A spatio-temporal topology data denoising method, system, device and medium based on multi-dimensional feature mapping and deep heterogeneous network

Abstract

The application provides a kind of based on multi-dimensional feature mapping and deep heterogeneous network's spatiotemporal topology data denoising method, system, equipment and medium, it is related to artificial intelligence and multimodal data processing technical field.This method includes: collecting original spatiotemporal fusion data and performing block processing, obtaining one-dimensional time series signal of each spatial channel;The one-dimensional time series signal of each channel is subjected to time-frequency transform and the mel filter bank is extracted to extract mel cepstrum coefficient features, and a cross-modal time-frequency feature map is constructed;Data acquisition nodes are abstracted as graph nodes and a weighted spatial adjacency matrix is constructed according to the spatial position;The cross-modal time-frequency feature map and the spatial adjacency matrix are jointly input into the spatiotemporal cross-modal deep auto-encoding network for feature decoupling and noise filtering;Inverse time-frequency transform is performed on the denoised feature map, and the denoised spatiotemporal domain signal is reconstructed and output.The application realizes spatiotemporal cross-modal collaborative denoising, has the advantages of low computational delay, complete noise suppression, strong generalization ability, etc.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, multimodal data processing, and deep neural network technology, specifically to a cross-modal intelligent denoising method and system for time-series and spatial geometrically heterogeneous data, and more particularly to a spatiotemporal topology data denoising processing method and system that integrates nonlinear acoustic feature transformation, graph neural networks, and a time-domain-aware autoencoder architecture. This invention decouples the feature redundancy of multidimensional heterogeneous sensing streams in the latent space and can be widely applied to real-time, high-precision optimization processing scenarios for massive and complex heterogeneous data, such as various distributed networked sensing systems and dynamic reconstruction of large-scale topology array signals. Background Technology

[0002] In long-distance, high-density, large-scale distributed continuous monitoring scenarios, dense topological sensing arrays continuously and frequently generate massive amounts of multidimensional heterogeneous sensing streams during operation. This type of data typically possesses both one-dimensional microscopic dynamic temporal attributes and macroscopic irregular spatial topological geometry, exhibiting typical "temporal-spatial geometric heterogeneity" characteristics. In actual physical operating environments, due to complex cross-interference from multiple sources, non-uniform evolution of the transmission medium, and the inherent electronic noise of the sensing channels themselves, the final acquired raw feature matrix generally suffers from three common industry-wide technical bottlenecks: deep modal interleaving, sudden random nonlinear strong noise contamination, and exponential expansion of the global data volume. This significantly obscures the valuable, weak, and transient microscopic target features at the underlying level, restricting the accuracy of subsequent intelligent analysis algorithms such as feature recognition, convergence domain evaluation, and trend prediction.

[0003] Traditional signal processing algorithms (such as bandpass filtering, classical wavelet transform, and empirical mode decomposition (EMD)) mainly rely on artificial prior rules or fixed basis function evolution, and are essentially linear or quasi-linear mappings. When faced with highly nonlinear, non-stationary spatiotemporally interwoven strong noise, these classical methods often fail to achieve deep decoupling of features in the time domain, frequency domain, and topological space dimensions. This often results in severe distortion and loss of weak and effective abrupt features while suppressing background clutter, leading to extremely low feature separation accuracy and fidelity.

[0004] In recent years, although two-dimensional convolutional neural networks (CNNs) or recurrent neural networks based on artificial intelligence have been introduced into the field of information denoising, standard two-dimensional convolution operators, due to the limitation of their regular rectangular local receptive fields, can only process regular grid-like spatiotemporal matrices. They are completely unable to adaptively match and model dynamically changing, asymmetric graph-like spatial nonlinear geometric relationships. In addition, modern distributed topology arrays have extremely high sampling frequencies and extremely dense dynamic channels, and the scale of data generated by the system per day usually reaches TB or even PB levels. Existing single-network deep models generally lack cross-modal collaborative alignment and streaming constraint mechanisms for high-dimensional temporal features and nonlinear spatial topological features, resulting in high overall computational power consumption and excessive inference latency, which cannot meet the requirements of low-latency, streaming, online real-time dynamic feature reconstruction processing.

[0005] In summary, the fields of data processing and intelligent computing currently lack a deep cross-modal collaborative decoupling and noise reduction architecture that can simultaneously accommodate one-dimensional nonlinear dynamic temporal features and multi-dimensional heterogeneous topological features, while also possessing high feature compression ratios and strong spatial topology generalization capabilities. This has become a technical bottleneck restricting the intelligent and in-depth application of massive heterogeneous sensing data. Summary of the Invention

[0006] In view of this, this application provides a spatiotemporal topology data denoising method and system based on multidimensional feature mapping and deep heterogeneous networks. This solution addresses the common technical bottleneck of difficulty in deeply decoupling temporal acoustic features and spatial geometric topology features. First, it transforms and highly compresses massive one-dimensional original time-series sequences into low-dimensional, highly representative cross-modal time-frequency feature maps through nonlinear feature mapping (Mel cepstral analysis), unloading white noise redundancy at the source and significantly reducing the computational load of subsequent networks. Then, it uses graph convolutional networks to adaptively model the spatial geometric associations and nonlinear topological adjacency weights of heterogeneous sensing nodes. Based on this, a spatiotemporal cross-modal deep autoencoder network (TSCM-AE) based on a cross-attention mechanism is constructed, forcibly introducing spatial geometric continuity constraints into the latent space to achieve high-precision decoupling of weak and effective transient signals, multi-scale strong noise suppression, and inverse physical time-domain reconstruction. This invention has significant advantages such as low computational latency, thorough noise suppression, and seamless adaptation to arbitrarily complex topology sensing arrays.

