Lightweight microseismic signal seismic phase identification method and system suitable for edge deployment

Abstract

The invention provides a lightweight micro-seismic signal seismic phase identification method and system suitable for edge deployment, and the method comprises the steps: obtaining the waveform data of a micro-seismic signal, and carrying out the preprocessing of the waveform data, and obtaining a micro-seismic waveform data set; constructing a lightweight micro-seismic seismic phase identification model based on wavelet convolution, and training and verifying the constructed model by using the data of the micro-seismic waveform data set to obtain a trained lightweight micro-seismic seismic phase identification model based on wavelet convolution; and inputting waveform data of a micro-seismic signal to be tested into the trained lightweight micro-seismic seismic phase identification model based on wavelet convolution to obtain a micro-seismic seismic phase identification result.

Description

Technical Field

[0001] The present invention belongs to the technical field of microseismic signal phase identification and earthquake early warning, and particularly relates to a lightweight microseismic signal phase identification method and system suitable for edge deployment. Background Technique

[0002] The statements in this part merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] Traditional phase identification methods mainly include algorithms such as STA / LTA (Short-Term Average / Long-Term Average method), AIC (Akaike Information Criterion) method, wavelet transform, etc. These methods rely on artificially set characteristic parameters, are sensitive to noise, and have low identification accuracy and insufficient robustness under complex geological backgrounds or low signal-to-noise ratio conditions. At the same time, traditional algorithms often have difficulty meeting the automatic identification requirements of multi-channel and multi-event signals, resulting in low efficiency in actual engineering applications.

[0004] With the development of deep learning technology, phase identification methods based on neural networks have gradually become a research hotspot. For example, models such as PhaseNet and EQTransformer perform phase identification on microseismic signals in an end-to-end manner and have made remarkable progress. These models are generally based on architectures such as convolutional neural networks, recurrent neural networks, and Transformer, and have achieved high identification accuracy on public datasets. However, such models have a large number of parameters and high computational complexity, require high computing resources, usually rely on GPUs or high-performance servers for training and inference, and are not suitable for edge device deployment scenarios with limited resources.

[0005] As an emerging computing mode, edge computing has broad application prospects in scenarios such as intelligent manufacturing, intelligent mines, and emergency rescue. Edge devices usually have characteristics such as low power consumption, small volume, and strong real-time performance, and can process information in real time near the data source end, reducing latency and improving response speed. However, limited by the computing power and storage resources of edge devices, existing phase identification models are difficult to be directly deployed at the edge. Especially when facing problems such as long time series, low signal-to-noise ratio, and diverse signal types required in actual microseismic monitoring tasks, traditional models are difficult to balance identification accuracy and inference efficiency.

[0006] In addition, the spectral characteristics of microseismic signals and the structural complexity pose higher requirements for the expression ability of the model. Microseismic signals often contain obvious low-frequency and high-frequency characteristics, and the arrival time information of seismic phases is often buried in complex waveform patterns. Extracting effective time-frequency characteristics has become a key difficulty in model design. Some studies have attempted to combine signal processing methods such as wavelet transform and Fourier transform with deep learning models to improve the performance of seismic phase identification. However, most of these hybrid methods are still limited to the server environment and lack systematic consideration of lightweight and deployability.

[0007] In actual deployment, especially in environments such as field mining areas, offshore platforms, and mountain tunnels, the network is unstable or the communication cost is high, which poses strict requirements for the independent operation ability and energy consumption control of equipment. Edge devices with local computing capabilities, such as embedded motherboards and AI acceleration chips, have become important supports for the evolution of seismic monitoring systems towards intelligence, distribution, and real-time. Therefore, designing a lightweight seismic phase identification method and system that not only has good identification performance but also adapts to the constraints of edge deployment has important engineering value and practical significance.

[0008] To sum up, the current seismic phase identification methods have the following main problems: (1) Traditional methods are sensitive to noise and highly dependent on humans, making it difficult to meet the automatic identification requirements in complex scenarios; (2) Although deep learning models have high identification accuracy, they have high computational costs and deployment costs and are not suitable for use in embedded or edge scenarios; (3) Existing methods generally lack systematic optimization of model lightweight design and edge deployment adaptability. Summary of the Invention

[0009] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a lightweight microseismic signal seismic phase identification method suitable for edge deployment, a lightweight seismic phase identification network LSPIN-FFF (Lightweight Seismic Phase Identification Network with Fourier Feature Fusion) model based on Fourier feature fusion, which is used for microseismic signal seismic phase identification, and a seismic phase identification method and system with high identification accuracy, low computational resource consumption, and embeddable deployment to meet the needs of intelligent seismic monitoring and disaster warning systems.

[0010] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0011] In the first aspect, a lightweight microseismic signal seismic phase identification method suitable for edge deployment is disclosed, including:

[0012] Obtain the waveform data of microseismic signals and preprocess the waveform data to obtain a microseismic waveform data set;

[0013] Construct a lightweight microseismic phase identification model based on wavelet convolution. Use the data of the microseismic waveform dataset to train and validate the constructed model to obtain a trained lightweight microseismic phase identification model based on wavelet convolution, and deploy the trained lightweight microseismic phase identification model on edge devices;

[0014] Input the waveform data of the microseismic signal to be tested into the trained lightweight microseismic phase identification model based on wavelet convolution to obtain the microseismic phase identification result;

[0015] Among them, the lightweight microseismic phase identification model based on wavelet convolution includes: a multi-scale Fourier transform module, a lightweight amplitude spectrum feature extraction network, a phase-aware lightweight network, and an amplitude-phase fusion module;

[0016] The multi-scale Fourier transform module is used to perform multi-scale processing on the microseismic signal to generate signal segments at multiple time scales, convert the signal segments at different time scales into complex frequency spectra, and decompose the complex frequency spectra into amplitude spectra and phase spectra;

[0017] The lightweight amplitude spectrum feature extraction network processes the input amplitude spectrum to obtain lightweight amplitude spectrum features;

[0018] The phase-aware lightweight network processes the input phase spectrum to obtain phase features;

[0019] The feature fusion unit is used to receive lightweight amplitude spectrum features and phase spectrum features and perform feature fusion processing.

