Seismic data phase picking method and device based on low-bit quantization of neural network

Abstract

This invention discloses a method for phase picking of low-depth quantized seismic data based on neural networks. The method trains an independent and specially optimized QDNN model for each target depth, performing independent training processes for full-precision floating-point data and quantized data with different compression levels. At extremely low depths, the absolute amplitude information of the data is almost completely lost; the effective information is mainly contained in the zero-crossing rate, polarity changes, and time patterns of the signal. The model learns how to identify seismic events solely from these phase-related features. At higher depths, the data retains rich relative amplitude information, and the model can learn to accurately pick the first arrivals of P-waves and S-waves using subtle amplitude increases and signal-to-noise ratio changes. The method can compress data to 1-2 bits, significantly reducing transmission pressure, while maintaining P-wave / S-wave picking accuracy comparable to full-precision data.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration and artificial intelligence technology, specifically relating to a method for picking phases of low-seismic-depth quantitative seismic data based on neural networks, and also to a computer device. Background Technology

[0002] With the explosive popularization of dense seismic arrays and distributed acoustic sensing (DAS) technologies, the amount of seismic data collected has increased exponentially, posing a severe challenge to data transmission bandwidth and storage costs. In particular, in scenarios with extremely limited bandwidth, such as remote stations that rely on satellite communication and deep-sea observation networks, the inability to transmit massive amounts of data back in real time severely restricts the timeliness of earthquake early warning and aftershock monitoring.

[0003] Existing solutions have significant bottlenecks: lossless compression (such as Steim1 / 2) typically achieves a compression ratio of only 2-3 times, failing to fundamentally address the order-of-magnitude bandwidth gap; traditional lossy compression (such as wavelet transform and DCT) achieves higher compression ratios, but its encoding and decoding computational overhead is large, significantly increasing the power consumption of front-end sensors and easily causing waveform distortion in weak seismic signals; while edge computing triggering mechanisms can transmit only event data, they are susceptible to false triggering leading to missed reports, and the high power consumption of high-performance edge computing chips contradicts the low-energy consumption requirements of long-term deployment in the field.

[0004] Existing research focuses on reducing the sampling rate or the number of data points, but neglects the huge compression potential brought by reducing bit depth. The core difficulty is that drastically reducing bit depth (such as from 24 bits to 1-2 bits) will lead to serious loss of amplitude information, and traditional algorithms will almost fail at extremely low signal-to-noise ratios. Therefore, there is an urgent need for a lightweight algorithm that can directly perform high-precision phase identification on data streams with extremely low bit depth. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems in existing technologies by providing a neural network-based method for seismic phase picking in low-bit-depth quantized seismic data. This method employs non-uniform quantization preprocessing to introduce a signed logarithmic transformation (Symlog), effectively preserving low-amplitude background noise statistical features and weak signal morphology at extremely low bit depths (e.g., 1-2 bits). A dedicated network architecture based on the U-Net architecture is used, incorporating residual connections and a squeeze-and-excitation mechanism to enhance the network's ability to extract phase, frequency, and waveform envelope features even when amplitude information is missing. Dedicated models are trained for different quantization levels (1-8 bits) to allow the network to fully adapt to signal statistical distributions under varying compression ratios. The method utilizes deep neural networks for real-time seismic event detection and P-wave / S-wave phase arrival picking in highly compressed quantized data streams. This invention is widely applicable to natural earthquake monitoring networks, oil and gas exploration, geotechnical engineering monitoring, and critical infrastructure security early warning based on distributed optical fiber sensing (DAS).

[0006] The above-mentioned objectives of the present invention are achieved by the following technical means: A neural network-based method for phase picking in low-seismic-depth quantized seismic data includes the following steps: Step 1: Obtain the earthquake dataset, divide the earthquake dataset into multiple different target depths, and preprocess the earthquake datasets of each different target depth to obtain the preprocessed earthquake datasets of each different target depth. Step 2: Construct the preprocessed seismic datasets for each target depth into sample sets for the corresponding target depths, and then divide them into training sets, validation sets, and test sets for each target depth. The sample includes input samples and ground truth labels. The input samples are preprocessed seismic waveform slices including vertical, north-south, and east-west components. The ground truth labels are the probabilities that the seismic waveform at each time point in the seismic data belongs to P-wave, S-wave, or background noise. Step 3: Construct the QDNN model; Step 4: Build the optimizer, loss function, and learning strategy; Step 5: Train the QDNN model according to the training set for each target bit depth to obtain the trained QDNN model corresponding to each target bit depth. Step 6: Obtain the seismic data to be processed, and perform the target depth segmentation and preprocessing in Step 1. Input the preprocessed seismic data to be processed at each target depth into the trained QDNN model at the corresponding target depth to obtain the prediction results corresponding to the seismic data to be processed at each target depth.

[0007] As described above, the preprocessing includes high-pass filtering and amplitude quantization, and the amplitude quantization includes signed logarithmic transformation, normalization, and discretization.