[0007] This application provides the following technical solution: a spatiotemporal topology data denoising method based on multidimensional feature mapping and deep heterogeneous networks, comprising:

[0008] Based on the collected raw spatiotemporal fusion data, a multidimensional spatiotemporal data matrix is ​​constructed. The multidimensional spatiotemporal data matrix includes at least a spatial dimension and a time dimension. The multidimensional spatiotemporal data matrix is ​​divided into blocks along the spatial dimension and the time dimension respectively to obtain independent one-dimensional time-series signals corresponding to each spatial channel. For the one-dimensional time-series signal of each spatial channel, frequency domain transformation, Mel filter bank nonlinear frequency domain mapping and discrete cosine transform are performed sequentially to extract Mel cepstral coefficients; the Mel cepstral coefficients and the corresponding dynamic difference features are combined along the feature dimension to construct the cross-modal time-frequency feature map corresponding to each spatial channel; Each data acquisition node is abstracted as a graph network node. The spatial association weight between nodes is calculated based on the spatial location information of each node, and a weighted spatial adjacency matrix is ​​constructed to represent the spatial geometric heterogeneous topological relationship. The cross-modal time-frequency feature map and the weighted spatial adjacency matrix are used as joint inputs of heterogeneous modalities and fed into a pre-constructed spatiotemporal cross-modal deep autoencoder network. The spatiotemporal cross-modal deep autoencoder network extracts time-dimensional features and spatial-dimensional features through parallel temporal coding branches and spatial coding branches, respectively, and performs feature decoupling and noise reduction on the time-dimensional features and spatial-dimensional features to output a denoised target feature map. The target feature map after noise reduction is subjected to inverse time-frequency transformation, and after signal reconstruction, the noise-reduced spatiotemporal domain signal is output.

[0009] According to one embodiment of this application, the block segmentation operation specifically includes: dividing the multidimensional spatiotemporal data matrix into channels along the spatial dimension to obtain the original time sequence of each spatial channel; and performing overlapping frame segmentation of the original time sequence of each spatial channel along the time dimension with a preset frame length and a preset frame shift to obtain the independent one-dimensional time sequence signal corresponding to each spatial channel.

[0010] According to one embodiment of this application, extracting Mel-frequency cepstral coefficients includes: The one-dimensional time-series signal is sequentially pre-emphasized, framed, and windowed. Perform a short-time Fourier transform on each frame of the processed signal to calculate the power spectrum; A Mel filter bank containing multiple triangular filters is constructed, and the power spectrum is linearly smoothed in the frequency domain by passing it through the Mel filter bank. The logarithmic energy of each Mel frequency domain channel is then calculated. The logarithmic energy is subjected to a discrete cosine transform, and the first n-dimensional low-frequency coefficients are retained as the Mel-frequency cepstral coefficients.

[0011] According to one embodiment of this application, constructing a weighted spatial adjacency matrix includes: Obtain the three-dimensional spatial coordinates of each data acquisition node in the physical space, and calculate the spatial Euclidean distance between any two nodes; Spatial similarity between nodes is calculated based on the Gaussian kernel function, where the bandwidth parameter of the Gaussian kernel function is determined according to the propagation attenuation characteristics of the signal in space. Neighborhood filtering is performed using a preset spatial association radius: when the spatial Euclidean distance is not greater than the spatial association radius, the calculation result of the Gaussian kernel function is used as the spatial association weight between the corresponding node pairs; when the spatial Euclidean distance is greater than the spatial association radius, the spatial association weight between the corresponding node pairs is assigned to zero; the weighted spatial adjacency matrix is ​​constructed from the spatial association weights between all node pairs.

[0012] According to one embodiment of this application, the spatiotemporal cross-modal deep autoencoder network includes: The temporal encoder is composed of multiple stacked temporal convolutional layers and is used to perform temporal convolutional encoding on the cross-modal time-frequency feature map to extract the time-dimensional hidden layer features. The spatial encoder, composed of multiple stacked graph convolutional layers, is used to perform graph convolutional encoding on the weighted spatial adjacency matrix to extract spatial geometric hidden layer features. The cross-modal feature fusion module uses a cross-attention mechanism to perform cross-modal correlation modeling on the temporal hidden layer features and the spatial geometric hidden layer features, and outputs fused hidden layer features; The decoder is used to decode and reconstruct the fused hidden layer features and output the denoised target feature map.

[0013] According to one embodiment of this application, the spatiotemporal cross-modal deep autoencoder network is trained end-to-end using a joint loss function, which includes a temporal fidelity loss term and a spatial continuity constraint loss term. The temporal fidelity loss term is used to constrain the waveform fidelity between the denoised feature map and the clean target feature map. The spatial continuity constraint loss term is used to penalize discontinuous abrupt changes in the reconstructed signal in the spatial neighborhood to suppress isolated spatial noise.

[0014] According to one embodiment of this application, performing an inverse time-frequency transform on the denoised target feature map and outputting a denoised spatiotemporal domain signal after signal reconstruction includes: Perform an inverse discrete cosine transform on the denoised target feature map to recover the log-Mel spectrum energy of each frame; After performing an exponential operation on the logarithmic Mel spectrum energy, the logarithmic Mel spectrum energy is mapped from the Mel frequency domain back to the linear frequency domain using the generalized inverse matrix of the Mel filter bank to obtain the reconstructed amplitude spectrum; The phase spectrum of the original noisy signal is obtained, and the reconstructed amplitude spectrum is combined with the phase spectrum of the original noisy signal to obtain a complex spectrum. Perform an inverse short-time Fourier transform on the complex spectrum to obtain the denoised time-domain framed signal; Based on the frame shift during framing, the time-domain framed signals are overlapped and added to eliminate the windowing effect and splice together to reconstruct a complete multi-channel spatiotemporal signal.