[0020] As a further technical solution, the multi-scale Fourier transform module includes: a multi-scale construction unit and a spectrum feature extraction unit;

[0021] The multi-scale construction unit is used to perform multi-scale processing on the original microseismic signal, and generate signal segments at three time scales by downsampling the original time series signal to retain the phase features within different scale ranges;

[0022] The spectrum feature extraction unit includes multiple Fourier transform units, and each Fourier transform unit acts on the signal segments at different scales respectively, and is used to convert the signal at the corresponding scale into a complex frequency spectrum, and decompose the complex frequency spectrum into an amplitude spectrum and a phase spectrum; the amplitude spectrum is used to reflect the energy distribution characteristics of the phases at different scales, and the phase spectrum is used to reflect the structural change information of the microseismic signal; the amplitude spectra at multiple scales are input into the lightweight amplitude feature extraction network, and the phase spectra at multiple scales are input into the phase-aware lightweight network, respectively for subsequent phase identification tasks.

[0023] As a further technical solution, the lightweight amplitude spectrum feature extraction network includes: an encoding module, a decoding module, and a modulation enhancement module;

[0024] The encoding module includes a plurality of cascaded encoding units. Each encoding unit includes a combined structure of a convolution operation with a set stride, instance normalization, and a non-linear activation function, which is used to gradually extract the deep features of the input amplitude spectrum. Preferably, two layers of 1×3 convolutions are set in each encoding unit, both of which are equipped with instance normalization and activation functions to ensure the stability and lightweight of feature extraction.

[0025] The decoding module is symmetric to the encoding module in structure and is used to perform upsampling reconstruction on the encoded features to restore the detailed information of the seismic phase region. Each decoding unit in the decoding module includes a combined structure of a bilinear upsampling operation, a spatial adaptive normalization module, and a 1×3 convolution operation. The spatial adaptive normalization module is used to dynamically adjust the normalization method of the lower-layer features according to the upper-layer feature responses, thereby enhancing the model's structural perception ability of the seismic phase region.

[0026] The modulation enhancement module is used to introduce the original amplitude spectrum features into the network backbone to enhance the response of the seismic phase-related features. This module includes: three cascaded 1×3 convolution operations, batch normalization, and non-linear activation, which are used to extract low-level amplitude local features. The extracted amplitude features are respectively generated into modulation factors α and β through two independent convolutions, which are respectively used for element-wise multiplication and element-wise addition of the upstream features to achieve significant enhancement of the backbone features.

[0027] As a further technical solution, the phase-aware lightweight network includes: a sine-cosine encoding module, a stitching window module, a shape-changing module, a linear mapping module, a low-rank projection module, a scaled dot-product attention mechanism module, a stitched multi-head module, and a linear layer output module, where:

[0028] The sine-cosine encoding module is used to receive the spectral results output by the multi-scale Fourier transform module, extract the phase information of the complex spectra at each scale, and then perform sine and cosine encoding respectively to obtain two-dimensional phase structure features. This processing method can enhance the expression stability of the periodic jump features in the phase spectrum and improve the network's recognition ability of microseismic phases.

[0029] The stitching window module is used to stitch the encoded phase features at each scale along the window dimension to form a four-dimensional tensor in a unified format, ensuring the integrated input of multi-scale information.

[0030] The shape-changing module performs a dimensional transformation on the above stitching result, adjusting the four-dimensional tensor into a three-dimensional input tensor adapted to the Transformer structure to meet the requirements of the subsequent structure modeling module for the input format.

[0031] The linear mapping module is used to perform a linear mapping on the dimension of the phase feature channels, mapping low-dimensional features to a higher-dimensional feature space. Optionally, a learnable position encoding can be added to enhance the model's perception ability of the phase sequence structure in the time series.

[0032] The low-rank projection module, as a core component of the Linformer structure, compresses and maps the key-value vectors, reduces the computational complexity of the attention mechanism, and improves the model efficiency and adaptability for edge deployment.

[0033] The scaled dot-product attention mechanism module adopts a simplified Transformer structure, performs efficient attention calculations through the compressed key-value vectors and the query vector, and combines a feed-forward network and a normalization layer to achieve in-depth modeling of the input sequence.

[0034] The concatenated multi-head module is used to concatenate and fuse the output results of multiple attention heads to enhance the model's feature expression ability.

[0035] The linear layer output module is used to output the phase classification probability corresponding to each time point, which serves as an important basis for phase type recognition and arrival time point positioning.

[0036] This module can effectively model the phase structure changes in the phase spectrum, and has the characteristics of high efficiency, light weight, and deployment friendliness, and is suitable for phase recognition and time series modeling tasks in microseismic signals.

