Diffusion model method and system for microseismic signal noise reduction

Abstract

The invention discloses a diffusion model method and system for microseismic signal noise reduction, and relates to the technical field of engineering monitoring signal processing. According to the technical scheme, the method comprises the following steps: obtaining pure microseismic signals and multiple types of real engineering noise signals, and generating a pure-noisy signal pair data set; a diffusion denoising neural network is designed, a residual block and an attention module are integrated, and a noise distribution feature coding and time step condition dual-control mechanism is introduced; a distributed data parallel strategy is adopted to optimize the network, and parameter updating is achieved by minimizing mean square error loss of a prediction fractional function and real probability density; and based on stochastic differential equation reverse solution, noise is gradually removed through multi-step correction-prediction iteration. The method has the advantages that full-automatic real-time noise reduction is achieved, noise adaptability is remarkably improved, P / S wave arrival time characteristics and micro-seismic energy are accurately reserved, and time-frequency domain fidelity is excellent. The method has strong robustness for complex noise such as blasting, current, whistling and the like, and provides a high-quality data basis for rock mass stability analysis, seismic source positioning and rockburst early warning.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of engineering monitoring signal processing, and relates to processing and analysis of microseismic monitoring signals, in particular to a diffusion model method and system for microseismic signal noise reduction. BACKGROUND

[0002] Microseismic monitoring is a key technology for evaluating the stability of rock mass in rock engineering, which can capture the micro-fracture signals inside the rock mass, and invert the event occurrence time, location and energy. This technology can analyze the spatio-temporal evolution characteristics of microseismic events, realize effective prediction and early warning of rock burst disasters, and significantly reduce the safety risk of engineering.

[0003] However, the engineering site environment is complex and changeable, and there is strong construction noise interference (mechanical vibration, blasting vibration, personnel activity, electromagnetic interference, etc.), and the microseismic sensor is highly sensitive to vibration, resulting in extremely low signal-to-noise ratio of the collected signals. Noise pollution causes two key problems: (1) noise covers the effective microseismic signal, making it difficult to quantify the microseismic energy; (2) noise blurs the P / S wave arrival time characteristics, which is difficult for experts to accurately pick up, and seriously reduces the reliability of the source location and rock burst prediction. Therefore, it is necessary to perform noise reduction processing on the original microseismic signal to improve the signal quality and provide necessary conditions for accurate source location and accurate rock burst prediction.

[0004] The existing noise reduction methods can be divided into two categories, both of which have significant defects. The traditional signal processing algorithm decomposes and filters the time-frequency domain of the noisy signal to separate the noise part and suppress the noise content in the signal. However, the traditional noise reduction method has three limitations: (1) highly dependent on manual intervention, requiring manual threshold setting; (2) weak generalization ability, parameter adjustment depends on experience, and difficult to adapt to complex and changeable engineering noise environment; (3) low signal fidelity, easy to cause distortion of effective microseismic signal. The noise reduction method based on machine learning can overcome the problem of manual intervention to a certain extent, but still has some defects: existing researches mostly rely on synthetic noise (such as Gaussian noise) or seismic wave data to train the model, and the performance drops sharply when migrated to real engineering noise; Therefore, there is an urgent need for a new noise reduction solution in the field of rock engineering applications that can meet the conditions of high automation, strong noise adaptability, and excellent feature fidelity. The present application is proposed based on this purpose. SUMMARY

[0005] In order to solve the technical problems existing in the prior art, the present application provides a diffusion model method and system for microseismic signal noise reduction, which aims to solve the problems of low signal-to-noise ratio, distortion of effective signal and fuzzy P / S wave characteristics of microseismic monitoring signals caused by strong noise interference in complex engineering environment; the scheme has high automation, strong noise adaptability and excellent feature fidelity.

[0006] To achieve the above object, the present application provides the following technical solutions: A diffusion model method and system for microseismic signal noise reduction, the steps are as follows: S1, data set construction, through artificial screening and machine aided classification to obtain pure microseismic signal and multi-class real engineering noise signal, pure-noise signal pair data set is generated by mixing according to random signal-to-noise ratio; S2, diffusion model construction, design diffusion denoising neural network, integrate residual block and attention module, introduce noise distribution feature coding and time step condition double control mechanism; S3, model training, optimize the network by using distributed data parallel strategy, realize parameter update by minimizing the mean square error loss of the predicted score function and the real probability density; S4: reverse diffusion reasoning, based on random differential equation inverse solution, remove noise step by step through multi-step correction-prediction iteration.