[0015] This application also provides a spatiotemporal topology data denoising system based on multidimensional feature mapping and deep heterogeneous networks, including: The data acquisition and segmentation module is used to acquire raw spatiotemporal fusion data and construct a multidimensional spatiotemporal data matrix. The multidimensional spatiotemporal data matrix includes at least a spatial dimension and a time dimension. The module performs segmentation operations on the multidimensional spatiotemporal data matrix along the spatial dimension and the time dimension respectively to obtain independent one-dimensional time-series signals corresponding to each spatial channel. The cross-modal time-frequency conversion module is used to sequentially perform frequency domain transformation, Mel filter bank nonlinear frequency domain mapping, and discrete cosine transform on the one-dimensional time-series signal of each spatial channel to extract Mel cepstral coefficients; and to combine the Mel cepstral coefficients with the corresponding dynamic difference features along the feature dimension to construct the cross-modal time-frequency feature map corresponding to each spatial channel. The spatial topology graph construction module is used to abstract each data acquisition node into a graph network node, calculate the spatial association weight between nodes based on the spatial location information of each node, and construct a weighted spatial adjacency matrix to represent the spatial geometric heterogeneous topological relationship. The deep spatiotemporal neural network denoising module is used to perform denoising processing with the cross-modal time-frequency feature map and the weighted spatial adjacency matrix as joint inputs of heterogeneous modalities, and output the denoised target feature map. The signal inverse transformation and reconstruction module is used to perform inverse time-frequency transformation on the denoised target feature map and output the denoised spatiotemporal domain signal after signal reconstruction.

[0016] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0017] This application also provides a computer-readable storage medium storing a computer program that performs the above-described methods.

[0018] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: (1) Cross-modal feature compression to break through the bottleneck of data storage and computation: By converting the original massive time series signals into MFCC features, a large amount of useless background white noise redundancy is removed, and the feature data volume is compressed by more than 80%, which greatly reduces the power consumption of training and inference computation of deep learning networks.

[0019] (2) Spatiotemporal co-representation significantly improves the separation accuracy of weak crack signals: This invention combines nonlinear psychoacoustic features (MFCC) in the time dimension with geometric topology maps (GCN) in the spatial dimension, solving the problem that conventional algorithms cannot handle "time-series-spatial geometric heterogeneous data". Even under strong pumping noise with a signal-to-noise ratio (SNR) as low as -15dB, it can still accurately extract weak initial arrival signals of crack rupture.

[0020] (3) Strong generalization ability and adaptability to various complex topologies: The introduction of graph neural network to process spatial geometry makes this method not only applicable to regular arrays (vertical wells), but also applicable to DAS fiber optic networks with various spatial geometric heterogeneous arrangements such as horizontal wells and highly inclined wells.

[0021] (4) End-to-end joint optimization to ensure signal fidelity: By designing a joint loss function with dual constraints of temporal fidelity and spatial continuity, the original effective signal morphological characteristics are preserved to the maximum extent while effectively suppressing noise. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic flowchart of the spatiotemporal topology data noise reduction method according to an embodiment of the present invention; Figure 2 This is a diagram illustrating the overall architecture of the cross-modal spatiotemporal noise reduction method according to an embodiment of the present invention. Figure 3 This is a diagram of the internal structure of the spatiotemporal cross-modal deep autoencoder network according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the spatiotemporal topology data noise reduction system according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the computer device of the present invention. Detailed Implementation

[0024] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0025] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] like Figure 1 As shown, this embodiment of the invention provides a spatiotemporal topology data denoising method based on multidimensional feature mapping and deep heterogeneous networks, including: S101. Based on the collected original spatiotemporal fusion data, a multidimensional spatiotemporal data matrix is ​​constructed. The multidimensional spatiotemporal data matrix includes at least a spatial dimension and a time dimension. The multidimensional spatiotemporal data matrix is ​​divided into blocks along the spatial dimension and the time dimension respectively to obtain independent one-dimensional time-series signals corresponding to each spatial channel. S102. For the one-dimensional time-series signal of each spatial channel, frequency domain transformation, Mel filter bank nonlinear frequency domain mapping and discrete cosine transform are performed sequentially to extract Mel cepstral coefficients; the Mel cepstral coefficients and the corresponding dynamic difference features are combined along the feature dimension to construct the cross-modal time-frequency feature map corresponding to each spatial channel. S103. Abstract each data acquisition node into a graph network node, calculate the spatial association weight between nodes based on the spatial location information of each node, and construct a weighted spatial adjacency matrix to represent spatial geometric heterogeneous topological relationships. S104. The cross-modal time-frequency feature map and the weighted spatial adjacency matrix are used as joint inputs of heterogeneous modalities and fed into a pre-constructed spatiotemporal cross-modal deep autoencoder network. The spatiotemporal cross-modal deep autoencoder network extracts time-dimensional features and spatial-dimensional features through parallel time-domain coding branches and spatial coding branches, respectively, and performs feature decoupling and noise reduction on the time-dimensional features and the spatial-dimensional features to output a denoised target feature map. S105. Perform inverse time-frequency transformation on the denoised target feature map, and output the denoised spatiotemporal domain signal after signal reconstruction.

[0027] This invention provides a cross-modal intelligent denoising technology solution for time-series and spatial geometric heterogeneous data. The overall process consists of five core steps: data preprocessing, cross-modal feature mapping, spatial topology modeling, deep network denoising, and feature inverse reconstruction.

[0028] In specific implementation, step S101 includes: construction and block processing of multidimensional spatiotemporal data matrix. Raw spatiotemporal fusion data is collected, a two-dimensional spatiotemporal feature matrix is ​​constructed, and overlapping block operations are performed along the spatial and temporal dimensions respectively to split it into multiple independent single-channel one-dimensional time series feature sequences.

[0029] In this embodiment, the block segmentation operation specifically includes: dividing the multidimensional spatiotemporal data matrix into channels along the spatial dimension to obtain the original time sequence of each spatial channel; and performing overlapping frame segmentation of the original time sequence of each spatial channel along the time dimension with a preset frame length and a preset frame shift to obtain the independent one-dimensional time sequence signal corresponding to each spatial channel.

[0030] Step S102 includes: extraction of cross-modal time-frequency features from one-dimensional time-series data. A short-time Fourier transform is performed on the one-dimensional time-series sequence of each channel. Combined with the Mel filter bank and discrete cosine transform, Mel cepstral coefficients (MFCCs) are extracted, and first-order and second-order dynamic difference features are added. This transforms the one-dimensional time-series data into a low-dimensional, highly representative cross-modal time-frequency feature map, completing data dimensionality reduction and feature enhancement.