[0008] As mentioned above, the QDNN model adopts the symmetric encoder-decoder architecture of the U-Net fully convolutional network. The QDNN model includes an input layer, encoder, central bottleneck layer, decoder and output layer. The input layer includes a residual convolutional attention module RCSEa; The encoder includes 2n+1 cascaded residual convolutional attention modules (RCSEs). The 2n+1 residual convolutional attention modules (RCSEs) are alternately set with residual convolutional attention modules (RCSEa) and residual convolutional attention modules (RCSEb). The first residual convolutional attention module (RCSE) of the encoder is the residual convolutional attention module (RCSEa). The feature map output by the residual convolutional attention module (RCSEa) of the input layer is input into the first residual convolutional attention module (RCSEa) of the encoder. The encoder's residual convolutional attention module RCSEa is used to downsample while keeping the time dimension unchanged, increasing the output dimension of the input feature map to twice the input dimension; the encoder's residual convolutional attention module RCSEb is used to compress the time dimension of the input feature map to one-quarter of its original value. The central bottleneck layer includes two cascaded residual convolutional attention modules (RCSEa). The feature map output by the last residual convolutional attention module (RCSEa) of the encoder is input into the first residual convolutional attention module (RCSEa) of the central bottleneck layer. The residual convolutional attention module (RCSEa) of the central bottleneck layer expands the receptive field by stacking convolutional layers without reducing the spatial and temporal dimensions. The decoder includes 2n-1 cascaded modules, which alternate between transposed convolutional attention modules (CTSE) and residual convolutional attention modules (RCSEa). The first module of the decoder is the transposed convolutional attention module (CTSE). The feature map output by the second residual convolutional attention module (RCSEa) of the central bottleneck layer is input into the first transposed convolutional attention module (CTSE) of the decoder. The transposed convolutional attention module CTSE is used to concatenate the input feature map channels to complete upsampling; the residual convolutional attention module RCSEa of the decoder is used to perform nonlinear transformation and feature refinement on the upsampled feature map while keeping the time dimension of the feature map unchanged, and to fuse the feature map output by the residual convolutional attention module RCSEa of the encoder with the same output dimension. The residual convolutional attention modules RCSEa in the encoder and decoder, which have the same output dimension, are residually connected. The output layer includes a convolutional layer with an output dimension of 3 and a kernel dimension of 1x1, and a Softmax activation function. The feature map output by the last residual convolutional attention module RCSEa of the decoder is input into the output layer. The output layer outputs the probability distribution of the seismic waveform at each time point in the seismic data as belonging to P-wave, S-wave, and noise.

[0009] As described above, the Residual Convolutional Attention Module (RCSE) includes a main convolutional path and a residual connection path. The main convolutional path includes a two-dimensional convolutional layer, a batch normalization layer, a LeakyReLU activation layer, a Dropout layer, and a squeeze-activation module connected in sequence. The residual connection path adds the input feature map to the feature map output by the squeeze-activation module element by element to obtain the feature map output by the Residual Convolutional Attention Module (RCSE). The input feature map of the Residual Convolutional Attention Module (RCSE) is input into the two-dimensional convolutional layer. The difference between the residual convolutional attention module RCSEa and the residual convolutional attention module RCSEb of the encoder is that the stride is different. The stride of the residual convolutional attention module RCSEa of the encoder is 1, and the stride of the residual convolutional attention module RCSEb of the encoder is 4. The difference between the transposed convolutional attention module CTSE and the residual convolutional attention module RCSE is that the two-dimensional convolutional layer of the transposed convolutional attention module CTSE uses a transposed convolution with a stride of 4; and neither of the transposed convolutional attention modules CTSE includes residual connection paths.

[0010] As described above, the squeeze-excitation module SEBlock includes, in sequence, global adaptive average pooling, reconstruction, the first fully connected layer, the second fully connected layer, reconstruction, and broadcast multiplication.

[0011] As described above, the feature map input to the squeeze-excitation module undergoes global adaptive average pooling, compressing the feature map of dimension [b,c,h,w] to [b,c,1,1], and then reconstructing it into a single-channel feature map of dimension [b,c]. The single-channel feature map then enters a bottleneck structure consisting of two fully connected layers. The first fully connected layer compresses the data from [b,c] to [b,c / 16] and passes it through the ReLU activation function. The second fully connected layer restores the channel dimension from c / 16 to c and passes it through the Sigmoid activation function to obtain the weight vector. The weight vector is then reconstructed from [b,c] to the channel weights of [b,c,1,1]. Finally, the channel weights are multiplied by the original input feature map through residual connections to obtain the feature map output by the squeeze-excitation module. Where b, c, h, and w represent the batch size, channel dimension, feature map height, and feature map width, respectively.

[0012] As mentioned above, the residual convolutional attention module RCSEa in the input layer has an input dimension of 3, an output dimension of 16, a kernel size of 7, and a stride of 1. The kernel size of both the residual convolutional attention module RCSEa and the residual convolutional attention module RCSEb of the encoder is 7. The residual convolutional attention module RCSEa in the central bottleneck layer has a kernel size of 3 and a stride of 1. The transposed convolutional attention module CTSE of the decoder has a kernel size of 7 and a stride of 4; the residual convolutional attention module RCSEa of the decoder has a kernel size of 7 and a stride of 1.

[0013] As mentioned above, the optimizer used is the Adam optimizer, the loss function is the cross-entropy loss function, and the learning strategy is the early stopping strategy.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the low-bit depth quantization seismic data phase picking method based on neural networks as described above.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the low-bit deep quantization seismic data phase picking method based on neural networks as described above.

[0016] Compared with the prior art, the present invention has the following advantages: (1) Extreme compression and real-time transmission: It can compress data to 1-2 bits with a compression rate of 16-32 times, significantly reducing transmission pressure.

[0017] (2) Low power consumption and high precision coexist: The algorithm has low computational load and is suitable for edge deployment; while greatly compressing data, it still maintains P-wave / S-wave picking accuracy comparable to full-precision data.