[0007] Further, in the step S1: the pure microseismic signal and noise signal include at least one of blue noise, current noise, blasting noise, siren noise, impulse noise and Gaussian noise; the signal-to-noise ratio ranges from-10 dB to 10 dB; the length of each microseismic signal is 4000 sampling points, and the sampling frequency is 10 kHz.

[0008] Further, in the step S2: the diffusion denoising neural network adopts U-Net architecture, including symmetrical encoder and decoder, which are respectively composed of multiple residual blocks and attention modules; the residual block includes group normalization layer, convolution layer, activation function and dropout layer; the attention module is used to capture long-distance dependence relationship, including multiple attention layers and linear mapping layers.

[0009] Further, the input of the diffusion denoising neural network is the time-frequency graph obtained by short-time Fourier transform, including two channels of real part and imaginary part; the input data is further spliced with noise signal to form four-channel data; the time step condition and noise distribution feature coding are embedded into the model through Gaussian Fourier projection and light neural network respectively.

[0010] Further, in the step S3: the distributed data parallel strategy is used to train on multiple GPUs synchronously; the training loss function is the mean square error of the gradient of the predicted score function and the real probability density; the Adam optimizer is used in the training process, the learning rate is 8x10 -4 , and the batch size is 512.

[0011] Further, in the step S4: the reverse diffusion process includes a correction step and a prediction step; the correction step uses an annealing Langevin dynamics corrector to move the sample in the predicted score direction; the prediction step uses the Euler-Maruyama method to solve the inverse stochastic differential equation.

[0012] Further, the diffusion model is constructed based on a stochastic differential equation, and the forward diffusion process is defined as: (1) wherein is a drift term, is a diffusion term, is a standard Wiener process; the drift term determines the diffusion direction of the sample, and the diffusion term avoids falling into a local optimum by adding random jitter; The reverse diffusion process is defined as: (2) wherein is a learnable parameter, i.e., a score function in SDE, which reflects the diffusion probability gradient of the current sample.

[0013] The present application also includes a microseismic signal denoising system for implementing the above method, comprising: A data acquisition and preprocessing module for acquiring microseismic signals and performing classification, mixing and normalization processing; A diffusion denoising neural network module for estimating the score function and performing reverse diffusion; A training module for training the neural network based on a distributed data parallel strategy; An inference module for performing multi-step reverse diffusion on the noisy signal to output a pure signal.

[0014] Further, the diffusion denoising neural network module includes an encoder, a decoder, a residual block, an attention module, a time step encoder and a noise feature encoder; Further, it also includes a short-time Fourier transform module for converting time-domain signals into time-frequency diagrams.

[0015] The present application has the following advantages: Without human intervention, real-time automatic efficient noise reduction can be realized; the time-frequency domain features of effective microseismic signals can be recovered to the maximum extent, and the key advantages lie in accurately maintaining the relevant features of P waves and S waves and microseismic energy; good robustness and adaptability are shown to different types of noise (blasting, siren, current, etc.); the accuracy and reliability of subsequent microseismic event automatic identification, accurate positioning, focal mechanism analysis and other tasks are greatly improved after noise reduction; a higher quality data basis is provided for rock engineering safety monitoring and disaster warning. Compared with the prior art, the diffusion model method and system for microseismic signal noise reduction have the following technical features and beneficial effects: (1) Realize full-automatic real-time noise reduction, significantly reduce human dependence: the present application constructs an end-to-end noise reduction process, without the cumbersome manual threshold setting and parameter adjustment in traditional methods, realizing full automation from signal input to pure signal output, greatly improving the efficiency and convenience of engineering application; (2) Excellent noise adaptability and strong robustness: by constructing a large-scale data set containing various real engineering noises for training, and combining with noise distribution feature coding mechanism, the model can effectively distinguish and remove complex and variable real noises, overcoming the problem of performance drop in real scene of traditional methods and models using only synthetic noise; (3) Accurate preservation of effective signal features, high time-frequency domain fidelity: the model integrates attention modules and residual blocks, and introduces a double-condition control mechanism, which can finely preserve the key features of microseismic signals during the noise reduction process, especially the arrival time information and waveform form of P waves and S waves, the time-frequency spectrum shows that the noise-reduced signal is highly close to the original pure signal, effectively avoiding distortion of effective signals; (4) Significantly improve the accuracy and reliability of subsequent analysis tasks: the signal-to-noise ratio of the noise-reduced signal is greatly improved, and the P / S wave features are clear, providing a high-quality data basis for subsequent key tasks such as microseismic event automatic identification, accurate positioning of the source, energy calculation and rock burst warning, directly enhancing the reliability of rock mass engineering safety monitoring; (5) Excellent training efficiency and model generalization ability: distributed data parallel strategy is used for training, fully utilizing multi-GPU computing power to achieve efficient training on large-scale data sets; the training and validation loss decreases and stabilizes quickly, indicating that the model has good generalization ability and does not overfit, suitable for engineering deployment; BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the present application will be described in detail below with reference to the drawings and detailed embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor. Among them: Figure 1 This is a schematic diagram of noise reduction for the diffusion model of the present invention; Figure 2 This is a diagram of the neural network framework of the present invention; Figure 3 This is a schematic diagram of the reverse diffusion (noise reduction) process of the present invention; Figure 4 This is a comparison diagram of noise reduction in Embodiment 1 of the present invention and traditional methods; Figure 5 These are some noise reduction examples from Embodiment 1 of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The following description, in conjunction with the accompanying drawings... Figures 1-5 The diffusion model method and system for denoising microseismic signals are further explained.