[0031] In this embodiment, the extraction of Mel-Cepstral Coefficients includes: pre-emphasizing the one-dimensional time-series signal to compensate for high-frequency signal attenuation; performing frame segmentation and windowing on the pre-emphasized signal, and applying a window function to suppress spectral leakage; performing a Fast Fourier Transform on each frame of the signal to calculate the power spectrum; constructing a Mel-Cepstral filter bank containing multiple triangular filters, smoothing the power spectrum in the frequency domain through the Mel-Cepstral filter bank, and calculating the logarithmic energy of each Mel-Cepstral frequency channel; performing a Discrete Cosine Transform on the logarithmic energy, retaining the first n low-frequency coefficients as Mel-Cepstral Coefficients, and discarding high-frequency detail coefficients to achieve noise reduction and dimensionality reduction.

[0032] Finally, the first-order and second-order difference features of the Mel-Cepstral Coefficients are calculated, and the Mel-Cepstral Coefficients, the first-order difference features, and the second-order difference features are concatenated along the feature dimension to construct the cross-modal time-frequency feature map.

[0033] In specific implementation, the core of this embodiment for processing spatiotemporally heterogeneous data lies in: using Mel-Cepstral Analysis to reduce and align the one-dimensional nonlinear acoustic time series (the one-dimensional time series signal), using a Graph Convolutional Network (GCN) to adaptively match the spatial geometric topology of the wellbore and fractures, and achieving decoupling and noise reduction of spatiotemporal cross-modal features through a cross-attention mechanism. The specific implementation steps and mathematical mapping relationships are as follows: (1): Nonlinear frequency domain feature mapping of time series features Let the original input two-dimensional DAS data matrix be... Where N is the number of spatial channels, T is the number of sampling points within a single time block, and R represents a matrix. For any single channel One-dimensional time-series signal : ① Pre-emphasis and frame-by-frame windowing: Improve the signal-to-noise ratio of the high-frequency crack rupture initial arrival signal through a high-pass filter:

[0034] Where, x i (t) represents the amplitude of the i-th spatial channel at the original discrete time point t; x i (t 1) indicates that the same spatial channel at the previous adjacent time (t) 1) Amplitude; α is the pre-emphasis coefficient, used to control the intensity of high-frequency boost and compensate for high-frequency attenuation of the signal during transmission; This represents the output signal value of the i-th channel at time t after pre-emphasis processing.

[0035] Then by frame length Frames are divided using a frame shift of S (e.g., 2560 points) and a frame shift of S (e.g., 640 points), and a Hamming window is added to eliminate spectral leakage, resulting in the first frame. Windowed amplitude of frame signal .

[0036] ② Time-frequency power spectrum calculation: Perform the following calculations on each frame of the signal: Point Fast Fourier Transform, calculate its power spectrum :

[0037] Where k is the frequency point, corresponding to the k-th frequency point in the discrete frequency domain after Fast Fourier Transform (FFT); K is the number of FFT transform points, used to control the frequency domain resolution; t is the time domain sampling point; and L represents the frame length. This represents the windowed amplitude of the m-th frame signal at time-domain sampling point t; The twitch factor (complex kernel) represents the Fourier transform, where j is the imaginary unit, used to transform a time-domain signal to the frequency domain.

[0038] ③ Mel frequency domain logarithmic energy mapping: construction A filter bank consisting of triangular filters. Actual physical frequency. With Mel frequency The conversion formula is:

[0039] power spectrum Through Mel filter bank Perform linear frequency domain smoothing and calculate each Mel frequency domain channel. logarithmic energy :

[0040] ④ Discrete Cosine Transform (DCT) and Dynamic Difference: Perform DCT on the logarithmic energy, retaining the previous... Low-frequency coefficients are used to achieve feature dimensionality reduction and remove nonlinear noise redundancy.

[0041] in, Let be the d-th dimension MFCC coefficients of the m-th frame signal; D is the total dimension of the MFCC coefficients. This represents the total number of filters in the Mel filter bank.

[0042] calculate First-order difference ( ) and second-order difference ( (The images are spliced ​​together to form a single-channel dynamic cross-modal time-frequency feature map.) ( (Total number of frames). All channels are stacked to form the global input tensor. .

[0043] Step S103 includes: constructing a spatial geometric topology graph model. Distributed data acquisition nodes are abstracted as graph network nodes. Based on the three-dimensional spatial coordinates of the nodes and the feature propagation law, the node association weights are calculated, a weighted spatial adjacency matrix is ​​constructed, and the spatial geometric heterogeneous relationship modeling is completed.

[0044] In this embodiment, constructing a weighted spatial adjacency matrix includes: obtaining the three-dimensional spatial coordinates of each data acquisition node in the physical space, and calculating the spatial Euclidean distance between any two nodes; calculating the spatial similarity between nodes based on a Gaussian kernel function, wherein the bandwidth parameter in the Gaussian kernel function is determined according to the propagation attenuation characteristics of the signal in space; performing neighborhood filtering with a preset spatial association radius: when the spatial Euclidean distance is not greater than the spatial association radius, the calculation result of the Gaussian kernel function is used as the spatial association weight between the corresponding node pairs; when the spatial Euclidean distance is greater than the spatial association radius, the spatial association weight between the corresponding node pairs is assigned to zero; and constructing the weighted spatial adjacency matrix from the spatial association weights between all node pairs.

[0045] In practice, this includes the process of modeling a nonlinear spatial geometric topology graph: In hydraulic fracturing, the wellbore topology and fracture strain propagate nonlinearly along three-dimensional space. This embodiment will... Each DAS fiber channel is abstracted as a graph node. Construct an undirected weighted graph ,in It is a spatial geometric adjacency matrix.

[0046] ① 3D geometric distance weight calculation: based on the actual 3D spatial coordinates of the DAS channel downhole Calculation channel With channel Euclidean distance between .

[0047] ②Establishing a spatial similarity matrix based on Gaussian kernel

[0048] in To control the hyperparameters of spatial association width, This is the maximum cutoff radius for the spatial propagation of stress waves from the crack. The resulting... The matrix explicitly characterizes the geometric heterogeneity correlation of space.

[0049] Step S104 includes: cross-modal deep network denoising inference. A spatiotemporal cross-modal deep autoencoder network integrating temporal convolution, graph convolution, and cross-attention mechanisms is constructed. The time-frequency feature map and the spatial topological adjacency matrix are used as joint inputs. The network completes multimodal feature decoupling and noise filtering, and outputs a clean target feature map.