[0018] (3) Strong generalization ability: It is not only applicable to traditional seismographs, but also achieves "zero sample" cross-domain application on DAS data, demonstrating strong adaptability to different sensing principles. Attached Figure Description

[0019] Figure 1 The original seismic signal before amplitude quantization; Figure 2 This is the seismic signal after 2-bit standard amplitude quantization; Figure 3 To perform step 1.2.2 of this invention, the seismic signal is quantized again after symlog transformation; Figure 4 This is a schematic diagram of the structure of the QDNN model of the present invention; Figure 5 This is a schematic diagram of the process by which the Residual Convolutional Attention Module (RCSE) of the present invention processes the input features. Figure 6 This is a schematic diagram of the process by which the transposed convolutional attention module CTSE of the present invention processes the input features. Figure 7 This is a schematic diagram of the process by which the SEBlock extrusion module of the present invention processes the input features. Figure 8 This is a schematic diagram of quantized seismic waveform data with a depth of 1 bit in an embodiment of the present invention; Figure 9 This is a schematic diagram of the probability prediction results of P-waves and S-waves obtained by a QDNN model trained using seismic data with a depth of 1. Figure 10 This is a schematic diagram of quantized seismic waveform data with a depth of 2 bits in an embodiment of the present invention (where the horizontal axis is time and the vertical axis is amplitude). Figure 11 This is a schematic diagram of the probability prediction results of P-waves and S-waves obtained by a QDNN model trained using seismic data with a depth of 2. Figure 12 This is a schematic diagram of quantized seismic waveform data with a depth of 3 bits in an embodiment of the present invention; Figure 13 This is a schematic diagram of the probability prediction results of P-waves and S-waves obtained by a QDNN model trained using seismic data with a depth of 3. Figure 14 This is a schematic diagram of quantized seismic waveform data with a depth of 4 bits in an embodiment of the present invention; Figure 15 This is a schematic diagram of the probability prediction results of P-waves and S-waves obtained by a QDNN model trained using seismic data with a depth of 4. Figure 16 This is a schematic diagram of quantized seismic waveform data with a depth of 6 bits in an embodiment of the present invention; Figure 17 This is a schematic diagram of the probability prediction results of P-waves and S-waves obtained by a QDNN model trained using seismic data with a depth of 6. Figure 18 This is a schematic diagram of quantized seismic waveform data with a depth of 8 bits in an embodiment of the present invention; Figure 19 This is a schematic diagram of the probability prediction results of P-waves and S-waves obtained by a QDNN model trained using seismic data with a depth of 8 depths. Figure 20The training QDNN models for each target bit depth (1 bit, 2 bits, 3 bits, 4 bits, 6 bits, 8 bits) are plotted on the test set at the corresponding target bit depth for P-wave arrival error and S-wave arrival error. Figure 21 The arrival errors are defined as follows: the arrival times of P-waves and S-waves predicted by the QDNN model trained according to this invention on a test set with a target depth of 1 bit, and the arrival times of P-waves and S-waves picked manually. The horizontal axis represents the arrival error, with positive values ​​indicating that the model prediction is later than the manual annotation (lagging) and negative values ​​indicating that the prediction is earlier than the manual annotation (leading); the frequency on the vertical axis represents the number of samples in each arrival error interval. Figure 22 The arrival errors of P-wave arrival time and S-wave arrival time obtained by using the QDNN model trained according to the present invention on a test set with a target depth of 2 bits are compared with the manually picked P-wave arrival time and S-wave arrival time, respectively. Figure 23 The arrival errors of P-wave arrival time and S-wave arrival time obtained by using the QDNN model trained according to the present invention on a test set with a target depth of 3 bits are compared with the manually picked P-wave arrival time and S-wave arrival time, respectively. Figure 24 The arrival errors of P-wave arrival time and S-wave arrival time obtained by using the QDNN model trained according to the present invention on a test set with a target bit depth of 4 bits are compared with the manually picked P-wave arrival time and S-wave arrival time, respectively. Figure 25 The arrival errors of P-wave arrival time and S-wave arrival time obtained by using the QDNN model trained according to the present invention on a test set with a target bit depth of 6 bits are compared with the manually picked P-wave arrival time and S-wave arrival time, respectively. Figure 26 The arrival errors of P-wave arrival time and S-wave arrival time obtained by using the QDNN model trained according to the present invention on a test set with a target bit depth of 8 bits are compared with the manually picked P-wave arrival time and S-wave arrival time, respectively. Figure 27 This is the original DAS data; Figure 28 The image shows the results of applying a QDNN trained on seismic data directly to untrained DAS (strain rate) data at a depth of 1 bit (including P-wave arrival time and S-wave arrival time). Figure 29 The image shows the results of applying a QDNN trained on seismic data directly to untrained DAS (strain rate) data at a depth of 2 digits for quantitative detection (including P-wave arrival time and S-wave arrival time). Figure 30The image shows the results of applying a QDNN trained on seismic data directly to untrained DAS (strain rate) data at a depth of 3 digits for quantitative detection (including P-wave arrival time and S-wave arrival time). Figure 31 The image shows the results of applying a QDNN trained on seismic data directly to untrained DAS (strain rate) data at a depth of 4 bits for quantitative detection (including P-wave arrival time and S-wave arrival time). Figure 32 The image shows the results of applying a QDNN trained on seismic data directly to untrained DAS (strain rate) data at a depth of 6 bits for quantization detection (including P-wave arrival time and S-wave arrival time). Figure 33 The image shows the results of applying a QDNN trained on seismic data directly to untrained DAS (strain rate) data at a depth of 8 bits for quantitative detection (including P-wave arrival and S-wave arrival). Detailed Implementation

[0020] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to embodiments. The embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Example 1: A neural network-based method for phase picking in low-seismic-depth quantized seismic data includes the following steps: This invention uses STEAD (STanford EArthquake Dataset) as the core training data source. STEAD is a large-scale, high-quality global earthquake waveform dataset containing approximately 1.2 million time series, totaling over 19,000 hours. This dataset consists of two key waveform types: records of near-field events within 350 kilometers of the epicenter, and clean background noise signals. Its systematic quality control and data processing procedures ensure the high reliability of the annotation information, providing a solid ground truth foundation for the supervised learning of the model.