[0037] As a further technical solution, the amplitude-phase fusion module includes: a feature fusion unit, a time-domain reconstruction unit, and a reconstructed feature extraction unit;

[0038] The feature fusion unit is used to receive the amplitude spectrum features output by the amplitude feature extraction network and the phase spectrum features output by the phase perception lightweight network, and perform joint expression of the two in the frequency domain; among them, the amplitude spectrum features represent the frequency energy distribution of the microseismic signal, and the phase spectrum features represent the phase structure information. The two are combined into a complex spectrum in the frequency domain through a complex exponential form;

[0039] The time-domain reconstruction unit is used to perform an inverse Fourier transform on the fused complex spectrum to restore it to a time-domain signal containing energy and structure information, so as to retain phase jumps and local features; preferably, the inverse Fourier transform operation is to perform an inverse Fourier transform on the complex signal combined by the amplitude spectrum and the phase spectrum along the time dimension;

[0040] The reconstruction feature extraction unit includes a set of convolution operation modules, which perform multi-scale convolution extraction on the reconstructed time-domain signal to capture temporal features such as abrupt edges and continuous undulations in the time-domain seismic phase waveform; the output of the convolution module is used as the final output of the network of the present invention to realize the automatic recognition of seismic phases or the prediction of arrival time points in microseismic signals.

[0041] As a further technical solution, the preprocessing of waveform data includes cropping, translation, filtering, and normalization of microseismic waveform data.

[0042] In a second aspect, a lightweight microseismic signal seismic phase recognition system suitable for edge deployment is disclosed, including:

[0043] A microseismic waveform dataset construction module, configured to: obtain waveform data of microseismic signals, and perform preprocessing on the waveform data to obtain a microseismic waveform dataset;

[0044] A lightweight microseismic seismic phase recognition model construction module, configured to: construct a lightweight microseismic seismic phase recognition model based on wavelet convolution, train and verify the constructed model using the data in the microseismic waveform dataset to obtain a trained lightweight microseismic seismic phase recognition model based on wavelet convolution, and deploy the trained lightweight microseismic seismic phase recognition model based on wavelet convolution on edge devices;

[0045] Among them, the lightweight microseismic seismic phase recognition model based on wavelet convolution includes: a multi-scale Fourier transform module, a lightweight amplitude spectrum feature extraction network, a phase-aware lightweight network, and an amplitude-phase fusion module;

[0046] The multi-scale Fourier transform module is used to perform multi-scale processing on microseismic signals to generate signal segments at multiple time scales, convert the signal segments at different time scales into complex frequency spectra, and decompose the complex frequency spectra into amplitude spectra and phase spectra;

[0047] The lightweight amplitude spectrum feature extraction network processes the input amplitude spectrum to obtain lightweight amplitude spectrum features;

[0048] The phase-aware lightweight network processes the input phase spectrum to obtain phase features;

[0049] The feature fusion unit is used to receive lightweight amplitude spectrum features and phase spectrum features and perform feature fusion processing;

[0050] An identification module, configured to: input the waveform data of the microseismic signal to be tested into the trained lightweight microseismic seismic phase recognition model based on wavelet convolution to obtain a microseismic seismic phase recognition result.

[0051] The above one or more technical solutions have the following beneficial effects:

[0052] The lightweight microseismic signal phase identification system proposed by the present invention takes frequency-domain features as the core, fuses amplitude and phase information, and solves the problems that traditional convolutional networks rely on time-domain local structures and are difficult to capture sudden changes in phase structures; by constructing spectral features at multiple time resolutions through multi-scale Fourier transform, the multi-scale perception ability of the model is effectively enhanced.

[0053] In the system of the embodiment of the present invention, the amplitude feature extraction network adopts a lightweight convolutional structure, combined with the original amplitude modulation mechanism, which greatly reduces the computational cost while ensuring the ability to extract high-frequency spectral features; the phase modeling branch introduces a lightweight Transformer (Linformer) structure, breaking through the problem of weak phase spectrum modeling ability of conventional CNNs and improving the expression accuracy of the model for phase structure changes.

[0054] The present invention performs inverse Fourier transform after frequency-domain fusion and extracts the final features in the time domain, establishing an effective mapping from the frequency-domain structure to the time-domain change, enabling the model to have high fitting ability and physical interpretability for real microseismic signals.

[0055] The entire model has the characteristics of few parameters, fast inference speed, and high recognition accuracy, and is especially suitable for the rapid identification and early warning applications of microseismic phases under the conditions of edge device deployment, which can significantly improve the real-time performance and stability of phase identification. The advantages of the additional aspects of the present invention will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present invention. Brief Description of the Drawings

[0056] The specification drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0057] Figure 1 It is a schematic structural diagram of a lightweight microseismic signal phase identification system suitable for edge deployment provided by an embodiment of the present invention, including a multi-scale Fourier transform module, an amplitude feature extraction branch, a phase feature modeling branch, and a fusion output module;

[0058] Figure 2 It is a schematic structural diagram of the amplitude spectrum feature extraction branch, showing the specific connection relationship between the encoding-decoding structure and the modulation enhancement module. Detailed Description of the Embodiment

[0059] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0060] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0061] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0062] Embodiment 1

[0063] See the attached Figure 1 As shown, this embodiment discloses a lightweight microseismic signal phase identification method for edge deployment that integrates Fourier information, including:

[0064] Step 1: Collect the waveform data of the microseismic signal, and preprocess the waveform data to obtain a microseismic waveform data set; the waveform data of the microseismic signal refers to the original data of the microseismic waveform; the length of the waveform data of the microseismic signal is 60s.

[0065] Among them, preprocessing the waveform data means performing cropping, translation, filtering, and normalization on the microseismic waveform data; after the microseismic waveform data is cropped, the length becomes 30s, and the sampling rate is 100HZ.

[0066] Step 2: Divide the preprocessed microseismic data into a training set, a validation set, and a test set, and preferably the division ratio is 8:1:1. Each piece of data is equipped with labeled P-wave, S-wave arrival time points or phase probability labels, and the labels are uniformly encoded in the format of a time series probability distribution with a length of 3000.

[0067] Step 3: Build a lightweight microseismic phase identification model that integrates Fourier information.