[0050] In this embodiment, the spatiotemporal cross-modal deep learning denoising network (i.e., the spatiotemporal cross-modal deep autoencoder network) adopts an autoencoder structure, which internally includes a parallel integrated temporal convolutional network (TCN) branch, a graph convolutional network (GCN) branch, and a cross-attention mechanism for feature fusion between the two branches. The spatiotemporal cross-modal deep autoencoder network specifically includes: The temporal encoder is composed of multiple stacked temporal convolutional layers and is used to perform temporal convolutional encoding on the cross-modal time-frequency feature map to extract the time-dimensional hidden layer features. The spatial encoder, composed of multiple stacked graph convolutional layers, is used to perform graph convolutional encoding on the weighted spatial adjacency matrix to extract spatial geometric hidden layer features. The cross-modal feature fusion module employs a cross-attention mechanism to perform cross-modal association modeling on the temporal hidden layer features and the spatial geometric hidden layer features, and outputs fused hidden layer features. The cross-attention mechanism uses the temporal hidden layer features as the query vector and the spatial geometric hidden layer features as the key vector and value vector, and calculates the attention weight of the temporal dimension features on the spatial dimension features, thereby realizing the guiding constraint of spatial geometric information on temporal features. The decoder is used to decode and reconstruct the fused hidden layer features and output the denoised target feature map.

[0051] In specific implementation, the spatiotemporal cross-modal decoupling and fusion noise reduction of the spatiotemporal cross-modal deep autoencoder network includes the following: The noise reduction network adopts an autoencoder architecture, and its core forward propagation algorithm steps are as follows: Operator 1: Nonlinear time-frequency cross-modal cepstral operator mapping (1D signal feature dimension cascade compression) Input the original high-order matrix: — Extracting a single-channel one-dimensional discrete high-frequency flow: x(t); (1) Time-domain high-frequency compensation domain frame-splitting and windowing operator; (2) The Fourier frequency domain power spectrum is converted into the log-Mel energy operator; (3) Reduction of discrete cosine feature space and dynamic order splicing.

[0052] Operator 2: Operator for modeling topological similarity in non-Euclidean geometric spaces Input node 3D topological real dataset: P={p1,p2,…,p N},p i ∈R 3 ; (1) Calculation of correlation weights based on Gaussian kernel in three-dimensional space nonlinear Euclidean distance.

[0053] Operator 3: Spatiotemporal cross-modal neural network forward autoencoder-decoder propagation term (1) Time-axis causal multi-scale temporal convolutional encoding of feature terms; (2) Spatial geometric virial laplace normalized adaptive graph convolutional coding term; (3) Cross-Attention cross-modal fusion and alignment operator in the bottleneck layer; (4) Inverse transform term of multi-branch joint dual decoder.

[0054] Operator 4: Optimization term of the full dual spatiotemporal lossy joint objective loss function The first term: The loss term strongly constrains the network, ensuring that the time-frequency features restored by the inverse transform remain completely undistorted, maximizing fidelity; The second item: the spatial topology graph adversarial constraint term provides global topological smoothness by penalizing discontinuous mutations and forcibly filtering out isolated bursts of spatial heterogeneous mutation noise.

[0055] In this embodiment, the spatiotemporal cross-modal deep autoencoder network is trained end-to-end using a joint loss function, which includes a temporal fidelity loss term and a spatial continuity constraint loss term. The temporal fidelity loss term is used to constrain the waveform fidelity between the denoised feature map and the clean target feature map. The spatial continuity constraint loss term is used to penalize discontinuous abrupt changes in the reconstructed signal in the spatial neighborhood to suppress isolated spatial noise.

[0056] In practice, the spatiotemporal cross-modal deep autoencoder network is trained using an end-to-end joint loss function optimization. The loss function consists of two parts, ensuring that the denoised features maintain both the acoustic waveform fidelity of a single channel and the spatial geometric continuity across channels:

[0057] in, This represents the total loss value during network training. To predict the mean square error of MFCCs and pure real MFCCs, and to ensure that the time domain morphology is not distorted; To mitigate spatial geometrical resistance loss, sudden isolated spatial noise is forcibly filtered out by penalizing abrupt changes in spatial discontinuities. The MFCC feature map is the output of the network prediction. This is a true MFCC feature map without noise; The weighted spatial adjacency matrix for network prediction reconstruction; It is a real weighted spatial adjacency matrix; and To adjust the hyperparameters of the weights (preferably) ).

[0058] Step S105 includes: feature inverse reconstruction and temporal data restoration. Inverse operations such as inverse discrete cosine transform and inverse Mel filtering are performed on the denoised target feature map, and the signal is reconstructed by combining the original phase information, finally outputting high signal-to-noise ratio original spatiotemporal data.

[0059] In this embodiment, the inverse time-frequency transform is performed on the denoised target feature map, and the denoised spatiotemporal domain signal is output after signal reconstruction. This includes: performing an inverse discrete cosine transform on the denoised target feature map to recover the logarithmic Mel spectrum energy of each frame; performing an exponential operation on the logarithmic Mel spectrum energy, and then using the generalized inverse matrix of the Mel filter bank to map the logarithmic Mel spectrum energy from the Mel frequency domain back to the linear frequency domain to obtain the reconstructed amplitude spectrum; obtaining the phase spectrum of the original noisy signal, and combining the reconstructed amplitude spectrum with the phase spectrum of the original noisy signal to obtain a complex spectrum; performing an inverse short-time Fourier transform on the complex spectrum to obtain the denoised time-domain segmented signal; and performing overlapping and addition processing on each time-domain segmented signal according to the frame shift during segmentation to eliminate the windowing effect and splice and reconstruct a complete multi-channel spatiotemporal domain signal.

[0060] In practice, the denoised pure MFCC matrix is ​​obtained through network inference. Then, the following inverse physical transformation is performed to restore the standard DAS time-domain signal: ①Inverse Discrete Cosine Transform: The previous... MFCC coefficients are padded to zero. In this dimension, perform IDCT to recover the log-Mel spectrum energy of each frame. .