[0022] Step 1: Obtain the seismic dataset. Divide the seismic dataset into multiple target depths and preprocess each target depth to obtain the preprocessed seismic dataset for each target depth. This includes the following steps: Step 1.1: Obtain the original Stanford seismic waveform dataset (STEAD) and divide the original Stanford seismic waveform dataset into multiple target depths to obtain multiple Stanford seismic waveform datasets with different target depths.

[0023] The Stanford earthquake waveform dataset contains approximately 1.2 million three-component earthquake records from around the world, each 60 seconds long, divided into two main categories: local earthquake events and pure background noise. The data is stored in HDF5 format and accompanied by a CSV file providing detailed metadata. It includes high-quality manually picked P-wave and S-wave arrival information (i.e., the precise arrival times of manually labeled P-waves and S-waves), providing reliable ground truth for supervised learning. This involves converting the precise arrival times of manually labeled P-waves and S-waves into a probability distribution aligned with the model output, thus obtaining the probability that the earthquake waveform at each time point in the earthquake data belongs to a P-wave, S-wave, or background noise.

[0024] In this embodiment, the target bit depths are 1 bit, 2 bits, 3 bits, 4 bits, 6 bits, and 8 bits.

[0025] Step 1.2: Preprocess the Stanford seismic waveform datasets at different depths. Preprocessing includes high-pass filtering and amplitude quantization, specifically including the following steps: Step 1.2.1: To suppress low-frequency background noise introduced by instrument drift and environmental factors and to highlight the effective seismic signal, high-pass filtering is performed on the Stanford seismic waveform dataset, which can effectively enhance the signal-to-noise ratio and obtain multiple Stanford seismic waveform datasets with different bit depths after high-pass filtering.

[0026] This embodiment uses a Butterworth filter. The flat passband characteristics of the Butterworth filter can ensure that noise is effectively filtered out while minimizing the introduction of distortion into the seismic waveform.

[0027] Step 1.2.2: Perform amplitude quantization on the Stanford seismic waveform datasets after high-pass filtering at different depths. Amplitude quantization includes signed logarithmic transform ( ), normalization ( Discretization This generates preprocessed Stanford seismic waveform datasets for training at various depths.

[0028] The signed logarithmic transformation, normalization, and discretization are based on the following formulas: (1); (2); (3); In the formula, The index is the time sampling point number, ranging from 1 to T (where T is the time series length of the seismic waveform record). This is the Stanford earthquake waveform dataset after high-pass filtering; It is a symbolic function; This is the floor function; and These are optional parameters, and are all set to 1 in this embodiment; To quantify the total number of levels, based on the target bit depth (8 bits, 6 bits, 4 bits, 2 bits, 1 bit) determines the number of bits, satisfying the formula. For example, the total number of quantization levels for 8-bit data is 256.

[0029] like Figures 1-3 The original seismic signal and the seismic signals after different quantizations shown are illustrated. Due to the extremely high dynamic range of seismic signals, traditional uniform quantization (such as...) Figure 2 As shown, this can cause a large amount of low-amplitude background noise and weak signals to fall into the same quantization step (usually 0), resulting in a "dead zone" phenomenon and complete loss of information. To solve this problem, this invention introduces a signed logarithmic transform (Symlog) before quantization. The signed logarithmic transform can effectively balance the dynamic range of the signal, preserve noise details, and not affect the identification of strong vibration signals.

[0030] Amplitude quantization of the Stanford earthquake waveform dataset can greatly compress the data volume. This invention repeats this quantization process for different target depths, generating a dedicated dataset for each data compression level to train the corresponding optimization model. Figures 1-3 The original seismic signal, the standard 2-bit uniformly quantized signal, and the 2-bit quantized signal using logarithmic transformation (Symlog) are shown. It can be seen that after introducing logarithmic transformation, the quantized data can more effectively preserve the subtle dynamic characteristics of the original waveform, especially in the low-amplitude noise range.

[0031] Step 2: Create sample sets for each target depth from the preprocessed seismic datasets. Then, divide the sample sets for each target depth into training sets, validation sets, and test sets.

[0032] The input samples are constructed as follows: The input samples are preprocessed seismic waveform slices including vertical, north-south, and east-west components. In this embodiment, the truncation time window length is 30.72 seconds and the sampling rate is 100Hz, so each channel contains 3072 sampling points. The tensor dimension of the input samples is [Batch_Size, 3, 3072, 1], where the four dimensions correspond to the batch size, the three component channels (vertical component channel, north-south component channel, and east-west component channel), the time dimension, and the spatial height dimension, respectively.

[0033] The true labels corresponding to the input samples are constructed in the following way: the precise time of manually labeled P-waves and S-waves is converted into the category probability distribution of seismic waveforms at each time point in the seismic data.

[0034] As one possible implementation method, the arrival time of manually marked seismic phases is used as the first step. A Gaussian probability distribution window is generated centered on the time point. Its probability value The calculation formula is ,in, The standard deviation is set to 1s. The Gaussian probability distribution transforms the discrete arrival time into a continuous time probability distribution, which is the probability that the earthquake waveform at each time point in the earthquake data belongs to P-wave, S-wave, or background noise, and serves as the true label.

[0035] Step 3: Construct the QDNN (Quantized Data Neural Network) model, which includes the following steps: like Figure 4 The diagram shows the structure of the QDNN model. The QDNN model of this invention adopts a symmetric encoder-decoder architecture of the U-Net fully convolutional network, including an input layer, encoder, central bottleneck layer, decoder, and output layer. The input dimension (i), output dimension (o), convolutional kernel size (k), and stride (s) of each layer are all within [specific parameters]. Figure 4 Mark it.