[0068] The specific model in this step adopts a branch structure and is composed of a multi-scale Fourier transform module, a lightweight amplitude spectrum feature extraction network, a phase-aware lightweight network, and an amplitude-phase fusion module.

[0069] Step (3-1): Build a multi-scale Fourier transform module for extracting the frequency domain representation of the microseismic signal, and its structure includes: a multi-scale construction unit and a spectrum feature extraction unit; it can realize the spectrum modeling of the original seismic waveform at different time scales, and is used to enhance the model's perception ability of the phase energy structure and jump characteristics.

[0070] The multi-scale Fourier transform module specifically includes the following parts:

[0071] Multi-scale construction unit: used to receive the original seismic waveform data, and generate sub-waveform sequences at multiple time resolutions based on the time window scaling method to form a multi-scale input. Each scale corresponds to a specific downsampling or truncation strategy, so that the waveforms at different scales can cover the short-term mutation and long-term energy distribution characteristics of the phase. Let the original seismic signal be: Perform sliding window cropping and scaling on it to construct M scales:

[0072]

[0073] where r m represents the scaling / downsampling rate of the m-th scale.

[0074] Spectrum feature extraction unit: Perform Fourier transform operations on the sub-waveforms at each scale respectively:

[0075]

[0076] Generate the corresponding complex spectrum representation using the above transformation operations, retaining the overall frequency domain structure information of the signal. This transformation step ensures that the model can capture the energy and phase structure from the frequency domain perspective. And further decompose each complex spectrum into two parts: the phase spectrum and the amplitude spectrum; among them, the phase spectrum is used to reflect the periodic changes and jump information in the seismic phase structure, and the amplitude spectrum is used to represent the energy distribution law under different frequency components. The phase spectra and amplitude spectra at all scales will be sent to the corresponding feature modeling branches respectively for structure modeling and seismic phase extraction.

[0077] The multi-scale Fourier transform module can enrich the time-frequency expression of the input features and improve the subsequent network's ability to capture seismic phase features by jointly using the spectrum analysis under different time windows, providing a multi-scale frequency domain basis for amplitude and phase modeling.

[0078] Step (3-2): Construct a lightweight amplitude spectrum feature extraction network for extracting the frequency energy information of the amplitude spectrum. See the appendix Figure 2 As shown, its network structure includes: an initial input module, an encoding module (this module is based on downsampling), a decoding module (this module is based on upsampling), an output module, and a modulation enhancement module based on amplitude modulation. The network structure adopts a symmetric design, with good seismic phase positioning ability and adaptability for edge deployment.

[0079] This branch specifically includes the following parts:

[0080] Input module: It is composed of a 1×1 convolutional layer, an instance normalization layer, and a rectified linear activation function, performing channel mapping and preliminary normalization on the amplitude spectrum features.

[0081] Encoding module: This module is composed of four cascaded feature extraction units. Each feature extraction unit contains: a 1×3 convolution (stride of 2) for spatial downsampling; two subsequent 1×3 convolutions for multi-scale energy feature extraction; after all convolution operations, an instance normalization layer and a rectified linear activation function are connected.

[0082] Decoding module: Corresponding to four upsampling units, each unit includes: a bilinear interpolation upsampling module for restoring the spatial resolution; two 1×3 convolution operations for fusing high-level features during the decoding process; after each convolution operation, spatial adaptive normalization and a rectified linear activation function are connected; meanwhile, in each level of upsampling, feature maps corresponding to the corresponding stage of the encoding layer are introduced, and feature fusion is performed with the current layer through skip connection.

[0083] Output module: Consisting of a 1×3 convolutional layer and a 1×1 convolutional layer, it is used to output the amplitude spectrum feature map with spatial localization ability after restoration.

[0084] In addition, to enhance the sensitivity of the model to the amplitude energy region, a modulation enhancement module based on amplitude modulation is introduced, which specifically includes:

[0085] Using the intermediate features in the backbone network as input, first passing through batch normalization as the backbone features;

[0086] Inputting the original amplitude spectrum obtained by Fourier transform into a modulation generation network containing 1×3 convolution and rectified linear activation function to generate two modulation factors α and β;

[0087] Multiplying α with the backbone features element-wise and adding β to the multiplication result element-wise to obtain the modulated enhanced features;

[0088] The modulation enhancement structure is used to highlight the high-response region of seismic amplitude and improve the accuracy of identifying the phase boundary.

[0089] Step (3-3): Construct a lightweight phase-aware network for extracting phase spectrum structure features, specifically modeling the jump structure characteristics of the frequency-domain phase information of seismic signals. This network is constructed based on a simplified variant of Transformer, Linformer, which is suitable for efficient edge deployment. The structure includes: a sine-cosine encoding module, a splicing window module, a shape-changing module, a linear mapping module, a low-rank projection module, a scaled dot-product attention mechanism module, a splicing multi-head module, and a linear layer output module.

[0090] Sine-cosine encoding module: Using the phase spectrum obtained by Fourier transform as input, where φ(f) is the phase spectrum, representing the phase distribution of the signal in the frequency domain, with dimensions of where B is the batch size and L is the length of the frequency-domain sequence. It is converted into a two-dimensional input representation using the sine-cosine encoding method:

[0091]

[0092] where X phase-encis the phase sine-cosine coding result. The phase spectrum is converted into a two-dimensional input representation through sin(·) and cos(·), and the dimension is

[0093] Then the splicing window module slices the phase spectrum into segments (window processing) and splices them along the time dimension to obtain a complete sequence:

[0094]

[0095] where X windowed is the tensor after the sliding window slicing and time splicing, and the dimension is where N win is the number of windows, and L win is the length of each window.