[0061] ② Inverse Mel filtering and exponential restoration: deriving the exponent from the logarithmic energy The generalized inverse matrix of the Mel filter bank is used to map it back to the standard linear frequency domain to obtain the reconstructed amplitude spectrum. .

[0062] ③ Phase Combining and Inverse Fast Fourier Transform: Preserving the Phase Spectrum of the Original Noisy Signal Unchanged, synthesized complex spectrum Perform an IFFT on it to obtain the denoised time-domain framed signal. .

[0063] ④ Overlapping and Adding: Based on the frame shift S during framing, adjacent frames are overlapped and added together to remove windowing effects, and finally, a high signal-to-noise ratio two-dimensional spatiotemporally clean signal matrix for the entire channel is reconstructed. .

[0064] Figure 2 The overall architecture diagram of the cross-modal spatiotemporal noise reduction method is as follows: Figure 2 As shown: (1) Input end (left of the figure): The original two-dimensional DAS spatiotemporal matrix acquired.

[0065] (2) Feature transformation (in the figure): In the cross-modal mapping / feature extraction stage, the original one-dimensional time series is transformed into a two-dimensional Mel-Cepstral coefficient spectrum reflecting nonlinear psychoacoustic features. This step achieves strong-sensory feature representation and significantly compresses the amount of data.

[0066] (3) Core denoising (bottom of the figure): The MFCC feature map and the graph structure reflecting the spatial heterogeneity of optical fiber are jointly input into the spatiotemporal cross-modal deep learning denoising network.

[0067] (4) Output end (right side of the figure): The network outputs a clean denoised MFCC feature, which is then reconstructed / inverse Mel spectrum transform, including IDCT and inverse Mel filtering, to finally obtain a high signal-to-noise ratio time-domain continuous signal.

[0068] Figure 3 This is a diagram of the internal structure of a spatiotemporal cross-modal deep autoencoder network (TSCM-AE), as shown below. Figure 3 As shown: (1) Encoder branch (left side of the figure): The input is divided into two paths. The upper path input MFCC tensor enters the time encoder module to extract the temporal latent features. The lower path fiber geometry structure enters the spatial geometry encoder module to extract the spatial geometric latent features using GCN.

[0069] (2) Bottleneck layer fusion (in the figure): In the cross-modal feature fusion and alignment module, the cross attention mechanism is used to establish a connection between the temporal layer and the spatial geometry layer, so as to achieve collaborative decoupling and noise reduction of different modal features.

[0070] (3) Decoder branch (right side of the figure): The fused denoised latent features are used to restore the pure MFCC through the temporal decoder and to reconstruct the geometric topology through the spatial decoder (to ensure spatial continuity). The final output is the denoised MFCC feature that meets the requirements.

[0071] This embodiment takes the real-time monitoring data processing of DAS in the hydraulic fracturing section of a coal mine as an example. The processing method includes the following steps: 1. Data preparation: Read the raw DAS data during the single-stage fracturing of the working face, with a sampling rate of 10kHz, a spatial channel spacing of 1m, and a total of 4000 channels.

[0072] 2. Feature Transformation (Cross-Modal Mapping): The time window length is set to 256ms (2560 sampling points), and the step size is 64ms. Hamming windowing and FFT transformation are performed on the signal within the window for each channel. Logarithmic energy is calculated using a Mel filter bank consisting of 40 filters, followed by DCT transformation, retaining the first 13 dimensions of MFCC coefficients. Simultaneously, the first and second order differences are calculated, and finally, the time window for each channel is characterized as a... Dimensional cross-modal time-frequency image.

[0073] 3. Graph Adjacency Matrix Construction: Based on the three-dimensional coordinates of the horizontal wellbore trajectory, calculate the actual spatial Euclidean distance between any two channels, and construct the graph adjacency matrix. Spatial topological adjacency matrix .

[0074] 4. Network Inference Denoising: The features from steps 2 and 3 are jointly input into the pre-trained TSCM-AE network. The network uses the GCN operator to filter out isolated spatial random noise unrelated to the wellbore, and the TCN operator to filter out continuous mechanical harmonic noise of the pump unit.

[0075] 5. Inverse Transform Reconstruction: The network outputs a clean 13-dimensional MFCC matrix. The log-Mel spectrum is restored using the inverse discrete cosine transform. Combined with the original phase information, the Overlap-Add algorithm is used to reconstruct a continuous time-domain microseismic signal with 4000 channels. In the reconstructed signal, the initial arrival of the microseismic event at the moment of hydraulic fracturing fracture opening is clearly visible, and the signal-to-noise ratio is improved from -12dB to 18dB.

[0076] like Figure 4 As shown, this application also provides a spatiotemporal topology data denoising system 200 based on multidimensional feature mapping and deep heterogeneous networks, comprising: The data acquisition and segmentation module 201 is used to acquire raw spatiotemporal fusion data and construct a multidimensional spatiotemporal data matrix. The multidimensional spatiotemporal data matrix includes at least a spatial dimension and a time dimension. The module performs segmentation operations on the multidimensional spatiotemporal data matrix along the spatial dimension and the time dimension respectively to obtain independent one-dimensional time-series signals corresponding to each spatial channel. The cross-modal time-frequency conversion module 202 is used to sequentially perform frequency domain transformation, Mel filter bank nonlinear frequency domain mapping, and discrete cosine transform on the one-dimensional time-series signal of each spatial channel to extract Mel cepstral coefficients; and combine the Mel cepstral coefficients with the dynamic difference features corresponding to the coefficients along the feature dimension to construct the cross-modal time-frequency feature map corresponding to each spatial channel. The spatial topology graph construction module 203 is used to abstract each data acquisition node into a graph network node, calculate the spatial association weight between nodes based on the spatial location information of each node, and construct a weighted spatial adjacency matrix to represent the spatial geometric heterogeneous topological relationship. The deep spatiotemporal neural network denoising module 204 is used to perform denoising processing with the cross-modal time-frequency feature map and the weighted spatial adjacency matrix as joint inputs of heterogeneous modalities, and output the denoised target feature map. The signal inverse transformation and reconstruction module 205 is used to perform inverse time-frequency transformation on the denoised target feature map and output the denoised spatiotemporal domain signal after signal reconstruction.