[0018] Example 1 The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description. The following technical paths are only preferred solutions of the embodiments of the present invention. The scope of protection of the present invention covers all microseismic signal denoising implementations that include the core architecture of the diffusion model (including the forward diffusion process and the reverse denoising inference), including but not limited to network structure adjustment, conditional coding mechanism optimization, or equivalent replacement of probability gradient calculation methods.

[0019] The steps of this invention include: The S1 dataset was constructed by manually screening and machine-assisted classification to obtain pure microseismic signals and multiple types of real engineering noise signals, and then mixed according to random signal-to-noise ratios to generate a pure-noise signal pair dataset. S2 diffusion model construction, design of diffusion denoising neural network (DMSD), integration of residual block and attention module, introduction of noise distribution feature encoding and time step condition dual control mechanism; S3 model training employs a distributed data parallel strategy to optimize the network, and parameter updates are achieved by minimizing the mean squared error loss between the predicted score function and the true probability density. S4 backward diffusion inference is based on the inverse solution of stochastic differential equations and gradually removes noise through multi-step correction-prediction iteration.

[0020] Preferably, in the S1 step, first, the microseismic data of the rock mass engineering is collected by using a microseismic monitoring system, and the system mainly includes a digital signal acquisition system, an accelerometer, a cable optical fiber and a digital signal processing system. After data collection, the classification neural network is used to preliminarily screen and classify the monitoring signals, and then the technical personnel manually review and finely divide the waveforms.

[0021] After machine screening and manual fine screening, the monitoring signals are classified into six categories, including pure microseismic signals and blue noise, current noise, blasting noise, whistle noise, pulse noise and other noise. Among them, there are 4500 pure microseismic signals, 350 representative samples are selected from each of the six categories of noise signals, and 350 Gaussian noise samples are additionally synthesized to construct a comprehensive noise data set. Then, each pure microseismic signal is mixed with each noise signal one by one to form a noisy microseismic data. In the mixing process, the signal-to-noise ratio (SNR) is randomly set, and the range is -10 dB to 10 dB, and finally a total of 11,025,000 pairs of pure-noisy microseismic sample data sets are constructed. Each microseismic signal has a fixed length of 4000 sampling points, and the sampling frequency is 10 kHz.

[0022] Further, in the S2 step, a microseismic signal denoising diffusion model is constructed. In this embodiment, the training framework is realized based on PyTorch 2.1.0 and Python 3.8. As shown in Figure 1 , the forward diffusion process is used to add Gaussian noise to the clean microseismic training sample step by step, and the noise amount added at each step is controlled according to the noise schedule. After T steps, it becomes pure noise . The present application uses a diffusion model based on stochastic differential equation (SDE) to describe the forward and reverse diffusion. The steps of the forward diffusion are: (1) where is the drift term, is the diffusion term, and is the standard Wiener process. The drift term determines the diffusion direction of the sample, and the diffusion term avoids falling into local optimum by adding random jitter. The formula of the reverse SDE diffusion is: (2) where is a learnable parameter, that is, the score function in SDE, which reflects the diffusion probability gradient of the current sample. Since other parameters are known, the score function estimate of the input is predicted by a neural network, and the next step of the reverse diffusion sample is obtained.