[0096] The shape-changing module reshapes and converts it into the input format:

[0097]

[0098] where X flat is the two-dimensional vector sequence obtained after reshaping, and the dimension is where L total = N win ·L win .

[0099] The linear mapping module uses a linear transformation to map the two-dimensional encoded vector to a high-dimensional representation space:

[0100]

[0101] where, X in is the high-dimensional representation after the linear transformation, and the dimension is d model is the model dimension.

[0102] The Linformer encoder structure contains two layers of lightweight attention structures, and each layer includes a low-rank projection attention mechanism and a feed-forward network.

[0103] Construct Q / K / V:

[0104] Q = X input W Q , K = X input W K , V = X input W V

[0105] X input is the input of the attention mechanism, as the original sequence for generating Q, K, and V, and the projection matrices of the query (Q), key (K), and value (V) vectors.

[0106] In the above low-rank projection module, the projection matrix:

[0107]

[0108] where E K , E V is a low-rank projection matrix for projecting and compressing the sequence dimension L of K and V to k, where k << L.

[0109] The specific scaled dot-product attention in the above scaled dot-product attention mechanism module:

[0110]

[0111] where is the output of the scaled dot-product attention.

[0112] Residual connections and normalization are added to each layer:

[0113] X attn = LayerNorm(Q + Attention)

[0114] where X attn is the result of normalization after the first-layer residual connection.

[0115] X out = LayerNorm(X attn + FFN(X attn ))

[0116] where X out is the result of the second-layer residual connection and normalization. FFN is the output of a feed-forward neural network (Feed-Forward Network), usually consisting of two linear transformations plus activation, with the dimension remaining unchanged.

[0117] Concatenate the multi-head module and the linear layer output module: After concatenating the outputs of multiple heads, perform a linear transformation again:

[0118] X final = Concat(head1,..., head h )W O

[0119] where head i is the output of the i-th attention head. Usually, each dimension has a total of h heads. W O is the linear transformation matrix that maps the concatenated multi-head output back to the model dimension. Concat concatenates all attention heads along the last dimension. X final is the final output of the module.

[0120] Step (3-4): Construct an amplitude-phase fusion module for integrating frequency-domain structural features and realizing seismic phase output. Its structure includes: a feature fusion unit, a time-domain reconstruction unit, and a reconstructed feature extraction unit. The amplitude-phase fusion module fuses phase structure information and amplitude energy information, restores the time-domain signal representation with real physical meaning, and is used for the final seismic phase type discrimination and arrival time point positioning.

[0121] The above-mentioned feature fusion unit: is used to receive the amplitude spectrum feature and the phase spectrum feature, and perform frequency-domain fusion in complex form. The amplitude spectrum and the phase spectrum are encoded and then reconstructed into a complex frequency spectrum:

[0122]

[0123] Among them, A(f) is the amplitude spectrum (amplitude spectrum), which is obtained by processing the modulus value of the original signal through Fourier transform and reflects the frequency-domain energy intensity of the signal. φ(f) is the phase spectrum, which is obtained by taking the angle after Fourier transform and reflects the phase information of the frequency-domain structure. The reconstructed complex frequency spectrum: the complex frequency spectrum expression obtained by combining A(f) with the sine and cosine phase spectra.

[0124] Construct a complete complex frequency spectrum for frequency-domain reconstruction.

[0125] Time-domain reconstruction unit: Perform an inverse Fourier transform operation on the constructed complex frequency spectrum to convert it into a time-domain waveform with physical time sequence meaning:

[0126]

[0127] Among them, The reconstructed time-domain signal, which is used for the subsequent time-domain feature extraction module, is the inverse Fourier transform.

[0128] Reconstructed feature extraction unit: Input the restored time-domain signal into a lightweight convolutional network to extract its seismic phase jump features and structural patterns. The modeling structure includes multiple one-dimensional convolutional layers, and each layer includes:

[0129] DSConv1D→InstanceNorm→LeakyReLU.

[0130] DSConv1D is Depthwise SeparableConv1D, which is used to reduce the number of parameters and computational overhead. InstanceNorm is instance normalization, which performs mean and variance normalization independently for each sample to improve training stability. LeakyReLU is a leaky rectified linear activation function.

[0131] Seismic phase output module:

[0132] Used to output the final seismic phase recognition results, including the classification probability or arrival time index corresponding to each time point. Two output forms are supported:

[0133] Classification output (multi-class Softmax):

[0134] P(t) = Softmax(Conv 1×1 (F time ))。

[0135] P(t) is the classification probability distribution at each time point.

[0136] Regression output (seismic phase arrival prediction):

[0137] T arrival = Linear(F time )

[0138] T arrival is the regression value of the seismic phase arrival time point. The Linear() linear layer acts on the time domain features to extract arrival time information, which can be understood as attention weighted average or time point score integration.

[0139] This module can further compress features, remove noise, and highlight the variation characteristics of the seismic phase region. This module is directly connected to the final inference layer and serves as the core output link for seismic phase recognition and time series judgment in the system.

[0140] Step 4: Use the training set data to train the constructed model. The training uses the Pytorch framework, the optimizer is Adam, and the loss function is MSELoss. The initial learning rate is 0.0001, and it is set to terminate training early whenever the validation set loss does not decrease within 440 consecutive Epochs. The maximum number of epochs is set to 500.

[0141] Step 5: Input the data of the test set into the trained lightweight microseismic phase recognition model LWCU-Net to obtain the microseismic phase recognition results.

[0142] In this embodiment, the design of the lightweight seismic phase recognition model integrating Fourier information is based on the following three core principles: enhancing the receptive field and structural modeling ability without increasing the model complexity; effectively extracting and expressing the spectral features of microseismic signals at a low computational cost; enhancing the model's ability to capture the seismic phase boundary structure, improving the model's robustness, and reducing the risk of overfitting.