[0036] The tensor dimension of the QDNN input data is [b, 3, T, 1], where b represents the batch size during training, i.e., the number of samples input to the QDNN model in one iteration; 3 represents the channel dimension of the seismic data, corresponding to the vertical component (Z), north-south component (N), and east-west component (E) of the three-component seismic detector; T represents the time dimension length of the seismic waveform record, i.e., the number of seismic waveform sampling points in a single input; and 1 represents the spatial height dimension. In this invention, the one-dimensional seismic time series is regarded as a two-dimensional feature map with a height of 1. Both RCSE and CTSE use the two-dimensional convolution operator (Conv2d) for processing, so as to facilitate subsequent use for high-dimensional data such as distributed fiber optic sensing data recording.

[0037] The input data first passes through an input layer, which uses an RCSE module with an input dimension of 3, an output dimension of 16, and a stride of 1. The input layer maps the original 3-channel input data to a 16-dimensional feature space while keeping the time dimension T unchanged. The output feature map of the input layer is then input into the encoder.

[0038] The encoder includes nine cascaded residual convolutional attention modules (RCSEs). These RCSEs are divided into two types, denoted as RCSEa and RCSEb. The difference between RCSEa and RCSEb lies in their stride: RCSEa has a stride of 1, while RCSEb has a stride of 4. The nine cascaded RCSEs are constructed by alternating between RCSEa and RCSEb. The first RCSE in the encoder is RCSEa. Therefore, the specific structure of the nine cascaded RCSEs is [RCSEa, residual convolutional attention module ... convolutional attention module, residual convolutional convolutional attention module, residual convolutional convolutional attention module, residual convolutional convolutional attention module, residual convolutional convolutional attention module, residual convolutional convolutional attention module, residual convolutional convolutional convolutional attention module, residual convolutional convolutional convolutional attention module, residual convolutional convolutional convolutional attention module, residual convolutional convolutional convolutional convolutional attention module, residual convolutional convolutional convolutional convolutional attention module, residual convolutional convolutional convolutional convolutional convolutional attention module, residual convolutional convolutional convolutional convolutional convolutional convolutional attention module The input layer's output feature map is fed into the encoder's first residual convolutional attention module RCSEa, and the feature map output by the encoder's last residual convolutional attention module RCSEa (i.e., the encoder's output feature map) is fed into the central bottleneck layer. The encoder's residual convolutional attention module RCSEa is used to downsample while maintaining the temporal dimension, increasing the output dimension to twice the input dimension, thereby deepening the network layers and refining the features. The encoder's residual convolutional attention module RCSEb is used to compress the temporal dimension of the input feature map to one-quarter of its original size, increasing the network's receptive field.

[0039] The central bottleneck layer consists of two cascaded residual convolutional attention modules (RCSEa). The feature map output by the last residual convolutional attention module (RCSEa) of the encoder is input into the first residual convolutional attention module (RCSEa) of the central bottleneck layer. The feature map output by the second residual convolutional attention module (RCSEa) of the central bottleneck layer (i.e., the feature map output by the central bottleneck layer) is input into the decoder. The residual convolutional attention modules (RCSEa) of the central bottleneck layer expand the receptive field by stacking convolutional layers without further reducing the spatial and temporal dimensions. The kernel size (k) and stride (s) of the residual convolutional attention modules (RCSEa) of the central bottleneck layer are all 3, and the input and output feature dimensions are consistent. They process the deep abstract feature map with the minimum temporal resolution and the maximum number of channels obtained after multiple downsampling by the encoder.

[0040] The decoder consists of seven cascaded modules, constructed by alternating between transposed convolutional attention modules (CTSE) and residual convolutional attention modules (RCSEa). The first module of the decoder is the transposed convolutional attention module (CTSE), and the specific order of the seven cascaded modules is [CTSE, residual convolutional attention module RCSEa, CTSE, residual convolutional attention module RCSEa, CTSE, residual convolutional attention module RCSEa, CTSE, residual convolutional attention module RCSEa]. The feature map output from the central bottleneck layer is input to the first transposed convolutional attention module (CTSE) of the decoder, and the feature map output from the last residual convolutional attention module (RCSEa) of the decoder is input to the output layer. The transposed convolutional attention module (CTSE) uses a stride of s=4, and each transposed convolutional attention module (CTSE) concatenates the input feature map channels. Cat) completes upsampling, which enables the decoder to utilize both deep semantic information and shallow detailed positional information, greatly improving the accuracy of phase picking; the residual convolutional attention module RCSEa of the decoder adopts convolution with stride s=1. The residual convolutional attention module RCSEa of the decoder is used to perform nonlinear transformation and feature refinement on the input feature map while keeping the time dimension of the input feature map unchanged, and fuses the feature map output from the residual convolutional attention module RCSEa of the encoder with the same output dimension.

[0041] Furthermore, the residual convolutional attention module RCSEa, which has the same output dimension in both the encoder and decoder, uses residual connections to establish a direct identity transformation (skip connection) between the input and output.

[0042] The output layer consists of a cascaded convolutional layer with an output dimension of 3 and a kernel dimension of 1x1, and a Softmax activation function. The output layer outputs the three-dimensional probability distribution vector of the seismic waveform at each time point in the seismic data. The three-dimensional probability distribution vector of the seismic waveform at each time point in the seismic data includes the probability that the seismic waveform at the corresponding time point belongs to P-wave, S-wave, and noise, respectively.

[0043] The core building blocks of the QDNN module of this invention include a residual-conv-SEBlock (RCSE), a transposed-conv-SEBlock (CTSE), and a squeeze-excitation block (SEBlock).

