U-shaped network head wave arrival time extraction method

Abstract

According to the U-shaped network head wave arrival time extraction method, a U-shaped coding and decoding network is constructed, forward and backward spatial-temporal characteristics of sound wave signals are captured in combination with a bidirectional convolution long and short time memory module, the modeling capability of a head wave propagation rule is enhanced, in addition, a multi-scale attention module is introduced into the U-shaped coding and decoding network, head wave key characteristics are adaptively focused, and the head wave arrival time extraction efficiency is improved. Interference of environmental noise is effectively suppressed, and the problems of large parameter setting error and poor noise immunity caused by traditional artificial experience are solved.

Description

TECHNICAL FIELD

[0001] The application relates to a U-shaped network first wave arrival time extraction method and belongs to the technical field of first wave arrival time extraction methods. BACKGROUND

[0002] Acoustic logging is an important technical means for accurately obtaining the acoustic velocity of the rock around a well in the field of oil exploration. In the processing of acoustic signals, the extraction of the first wave arrival time of the acoustic signal is a key step in acoustic logging data processing, and the accurate extraction of the first wave arrival time can provide an important basis for formation velocity calculation and fracture identification. At present, the first wave arrival time extraction methods mainly include a threshold-based method, an STA / LTA algorithm, an energy ratio method, an autocorrelation method and a waveform coherence superposition method. The threshold-based method and the STA / LTA algorithm are susceptible to noise interference, and the picking result is not accurate in a complex environment; the calculation of the energy ratio method involves the amplitudes of the samples before the first arrival point, and if the amplitudes of the samples before the first arrival point are close to zero, the denominator tends to zero, thereby producing a singular value; the autocorrelation method is largely dependent on the characteristics of the signal, and if the signal has a complex frequency spectrum structure or irregular time domain characteristics, the autocorrelation method may not be able to accurately extract the acoustic arrival time; the waveform correlation superposition method usually needs to search in a two-dimensional grid (one dimension is time and the other dimension is slowness), and this search process may involve a large amount of calculation, especially when processing long signals or high-resolution data, the calculation complexity will increase significantly.

[0003] Therefore, in view of the low extraction precision, low processing efficiency and poor noise resistance of the first wave arrival time of the acoustic signal in the prior art, it is necessary to develop a new first wave arrival time extraction method to solve the above problems existing in the prior art. SUMMARY

[0004] The purpose of the present application is to provide a U-shaped network first wave arrival time extraction method, which combines the time sequence feature extraction capability of the bidirectional convolution long short-term memory network and the feature refinement capability of the attention mechanism, and solves the problems of low extraction precision, low processing efficiency and poor noise resistance of the first wave arrival time of the acoustic signal in the prior art.

[0005] The technical scheme of the present application is: A U-shaped network first wave arrival time extraction method, comprising the following steps: Step one, acoustic data preprocessing and data enhancement: (1)Filtering: The first arrival time extraction is a key step in array acoustic signal processing, and its extraction accuracy directly affects the calculation results of formation parameters such as porosity and anisotropy. The array acoustic signal contains P-wave, S-wave and Stoneley wave, and the frequency range of these waves is relatively close. When filtering the data, the frequency band characteristics are required to be high, and the traditional filtering method is easy to cause the loss of signal details. The cosine band-pass filter avoids the problem of signal loss by constructing a smooth transition window function in the frequency domain, providing more complete information support for accurate first arrival time extraction, and its mathematical expression is: Where f represents the signal frequency, f1 represents the first cutoff frequency, signals below this frequency are completely filtered out, f2 represents the second cutoff frequency, f1 and f2 are connected by the cosine window function to realize low-pass filtering, f3 represents the third cutoff frequency, f2 and f3 form an ideal passband region, f4 represents the fourth cutoff frequency, f3 and f4 are connected by the cosine window function to realize high-pass filtering, and signals with frequency higher than f4 are completely filtered out.

[0006] (2)Data standardization: In order to eliminate the dimensional differences between different instruments and different channels, the acoustic signal is subjected to Z-score standardization processing, and the data distribution is adjusted to a normal distribution with a mean of 0 and a standard deviation of 1, and its mathematical expression is: s i represents the original acoustic data, s i ' represents the standardized data, N represents the total number of single-channel signal sampling points (1000 in the experiment), μ represents the signal mean, and σ represents the signal standard deviation.

[0007] (3)Data augmentation: Let the preprocessed standardized acoustic signal be z(t), and the signal data after data augmentation be z aug (t), and the data augmentation operation can be expressed as: z aug (t) = z(t) + n g (t) (5) n g (t) ~ N(0, σ g 2 ), σ g = α·max(|z(t)|) (6) Where n g (t) is Gaussian white noise, σ g is the noise standard deviation, and α is the noise intensity coefficient, whose value range satisfies 0.05≤α≤0.15.