[0143] To achieve the first goal, the present invention proposes a multi-scale Fourier transform module, which constructs the original time-domain microseismic signal into subsequences at multiple time scales, performs fast Fourier transform on them respectively, and extracts their amplitude spectra and phase spectra. Multi-scale modeling not only enhances the adaptability to the durations of different seismic phases, but also expands the equivalent receptive field of the model in the frequency domain.

[0144] To achieve the second goal, a dual-branch structure is proposed for amplitude and phase decoupling modeling. Among them, the amplitude spectrum branch uses a lightweight convolutional network to extract energy distribution features, which has a simple structure, few parameters, fast calculation, and is suitable for edge deployment; the phase spectrum branch uses Linformer for structural modeling, which effectively reduces the computational complexity of the attention mechanism while maintaining the long-sequence modeling ability.

[0145] To achieve the third goal, after the amplitude and phase features are fused, complex spectrum reconstruction is performed, and the signal is restored to the time domain through inverse Fourier transform, and then input into the lightweight convolution module to extract local mutation features. This method not only enhances the sensitivity of the model to the structural jump points of P waves / S waves, but also improves the physical interpretability and generalization ability of the model, and effectively alleviates the overfitting phenomenon.

[0146] The lightweight seismic phase recognition model integrating Fourier information of the present invention is applied to the automatic recognition task of the arrival times of P waves and S waves in microseismic signals, and has the advantages of high precision, strong robustness, and low calculation, and is suitable for being deployed in engineering environments such as mining areas, seismic stations, and tunnel monitoring where high real-time performance and accuracy requirements for seismic phase picking are required.

[0147] Verification example:

[0148] (1) Collect the waveform data of microseismic signals, and preprocess the waveform data to obtain a microseismic waveform dataset; the length of the waveform data of microseismic signals is 60s; where the preprocessing refers to cropping, translation, filtering, and normalization processing of the seismic waveform data; after the microwave waveform data is cropped, the length becomes 30s, and the sampling rate is 100HZ.

[0149] The seismic waveform dataset of this embodiment uses the data of the Shanxi monitoring system from 2011 to 2017 as the test microseismic dataset. The monitoring data comes from 9 stations arranged in coal mines, and the monitoring equipment is mainly broadband three-component seismographs. The spatial range of the 9 stations is about 6km×4km×1km, and the original sampling frequency is 5kHz. The paper selects 9459 microseismic waveforms from the monitoring data as the original data for model testing, and determines the confidence probability labels by comprehensively considering the picking results provided by multiple experts. The dataset includes clean microseismic waveforms, low signal-to-noise ratio waveforms, and microseismic waveforms with difficult-to-identify seismic phases. The length of the microseismic data is 60s. Figure 1 For the microseismic waveform data and its labels.

[0150] (1.1) Crop the 60s-long data to 30s-long (sampling rate 100Hz, each segment is 3000 in length), ensuring that the 30S-long data contains microseismic events.

[0151] (1.2) Preprocess the waveform data, including removing DC drift, Min-Max normalization, sliding window segmentation, and random augmentation; finally, construct a standardized microseismic waveform dataset.

[0152] (2) Divide the microseismic waveform dataset obtained in step (1) into a training set, a validation set, and a test set:

[0153] (2.1) Adopt an 80% / 10% / 10% division ratio, where the training set and the validation set together have 7658 samples, and the test set has 1801 samples.

[0154] (2.2) Convert the label into a time series probability distribution with a length of 3000 for the regression-based seismic phase arrival time prediction output.

[0155] (3) Construct a lightweight microseismic phase identification model that integrates Fourier information, including: a multi-scale Fourier transform module, an amplitude spectrum convolution feature extraction branch, a phase spectrum Transformer feature modeling branch, and a frequency domain reconstruction fusion and time domain output module. This model comprehensively utilizes the energy distribution of the amplitude spectrum and the structural mutation information of the phase spectrum to enhance the model's ability to identify microseismic phase features.

[0156] (4) Use the training set to train the model. The training is implemented based on the PyTorch framework. The loss function uses the mean squared error (MSELoss), the optimizer uses Adam, the initial learning rate is 0.0001, and a learning rate decay strategy is set; if the loss does not decrease for 440 consecutive Epochs, the training stops early, and the maximum number of training epochs is 500.

[0157] (5) Input the validation set into the trained model for validation, and evaluate its seismic phase identification robustness and time series fitting ability on data with different signal-to-noise ratios.

[0158] (6) Input the test set into the model and output the prediction results of the P-wave, S-wave, and background categories in each segment of data. To evaluate the model performance, six indicators, namely precision, recall, F1-score, mean absolute error (MAE), number of model parameters (Param), and inference time (Time), are used for comparison. The definitions of the relevant indicators are as follows:

[0159]

[0160] Wherein, TP is the number of times the arrival time of microseismic P / S waves is picked up, FP is the number of times non-P / S waves are mispicked as the arrival time of microseismic P / S waves, and FN is the number of times the arrival time of microseismic P / S waves is missed. A higher recall rate means fewer missed arrival times of microseismic P / S waves; a higher precision also means fewer missed arrival times of microseismic P / S waves. Since recall rate and precision are not compatible, the F1 score performs a weighted harmonic average of the two. The average time delay is the difference between the picked-up arrival time point and the actual arrival time point.

[0161] Table 1 Test results of the seismic phase identification model

[0162]

[0163] As can be seen from Table 1, the proposed model has high identification accuracy and robustness for both P-waves and S-waves, and the mean absolute error is less than 0.16 seconds, meeting the accuracy requirements for seismic phase identification.