[0044] like Figure 5As shown, the Residual Convolutional Attention Module (RCSE) is one of the main components of the entire QDNN module. The RCSE includes a main convolutional path and a residual connection path. The main path consists of a sequentially connected 2D convolutional layer (Conv2d), a batch normalization layer (BatchNorm2d), a LeakyReLU activation layer, a Dropout layer, and a squeeze-activation module. The 2D convolutional layer uses a k×1 kernel to extract local temporal correlations; the batch normalization layer normalizes the feature distribution; the LeakyReLU activation layer (with a negative slope of 0.2) introduces non-linearity; and the Dropout module... The layer (dropout layer, dropout rate can be set to 0.2) is used to prevent overfitting; the residual connection path directly adds the input feature map to the feature map output by the squeezing-excitation module SEBlock of the main path element-wise (ADD) to preserve the integrity of the original feature flow. When the dimension of the input feature map of the residual convolutional attention module RCSE is [b,c,h,w], the dimension of the feature map output by the residual convolutional attention module RCSE is [b,o,hs,ws].

[0045] like Figure 6 As shown, the difference between the transposed convolutional attention module CTSE and the residual convolutional attention module RCSE is as follows: (1) The two-dimensional convolutional layer of the transposed convolutional attention module CTSE uses a transposed convolution with a stride of s=4 (i.e., a two-dimensional transposed convolutional layer) to perform upsampling, which increases the time dimension of the input feature map to 4 times the original. (2) The transposed convolutional attention module CTSE does not include residual connection paths for channel splicing. When the dimension of the input feature map of the transposed convolutional attention module CTSE is [b,c,h,w], the dimension of the output feature map of the transposed convolutional attention module CTSE is [b,o,hs,ws].

[0046] like Figure 7 As shown, the squeeze-activation module SEBlock adaptively recalibrates the importance of channel features by explicitly modeling the interdependencies between channels. The squeeze-activation module SEBlock includes two key steps: "squeezing" and "activation". Specifically, the squeeze-activation module SEBlock includes global adaptive average pooling (AdaptiveAvgPool2d), reshape, the first fully connected layer (Linear), the second fully connected layer, reshape, and broadcast multiplication. Among them, global adaptive average pooling and reshape constitute the squeezing step, and the first fully connected layer, the second fully connected layer, and reshape constitute the activation step.

[0047] The SEBlock squeezing-excitation module processes the input feature map of dimension [b,c,h,w] as follows: First, the input feature map undergoes global adaptive average pooling (AdaptiveAvgPool2d), compressing the input feature map of dimension [b,c,h,w] along the time axis to [b,c,1,1], and then reshapes it into a single-channel feature map of dimension [b,c]. The single-channel feature map then enters a bottleneck structure consisting of two fully connected (Linear) layers. The first fully connected layer compresses the data from [b,c] to [b,c / 16], and then passes it through the ReLU activation function, introducing nonlinearity and reducing parameters. The first layer consists of a second fully connected layer: the channel dimension is restored from c / 16 to c, and a weight vector is obtained by passing it through a Sigmoid activation function (at this point, the values ​​of each element in the output vector are between (0,1), representing the importance weight of each feature channel); then the weight vector is reconstructed from [b,c] to channel weights of [b,c,1,1] for broadcast multiplication with the original feature map; finally, the channel weights are multiplied by channel weights with the original input feature map through residual connections (multiplication); through the above process, channels containing key P / S wave features are automatically enhanced, while irrelevant channels containing background noise are suppressed.

[0048] Step 4: Construct the optimizer, loss function, and learning strategy, which includes the following steps: To efficiently train the QDNN model and ensure its performance, this invention uses the Adam optimizer to train the QDNN model, and the loss function used for model training is the cross-entropy loss function.

[0049] Adam is an efficient and robust optimization algorithm that combines the advantages of momentum and root mean square propagation to dynamically calculate an adaptive learning rate for each parameter in the network. This feature enables it to efficiently handle large-scale datasets and complex loss functions commonly found in the field of seismology.

[0050] Given that the task of QDNN is to perform multi-class classification (P-wave, S-wave, noise) at each time point, the cross-entropy loss function is the standard choice for measuring the difference between the model's predicted probability distribution and the true label. It can effectively punish inaccurate and uncertain predictions and guide the model parameters to converge in the direction of minimizing the prediction error.

[0051] To prevent overfitting during training, i.e., when a model performs well on training data but poorly on unseen data, this invention introduces an early stopping mechanism. Specifically, by monitoring the loss function value of the model on the validation set, if the value fails to decrease further within 100 consecutive epochs, training is automatically terminated, and the weights of the model with the best validation performance are saved, which can effectively improve the model's generalization ability.

[0052] Step 5: Train the QDNN model based on training sets with different target bit depths. After training, save the parameter model to obtain the trained QDNN model for each target bit depth. This includes the following steps: The core strategy of this invention is to train an independent and specially optimized QDNN model for each target bit depth. Independent training processes are performed for full-precision floating-point data (32 bits) and quantized data with different compression levels (8 bits, 6 bits, 4 bits, 2 bits, and 1 bit). The theoretical basis of this strategy is that data at different bit depths have essential differences in information representation and statistical distribution: at extremely low bit depths (such as 1 or 2 bits), the absolute amplitude information of the data is almost completely lost. At this time, the effective information is mainly contained in the zero-crossing rate, polarity change, and time pattern of the signal. The model trained specifically for this type of data is tasked with learning how to identify seismic events only from these phase-related features. At higher bit depths (such as 8 bits), the data retains rich relative amplitude information, and the model can learn to use features such as subtle amplitude growth and signal-to-noise ratio changes to accurately pick up the first arrival of P waves and S waves.

