A tunnel advanced detection seismic wave signal denoising method and system

Abstract

The application provides a tunnel advanced detection seismic wave signal denoising method and system in the technical field of tunnel engineering. The method comprises the following steps: obtaining numerical simulation data of tunnel advanced detection and environmental vibration noise data of a construction site, and constructing clean superimposed seismic signals, noisy superimposed seismic signals and geological condition vectors, and extracting signal active section masks; in the training stage, taking the clean superimposed seismic signals as initial data in the forward diffusion process, taking the noisy superimposed seismic signals, seismic source wavelets in the numerical simulation data, the signal active section masks and the geological condition vectors as conditional guidance information, and training a conditional diffusion model; in the reasoning stage, performing noise prediction through the trained conditional diffusion model, and applying physical consistency gradient correction to the reverse sampling mean in the reverse diffusion process to obtain denoised seismic wave signals. The application enhances the generalization ability and adaptability of the model, and improves the denoising accuracy.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology, and in particular to a method and system for noise reduction of seismic wave signals in tunnel advance detection. Background Technology

[0002] Advanced geological prediction for tunnels is a crucial step in ensuring the safe construction of underground engineering projects. Among these methods, seismic wave detection has become one of the most widely used geophysical exploration techniques due to its advantages such as long detection range and sensitivity to large, adverse geological bodies. This method involves generating elastic waves at the tunnel face using artificial seismic sources (such as micro-blasting or mechanical impact), and then inverting the structure and mechanical parameters of the underlying medium based on the travel time and amplitude information of reflected waves generated at the interface with impedance differences. However, seismic wave signal acquisition for tunnel advance detection is typically conducted in complex and highly interfering construction environments: drilling and blasting operations at the tunnel face and mechanical tunneling processes generate high-energy transient vibrations, creating strong background interference. Furthermore, multiple reflected waves, surface waves, and industrial noise generated by equipment operation within the narrow tunnel space superimpose with geological reflection signals in both the spatiotemporal and frequency domains, leading to a significant decrease in the signal-to-noise ratio of the observed reflection signals. Since tunnel advance detection is highly dependent on reflection signals, the quality of the raw data directly affects the stability of subsequent wavefield imaging and medium parameter inversion. Therefore, acquiring high-fidelity, high-signal-to-noise-ratio seismic wave signals is fundamental to improving the accuracy and reliability of advance detection. Therefore, it is necessary to develop efficient and high-fidelity noise reduction technologies. Currently, engineering practice primarily employs two methods to address noise interference in seismic wave signals used in tunnel advance detection: offline detection and digital signal processing. Offline detection reduces environmental noise by halting the operation of construction equipment, improving data quality from the source. However, it disrupts construction continuity and reduces tunneling efficiency, making it difficult to implement frequently under the time constraints of deep-buried, long tunnels. Digital signal processing methods, on the other hand, preprocess the acquired data using techniques such as spectral analysis, time-frequency transformation, and adaptive filtering to suppress noise and enhance effective signals. However, traditional digital signal processing methods mainly rely on deterministic mathematical transformations for signal separation. When noise and effective reflected waves highly overlap in the frequency and time domains, it can easily lead to energy loss in weak reflected signals and attenuation of high-frequency detail information.

[0003] In recent years, with the development of artificial intelligence technology, deep learning-based signal denoising methods have been gradually applied to the field of seismic data processing. By learning the statistical characteristics of signals and noise, automated denoising is achieved, providing a new technical approach for signal extraction in complex environments. However, the optimization objective of this type of purely data-driven neural network training usually focuses on signal reconstruction error, lacks constraints on the seismic wave propagation mechanism, and makes it difficult to guarantee the consistency between the denoising results and the actual wavefield dynamics.

[0004] Therefore, under the background of complex construction interference, how to effectively suppress the noise of seismic wave signals and maintain the consistency between the noise reduction results and the actual wave field dynamic characteristics is one of the keys to improving the accuracy and reliability of tunnel advance detection. Summary of the Invention

[0005] This invention provides a method and system for denoising seismic wave signals in tunnel advance detection, which addresses the shortcomings of existing seismic wave signal denoising methods that lack constraints on seismic wave propagation mechanisms, and achieves consistency between the denoising results and the actual wave field dynamics characteristics, thereby effectively preserving the physical characteristics of seismic wave propagation and geological response information.

[0006] In a first aspect, the present invention provides a method for denoising seismic wave signals in tunnel advance detection, comprising: Obtain numerical simulation data for tunnel advance detection and environmental vibration and noise data at the construction site; A clean superimposed seismic signal is constructed based on the numerical simulation data, and a noisy superimposed seismic signal is constructed based on the clean superimposed seismic signal, the environmental vibration noise data, and standard white noise. The active segment mask of the noisy superimposed seismic signal is then extracted. Construct a geological condition vector based on the numerical simulation data; During the training phase, the clean superimposed seismic signal is used as the initial data in the forward diffusion process, and the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active segment mask of the signal, and the geological condition vector are used as conditional guidance information to train the conditional diffusion model. During the inference phase, noise prediction is performed using the trained conditional diffusion model, and a physical consistency gradient correction is applied to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal.

[0007] According to the present invention, a method for denoising seismic wave signals for tunnel advance detection is provided, wherein the numerical simulation data includes clean received signals, source wavelets, source coordinates, receiver coordinates, and background medium parameters under different geological models; A noisy superimposed seismic signal is constructed based on the clean superimposed seismic signal, the environmental vibration noise data, and standard white noise, including: The clean received signals under each geological model are normalized and averaged one channel at a time to obtain clean stacked seismic signals. The environmental vibration noise data is mixed with standard white noise at a random ratio to obtain mixed noise data. Equal-length segments are randomly extracted from the mixed noise data, root-normalized one channel at a time, and then superimposed and averaged to obtain superimposed noise. The signal-to-noise ratio coefficients are randomly sampled according to a mirror-flipped logarithmic distribution. The superimposed noise is then scaled according to the signal-to-noise ratio coefficients and superimposed onto the clean superimposed seismic signal to obtain a noisy superimposed seismic signal.

[0008] According to the present invention, a method for denoising seismic wave signals in tunnel advance detection is provided, which constructs a geological condition vector based on the numerical simulation data, including: The amplitude ratio before and after normalization is determined based on the clean superimposed seismic signal and the noisy superimposed seismic signal. A geological condition vector is constructed based on the background P-wave velocity, background S-wave velocity, and background density in the background medium parameters, as well as the source coordinates, the receiver coordinates, and the amplitude ratio before and after normalization.

[0009] According to the present invention, a method for denoising seismic wave signals in tunnel advance detection includes, during the training phase, using the clean superimposed seismic signal as initial data in the forward diffusion process, and using the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active signal segment mask, and the geological condition vector as conditional guidance information to train a conditional diffusion model, including: Based on the environmental vibration and noise data, the noise power spectral density of the construction site is estimated, and colored Gaussian noise is generated. During the forward diffusion process, the colored Gaussian noise and standard white noise are mixed in a random ratio to obtain mixed diffusion noise. The mixed diffusion noise is then gradually added to the clean superimposed seismic signal according to the diffusion time step to form intermediate diffusion states under different noise levels. Conditional guidance information for the training phase is constructed based on the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active segment mask of the signal, and the geological condition vector. The conditional guidance information of the training phase, the intermediate diffusion state, and the diffusion time step are input into the noise prediction network in the conditional diffusion model to predict noise. The network parameters of the noise prediction network are then optimized through backpropagation to obtain the trained conditional diffusion model.