[0164] Local deployment test

[0165] To verify the applicability of this model in the actual task of picking up the arrival time of microseismic P / S waves and ensure that it can accurately pick up the starting points of P-waves and S-waves, this model is deployed on a Jetson Nano processor for testing. Test process: After the processor receives the waveform, it adjusts the waveform size, crops it point by point to match the requirements, and then inputs the waveform into the model to obtain the result of picking up the arrival time of microseismic P / S waves.

[0166] To meet the actual requirements, TensorRT is used for accelerated compilation and inference testing. The model is converted into the.engine format, and the response speeds under float32 and float16 precisions are tested respectively. The results are shown in Table 2. There is no significant difference in the picking performance between the two modes. However, on the premise of maintaining picking accuracy, using the float16 precision mode has a significant computational advantage in processing speed.

[0167] Table 2 Results of actual tests

[0168]

[0169] The results show that the model deployed using TensorRT has an average inference time of less than 10 ms under float16 precision, meeting the real-time requirement (<10 ms) of the seismic phase identification system, providing engineering feasibility support for deployment on edge devices (such as portable seismographs and intelligent monitoring terminals in mining areas).

[0170] In summary, the lightweight seismic phase identification model integrating Fourier information described in the present invention exhibits good performance in terms of accuracy, speed, and deployability, and has the engineering application value of real-time identifying the arrival time points of P / S waves in an actual microseismic monitoring system.

[0171] Example Two

[0172] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it is used to implement all steps of the above-mentioned lightweight microseismic signal phase identification method applicable to edge deployment. The method includes: obtaining mine seismic signal waveform data, performing multi-scale construction and Fourier transform, and extracting the amplitude spectrum and phase spectrum; inputting the amplitude spectrum into a lightweight amplitude feature extraction network, inputting the phase spectrum into a phase-aware lightweight network, after frequency domain fusion and inverse Fourier transform, then using a convolutional network to extract final features and output the phase identification result.

[0173] Example Three

[0174] The purpose of this embodiment is to provide a computer-readable storage medium. A computer-readable storage medium stores a computer program, which when executed by a processor, executes the method steps described in Example One, including: extracting multi-scale spectral information from the original microseismic signal, separately modeling and fusing the amplitude and phase, and further realizing the functions of phase identification and arrival time point extraction.

[0175] Example Four

[0176] The purpose of this embodiment is to provide a lightweight microseismic signal phase identification system applicable to edge deployment, including:

[0177] A mine seismic waveform dataset construction module, configured to: obtain the waveform data of mine seismic signals, and perform preprocessing operations such as normalization, denoising, and segmentation on the waveform data to construct a labeled mine seismic waveform dataset;

[0178] A lightweight microseismic phase identification model construction module, configured to: construct a lightweight phase identification model integrating Fourier information, and use the mine seismic waveform dataset to train and verify the model to obtain a trained phase identification model;

[0179] Among them, the lightweight microseismic phase identification model integrating Fourier information includes:

[0180] A multi-scale Fourier transform module, configured to: construct waveform segments of multiple scales for the input microseismic signal through downsampling operations, and perform Fourier transform operations respectively to extract the amplitude spectrum and phase spectrum at each scale;

[0181] An amplitude spectrum feature extraction branch, configured to: receive the amplitude spectrum features at each scale, and perform feature extraction through a lightweight convolutional network, extract the frequency energy structure and encode it as a phase energy representation;

[0182] The phase spectrum feature modeling branch is configured to: receive the phase spectra at various scales (represented by sine / cosine encoding), model its structural features through the lightweight Transformer model Linformer, and capture phase jumps and global structural information;

[0183] The amplitude-phase fusion module is configured to: construct the outputs of the two branches into the form of complex spectra, perform Fourier inverse transform after fusing in the frequency domain, and reconstruct them into a time-domain signal containing energy and structural information; and use a set of convolutional operations to extract the reconstructed signal, and directly output the phase identification result or the arrival time point probability distribution;

[0184] The microseismic phase identification module is configured to: input the to-be-detected mine earthquake signal into the trained model, and output the corresponding P-wave, S-wave or background category, or output the phase arrival probability map.

[0185] Embodiment 5

[0186] The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any one of the above embodiments.

[0187] Each step involved in the device of the above embodiments corresponds to the first method embodiment, and the specific implementation manner can refer to the relevant description part of the first embodiment. The term "computer-readable storage medium" should be understood to include a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0188] Those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately made into individual integrated circuit modules, or multiple modules or steps among them can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0189] Although the specific implementation manner of the present invention has been described above in conjunction with the drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made without creative labor by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A lightweight microseismic signal phase identification method applicable to edge deployment, characterized in that, Including: Obtain the waveform data of microseismic signals, and preprocess the waveform data to obtain a microseismic waveform dataset; Construct a lightweight microseismic phase identification model based on wavelet convolution, use the data in the microseismic waveform dataset to train and verify the constructed model to obtain a trained lightweight microseismic phase identification model based on wavelet convolution, and deploy the trained lightweight microseismic phase identification model based on wavelet convolution on edge devices; Input the waveform data of the microseismic signal to be tested into the trained lightweight microseismic phase identification model based on wavelet convolution to obtain a microseismic phase identification result; Among them, the lightweight microseismic phase identification model based on wavelet convolution includes: a multi-scale Fourier transform module, a lightweight amplitude spectrum feature extraction network, a phase-aware lightweight network, and an amplitude-phase fusion module; The multi-scale Fourier transform module is used to perform multi-scale processing on the microseismic signal to generate signal segments at multiple time scales, convert the signal segments at different time scales into complex frequency spectra, and decompose the complex frequency spectra into amplitude spectra and phase spectra; The lightweight amplitude spectrum feature extraction network processes the input amplitude spectrum to obtain lightweight amplitude spectrum features; The phase-aware lightweight network processes the input phase spectrum to obtain phase features; The feature fusion unit is used to receive the lightweight amplitude spectrum features and phase spectrum features and perform feature fusion processing.