[0023] Specifically, a U-Net is used as the main architecture of the DMSD. The architecture of the DMSD is as follows: Figure 2As shown, mainly contains a U-Net architecture consisting of residual blocks and attention modules. U-Net contains two parts of encoder (down-sampling) and decoder (up-sampling). It should be noted that due to the difference between prediction and training target, the diffusion model is essentially different from the encoder-decoder. The down-sampling and up-sampling process of the model has a symmetrical resolution change, a total of 6 down-sampling layers and 6 up-sampling layers, and uses finite impulse response filter (FIR) to realize smooth sampling; After each down-sampling, the resolution is reduced to 1 / 2 of the original (blue arrow represents the down-sampling process, and the red arrow represents the up-sampling process).

[0024] The input data of the neural network is a spectrum with a size of 4@128x128 (4 channels 128*128 resolution). In order to obtain this input, the original waveform data needs to be processed. In order to reduce the influence of signal energy on the estimation of fractional function, the signal is first normalized. Then, the short-time Fourier transform (STFT) is used to extract the time-frequency features of the signal. The specific configuration is: FFT point number , sliding step , Window function type , number of time frames contained in each spectrum . The original length of the microseismic signal used is 4000 sampling points, which is expanded to 4064 sampling points after zero padding at the end. After STFT of the signal, a 128x128 complex spectrogram is obtained. The real and imaginary parts of the complex spectrum are separated to obtain two channels of data (x Figure 2 ). The noise signal y and the signal at time t based on the edge distribution are spliced together to form the input data. The method for calculating the edge distribution based on is as follows: In the forward diffusion, given the pure signal , the noisy signal and the random time step , the edge distribution is estimated as follows: (3) That is: , (4) Where: (5) (6) (7) Where and represent the mean and variance of the sample at time t, ​This indicates the drift intensity. and These represent the maximum and minimum values ​​of the diffusion coefficient, respectively. They are obtained based on a grid hyperparameter search.

[0025] In back diffusion, it is necessary to construct an initial state. ,at this time: (8) Where T=1. The obtained... It is concatenated with the noisy signal y to form a 128×128 4-channel data input DNN.

[0026] To effectively preserve the time-frequency domain details of the microseismic signal and finely remove noise, a noise distribution and feature encoding condition is introduced and injected into the neural network together with the time step condition. Figure 2 In this invention, the upper part of the DNN consists of the aforementioned biconditional encoding and progressive growth module. The current time step t and the noise distribution characteristics are encoded and embedded into the network structure. Specifically, this invention pre-trains a lightweight neural network to estimate the noise distribution characteristics. This lightweight neural network is a simplified version of the DNN, containing only three resolution layers, with the input... The spectrogram first undergoes self-attention position encoding. Before inputting the spectrogram into the self-attention module, sine-cosine position encoding is required. (9) (10) (11) in , where i represents the i-th position in the sequence. Indicates the encoding dimension. This represents half the index of a dimension pair. This invention encodes each position as two channels, so... The data is then merged with the original channels to form 4@128×128. Next, a 1×1 convolution is used for channel fusion to obtain 2@128×128 data. Then, a 3×3 2D convolution is used to expand it to 32@128×128, which is then downsampled by 1 / 4 to obtain 64@32×32 and 128@8×8 data. An attention module is included between the downsampled and upsampled 128@8×8 layers.

[0027] The network is pre-trained, and these parameters are frozen during the training of DMSD to achieve joint training. By introducing this condition, the DMSD's ability to perceive noise is improved, focusing on preserving the time-frequency local structural information of the microseismic signal, thus achieving fine noise denoising.

[0028] In terms of time-conditioned encoding, a common practice is to use Gaussian Fourier Projection (GFP), a learned projection that maps the scalar time coordinate into a multi-dimensional vector. Ultimately, the time-step conditioned is encoded as a tensor of size 2@128x128, which is merged with the noise-conditioned encoding and embedded into the residual block of the DNN, as shown in the top-left part of Figure 2 .

[0029] The residual block contains two main convolutional sub-layers, each preceded by group normalization and an activation function (SiLU). The first layer reduces the input channels from 128 to 64 and adds a linear layer processed time-noise conditioned embedding vector as a modulation term. The second layer maintains the channel number as 128 and uses dropout to enhance the generalization ability. At the same time, in order to realize the residual connection, a 1x1 convolution is used to map the original input from 256 channels to 128 channels, which is convenient for addition with the main branch output.