[0010] According to the seismic wave signal denoising method for tunnel advance detection provided by the present invention, the loss function used to optimize the network parameters of the noise prediction network through backpropagation includes: noise prediction loss term, amplitude scale constraint term, and event window structure constraint term; Wherein, the noise prediction loss term is the sample-weighted noise prediction mean square error loss; the weight of the sample weighting is inversely proportional to the signal-to-noise ratio of the sample; The amplitude scale constraint term is used to constrain the amplitude scale consistency between the denoised seismic signal reconstructed based on the noise prediction results of the noise prediction network and the clean superimposed seismic signal. The event window structure constraint terms are determined as follows: one or more reflection event windows are determined based on the signal envelope of the noisy superimposed seismic signal; adjacent or overlapping reflection event windows are merged to generate an event window mask; and the local waveform, local envelope, or local energy ratio loss is weighted based on the event window mask.

[0011] According to the present invention, a method for denoising seismic wave signals in tunnel advance detection is provided. This method uses a trained conditional diffusion model to predict noise in the seismic wave signal to be denoised, and applies a physical consistency gradient correction to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal. The method includes: Based on the estimation of the active signal segment mask during the inference stage of the seismic wave signal to be denoised; Based on the seismic wave signal to be denoised, the signal active segment mask of the inference stage, the source wavelet corresponding to the seismic wave signal to be denoised, and the geological condition vector, the condition guidance information of the inference stage is constructed. The seismic wave signal to be denoised is subjected to scale correction and Fourier transform to obtain the frequency domain reference of the received signal; Based on the source wavelet corresponding to the seismic wave signal to be denoised, a source spectrum reliability analysis is performed to obtain the comprehensive weight of the source spectrum. At each time step of the back diffusion process, noise prediction is performed based on the condition guidance information of the inference stage, and a physical consistency gradient correction is applied to the back sampling mean in the back diffusion process based on the signal active segment mask of the inference stage, the received signal frequency domain reference, and the source reliability weight, to obtain the denoised seismic wave signal.

[0012] According to the present invention, a method for denoising seismic wave signals in tunnel advance detection includes, at each time step of the back-diffusion process, noise prediction is performed based on the conditional guidance information of the inference stage, and a physical consistency gradient correction is applied to the back-sampling mean during the back-diffusion process based on the signal active segment mask of the inference stage, the received signal frequency domain reference, and the source reliability weight, to obtain the denoised seismic wave signal, including: At each time step of the back diffusion process, the following denoising steps are performed: The intermediate diffusion state at the current time step and the conditional guidance information of the inference stage are fed into the noise prediction network to perform noise prediction and obtain the noise prediction value. Based on the noise prediction value, the clean seismic signal after denoising of the seismic wave signal to be reduced is estimated. And through Fourier transform, a clean seismic signal estimate in the frequency domain is obtained; A physical guidance gradient is constructed based on the frequency domain clean seismic signal estimation, the signal active segment mask in the inference stage, the received signal frequency domain reference, and the source spectrum comprehensive weight; The physical consistency gradient correction is applied to the backward sampling mean of the current time step based on the physical guidance gradient to obtain the intermediate diffusion state of the next time step. Repeat the denoising steps until the preset iteration termination condition is met to obtain the denoised seismic wave signal.

[0013] According to the present invention, a method for denoising seismic wave signals in tunnel advance detection is provided, which constructs a physical guidance gradient based on the frequency domain clean seismic signal estimation, the signal active segment mask in the inference stage, the received signal frequency domain reference, and the source spectrum comprehensive weight, including: The frequency domain physical consistency constraint is constructed based on the frequency domain clean seismic signal estimation, the received signal frequency domain reference, and the source spectrum comprehensive weight, and is expressed as follows:

[0014] in, For frequency domain physical consistency constraints; For inverse fast Fourier transform, These are frequency-domain clean seismic signal estimation and received signal frequency-domain reference, respectively. , These are the source reliability weight and the source spectrum comprehensive weight, respectively. μ The weak reliability band leakage penalty weight for the seismic source signal; Calculate the gradient of the frequency domain physical consistency constraint with respect to the intermediate diffusion state, the frequency domain physical consistency gradient; The physical guidance gradient is obtained by performing time-domain weighting and clipping on the frequency domain physical consistency gradient based on the signal active segment mask of the inference phase.

[0015] According to the seismic wave signal denoising method for tunnel advance detection provided by the present invention, the expression of the physical guiding gradient is:

[0016] in, To guide the gradient in physics, For the frequency domain physical consistency gradient, For the intermediate diffusion state at the current time step t x t Operators for finding the gradient; For masking the active signal segments during the inference phase; m min The minimum gradient weights are reserved for the inactive regions of the seismic wave signal to be denoised.

[0017] Secondly, the present invention also provides a seismic wave signal noise reduction system for tunnel advance detection, comprising: The data acquisition module is used to acquire numerical simulation data for tunnel advance detection and environmental vibration and noise data at the construction site; The sample construction module is used to construct a clean superimposed seismic signal based on the numerical simulation data, and to construct a noisy superimposed seismic signal based on the clean superimposed seismic signal, the environmental vibration noise data, and standard white noise; to extract the signal active segment mask of the noisy superimposed seismic signal; and to construct a geological condition vector based on the numerical simulation data. The model training module is used to train the conditional diffusion model during the training phase, using the clean superimposed seismic signal as the initial data in the forward diffusion process, and the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active segment mask of the signal, and the geological condition vector as conditional guidance information. The inference denoising module is used to perform noise prediction through the trained conditional diffusion model during the inference phase, and to apply a physical consistency gradient correction to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal.

[0018] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the seismic wave signal noise reduction method for tunnel advance detection as described above.

[0019] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the seismic wave signal noise reduction method for tunnel advance detection as described above.

[0020] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the seismic wave signal noise reduction method for tunnel advance detection as described above.

[0021] The beneficial effects of the technical solutions provided by some embodiments of the present invention include at least the following: 1) This invention provides a method and system for denoising seismic wave signals in tunnel advance detection. It constructs samples required for training a conditional diffusion model using numerical simulation data and environmental vibration and noise data from the construction site. The noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active signal segment mask, and the geological condition vector serve as guiding information. This allows the conditional diffusion model to simultaneously perceive the observed signal, source information, the time window of effective reflection events, and geological background / geometric constraints during denoising. This improves the accuracy of noise prediction while effectively suppressing data distribution shifts caused by complex construction environments, enhancing the model's generalization and adaptability in unknown or changing environments. Furthermore, during the inference stage, this invention applies a physical consistency gradient correction to the back-sampling mean during the back-diffusion process, compensating for the lack of seismic wave propagation mechanism constraints in existing seismic wave signal denoising methods.