[0053] The general strategy of "one model for all bit depths" is inefficient because the model will face the challenge of conflicting learning objectives, making it difficult to master decoding two completely different data paradigms simultaneously, ultimately resulting in suboptimal performance across all compression levels. Therefore, by training a dedicated model for each bit depth, we ensure that each QDNN instance is highly matched to the unique statistical characteristics and information content of its input data. This not only maximizes the accuracy of seismic phase recognition at various data compression levels but also provides high flexibility for practical applications—users can select and deploy the optimal model for the corresponding bit depth based on actual bandwidth or storage limitations.

[0054] Step 6: Obtain the seismic data to be processed, and perform the target depth segmentation and preprocessing in Step 1. Input the preprocessed seismic data to be processed at each target depth into the trained QDNN model at the corresponding target depth to obtain the prediction results corresponding to the seismic data to be processed at each target depth.

[0055] Example 2: A systematic evaluation was conducted on training models with different bit depths: such as Figures 8-19 As shown, even in the extreme case of seismic data with a depth of 1 bit, although the signal becomes a binary pulse, it can be seen that the prediction results of the QDNN model of the present invention can still lock the arrival area of ​​the P wave; when the depth is increased to 2 bits, the accuracy of the prediction results of the QDNN model is very close to that of the high depth results.

[0056] like Figures 20-26To compare the arrival errors of the P-wave and S-wave predicted by the QDNN model trained according to this invention with those obtained manually on test sets with different target depths, the pickup error statistics show that the S-wave error is relatively large (MAE 24.2 points) when using 1-bit quantization. This is because the S-wave is usually submerged in the tail of the P-wave and lacks amplitude information, making it difficult to distinguish. The error distribution of the 2-bit model narrows sharply (…). Figures 20-26 The P-wave MAE dropped to 5.7 points; this proves that with logarithmic processing, only 2 bits of data are needed to meet the needs of most earthquake monitoring, achieving the best balance between storage and accuracy.

[0057] As the bit depth continued to increase from 2 bits to 8 bits, the overall performance of the model showed high stability. Figures 20-26 The error of the P-wave stabilizes at about 5 sampling points, while the error of the S-wave stabilizes at about 9 sampling points.

[0058] This series of tests strongly demonstrates that the proposed QDNN model can maintain the accuracy and robustness of P-wave phases even when the data is extremely compressed to 1 bit; and in the case of 2 bits, it can maintain high accuracy and robustness for both P-wave and S-wave phases. This discovery reveals the optimal balance between data compression rate and model performance, which has significant practical implications for achieving reliable real-time earthquake monitoring in bandwidth-constrained scenarios.

[0059] Compared to traditional point seismographs, distributed acoustic sensing (DAS) technology generates massive amounts of data far exceeding those of traditional seismic arrays due to its ultra-high temporal and spatial resolution. However, DAS seismological applications still face two major challenges: first, high-quality labeled datasets are relatively scarce; second, the different fiber optic cable deployment methods at different sites lead to huge differences in signal characteristics, which places stringent requirements on the generalization ability of the model.

[0060] like Figures 27-33 As shown, to explore the applicability of the model of this invention in a novel data domain, we conducted a rigorous cross-domain generalization test. To verify its practicality, we directly applied the model trained on a traditional seismograph (velocity / accelerometer) to untrained raw DAS (strain rate) data (raw optical signal measurements acquired by a distributed acoustic sensing system without any processing) and evaluated its phase picking ability at different bit depths. The experimental results showed good generalization performance: under the extreme compression condition of 1 bit, the model was able to initially identify the clear arrival time of S-waves; when the bit depth was increased to 3 bits or more, the model achieved a stable and reliable level of arrival time detection for both P-waves and S-waves.

[0061] Given that the model has never learned the signal characteristics of any DAS data, the success of this "zero-sample" test is particularly important. It fully demonstrates that the method proposed in this invention not only performs well on a single data type, but also that the intrinsic physical laws of the seismic waveforms it learns have strong cross-domain generalization potential. This result strongly proves that low-depth data processing technology can be effectively applied to DAS data, providing a promising technical path for solving the problem of real-time transmission and intelligent processing of massive DAS data.

[0062] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0063] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0064] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0065] It should be noted that the embodiments described in this invention are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for picking phases from low-level, deep-quantized seismic data based on neural networks, characterized in that, Includes the following steps: Step 1: Obtain the earthquake dataset, divide the earthquake dataset into multiple different target depths, and preprocess the earthquake datasets of each different target depth to obtain the preprocessed earthquake datasets of each different target depth. Step 2: Construct the preprocessed seismic datasets for each target depth into sample sets for the corresponding target depths, and then divide them into training sets, validation sets, and test sets for each target depth. The sample includes input samples and ground truth labels. The input samples are preprocessed seismic waveform slices including vertical, north-south, and east-west components. The ground truth labels are the probabilities that the seismic waveform at each time point in the seismic data belongs to P-wave, S-wave, or background noise. Step 3: Construct the QDNN model; Step 4: Build the optimizer, loss function, and learning strategy; Step 5: Train the QDNN model according to the training set for each target bit depth to obtain the trained QDNN model corresponding to each target bit depth. Step 6: Obtain the seismic data to be processed, and perform the target depth segmentation and preprocessing in Step 1. Input the preprocessed seismic data to be processed at each target depth into the trained QDNN model at the corresponding target depth to obtain the prediction results corresponding to the seismic data to be processed at each target depth.

2. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 1, characterized in that, The preprocessing includes high-pass filtering and amplitude quantization, and the amplitude quantization includes signed logarithmic transformation, normalization, and discretization.

3. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 2, characterized in that, The QDNN model adopts the symmetric encoder-decoder architecture of the U-Net fully convolutional network. The QDNN model includes an input layer, an encoder, a central bottleneck layer, a decoder, and an output layer. The input layer includes a residual convolutional attention module RCSEa; The encoder includes 2n+1 cascaded residual convolutional attention modules (RCSEs). The 2n+1 residual convolutional attention modules (RCSEs) are alternately set with residual convolutional attention modules (RCSEa) and residual convolutional attention modules (RCSEb). The first residual convolutional attention module (RCSE) of the encoder is the residual convolutional attention module (RCSEa). The feature map output by the residual convolutional attention module (RCSEa) of the input layer is input into the first residual convolutional attention module (RCSEa) of the encoder. The encoder's residual convolutional attention module RCSEa is used to downsample while keeping the time dimension unchanged, increasing the output dimension of the input feature map to twice the input dimension; the encoder's residual convolutional attention module RCSEb is used to compress the time dimension of the input feature map to one-quarter of its original value. The central bottleneck layer includes two cascaded residual convolutional attention modules (RCSEa). The feature map output by the last residual convolutional attention module (RCSEa) of the encoder is input into the first residual convolutional attention module (RCSEa) of the central bottleneck layer. The residual convolutional attention module (RCSEa) of the central bottleneck layer expands the receptive field by stacking convolutional layers without reducing the spatial and temporal dimensions. The decoder includes 2n-1 cascaded modules, which alternate between transposed convolutional attention modules (CTSE) and residual convolutional attention modules (RCSEa). The first module of the decoder is the transposed convolutional attention module (CTSE). The feature map output by the second residual convolutional attention module (RCSEa) of the central bottleneck layer is input into the first transposed convolutional attention module (CTSE) of the decoder. The transposed convolutional attention module CTSE is used to concatenate the input feature map channels to complete upsampling; the residual convolutional attention module RCSEa of the decoder is used to perform nonlinear transformation and feature refinement on the upsampled feature map while keeping the time dimension of the feature map unchanged, and to fuse the feature map output by the residual convolutional attention module RCSEa of the encoder with the same output dimension. The residual convolutional attention modules RCSEa in the encoder and decoder, which have the same output dimension, are residually connected. The output layer includes a convolutional layer with an output dimension of 3 and a kernel dimension of 1x1, and a Softmax activation function. The feature map output by the last residual convolutional attention module RCSEa of the decoder is input into the output layer. The output layer outputs the probability distribution of the seismic waveform at each time point in the seismic data as belonging to P-wave, S-wave, and noise.

4. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 3, characterized in that, The Residual Convolutional Attention Module (RCSE) includes a main convolutional path and a residual connection path. The main convolutional path includes a two-dimensional convolutional layer, a batch normalization layer, a LeakyReLU activation layer, a Dropout layer, and a squeeze-excitation module connected in sequence. The residual connection path adds the input feature map to the feature map output by the squeeze-excitation module element by element to obtain the feature map output by the Residual Convolutional Attention Module (RCSE). The input feature map of the Residual Convolutional Attention Module (RCSE) is input into the two-dimensional convolutional layer. The difference between the residual convolutional attention module RCSEa and the residual convolutional attention module RCSEb of the encoder is that the stride is different. The stride of the residual convolutional attention module RCSEa of the encoder is 1, and the stride of the residual convolutional attention module RCSEb of the encoder is 4. The difference between the transposed convolutional attention module CTSE and the residual convolutional attention module RCSE is that the two-dimensional convolutional layer of the transposed convolutional attention module CTSE uses a transposed convolution with a stride of 4; and neither of the transposed convolutional attention modules CTSE includes residual connection paths.

5. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 4, characterized in that, The squeeze-excitation module, SEBlock, includes global adaptive average pooling, reconstruction, a first fully connected layer, a second fully connected layer, reconstruction, and broadcast multiplication.

6. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 5, characterized in that, The feature map input to the squeeze-excitation module undergoes global adaptive average pooling, compressing the feature map of dimension [b,c,h,w] to [b,c,1,1], and then reconstructing it into a single-channel feature map of dimension [b,c]. This single-channel feature map then enters a bottleneck structure consisting of two fully connected layers. The first fully connected layer compresses the data from [b,c] to [b,c / 16] and passes it through a ReLU activation function. The second fully connected layer restores the channel dimension from c / 16 to c and passes it through a Sigmoid activation function to obtain a weight vector. This weight vector is then reconstructed from [b,c] back into channel weights of dimension [b,c,1,1]. Finally, a residual connection is used to perform a channel-weighted multiplication of the channel weights and the original input feature map to obtain the feature map output by the squeeze-excitation module. Where b, c, h, and w represent the batch size, channel dimension, feature map height, and feature map width, respectively.

7. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 6, characterized in that, The residual convolutional attention module RCSEa in the input layer has an input dimension of 3, an output dimension of 16, a kernel size of 7, and a stride of 1. The kernel size of both the residual convolutional attention module RCSEa and the residual convolutional attention module RCSEb of the encoder is 7. The residual convolutional attention module RCSEa in the central bottleneck layer has a kernel size of 3 and a stride of 1. The transposed convolutional attention module (CTSE) of the decoder has a kernel size of 7 and a stride of 4. The residual convolutional attention module RCSEa of the decoder has a kernel size of 7 and a stride of 1.

8. The method for picking phases of low-position deep quantized seismic data based on neural networks according to claim 7, characterized in that, The optimizer used is the Adam optimizer, the loss function is the cross-entropy loss function, and the learning strategy is the early stopping strategy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the low-position deep quantization seismic data phase picking method based on neural networks as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the low-position deep quantization seismic data phase picking method based on neural networks as described in any one of claims 1 to 8.

Images