[0077] In specific implementation, the cross-modal time-frequency conversion module 202 includes: A pre-emphasis unit is used to perform high-pass filtering and pre-emphasis on the one-dimensional time-series signal; The framing and windowing unit is used to frame the pre-emphasized signal and apply a window function. The Fourier transform unit is used to perform fast Fourier transform on each frame of signal and calculate the power spectrum; The Mel filter unit, with a built-in Mel filter bank, is used to map the power spectrum into the logarithmic energy of each Mel frequency domain channel; The discrete cosine transform unit is used to extract Mel-frequency cepstral coefficients by performing a discrete cosine transform on the logarithmic energy. The dynamic difference unit is used to calculate the first-order and second-order difference features of the Mel-frequency cepstral coefficients; The feature combination unit is used to combine the Mel cepstral coefficients, first-order difference features, and second-order difference features into the cross-modal time-frequency feature map.

[0078] In specific implementation, the spatial topology graph construction module 203 performs the following operations: acquiring the three-dimensional spatial coordinates of each data acquisition node in the physical space, calculating the spatial Euclidean distance between any two nodes; calculating the spatial similarity between nodes based on the Gaussian kernel function, wherein the bandwidth parameter in the Gaussian kernel function is determined according to the signal propagation attenuation characteristics in space; performing neighborhood filtering with a preset spatial association radius: when the spatial Euclidean distance is not greater than the spatial association radius, the calculation result of the Gaussian kernel function is used as the spatial association weight between the corresponding node pairs; when the spatial Euclidean distance is greater than the spatial association radius, the spatial association weight between the corresponding node pairs is assigned to zero; and constructing the weighted spatial adjacency matrix from the spatial association weights between all node pairs.

[0079] In specific implementation, the spatiotemporal cross-modal deep autoencoder network in the deep spatiotemporal neural network denoising module 204 includes: The temporal encoder is composed of multiple stacked temporal convolutional layers and is used to perform temporal convolutional encoding on the cross-modal time-frequency feature map to extract the time-dimensional hidden layer features. The spatial encoder, composed of multiple stacked graph convolutional layers, is used to perform graph convolutional encoding on the weighted spatial adjacency matrix to extract spatial geometric hidden layer features. The cross-modal feature fusion module uses a cross-attention mechanism to perform cross-modal correlation modeling on the temporal hidden layer features and the spatial geometric hidden layer features, and outputs fused hidden layer features; The decoder is used to decode and reconstruct the fused hidden layer features and output the denoised target feature map.

[0080] In specific implementation, the signal inverse transformation and reconstruction module 205 performs the following operations: performing an inverse discrete cosine transform on the denoised target feature map to recover the logarithmic Mel spectrum energy of each frame; performing an exponential operation on the logarithmic Mel spectrum energy and then using the generalized inverse matrix of the Mel filter bank to map the logarithmic Mel spectrum energy from the Mel frequency domain back to the linear frequency domain to obtain the reconstructed amplitude spectrum; acquiring the phase spectrum of the original noisy signal and combining the reconstructed amplitude spectrum with the phase spectrum of the original noisy signal to obtain a complex spectrum; performing an inverse short-time Fourier transform on the complex spectrum to obtain the denoised time-domain segmented signal; and performing overlapping and addition processing on each time-domain segmented signal according to the frame shift during segmentation to eliminate the windowing effect and splice and reconstruct a complete multi-channel spatiotemporal domain signal.

[0081] In one embodiment, a computer device is provided, such as Figure 5 As shown, it includes a memory 301, a processor 302, and a computer program stored in the memory 301 and executable on the processor 302. When the processor 302 executes the computer program, it implements the above-mentioned spatiotemporal topology data denoising method based on multidimensional feature mapping and deep heterogeneous networks.

[0082] Specifically, the computer device can be a computer terminal, a server, or a similar computing device.

[0083] In this embodiment, a computer-readable storage medium is provided, which stores a computer program that performs the above-described spatiotemporal topology data denoising method based on multidimensional feature mapping and deep heterogeneous networks.

[0084] Specifically, computer-readable storage media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media does not include transient media, such as modulated data signals and carrier waves.

[0085] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A spatiotemporal topology data denoising method based on multidimensional feature mapping and deep heterogeneous networks, applied to a distributed fiber optic acoustic wave sensing system, characterized in that... include: Based on the collected raw spatiotemporal fusion data, a multidimensional spatiotemporal data matrix is ​​constructed. The multidimensional spatiotemporal data matrix includes at least a spatial dimension and a time dimension. Block operations are performed on the multidimensional spatiotemporal data matrix along the spatial dimension and the time dimension, respectively, to obtain independent one-dimensional time-series signals corresponding to each spatial channel. The raw spatiotemporal fusion data is sensor data reflecting the spatiotemporal distribution of physical vibration / acoustic signals collected by a distributed fiber optic acoustic wave sensing system. For the one-dimensional time-series signal of each spatial channel, frequency domain transformation, Mel filter bank nonlinear frequency domain mapping and discrete cosine transform are performed sequentially to extract Mel cepstral coefficients; the Mel cepstral coefficients and the corresponding dynamic difference features are combined along the feature dimension to construct the cross-modal time-frequency feature map corresponding to each spatial channel; Each data acquisition node is abstracted as a graph network node. The spatial association weight between nodes is calculated based on the spatial location information of each node, and a weighted spatial adjacency matrix is ​​constructed to represent the spatial geometric heterogeneous topological relationship. The cross-modal time-frequency feature map and the weighted spatial adjacency matrix are used as joint inputs to the heterogeneous modalities and fed into a pre-constructed spatiotemporal cross-modal deep autoencoder network. The spatiotemporal cross-modal deep autoencoder network extracts temporal and spatial features through parallel temporal and spatial coding branches, respectively, and performs feature decoupling and noise reduction on the temporal and spatial features to output a denoised target feature map. The target feature map after noise reduction is subjected to inverse time-frequency transformation, and after signal reconstruction, the noise-reduced spatiotemporal domain signal is output. The spatiotemporal domain signal is used for seismic exploration data processing and geological structure interpretation.