[0022] 2) Based on clean superimposed seismic signals, this invention constructs noisy superimposed seismic signals by using randomly extracted environmental vibration noise segments and randomly sampled signal-to-noise ratio coefficients. This allows the training samples to cover a wide range from low noise to high noise, and makes low signal-to-noise ratio samples appear more frequently during training, thereby improving the generalization ability of the model.

[0023] 3) This invention improves the model's adaptability to different working conditions by fusing geological background parameters, source receiving geometric parameters, and signal amplitude ratios through geological condition vector fusion.

[0024] 4) This invention constructs colored Gaussian noise that matches the spectral characteristics of vibration and noise in the construction site environment. During the forward diffusion process, colored Gaussian noise is used to simulate the disturbance process, making the forward diffusion process closer to the actual engineering noise. This avoids the problem that traditional diffusion models based solely on the assumption of white noise are difficult to adapt to complex engineering noise environments.

[0025] 5) This invention constructs the loss function of the noise prediction network by using a noise prediction loss term, an amplitude scale constraint term, and an event window structure constraint term. Combined with a low signal-to-noise ratio sample weighting strategy and a time-step gating mechanism of the event window structure loss, it achieves the enhancement and fidelity reconstruction of weak reflection events.

[0026] 6) In the backdiffusion process, this invention constructs a frequency domain physical consistency constraint, which includes the frequency domain consistency constraint of the received signal and the frequency domain reliability constraint of the source spectrum. This ensures that the clean superimposed seismic signal estimation is consistent with the frequency domain reference of the received signal within the reliable frequency band of the source wavelet, and suppresses spectral leakage that does not have source energy support within the weakly reliable frequency band of the source wavelet. Furthermore, by guiding the observation consistency within the effective frequency band of the source through physical gradient, the invention embeds the observation consistency within the effective frequency band of the source into the backdiffusion sampling process. This solves the problems of over-smoothing, amplitude distortion, and false reflection events that easily occur in the pure data-driven denoising process. It can further suppress noise while preserving the effective reflection event morphology, thereby improving the fidelity of the denoising results of the advanced detection seismic waves.

[0027] 7) In summary, the present invention can achieve high-fidelity noise reduction and effective information recovery of seismic advance detection signals in complex tunnels, providing reliable data support for the identification of adverse geological bodies in front of the tunnel face and early warning of construction risks. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is one of the flowcharts illustrating the seismic wave signal noise reduction method for tunnel advance detection provided by the present invention; Figure 2 This is a comparative diagram of the estimated noise power spectral density, normalized power spectral density of colored Gaussian noise, and normalized power spectral density of white Gaussian noise provided by the present invention. Figure 3 This is a flowchart of mask construction for the training and inference phases provided by the present invention; Figure 4 This is a flowchart of the physical consistency gradient correction provided by the present invention; Figure 5 This is a schematic diagram comparing the low signal-to-noise ratio sample seismic wave signal to be denoised and its denoising result with a clean reference signal provided by the present invention. Figure 6 This is a schematic diagram comparing the seismic wave signal to be denoised and its denoising result with a clean reference signal, provided by the present invention, which has an extremely low signal-to-noise ratio. Figure 7 This is a schematic diagram comparing the local noise reduction curves of the ordinary conditional diffusion model and the conditional diffusion model of this invention. Figure 8 This is a schematic diagram of the seismic wave signal noise reduction system for tunnel advance detection provided by the present invention; Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] Please see Figure 1 , Figure 1 A flowchart illustrating a method for denoising seismic wave signals in tunnel advance detection, provided as an embodiment of the present invention, includes: S101. Obtain numerical simulation data for tunnel advance detection and environmental vibration and noise data at the construction site; S102. Construct a clean superimposed seismic signal based on numerical simulation data, and construct a noisy superimposed seismic signal based on the clean superimposed seismic signal, environmental vibration noise data and standard white noise, and extract the active segment mask of the noisy superimposed seismic signal. S103. Construct a geological condition vector based on numerical simulation data; S104. During the training phase, clean superimposed seismic signals are used as the initial data in the forward diffusion process, and noisy superimposed seismic signals, source wavelets in numerical simulation data, active signal segment masks, and geological condition vectors are used as conditional guidance information to train the conditional diffusion model. S105. In the inference phase, noise prediction is performed using the trained conditional diffusion model, and a physical consistency gradient correction is applied to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal.

[0032] This invention constructs the samples required to train the conditional diffusion model using numerical simulation data and environmental vibration and noise data from the construction site. It uses noisy superimposed seismic signals, source wavelets from the numerical simulation data, active signal segment masks, and geological condition vectors as guiding information. This allows the conditional diffusion model to simultaneously perceive the observed signal, source information, the time window of effective reflection events, and geological background / geometric constraints during denoising. This improves the accuracy of noise prediction while effectively suppressing data distribution shifts caused by complex construction environments, enhancing the model's generalization and adaptability in unknown or changing environments. Furthermore, during the inference phase, this invention applies a physical consistency gradient correction to the back-sampling mean during the back-diffusion process, compensating for the lack of seismic wave propagation mechanism constraints in existing seismic signal denoising methods.

[0033] like Figure 1 As shown, the method of the present invention generally includes three parts: a training data construction stage, a training stage, and an inference stage. The method of the present invention will be described below with reference to specific steps.

[0034] In the above implementation, for step S101, numerical simulation data of tunnel advance detection and environmental vibration and noise data monitored at the construction site are obtained.

[0035] The numerical simulation data for tunnel advance detection includes basic data and theoretical response data obtained by forward modeling the propagation process of seismic signals under known geological model conditions using numerical calculation methods. The environmental vibration and noise data consists of environmental vibration signals continuously recorded at the construction site using accelerometers.

[0036] In some embodiments, the numerical simulation data for tunnel advance detection includes clean received signals, source wavelets, source location, receiver location, and background medium parameters under different geological models.

[0037] Among them, the background medium parameters are pre-set indicators that reflect the physical properties of the normal rock mass around the tunnel. Taking the seismic wave method as an example, the background medium parameters may include P-wave velocity, S-wave velocity, rock mass density, quality factor (Q value), etc.

[0038] In embodiments of the present invention, training data will be constructed based on the above numerical simulation data and environmental vibration and noise data.

[0039] In the above embodiments, step S102 mainly involves constructing a noisy superimposed seismic signal based on the numerical simulation data and environmental vibration noise data obtained in step S101, and extracting its active signal segment mask.

[0040] In some embodiments, constructing a noisy superimposed seismic signal based on a clean superimposed seismic signal, environmental vibration noise data, and standard white noise includes: Clean stacked seismic signals are obtained by normalizing and stacking the clean received signals under various geological models one by one. The environmental vibration noise data is mixed with standard white noise at a random ratio to obtain mixed noise data. Equal-length segments are randomly extracted from the mixed noise data, root-normalized one channel at a time, and then superimposed and averaged to obtain superimposed noise. The signal-to-noise ratio coefficients are randomly sampled according to the mirror-flipped logarithmic distribution. The superimposed noise is scaled according to the signal-to-noise ratio coefficients and then superimposed onto the clean superimposed seismic signal to obtain the noisy superimposed seismic signal.