2. The lightweight microseismic signal phase identification method applicable to edge deployment according to claim 1, characterized in that, The multi-scale Fourier transform module includes: a multi-scale construction unit and a spectrum feature extraction unit; The multi-scale construction unit is used to perform multi-scale processing on the microseismic signal, and generate signal segments at multiple time scales by downsampling the original time series signal; The spectrum feature extraction unit includes a plurality of Fourier transform units, and each Fourier transform unit acts on the signal segment corresponding to the corresponding scale respectively, and is used to convert the signal of the corresponding scale into a complex frequency spectrum, and decompose the complex frequency spectrum into an amplitude spectrum and a phase spectrum.

3. The lightweight microseismic signal phase identification method applicable to edge deployment according to claim 1, characterized in that, The lightweight amplitude spectrum feature extraction network includes: an encoding module, a decoding module, and a modulation enhancement module; The encoding module includes a plurality of cascaded encoding units, and is used to extract the deep features of the input amplitude spectrum step by step; The decoding module has a symmetric structure with the encoding module, and is used to perform upsampling reconstruction on the encoded features to restore the detailed information of the phase region; The modulation enhancement module is used to introduce the original amplitude spectrum features into the network backbone to enhance the response of the phase-related features.

4. The lightweight microseismic signal phase identification method applicable to edge deployment according to claim 1, wherein, The phase-aware lightweight network includes: a sine-cosine encoding module, a stitching window module, a shape-changing module, a linear mapping module, a low-rank projection module, a scaled dot-product attention mechanism module, a stitched multi-head module, and a linear layer output module, where: The sine-cosine encoding module is used to receive the phase spectrum output by the multi-scale Fourier transform module, extract phase information from the phase spectra at each scale, and perform sine and cosine encoding respectively to obtain two-dimensional phase structure features; The stitching window module is used to stitch the encoded phase features at each scale along the window dimension to form a four-dimensional tensor in a unified format; The shape-changing module performs a dimension transformation on the above stitching result, and adjusts the four-dimensional tensor into a three-dimensional input tensor; The linear mapping module is used to perform linear mapping on the phase feature channel dimension, map low-dimensional features to a higher-dimensional feature space, and obtain key-value vectors; The low-rank projection module performs compression mapping on the key-value vectors; The scaled dot-product attention mechanism module performs efficient attention calculation through the compressed key-value vectors and query vectors, and combines a feed-forward network and a normalization layer to achieve deep modeling of the input sequence; The concatenated multi-head module is used to concatenate and fuse the output results of multiple attention heads; The linear layer output module is used to output the phase classification probability corresponding to each time point.

5. The lightweight microseismic signal phase identification method applicable to edge deployment according to claim 1, characterized in that, The amplitude-phase fusion module includes: a feature fusion unit, a time-domain reconstruction unit, and a reconstructed feature extraction unit; The feature fusion unit is used to receive the amplitude spectrum features output by the amplitude feature extraction network and the phase spectrum features output by the phase-aware lightweight network, perform joint expression on the two in the frequency domain, and combine them into a complex spectrum in the form of a complex exponential; The time-domain reconstruction unit is used to perform inverse Fourier transform on the fused complex spectrum to restore it to a time-domain signal containing energy and structural information, so as to retain phase jumps and local features; The reconstructed feature extraction unit includes a set of convolution operation modules, which perform multi-scale convolution extraction on the reconstructed time-domain signal to capture the mutation edges and continuous fluctuation time-series features in the time-domain seismic phase waveform, and are used to realize the automatic identification of seismic phases or the prediction of arrival time points in microseismic signals.

6. The lightweight microseismic signal phase identification method applicable to edge deployment according to claim 1, characterized in that Preprocessing the waveform data includes cropping, translating, filtering, and normalizing the microseismic waveform data.

7. A lightweight microseismic signal phase identification system applicable to edge deployment, characterized in that, Including: A microseismic waveform dataset construction module, configured to: obtain the waveform data of microseismic signals, preprocess the waveform data, and obtain a microseismic waveform dataset; A lightweight microseismic phase identification model construction module, configured to: construct a lightweight microseismic phase identification model based on wavelet convolution, train and verify the constructed model using the data in the microseismic waveform dataset to obtain a trained lightweight microseismic phase identification model based on wavelet convolution, and deploy the trained lightweight microseismic phase identification model based on wavelet convolution on edge devices; Among them, the lightweight microseismic phase identification model based on wavelet convolution includes: a multi-scale Fourier transform module, a lightweight amplitude spectrum feature extraction network, a phase-aware lightweight network, and an amplitude-phase fusion module; The multi-scale Fourier transform module is used to perform multi-scale processing on microseismic signals to generate signal segments at multiple time scales, convert the signal segments at different time scales into complex spectra, and decompose the complex spectra into amplitude spectra and phase spectra; The lightweight amplitude spectrum feature extraction network processes the input amplitude spectrum to obtain lightweight amplitude spectrum features; The phase-aware lightweight network processes the input phase spectrum to obtain phase features; The feature fusion unit is used to receive the lightweight amplitude spectrum features and phase spectrum features, and perform feature fusion processing; An identification module, configured to: input the waveform data of the microseismic signal to be tested into the trained lightweight microseismic phase identification model based on wavelet convolution to obtain a microseismic phase identification result.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 6.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-6 above.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it executes the steps of the method described in any one of claims 1-6 above.

Images