[0030] In U-Net, the receptive field of convolution operation is limited. In order to capture long-distance dependencies and strengthen structural information, attention modules are added at the middle resolution layer (16x16) and the bottom resolution layer (4x4). As shown in the lower part of Figure 2 , the attention module contains three attention layers NIN_0, NIN_1, NIN_2 and a mapping layer NIN_3. For a given input , it is first subjected to group normalization operation, and then transformed into query (q), key (k) and value (v) vectors by the attention layer, whose dimensions remain the same as the input tensor. Subsequently, attention calculation is performed: (12) where is a commonly used normalization function in classification tasks. Finally, the output is linearly mapped by NIN_3 to restore the original input dimension.

[0031] Preferably, in the S3 step, two neural network models need to be trained. Given the large size of the dataset (a total of 11,025,000 pairs of clean-noisy microseismic samples), to accelerate the training process and improve the model's generalization ability, this study uses the Distributed Data Parallel (DDP) strategy to train synchronously on 8 NVIDIA GeForce RTX4090 GPUs (single card memory 24GB). Each GPU is allocated an independent data subset, and global parameter updating is achieved through a gradient synchronization mechanism, thereby significantly improving training efficiency. During training, the single-card batch size is 64, and the total batch size is 512. The optimizer is Adam, and the learning rate is set to 8x10 -4The training framework is implemented based on PyTorch 2.1.0. The dataset is divided into training and testing sets with a ratio of 9:1. The loaded data needs to be zero-padded, normalized, and processed by short-time Fourier transform (STFT) according to the method described in Section 2.2.

[0032] The lightweight neural network for encoding noise conditions needs to be pre-trained before training the DNN. The model is also trained using the above dataset. The training loss function is: (13) where is the training loss, is the noise encoding output by the model, , and represent the noisy signal and the pure signal, respectively. The lightweight neural network is pre-trained by backpropagation to participate in the training of the subsequent DNN. It should be noted that during the DNN training phase, the weights of the lightweight network will be frozen to prevent affecting the training of the diffusion model.

[0033] The training of DMSD follows a time-step-driven conditional diffusion process. During training, a random time step is sampled, and is estimated according to the marginal distribution of the clean signal and the noisy signal y (calculated by formula 4). After concatenating and y channels, time-step embedding t and noise feature encoding are injected simultaneously to form a time-noise double-condition control. The loss function of the diffusion model is defined as: (14) where is Gaussian noise following a standard Gaussian distribution with mean . is the predicted score function of the DNN model. During training, since is known, the log gradient of as a Gaussian distribution is explicitly known. Therefore is the constructed gradient of the log probability density function at time t (

[0034] ). Figure 2The model is trained step by step by minimizing the loss function. The training set and validation set loss decrease synchronously, and after 5 epochs, the training loss decreases from 8321.6 to 159.6, and the validation loss decreases to 172.4. It should be noted that the model is a diffusion model, and the loss function contains random Gaussian noise, so it cannot be reduced to close to 0. When the model is stable, the training loss has been reduced to 1.9% of the initial loss. In addition, due to the large data set, the total batch size (512) is relatively small (the memory utilization of each graphics card is already greater than 95%), and the number of model parameter updates is large, so the loss has stabilized after the first 2 epochs, and there is no need for excessive training. In this study, a total of 5 epochs were trained (time-consuming 108 hours). From the training and validation loss, it can be seen that the model does not overfit.

[0035] Preferably, in the S4 step, the number of reverse diffusion steps StepN is set, and the signal state of the previous step is inversely solved according to the stochastic differential equation (SDE) of the diffusion process. After multiple reverse diffusion, the pure signal is obtained.

[0036] Further, the process of microseismic signal denoising is as shown in Figure 3 First, the input noisy signal is zero-padded, normalized, and STFT-processed according to the method described in Section 2.2. The initial state of reverse diffusion is constructed using formula 8 , and The processed noisy signal y is channel spliced and input into the DMSD model to obtain the predicted score Reverse diffusion mainly includes two steps of correction and prediction. The correction step uses an annealed Langevin dynamics corrector. The core principle is to move in the direction of the model prediction score at the current sample point, so as to approach the high probability density area, and make the sample more approximate to the true distribution. This stochastic process is described by the following formula: (15) Where represents the sample at the k-th step at time t, is the step size of the movement, which satisfies , and snr is a set hyperparameter, and the optimal value is obtained by grid hyperparameter search. is the noise standard deviation at the current time t, which is given by formula 6. According to the experiment, only one correction operation is needed in each reverse diffusion.