[0041] Specifically, firstly, channel-by-channel normalization is performed on the multi-channel received records in the numerical simulation data. This involves calculating the root mean square value of each received signal and unifying its amplitude scale. Then, the effective received channels are stacked and averaged to obtain a clean stacked seismic signal. Channel-by-channel normalization and stacking averaging can be calculated using the following formula:

[0042] in, The normalized simulated received signal is the i-th channel, where t is the sampling time point. For the i-th channel simulated received signal, Let N be the root mean square of the i-th channel simulated received signal. r The effective number of channels received; δ is a very small positive number, such as 10. -6 ; This is a clean overlay of seismic signals.

[0043] For environmental vibration and noise data, preprocessing is performed first. Specifically, environmental vibration and noise data can be obtained from a one-dimensional time series continuously recorded by an accelerometer at the construction site. For low-frequency monitoring data, rational number resampling processing is used, that is, first upsampling proportionally, for example, upsampling by a factor of L, then inserting L-1 zeros between every two original sampling points; then low-pass filtering is performed, the low-pass FIR filter will retain the components within the effective frequency band of the original signal, remove the high-frequency image caused by zero insertion, and at the same time smooth and complete the intermediate values ​​between the original sampling points; finally, downsampling is performed proportionally. After the low-pass filtering is completed, the signal is already at a sufficiently high temporary sampling rate, and then it is decimated by a downsampling factor M, that is, one sampling point is retained for every M sampling points, to obtain new environmental vibration and noise data at the target sampling rate.

[0044] Then, the new environmental vibration noise data at the target sampling rate is mixed with standard white noise at a random ratio to obtain mixed noise data. Equal-length segments are randomly extracted from the mixed noise data, and root mean square normalization and superposition averaging are performed on these noise segments to obtain superimposed noise.

[0045] Finally, random sampling according to a logarithmic distribution and mirror flipping are performed to obtain the signal-to-noise ratio (SNR) coefficients. The superimposed noise is then scaled according to the SNR coefficients and superimposed onto the clean superimposed seismic signal to obtain the noisy superimposed seismic signal, ensuring that the training samples cover different SNR conditions. The noisy superimposed seismic signal can be represented as:

[0046] in, This is a noisy superimposed seismic signal. To cleanly overlay seismic signals, n s ( t (This is) superimposed noise. αThis is the noise intensity coefficient.

[0047] In some embodiments, during the training phase, the active segment mask of the noisy superimposed seismic signal can be constructed from the envelope of the noisy superimposed seismic signal:

[0048]

[0049] in, m ( t ) represents the active segment mask of the noisy superimposed seismic signal, with a value range of [0,1]. This is the envelope of the noisy superimposed seismic signal. Let Hilbert transform function, τ low and τ high The confidence threshold for the envelope; γ The compression index is Smooth(·), which is a smoothing function, such as a moving average, used to suppress local spikes in the envelope; clip(·) is a cutoff function used to restrict the result to the range [0,1].

[0050] In some embodiments, during the training phase, the active signal segment mask can also be obtained by mixing the envelope mask of clean superimposed seismic signals and the envelope mask of noisy superimposed seismic signals in a preset ratio, so that the trained conditional diffusion model can adapt to the mask distribution during the inference phase.

[0051] In the above implementation, step S103 mainly involves constructing a geological condition vector based on numerical simulation data.

[0052] In some embodiments, constructing a geological condition vector based on numerical simulation data includes: The amplitude ratio before and after normalization is determined based on the clean stacked seismic signal and the noisy stacked seismic signal; A geological condition vector is constructed based on the background P-wave velocity, background S-wave velocity, background density, source coordinates, receiver coordinates, and amplitude ratio before and after normalization from the background medium parameters.

[0053] Specifically, the geological condition vector g can be represented as:

[0054] in, These are the background P-wave velocity, background S-wave velocity, and background density, respectively. The preset P-wave velocity, S-wave velocity, and density scales are used for normalization processing in this embodiment. 5000 m / s, 3000 m / s, and 3000 kg / m³ can be used respectively; , () represents the coordinates of the earthquake source. , ) represents the receiver coordinates, N x and N z This represents the mesh size of the numerical simulation model;

[0055] The amplitude ratio, Represents a noisy superimposed seismic signal. Represents a clean superimposed seismic signal, max( ) is the function for finding the maximum value.

[0056] Training samples are constructed based on the clean superimposed seismic signals, noisy superimposed seismic signals, source wavelets, active signal segment masks, and geological condition vectors, and used as training inputs for subsequent conditional diffusion models.

[0057] In the above implementation, step S104 is mainly based on training samples obtained in steps S102 to S103 to train the conditional diffusion model.

[0058] In some embodiments, during the training phase, a conditional diffusion model is trained using clean superimposed seismic signals as initial data in the forward diffusion process, and noisy superimposed seismic signals, source wavelets from numerical simulation data, active signal segment masks, and geological condition vectors as conditional guidance information, including: S104-1. Estimate the noise power spectral density of the construction site based on environmental vibration and noise data, and generate colored Gaussian noise; S104-2. During the forward diffusion process, colored Gaussian noise and standard white noise are mixed in a random ratio to obtain mixed diffusion noise. The mixed diffusion noise is then gradually added to the clean superimposed seismic signal according to the diffusion time step to form intermediate diffusion states under different noise levels. S104-3. Construct conditional guidance information for the training phase based on the noisy superimposed seismic signal, source wavelet, active signal segment mask, and geological condition vector in the numerical simulation data; S104-4. Input the conditional guidance information, intermediate diffusion states, and diffusion time steps from the training phase into the noise prediction network in the conditional diffusion model to perform noise prediction, and optimize the network parameters of the noise prediction network through backpropagation to obtain the trained conditional diffusion model.

[0059] Specifically, in step S104-1 above, the environmental vibration and noise data are preprocessed and the on-site noise power spectrum is estimated.

[0060] Specifically, the environmental vibration and noise data is high-pass filtered to remove low-frequency drift. The data is then framed according to the training sample length and a window function is applied. The power spectrum of each frame is calculated using Fast Fourier Transform, and then averaged and normalized to obtain the noise power spectral density. This power spectrum is used to describe the non-uniform frequency distribution of construction noise and provides statistical basis for the forward diffusion of colored noise. The Welch power spectrum estimation and normalization process can be expressed as:

[0061] in, P n ( f (This refers to environmental vibration and noise data at various frequencies) f Power spectral density estimation at; K The total number of frames after segmenting the scene noise. n k ( t ) is the first k Frame-based on-site noise time series, w ( t ) is a window function, such as the Hanning window, used to reduce spectral leakage; The normalized noise power spectral density, δ It is a very small positive number.

[0062] Colored Gaussian noise is generated based on the normalized noise power spectral density:

[0063] In the formula: The normalized noise power spectral density is estimated from environmental vibration and noise data. f For frequency variables, It is a complex white noise in the frequency domain, with its real and imaginary parts each independently following a standard Gaussian distribution; This is the inverse fast Fourier transform.

[0064] Colored Gaussian noise After being normalized to unit variance, it will be used as the colored noise component in the forward diffusion process to avoid the amplitude scale differences of different noise segments affecting training stability.

