the network's ability to capture key information. An attention feature extraction module is connected at the bridge connection of the network to effectively aggregate the context information under different receptive fields, significantly improving the ability to extract waveform features in microseismic data, as shown in Figure 5a and Figure 5b .Figure 5a A structure diagram of an attention mechanism module provided by an embodiment of the present specification is shown in the following figure, Figure 5b A structure diagram of an attention feature extraction module provided by an embodiment of the present specification is shown in the following figure.

[0052] In the process, the network can be trained using a weight decay strategy, effectively balancing model parameter regularization and gradient optimization process, and improving the data denoising effect of the linear model, as shown in the following figure. Figure 6 It can be seen that the network can realize the denoising process of low signal-to-noise ratio microseismic data, and largely retains the original waveform characteristics. That is, the trained encoder has the function of extracting the theoretical waveform features without noise, laying a foundation for subsequent high-precision microseismic source positioning.

[0053] The present application adopts a self-supervised and supervised hybrid learning strategy, retains the encoder part in the pre-trained convolutional denoising autoencoder, and splices it with the downstream task module to construct a multi-scale aggregation microseismic positioning network model integrating denoising prior, as shown in the following figure. Figure 7

[0054] The jointly constructed microseismic positioning network model innovatively uses the MSAA multi-scale aggregation module, as shown in the following figure. Figure 8 Through aggregating the feature information output by the encoder at different stages, the MSAA multi-scale aggregation module realizes better aggregation ability of microseismic data waveform features and improves the precision of microseismic source positioning.

[0055] In the fine-tuning training process, the weights in the network are adjusted and updated, and the following weight adjustment strategy is adopted to prevent the model from overfitting. The expression is as follows:

[0056]

[0057] Where w t is the weight at time step t; alpha is the learning rate; is the gradient of the loss function L with respect to the weight w t ; lambda is the weight decay coefficient; w t+1 is the updated weight.

[0058] The position error between the to-be-tested Gaussian heat map and the data label is counted in the following way:

[0059]

[0060] Where error represents the position error; n represents the number of sources; i represents the i-th source; x T , y T represent the x and y grid point positions of the real source respectively; x P , y P ​respectively represent the x, y grid point positions of the predicted seismic source contained in the Gaussian heat map; d represents the velocity model grid spacing.

[0061] By the above formula, the peak position in the Gaussian heat map of the network prediction output is read, the x, y grid point position coordinates of the network predicted seismic source are obtained, and the difference comparison with the real seismic source position coordinates is performed to judge the goodness of the network prediction result.

[0062] Further, the test set can be used for testing, and after the precision and accuracy of the model meet certain requirements, the model training is completed, and the model can be used for predicting the position of the microseismic data.

[0063] In addition, the inventors of the present application also carried out the following comparative test. As shown in Table 1, under the same experimental environment, the following network models were tested respectively, and the results showed that even under the condition of using only 50% training data, the network of the present application could still achieve good positioning effect.

[0064] Table 1 Comparative analysis of error precision of different network predicted seismic source positions

[0065]

[0066] In order to further verify the effect of the scheme of the present application, the inventors also carried out the following test:

[0067] The Marmousi2 complex velocity model was selected to further test the network of the present application, and part of the Marmousi2 velocity model was intercepted for experimental test (as shown in Figure 9 The smoothed velocity model was used to generate a training set for model training, and the real velocity model was used to generate a test set for model testing. The training set and the test set generated by the complex Marmousi2 velocity model are shown in Figure 10

[0068] Each type of seismic event test set generates 40, a total of 120 test sets, and the network model performance test is carried out. The overall error graph of the microseismic source positioning is shown in Figure 11

[0069] At the same time, the following comparative test was also carried out on the complex velocity model. As shown in Table 2, under the same experimental environment, the following network models were tested respectively, and the results showed that the network of the present application could still achieve good positioning results using 50% training data in the complex velocity model.

[0070] Table 2 Comparative analysis of error precision of different network predicted seismic source positions in complex velocity model

[0071] ​​

[0072] Figure 12 The image shows the localization prediction results of the network of this invention for different source events in a complex velocity model. It can be seen that, in the complex velocity model, the scheme of this application still has good prediction results for different source events.

[0073] Secondly, based on the same idea, one or more embodiments of this specification also provide devices corresponding to the above-described methods, such as... Figure 13 As shown. Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this specification. The device includes:

[0074] At least one processor; and,

[0075] A memory communicatively connected to the at least one processor; wherein,

[0076] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0077] Based on the same idea, embodiments of this specification also provide a non-volatile computer storage medium corresponding to the above method, which stores computer-executable instructions. When a computer reads the computer-executable instructions from the storage medium, the instructions cause one or more processors to execute the method described in the first aspect.

[0078] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0079] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0080] The above merely provides one or more embodiments of the present specification and is not intended to limit the present specification. One of ordinary skill in the art can make various modifications and changes to one or more embodiments of the present specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of the present specification should be included in the scope of claims of the present specification.

Claims

1. A method of fusing noise reduction priori and multi-scale aggregation microseismic positioning, comprising: constructing an autoencoder network model comprising an encoder; pre-training the autoencoder network model using a theoretical waveform dataset comprising noise to generate a trained encoder, wherein the trained encoder is used to extract noise-removed theoretical waveform features; splicing the trained encoder and a downstream task module to construct a microseismic positioning network model; fine-tuning the microseismic positioning network model using a microseismic dataset comprising position labels to generate a trained target model; performing seismic positioning on microseismic data using the target model.

2. The method of claim 1, wherein, The autoencoder network model further comprises a gated residual module, an attention mechanism module, and an attention feature extraction module, wherein the gated residual module is used to learn a gating mechanism to achieve dynamic calibration of features; the attention module is used to extract microseismic data waveform features of the dataset, and the attention feature extraction module is used to extract high-level semantic features of the dataset.

3. The method of claim 1, wherein, Splicing the trained encoder and the downstream task module to construct the microseismic positioning network model comprises: In each level, the output of the trained encoder is taken as the corresponding input of the downstream task module; In the downstream task module, a multi-scale attention aggregation module is used to aggregate multi-scale features output by the encoder at different levels; A downstream task module is connected at the output layer of the downstream task module to perform secondary enhancement on the output preliminary positioning Gaussian heat map, so as to suppress noise interference and sharpen the spatial position features of the source.

4. The method of claim 1, wherein, Fine-tuning the microseismic positioning network model using a microseismic dataset comprising position labels comprises: The loss function is as follows The trained encoder and downstream task module are fine-tuned synchronously, wherein N is the total number of training samples, and l i is the i th microseismic data source location information predicted by the network, represents the true data label of the i th microseismic data source.

5. The method of claim 4, wherein, Synchronously fine-tuning the trained encoder and the downstream task module, comprising: Synchronously adjusting the parameter weights in the trained encoder and the downstream task module in the following manner: where w t is the weight at time step t; a is the learning rate; is the gradient of the loss function L with respect to the weight w t ; l is the weight decay coefficient; w t+1 is the updated weight.

6. An electronic device, comprising: at least one processor; and a memory communicatively connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 5.

Images