2. The method according to claim 1, characterized in that, The block segmentation operation specifically includes: dividing the multidimensional spatiotemporal data matrix into channels along the spatial dimension to obtain the original time sequence of each spatial channel; and overlapping and framing the original time sequence of each spatial channel along the time dimension with a preset frame length and a preset frame shift to obtain the independent one-dimensional time sequence signal corresponding to each spatial channel.

3. The method according to claim 1, characterized in that, Extracting Mel-frequency cepstral coefficients, including: The one-dimensional time-series signal is sequentially pre-emphasized, framed, and windowed. Perform a short-time Fourier transform on each frame of the processed signal to calculate the power spectrum; A Mel filter bank containing multiple triangular filters is constructed, and the power spectrum is linearly smoothed in the frequency domain by passing it through the Mel filter bank. The logarithmic energy of each Mel frequency domain channel is then calculated. The logarithmic energy is subjected to a discrete cosine transform, and the first n-dimensional low-frequency coefficients are retained as the Mel-frequency cepstral coefficients.

4. The method according to claim 1, characterized in that, Constructing a weighted spatial adjacency matrix includes: Obtain the three-dimensional spatial coordinates of each data acquisition node in the physical space, and calculate the spatial Euclidean distance between any two nodes; Spatial similarity between nodes is calculated based on the Gaussian kernel function, where the bandwidth parameter of the Gaussian kernel function is determined according to the propagation attenuation characteristics of the signal in space. Neighborhood filtering is performed using a preset spatial association radius: when the spatial Euclidean distance is not greater than the spatial association radius, the calculation result of the Gaussian kernel function is used as the spatial association weight between the corresponding node pairs; when the spatial Euclidean distance is greater than the spatial association radius, the spatial association weight between the corresponding node pairs is assigned to zero; the weighted spatial adjacency matrix is ​​constructed from the spatial association weights between all node pairs.

5. The method according to claim 1, characterized in that, The spatiotemporal cross-modal deep autoencoder network includes: The temporal encoder is composed of multiple stacked temporal convolutional layers and is used to perform temporal convolutional encoding on the cross-modal time-frequency feature map to extract the time-dimensional hidden layer features. The spatial encoder, composed of multiple stacked graph convolutional layers, is used to perform graph convolutional encoding on the weighted spatial adjacency matrix to extract spatial geometric hidden layer features. The cross-modal feature fusion module uses a cross-attention mechanism to perform cross-modal correlation modeling on the temporal hidden layer features and the spatial geometric hidden layer features, and outputs fused hidden layer features; The decoder is used to decode and reconstruct the fused hidden layer features and output the denoised target feature map.

6. The method according to claim 1, characterized in that, The spatiotemporal cross-modal deep autoencoder network is trained end-to-end using a joint loss function, which includes a temporal fidelity loss term and a spatial continuity constraint loss term. The temporal fidelity loss term is used to constrain the waveform fidelity between the denoised feature map and the clean target feature map. The spatial continuity constraint loss term is used to penalize discontinuous abrupt changes in the reconstructed signal in the spatial neighborhood to suppress isolated spatial noise.

7. The method according to claim 1, characterized in that, Performing an inverse time-frequency transform on the denoised target feature map and reconstructing the signal to output the denoised spatiotemporal domain signal includes: Perform an inverse discrete cosine transform on the denoised target feature map to recover the log-Mel spectrum energy of each frame; After performing an exponential operation on the logarithmic Mel spectrum energy, the logarithmic Mel spectrum energy is mapped from the Mel frequency domain back to the linear frequency domain using the generalized inverse matrix of the Mel filter bank to obtain the reconstructed amplitude spectrum; The phase spectrum of the original noisy signal is obtained, and the reconstructed amplitude spectrum is combined with the phase spectrum of the original noisy signal to obtain a complex spectrum. Perform an inverse short-time Fourier transform on the complex spectrum to obtain the denoised time-domain framed signal; Based on the frame shift during framing, the time-domain framed signals are overlapped and added to eliminate the windowing effect and splice together to reconstruct a complete multi-channel spatiotemporal signal.

8. A spatiotemporal topology data denoising system based on multidimensional feature mapping and deep heterogeneous networks, characterized in that, include: The data acquisition and segmentation module is used to acquire raw spatiotemporal fusion data and construct a multidimensional spatiotemporal data matrix. The multidimensional spatiotemporal data matrix includes at least a spatial dimension and a time dimension. Segmentation operations are performed on the multidimensional spatiotemporal data matrix along the spatial dimension and the time dimension, respectively, to obtain independent one-dimensional time-series signals corresponding to each spatial channel. The raw spatiotemporal fusion data is sensor data reflecting the spatiotemporal distribution of physical vibration / acoustic signals acquired by a distributed fiber optic acoustic wave sensing system. The cross-modal time-frequency conversion module is used to sequentially perform frequency domain transformation, Mel filter bank nonlinear frequency domain mapping, and discrete cosine transform on the one-dimensional time-series signal of each spatial channel to extract Mel cepstral coefficients; and to combine the Mel cepstral coefficients with the corresponding dynamic difference features along the feature dimension to construct the cross-modal time-frequency feature map corresponding to each spatial channel. The spatial topology graph construction module is used to abstract each data acquisition node into a graph network node, calculate the spatial association weight between nodes based on the spatial location information of each node, and construct a weighted spatial adjacency matrix to represent the spatial geometric heterogeneous topological relationship. The deep spatiotemporal neural network denoising module is used to perform denoising processing with the cross-modal time-frequency feature map and the weighted spatial adjacency matrix as joint inputs of heterogeneous modalities, and output the denoised target feature map. The signal inverse transformation and reconstruction module is used to perform inverse time-frequency transformation on the denoised target feature map and output the denoised spatiotemporal domain signal after signal reconstruction. The spatiotemporal domain signal is used for seismic exploration data processing and geological structure interpretation.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that performs the method according to any one of claims 1 to 7.

Images