[0065] like Figure 2 The diagram shows a comparison of the normalized noise power spectral density of environmental vibration noise data, colored Gaussian noise, and Gaussian white noise. It can be seen that the noise power spectral density estimated based on environmental vibration noise data has a good correspondence with the sampled colored Gaussian noise power spectral density, which is better than that of Gaussian white noise. This indicates that the noise statistical characteristics in the forward diffusion process are consistent with the actual construction noise.

[0066] In step S104-2 above, colored Gaussian noise is... After normalization to unit variance, it is mixed with standard white noise at a random ratio to obtain mixed diffuse noise. The expression for mixed diffuse noise is:

[0067] in, For mixed diffused noise, It is colored Gaussian noise. The noise is standard white noise, and I is the identity matrix. The mixing coefficient is randomly sampled within a preset range.

[0068] During the forward diffusion process, the mixed diffused noise will be Clean, superimposed seismic signals are gradually added according to the diffusion time step to form intermediate diffusion states at different noise levels. The formula for this forward diffusion process is:

[0069]

[0070] in, For the first t The intermediate diffusion state at each time step The cumulative noise figure, β t For the first t Noise scheduling coefficient for each time step; For normalized mixed diffuse noise, std( ε mix (This refers to mixed diffused noise) The standard deviation reflects the magnitude of amplitude fluctuation in the noise sequence. ε 0 is a very small positive number; Since the spectral structure of the colored noise of this invention is derived from environmental vibration and noise data monitored on-site, the forward diffusion process can more closely approximate the noise distribution in the actual tunnel construction environment.

[0071] In step S104-3 above, based on the noisy superimposed seismic signal y Source wavelet s Signal active segment mask m Three types of time series conditions c=[ y , s , m The geological condition vector g is used to construct condition guidance information.

[0072] The signal activity segment mask is used to identify the time intervals in which reflection events may occur; the geological condition vector is used to describe the geological background and observational geometric differences. These conditional guidance information, along with the current diffusion state, are input into the conditional diffusion model, enabling the model to simultaneously perceive the observed signal, source information, effective event locations, and geological / geometric constraints during denoising.

[0073] In step S104-4 above, the noise prediction network of the conditional diffusion model is trained using conditional guidance information and intermediate diffusion states at each time step.

[0074] In this embodiment, a one-dimensional conditional U-Net is used as the noise prediction network. During training, the noise prediction network is input to the current diffusion state, conditional guidance information, and the current time step, and outputs the predicted noise.

[0075] Predicted noise It can be represented as , For network parameters θ Noise prediction network, θ These are the trainable network parameters of the neural network; Here, t represents the current diffusion state, and t represents the current time step; c = [ y , s , m ], y This is a noisy superimposed seismic signal. s For the source wavelet, m Mask for active signal segments; g This is a vector representing geological conditions.

[0076] In some embodiments, the loss function used to optimize the network parameters of the noise prediction network through backpropagation includes: a noise prediction loss term, an amplitude scale constraint term, and an event window structure constraint term; Among them, the noise prediction loss term is the sample-weighted mean square error loss of noise prediction; the weight of the sample weighting is inversely proportional to the signal-to-noise ratio of the sample. The amplitude scale constraint term is used to constrain the amplitude scale consistency between the denoised seismic signal reconstructed from the noise prediction results based on the noise prediction network and the clean superimposed seismic signal. The event window structure constraints are determined as follows: one or more reflection event windows are determined based on the signal envelope of the noisy superimposed seismic signal; adjacent or overlapping reflection event windows are merged to generate an event window mask; and the local waveform, local envelope, or local energy ratio loss is weighted based on the event window mask.

[0077] Specifically, the optimization objective of the noise prediction network includes a noise prediction loss term, an amplitude scaling constraint term, and an optional event window structure constraint term. Its loss function can be expressed as:

[0078]

[0079]

[0080] in, For the noise prediction loss term, a sample-weighted mean square error loss for noise prediction is used. For the noise intensity coefficient α Determined sample-level weights, α The larger the value, the lower the signal-to-noise ratio, and the higher the weight. The higher; For normalized mixed diffuse noise, To predict noise. For amplitude scale constraint term, Its corresponding weight coefficient, For the loss of event window structure, q i For the first i The weights of continuous events for each sample. For later time step gating functions, τ To preset the gate threshold, For indicator functions, only when Activate event window structure constraints.

[0081] Through the above training, the conditional diffusion model can learn the statistical regularity of recovering clean superimposed seismic signals from the intermediate diffusion state, given noisy observations, source wavelet, active signal segment mask, and geological condition vector.

[0082] In the above implementation, step S105 mainly involves backdiffusion based on the trained conditional diffusion model, and applying physical consistency constraints.

[0083] In some embodiments, noise prediction is performed on the denoised seismic wave signal using a trained conditional diffusion model, and a physical consistency gradient correction is applied to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal, including: S105-1. Estimate the active signal segment mask for the inference stage based on the observed seismic wave signal to be denoised; S105-2. Construct condition guidance information for the inference stage based on the seismic wave signal to be denoised, the active segment mask of the signal in the inference stage, the source wavelet corresponding to the seismic wave signal to be denoised, and the geological condition vector. S105-3. Perform scale correction and Fourier transform on the seismic wave signal to be denoised to obtain the frequency domain reference of the received signal; S105-4. Based on the source wavelet corresponding to the seismic wave signal to be denoised, perform source spectrum reliability analysis to obtain the comprehensive weight of the source spectrum. S105-5. At each time step of the back diffusion process, noise prediction is performed based on the condition guidance information of the inference stage, and physical consistency gradient correction is applied to the back sampling mean in the back diffusion process based on the signal active segment mask, received signal frequency domain reference, and source reliability weight of the inference stage to obtain the denoised seismic wave signal.

[0084] Specifically, in step S105-1 above, to reduce the mask distribution offset between the training and inference phases, this invention uses different methods to estimate the active signal segment mask in the training and inference phases, maintaining mask consistency between the training and inference phases, such as... Figure 3 The diagram shows the mask construction process during the training and inference phases.

[0085] During the training phase, the active signal segment mask can be obtained by mixing the envelope masks of clean stacked seismic signals and noisy stacked seismic signals at a preset ratio. However, during the inference phase, since clean stacked seismic signals are unavailable, the active signal segment mask is estimated from the seismic signal to be denoised after Hilbert envelope processing and smoothing. The Hilbert envelope of the seismic signal to be denoised is... This is used to calculate the signal active segment mask during the inference phase. m obs ( t ).

[0086] In step S105-2 above, a similar method to that in step S104-3 is used to utilize the seismic wave signal to be denoised. Signal active segment mask during the inference phase Source wavelet corresponding to the seismic wave signal to be denoised and geological condition vectors Constructing conditional guiding information for the reasoning phase.

[0087] like Figure 4 The flowchart shown is a physical consistency gradient correction flowchart provided by an embodiment of the present invention. Physical consistency gradient correction is achieved by constructing a received signal frequency domain reference and source spectrum reliability analysis and applying them to the back-diffusion estimation process.