[0037] The prediction step uses the Euler-Maruyama method (Peter E Kloeden and Eckhard Platen. Numerical solution of stochastic differential equations) to solve the inverse SDE. According to formula 2, the discrete form of the inverse process can be expressed as: (16) where: (17) (18) In the formula, is the drift term, is the diffusion term. is the discrete time step, is a hyperparameter that controls the strength of the drift, set to 1.5.

[0038] After StepN backward diffusion iterations, the noise in the noisy signal is gradually removed, and the clean microseismic signal is finally obtained. As the number of diffusion steps increases, the denoising time will increase linearly, and the denoising performance will rapidly rise and then stabilize after reaching a certain value. Therefore, if you want to achieve the best denoising result, you need to randomly select a small amount of data for testing to determine the optimal StepN before performing a large-scale denoising task. In this embodiment, the number of diffusion steps is set to 15.

[0039] Figure 4 To compare the denoising results of the traditional method with those of the present embodiment 1, the denoising case of real blasting noise is shown in the figure. The left column of the figure is the waveform diagram of the signal, and the right column is the time-frequency spectrum diagram of the signal. The gray level of each pixel in the time-frequency spectrum diagram represents the energy size of the current time period and frequency band, and the bright color part represents the stronger energy distribution. Each row of subgraphs represents a signal, the first row is the original pure microseismic signal, the second row is the noisy signal, and the remaining rows are the denoising results of DMSD, wavelet denoising (WT), empirical mode decomposition (EMD), variational mode decomposition (VMD), autoencoder (AutoEncoder), and a deep learning denoising method based on mask (DeepDenoise), respectively. From the figure, it can be seen that the denoising result of DMSD is close to the original pure signal in both time domain and frequency domain. There is no blasting noise left in the denoising waveform and time-frequency diagram.

[0040] Some denoising cases in this embodiment are as follows Figure 5As shown, the figure contains two examples above and below. The first row in each example is the original pure signal and the added noise, the second row is the microseismic signal after adding noise. The third row is the signal after noise reduction and the separated noise. By comparing the added noise with the separated noise, it can be seen that the model realizes accurate removal of noise.

[0041] Embodiment 2 The embodiment provides a diffusion model method and system for microseismic signal noise reduction, including the following core steps: S1, data set construction; after data acquisition, the waveform is first manually screened and classified by technical personnel. After screening, the monitoring signal is classified into pure microseismic signal and noise signal. Then, each pure microseismic signal is mixed with each noise signal one by one to form a noisy microseismic data. In the mixing process, the signal-to-noise ratio (SNR) is randomly set. Finally, a pure-noisy microseismic signal pair data set is formed.

[0042] S2, diffusion denoising model construction; build a diffusion denoising model (DMSD). The model contains a symmetrical network structure of upsampling and downsampling, which is composed of multiple residual blocks and attention modules. They are respectively constructed by normalization layer, convolution layer, full connection layer, dropout layer, activation layer and attention layer. The noisy signal is output after the diffusion model, which is the current time log probability density function gradient (score function), that is, the diffusion direction.

[0043] S3, model training; to speed up the training process and improve the generalization ability of the model, consider using distributed data parallel (DDP) strategy for parallel training. The pure-noisy microseismic data pair obtained in step S1 needs to be preprocessed before inputting the model. In order to reduce the influence of signal energy on the score function estimation, first, the signal is normalized. Then, the short-time Fourier transform (STFT) is used to extract the time-frequency features of the signal. Then, the noise signal and the t time signal based on the edge distribution are spliced together to form the input data. During training, since the pure signal is known, the log gradient of the data distribution at t time is explicitly known. According to the true pure signal, the log probability density function gradient at t time is constructed, and the score function output by the model in S2 is calculated to form the loss function. Update the model parameters by back propagation, minimize the mean square error of the predicted score function to train the model, and use the grid hyperparameter search method to determine the optimal model hyperparameter combination.