[0088] In step S105-3 above, a frequency domain reference for the received signal is constructed. Specifically, considering the amplitude scale variation characteristics of the tunnel advance detection seismic wave signal, an amplitude scaling coefficient is used to scale the seismic wave signal to be denoised, ensuring that the scale-corrected seismic wave signal to be denoised is at the same amplitude scale as the current clean seismic signal estimate of the seismic wave signal to be denoised. Subsequently, a Fourier transform is performed on the scale-corrected seismic wave signal to be denoised to obtain the frequency domain reference for the received signal. This scale correction and Fourier transform can be expressed as:

[0089]

[0090] in, These are the seismic wave signals to be denoised before and after scale correction, respectively. The scaling factor for scale correction; For Fourier transform, This serves as the frequency domain reference for the received signal.

[0091] In step S105-4 above, source spectrum reliability analysis is performed. Specifically, the source spectrum reliability is determined based on the source wavelet spectrum energy distribution, and a comprehensive frequency domain weight is obtained by combining it with the field noise power spectrum. This weight is used to enhance the consistency constraint of the received signal in the frequency band where the source energy is reliable and the field noise has a low impact, thereby guiding the back-diffusion sampling process to maintain the spectral structure and amplitude characteristics of the effective reflection event.

[0092] Among them, the source reliability weight and the comprehensive source spectrum weight can be obtained through source spectrum reliability analysis, which can be expressed as:

[0093]

[0094] in, The source wavelet spectrum, As the source reliability weight, η The regularization coefficient is . The comprehensive weight of the source spectrum, For the field noise power spectral density The obtained frequency confidence, δ It is a very small positive number.

[0095] In step S105-5 above, backdiffusion estimation is performed. During the backdiffusion process, the conditional diffusion model starts backdiffusion sampling from a state of random mixed colored Gaussian noise. In each backdiffusion time step, the conditional diffusion model predicts noise based on the current diffusion state, the conditional guidance information of the inference stage, and the geological condition vector, and inversely derives the current clean seismic signal estimate from the predicted noise; the clean seismic signal estimate is used as the physically verifiable intermediate state of the current time step, and together with the received signal frequency domain reference, the source spectrum reliability analysis results, and the signal active segment mask of the inference stage, physical consistency constraints are calculated, and physical guidance gradients are calculated to guide the subsequent backsampling mean correction.

[0096] In some embodiments, at each time step of the backdiffusion process, noise prediction is performed based on the conditional guidance information of the inference phase, and a physical consistency gradient correction is applied to the back-sampling mean during the backdiffusion process based on the signal active segment mask, the received signal frequency domain reference, and the source reliability weight, to obtain the denoised seismic wave signal, including: At each time step of the back diffusion process, the following denoising steps are performed: S105-5-1. Input the intermediate diffusion state of the current time step and the condition guidance information of the inference stage into the noise prediction network to perform noise prediction and obtain the noise prediction value. S105-5-2. Based on the noise prediction value, the clean seismic signal estimate after denoising of the seismic wave signal to be denoised is calculated, and the frequency domain clean seismic signal estimate is obtained through Fourier transform. S105-5-3. Construct a physical guidance gradient based on frequency domain clean seismic signal estimation, signal active segment masking during the inference stage, received signal frequency domain reference, and source spectrum comprehensive weighting. S105-5-4. Apply a physical consistency gradient correction to the backward sampling mean of the current time step based on the physical guiding gradient to obtain the intermediate diffusion state of the next time step. S105-5-5. Repeat the above denoising steps until the preset iteration end condition is met to obtain the denoised seismic wave signal.

[0097] Specifically, in step S105-5-1 above, noise prediction is performed based on the noise prediction network to obtain the noise prediction value. ,in, , For the seismic wave signal to be denoised, For the signal active segment mask in the inference phase, The source wavelet corresponding to the seismic wave signal to be denoised This represents the geological condition vector for the reasoning stage.

[0098] In step S105-5-2 above, to avoid over-smoothing or spurious reflections caused by purely data-driven sampling, the clean stacked seismic signal estimate is derived from the noise prediction value, which can be expressed as:

[0099] in, x t This represents the intermediate diffusion state at the current time step t; Estimation of clean seismic signals; This is the cumulative noise figure; This represents the predicted noise output by the noise prediction network at the current time step t.

[0100] With the received signal frequency domain reference Similarly, estimating clean seismic signals Source wavelet The corresponding frequency domain representation is obtained through Fourier transform:

[0101] in, This represents the Fourier transform, i.e. It is frequency domain clean seismic signal estimation, For the source wavelet The frequency domain representation of .

[0102] In step S105-5-3 above, based on the received signal frequency domain reference, frequency domain clean seismic signal estimation, signal active segment mask in the inference stage, and source wavelet spectrum reliability analysis results, frequency domain physical consistency guidance is constructed, and physical guidance gradient is built.

[0103] In some embodiments, a physical guidance gradient is constructed based on frequency-domain clean seismic signal estimation, a signal active segment mask during the inference phase, a received signal frequency-domain reference, and a comprehensive weighted source spectrum, including: Based on the estimation of clean seismic signals in the frequency domain, the frequency domain reference of the received signal, and the comprehensive weight of the source spectrum, a frequency domain physical consistency constraint is constructed, which includes the frequency domain consistency constraint of the received signal and the reliability constraint of the source spectrum, and is expressed as follows:

[0104] in, For frequency domain physical consistency constraints; For inverse fast Fourier transform, These are frequency-domain clean seismic signal estimation and received signal frequency-domain reference, respectively. The comprehensive weight of the source spectrum, , As the source reliability weight, Confidence level of on-site noise frequency; μ The weak reliability band leakage penalty weight for the seismic source signal; Calculate the gradient of the frequency domain physical consistency constraint with respect to the intermediate diffusion state, and the frequency domain physical consistency gradient; The physical guidance gradient is obtained by time-domain weighting and clipping of the frequency domain physical consistency gradient based on the signal active segment mask during the inference phase.

[0105] Specifically, This is a data consistency term within the reliable frequency band, i.e., a frequency domain consistency constraint for the received signal, used to ensure that the clean seismic signal estimation remains consistent with the frequency domain reference of the received signal within the reliable frequency band of the source wavelet; The source weak reliability frequency band leakage penalty term, namely the source spectral reliability constraint, is used to suppress spectral leakage that does not have source energy support in the source wavelet weak reliability frequency band.

[0106] The frequency-domain physical consistency loss is used to calculate the gradient of the current diffusion state, then time-domain weighted using a mask of the active signal segment. During the weighting of the physical guided gradient, a preset minimum weight is retained for inactive regions, rather than resetting their weights to zero. Finally, the frequency-domain weights are combined for smoothing and pruning to form the physical guided gradient.

[0107] In some embodiments, the expression for the physical guidance gradient is:

[0108] in, To guide the gradient in physics, For the frequency domain physical consistency gradient, The gradient operator represents the computation of... For the intermediate diffusion state at the current time step t x t The partial derivative at each sampling point; For masking the active signal segments during the inference phase; m min The minimum gradient weight is reserved for inactive regions of the seismic wave signal to be denoised, in order to avoid the gradient of inactive regions being completely set to zero, which would lead to discontinuities at the boundary.