[0044] S4, reverse diffusion (denoising) inference; the prediction target of the diffusion model is the diffusion probability gradient at time t, so when performing the denoising task (model inference), the pure microseismic signal cannot be directly obtained from the noisy signal. The diffusion probability gradient predicted by the model needs to be used to perform multi-step reverse diffusion and denoising on the noisy signal. The number of reverse diffusion steps StepN is set, and a time node from 0-1 is generated according to the diffusion step. In each diffusion step, the diffusion process of the previous time node is generated through the correction and prediction steps, and finally the pure microseismic signal is obtained through T-step reverse diffusion.

[0045] Obviously, the above embodiments are only examples for clearly illustrating, but not limiting the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments do not need to be exhausted, and the obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A method for denoising microseismic signals by diffusion model, characterized in that, The steps are as follows: S1, data set construction, through artificial screening and machine aided classification to obtain pure microseismic signals and multiple types of real engineering noise signals, pure-noise signal pair data set is generated by mixing according to random signal-to-noise ratio; S2, diffusion model construction, design diffusion denoising neural network, integrate residual block and attention module, introduce noise distribution feature coding and time step condition double control mechanism; S3, model training, optimize network by using distributed data parallel strategy, realize parameter update by minimizing the mean square error loss of predicted score function and real probability density; S4: reverse diffusion reasoning, based on random differential equation inverse solution, remove noise step by step through multi-step correction-prediction iteration.

2. The diffusion model method for microseismic signal denoising according to claim 1, characterized in that, In the step S1: the pure microseismic signal and noise signal include at least one of blue noise, current noise, blasting noise, siren noise, pulse noise, Gaussian noise; the signal-to-noise ratio ranges from-10 dB to 10 dB; the length of each microseismic signal is 4000 sampling points, and the sampling frequency is 10 kHz.

3. The diffusion model method for microseismic signal denoising according to claim 1, characterized in that, In the step S2: the diffusion denoising neural network adopts U-Net architecture, including symmetrical encoder and decoder, which are respectively composed of multiple residual blocks and attention modules; the residual block includes group normalization layer, convolution layer, activation function and dropout layer; the attention module is used to capture long-distance dependence, including multiple attention layers and linear mapping layers.

4. The diffusion model method for microseismic signal denoising according to claim 3, characterized in that, The input of the diffusion denoising neural network is the time-frequency diagram obtained by short-time Fourier transform, including two channels of real part and imaginary part; the input data is further spliced with noise signal to form four-channel data; the time step condition and noise distribution feature coding are embedded into the model through Gaussian Fourier projection and light neural network respectively.

5. The diffusion model method for microseismic signal denoising according to claim 1, characterized in that, In the step S3, a distributed data parallel strategy is used to synchronize training on multiple GPUs. The training loss function is the mean square error between the predicted score function and the true probability density gradient. During the training process, the Adam optimizer is used with a learning rate of 8x10 -4 and a batch size of 512.

6. The diffusion model method for microseismic signal denoising according to claim 1, characterized in that, In the step S4: the reverse diffusion process includes correction step and prediction step; the correction step uses annealing Langevin dynamics corrector to move the sample along the predicted score direction; the prediction step uses Euler-Maruyama method to solve the inverse random differential equation.

7. The diffusion model method for microseismic signal denoising according to claim 1, characterized in that, The diffusion model is constructed based on random differential equation, and the forward diffusion process is defined as: (1) wherein is a drift term, is a diffusion term, is a standard Wiener process; the drift term determines the sample diffusion direction, and the diffusion term avoids falling into local optimum by adding random jitter; The reverse diffusion process is defined as: (2) wherein is a learnable parameter, i.e. the score function in SDE, which reflects the diffusion probability gradient of the current sample.

8. A microseismic signal denoising system for implementing the method of any one of claims 1-7, characterized in that, It includes: Data acquisition and preprocessing module, used for collecting microseismic signals and performing classification, mixing and normalization processing; Diffusion denoising neural network module, used for estimating score function and performing reverse diffusion; Training module, used for training the neural network based on distributed data parallel strategy; Inference module, used for performing multi-step reverse diffusion on noisy signal to output pure signal.

9. The microseismic signal denoising system according to claim 8, characterized in that, The diffusion denoising neural network module includes encoder, decoder, residual block, attention module, time step encoder and noise feature encoder.

10. The microseismic signal denoising system according to claim 8, characterized in that, It also includes a short-time Fourier transform module for converting time domain signals into time-frequency diagrams.

Images