[0109] In step S105-5-4 above, a physical consistency gradient correction is applied to the backsampled mean of the current time step based on the physical guidance gradient to obtain the intermediate diffusion state of the next time step, including:

[0110] in, , These represent the intermediate diffusion states at the current time step t and the next time step t-1, respectively. The backward sampled mean at the current time step t. , , These represent the signal retention rate at the previous time step t and the added noise intensity, respectively. The physical guiding strength for annealing with diffusion steps; To guide the gradient in physics, For the frequency domain physical consistency gradient, For masking the active signal segments during the inference phase; It is randomly sampled Gaussian noise, where I represents the identity matrix; It is the variance of the current time step t.

[0111] In step S105-5-5 above, repeat the denoising steps from S105-5-1 to S105-5-4 until the preset iteration termination condition is met, and obtain the denoised seismic wave signal.

[0112] The effectiveness of the method of the present invention will be verified below with specific experimental data.

[0113] In this experiment, the advanced detection seismic waves obtained from numerical simulation were selected as clean reference signals. The environmental vibration noise collected on site or the equivalent noise generated based on the on-site noise power spectrum were superimposed on the clean reference signals to construct two types of seismic wave signals to be denoised: low signal-to-noise ratio and extremely low signal-to-noise ratio.

[0114] Figure 5 This is a schematic diagram comparing the denoised seismic wave signal with a low signal-to-noise ratio sample and its denoising result with a clean reference signal. When the input signal-to-noise ratio increases from 0.9dB to 17.8dB, the overall correlation coefficient is >0.99, the peak phase correlation coefficient is 1.00, and the peak amplitude correlation coefficient is >0.99, indicating that the method of the present invention can maintain the phase and amplitude shape of the reflection event under strong noise conditions.

[0115] Figure 6 This diagram illustrates the comparison between the seismic wave signal to be denoised and its denoising result with a clean reference signal, given an extremely low signal-to-noise ratio sample. When the input signal-to-noise ratio increases from -4.3dB to 13.0dB, the overall correlation coefficient is >0.97, the peak phase correlation coefficient is >0.99, and the peak amplitude correlation coefficient is >0.99. This demonstrates that the method of this invention can still recover the main reflection events and suppress random noise even in extremely noisy environments.

[0116] Furthermore, a set of typical test samples was selected, and the local noise reduction curves of the ordinary conditional diffusion model and the conditional diffusion model of this invention were compared within their effective event time windows. The results are as follows: Figure 7As shown. Among them, the ordinary conditional diffusion model is the comparative model after disabling physical consistency guidance. Its network structure, input conditions and sampling process are consistent with the method of this invention, but it does not use the received signal frequency domain reference, source spectrum reliability analysis, on-site noise frequency confidence, and signal active segment mask weighted physical guidance gradient.

[0117] Depend on Figure 7 It is evident that the seismic wave signal to be denoised contains significant random disturbances, and the peaks and troughs of effective reflection events are interfered with by noise. While the conventional conditional diffusion model can recover the main waveform structure, it still deviates from the clean reference signal in some peak amplitudes and local phase tracking. The method of this invention outputs a curve that is closer to the clean reference signal, especially at the main reflection events around approximately 50 ms to 55 ms, 70 ms to 75 ms, and 88 ms to 93 ms, where it better preserves the waveform structure and amplitude variations.

[0118] In this typical sample, the output signal-to-noise ratio (SNR) of the ordinary conditional diffusion model is 12.33 dB, while the SNR of the method of this invention is 16.61 dB, an improvement of 4.28 dB compared to the ordinary conditional diffusion model. This result demonstrates that this invention, by constraining the back-diffusion sampling process through physical consistency guidance, can further suppress noise while preserving the effective reflection event morphology, thereby improving the fidelity of the noise reduction results for advanced seismic waves.

[0119] Please see Figure 8 , Figure 8 A schematic diagram of a seismic wave signal noise reduction system for tunnel advance detection provided as an embodiment of the present invention, the system comprising: The data acquisition module 810 is used to acquire numerical simulation data of tunnel advance detection and environmental vibration and noise data of the construction site; The sample construction module 820 is used to construct clean superimposed seismic signals based on numerical simulation data, and to construct noisy superimposed seismic signals based on clean superimposed seismic signals, environmental vibration noise data and standard white noise, and to extract the signal active segment mask of the noisy superimposed seismic signals; and to construct geological condition vectors based on numerical simulation data. The model training module 830 is used to train the conditional diffusion model during the training phase, using clean superimposed seismic signals as the initial data in the forward diffusion process, and using noisy superimposed seismic signals, source wavelets in numerical simulation data, active signal segment masks, and geological condition vectors as conditional guidance information. The inference denoising module 840 is used to perform noise prediction through the trained conditional diffusion model during the inference phase, and to apply a physical consistency gradient correction to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal.

[0120] The seismic wave signal denoising system for tunnel advance detection described above and the seismic wave signal denoising method for tunnel advance detection described above can be referred to each other.

[0121] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions in the memory 930 to execute a seismic wave signal noise reduction method for tunnel advance detection provided in the above-described method embodiments.

[0122] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute a method for reducing the noise of seismic wave signals for tunnel advance detection provided in the above-described method embodiments.

[0124] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform a method for denoising seismic wave signals for tunnel advance detection provided by the methods described above.

[0125] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0126] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for denoising seismic wave signals in tunnel advance detection, characterized in that, include: Obtain numerical simulation data for tunnel advance detection and environmental vibration and noise data at the construction site; A clean superimposed seismic signal is constructed based on the numerical simulation data, and a noisy superimposed seismic signal is constructed based on the clean superimposed seismic signal, the environmental vibration noise data, and standard white noise. The active segment mask of the noisy superimposed seismic signal is extracted. Construct a geological condition vector based on the numerical simulation data; During the training phase, the clean superimposed seismic signal is used as the initial data in the forward diffusion process, and the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active segment mask of the signal, and the geological condition vector are used as conditional guidance information to train the conditional diffusion model. During the inference phase, noise prediction is performed using the trained conditional diffusion model, and a physical consistency gradient correction is applied to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal.

2. The method for denoising seismic wave signals in tunnel advance detection according to claim 1, characterized in that, The numerical simulation data includes clean received signals, source wavelets, source coordinates, receiver coordinates, and background medium parameters under different geological models; A noisy superimposed seismic signal is constructed based on the clean superimposed seismic signal, the environmental vibration noise data, and standard white noise, including: The clean received signals under each geological model are normalized and averaged one channel at a time to obtain clean stacked seismic signals. The environmental vibration noise data is mixed with standard white noise at a random ratio to obtain mixed noise data. Equal-length segments are randomly extracted from the mixed noise data, root-normalized one channel at a time, and then superimposed and averaged to obtain superimposed noise. The signal-to-noise ratio coefficients are randomly sampled according to a mirror-flipped logarithmic distribution. The superimposed noise is then scaled according to the signal-to-noise ratio coefficients and superimposed onto the clean superimposed seismic signal to obtain a noisy superimposed seismic signal.

3. The method for denoising seismic wave signals in tunnel advance detection according to claim 2, characterized in that, Constructing a geological condition vector based on the numerical simulation data includes: The amplitude ratio before and after normalization is determined based on the clean superimposed seismic signal and the noisy superimposed seismic signal. A geological condition vector is constructed based on the background P-wave velocity, background S-wave velocity, and background density in the background medium parameters, as well as the source coordinates, the receiver coordinates, and the amplitude ratio before and after normalization.

4. The method for denoising seismic wave signals in tunnel advance detection according to claim 1, characterized in that, During the training phase, the clean superimposed seismic signal is used as the initial data in the forward diffusion process, and the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active signal segment mask, and the geological condition vector are used as conditional guidance information to train the conditional diffusion model, including: Based on the environmental vibration and noise data, the noise power spectral density of the construction site is estimated, and colored Gaussian noise is generated. During the forward diffusion process, the colored Gaussian noise and standard white noise are mixed in a random ratio to obtain mixed diffusion noise. The mixed diffusion noise is then gradually added to the clean superimposed seismic signal according to the diffusion time step to form intermediate diffusion states under different noise levels. Conditional guidance information for the training phase is constructed based on the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active segment mask of the signal, and the geological condition vector. The conditional guidance information of the training phase, the intermediate diffusion state, and the diffusion time step are input into the noise prediction network in the conditional diffusion model to predict noise. The network parameters of the noise prediction network are then optimized through backpropagation to obtain the trained conditional diffusion model.

5. The method for denoising seismic wave signals in tunnel advance detection according to claim 4, characterized in that, The loss function used to optimize the network parameters of the noise prediction network through backpropagation includes: a noise prediction loss term, an amplitude scale constraint term, and an event window structure constraint term. Wherein, the noise prediction loss term is the sample-weighted noise prediction mean square error loss; the weight of the sample weighting is inversely proportional to the signal-to-noise ratio of the sample; The amplitude scale constraint term is used to constrain the amplitude scale consistency between the denoised seismic signal reconstructed based on the noise prediction results of the noise prediction network and the clean superimposed seismic signal. The event window structure constraint terms are determined as follows: one or more reflection event windows are determined based on the signal envelope of the noisy superimposed seismic signal; adjacent or overlapping reflection event windows are merged to generate an event window mask; and the local waveform, local envelope, or local energy ratio loss is weighted based on the event window mask.

6. The method for denoising seismic wave signals in tunnel advance detection according to claim 1, characterized in that, During the inference phase, noise prediction is performed using the trained conditional diffusion model, and a physical consistency gradient correction is applied to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal, which includes: The active signal segment mask for the inference stage is estimated based on the observed seismic wave signals to be denoised; Based on the seismic wave signal to be denoised, the signal active segment mask of the inference stage, the source wavelet corresponding to the seismic wave signal to be denoised, and the geological condition vector, the condition guidance information of the inference stage is constructed. The seismic wave signal to be denoised is subjected to scale correction and Fourier transform to obtain the frequency domain reference of the received signal; Based on the source wavelet corresponding to the seismic wave signal to be denoised, a source spectrum reliability analysis is performed to obtain the comprehensive weight of the source spectrum. At each time step of the back diffusion process, noise prediction is performed based on the condition guidance information of the inference stage, and a physical consistency gradient correction is applied to the back sampling mean in the back diffusion process based on the signal active segment mask of the inference stage, the received signal frequency domain reference, and the source reliability weight, to obtain the denoised seismic wave signal.

7. The method for denoising seismic wave signals in tunnel advance detection according to claim 6, characterized in that, At each time step of the backdiffusion process, noise prediction is performed based on the conditional guidance information of the inference stage, and a physical consistency gradient correction is applied to the back-sampling mean during the backdiffusion process based on the signal active segment mask of the inference stage, the received signal frequency domain reference, and the source reliability weight, to obtain the denoised seismic wave signal, including: At each time step of the back diffusion process, the following denoising steps are performed: The intermediate diffusion state at the current time step and the conditional guidance information of the inference stage are fed into the noise prediction network to perform noise prediction and obtain the noise prediction value. Based on the noise prediction value, the clean seismic signal estimate after denoising of the seismic wave signal to be denoised is calculated, and the frequency domain clean seismic signal estimate is obtained through Fourier transform. A physical guidance gradient is constructed based on the frequency domain clean seismic signal estimation, the signal active segment mask in the inference stage, the received signal frequency domain reference, and the source spectrum comprehensive weight; The physical consistency gradient correction is applied to the backward sampling mean of the current time step based on the physical guidance gradient to obtain the intermediate diffusion state of the next time step. Repeat the denoising steps until the preset iteration termination condition is met to obtain the denoised seismic wave signal.

8. The method for denoising seismic wave signals in tunnel advance detection according to claim 7, characterized in that, A physical guidance gradient is constructed based on the frequency-domain clean seismic signal estimation, the signal active segment mask in the inference stage, the received signal frequency-domain reference, and the source spectrum comprehensive weights, including: Based on the frequency-domain clean seismic signal estimation, the received signal frequency-domain reference, and the source spectrum comprehensive weights, a frequency-domain physical consistency constraint is constructed, including received signal frequency-domain consistency constraints and source spectrum reliability constraints, expressed as: in, For frequency domain physical consistency constraints; For inverse fast Fourier transform, These are frequency-domain clean seismic signal estimation and received signal frequency-domain reference, respectively. The comprehensive weight of the source spectrum, , As the source reliability weight, Confidence level of on-site noise frequency; μ The weak reliability band leakage penalty weight for the seismic source signal; Calculate the gradient of the frequency domain physical consistency constraint with respect to the intermediate diffusion state, the frequency domain physical consistency gradient; The physical guidance gradient is obtained by performing time-domain weighting and clipping on the frequency domain physical consistency gradient based on the signal active segment mask of the inference phase.

9. The method for denoising seismic wave signals in tunnel advance detection according to claim 8, characterized in that, The expression for the physical guidance gradient is: in, To guide the gradient in physics, For the frequency domain physical consistency gradient, For the intermediate diffusion state at the current time step t x t Operators for finding the gradient; For masking the active signal segments during the inference phase; m min The minimum gradient weights are reserved for the inactive regions of the seismic wave signal to be denoised.

10. A seismic wave signal noise reduction system for tunnel advance detection, characterized in that, include: The data acquisition module is used to acquire numerical simulation data for tunnel advance detection and environmental vibration and noise data at the construction site; The sample construction module is used to construct a clean superimposed seismic signal based on the numerical simulation data, and to construct a noisy superimposed seismic signal based on the clean superimposed seismic signal, the environmental vibration noise data, and standard white noise; to extract the signal active segment mask of the noisy superimposed seismic signal; and to construct a geological condition vector based on the numerical simulation data. The model training module is used to train the conditional diffusion model during the training phase, using the clean superimposed seismic signal as the initial data in the forward diffusion process, and the noisy superimposed seismic signal, the source wavelet in the numerical simulation data, the active segment mask of the signal, and the geological condition vector as conditional guidance information. The inference denoising module is used to perform noise prediction through the trained conditional diffusion model during the inference phase, and to apply a physical consistency gradient correction to the back-sampling mean during the back-diffusion process to obtain the denoised seismic wave signal.

Images