[0008] Step two, acoustic data label processing: For the longitudinal wave, the transverse wave and the Stoneley wave, a Gaussian probability density function is constructed with the manually picked first arrival time as the center of distribution, for a time sampling point t, which belongs to the wave type X, X can be P, S, ST; the probability P of the first arrival event X(t) is defined as: Where, t X represents the manually picked first arrival time (t P , t S , t ST ) of the wave form X, and σ X is the preset standard deviation parameter (σ P , σ S , σ ST ) of the wave form X, which is used to control the width of the probability distribution. At the picked first arrival time point t X , the probability P X(t) reaches the maximum value, at the adjacent time point t X , P X(t) smoothly decays with the increase of the distance t X , and at the time point far away from t X , the probability tends to zero.

[0009] Step three, network model: The present application takes U-Net network as the basic framework, and aims at the first arrival time extraction task of the acoustic wave signal, and realizes the core innovation by embedding the BiConvLSTM module and the multi-scale attention module, and the innovation points are as follows: the first layer convolution output of the encoder and the last layer up-sampling output of the decoder are fused through the BiConvLSTM, the directional dependence relationship of the acoustic wave propagation is captured by using the bidirectional space-time modeling capability, and the environmental noise interference is effectively suppressed; further, the multi-scale attention module focuses on the short-time high-frequency features and suppresses the subsequent noise, so that the signal-to-noise ratio of the key signal area is significantly improved, and the positioning accuracy and the robustness of the first arrival time are optimized, and the specific processing process is as follows: Let the preprocessed acoustic wave signal be X∈R B×C×T , B, C, T represent batch size, channel number and time sequence length respectively; the encoder takes the acoustic wave signal X∈R B×C×T as the input, and extracts multi-scale features through multi-level convolution block processing and alternating operation of down-sampling layers.

[0010] During the processing of convolutional blocks, a one-dimensional convolutional kernel slides along the time dimension to extract waveform features. The kernel size is fixed at 3×1, the stride is 1, and the padding is 1. The output sequentially passes through a batch normalization layer and a ReLU activation function. The number of channels in the convolutional block is configured according to a hierarchical progression rule, increasing to 64, 128, 256, 512, and 1024 channels, forming a multi-scale feature extraction structure. During downsampling, each convolutional block is followed by a max-pooling downsampling layer (MaxPool). 1d The kernel size is fixed at 2×1, and the step size is 2; this operation ensures that the output time dimension is always half of the input, i.e., T. out =T in / 2; The above "convolution block → downsampling" combination is repeated 6 times to form a loop iterative processing structure, which finally achieves a 6-level compression path in the time dimension: T→T / 2→T / 4→T / 8→T / 16→T / 32→T / 64.

[0011] E (l)↓ =MaxPool 1d (E (l) (9) In Equations 8 and 9, E (l-1) E represents the feature map input to the (l-1)th convolutional block. (l) E represents the feature map output by the l-th convolutional block. (l)↓ This represents the feature map output by the l-th layer after downsampling. represents a one-dimensional convolution kernel, * represents a one-dimensional convolution operation, BN represents batch normalization, and ReLU is the modified linear unit activation function.

[0012] The first convolutional block output feature E (1) With the decoder's last layer upsampled output D last After channel concatenation, the data are input into the BiConvLSTM module, where forward and backward propagation operations are performed simultaneously to form a new feature map with a complete temporal context.

[0013] F = W f *H f +W b *H b +b(12) E (1) D last These are the outputs of the first layer of the encoder and the last layer of the decoder, respectively. Represents feature concatenation, ConvLSTM f and ConvLSTM bH represents the forward convolution LSTM unit and the backward convolution LSTM unit, respectively. f H b These represent the forward hidden state sequence and the reverse hidden state sequence, respectively, used to record the forward / reverse propagation characteristics of sound waves. W f W b represents the forward weight kernel and the reverse weight kernel, respectively; * represents the convolution operation; b is the bias term; and F represents the final output feature.

[0014] The feature maps F output by the second and third convolutional blocks (2) F (3) The input is directly fed into the multi-scale attention module (ATT) with the feature map F. (3) Taking the processing procedure as an example, F (2) Similarly: by using 3×1 convolution kernels Conv1D in parallel 3×1 and a 5×1 convolution kernel Conv1D 5×1 The high-frequency feature F3 and waveform envelope feature F5 of the first wave transition are captured respectively, and finally the two features are adaptively fused for output, where α∈[0,1]; F3 = Conv1D 3×1 (F (3) (13) F5 = Conv1D 5×1 (F (3) (14) F ATT =α·F3+(1-α)·F5(15) During the deep feature initialization phase of the decoder, the deep features E output by the encoder are... (l) The input to the decoder is first compressed to half the number of channels using a 1×1 transposed convolutional kernel. Then, the feature distribution is standardized through a batch normalization layer, and a non-linear transformation is introduced through a ReLU activation function to obtain the output D. (l) .in, This represents the weights of the transposed convolution kernel of the decoder layer l.

[0015] During the upsampling process in the decoder, the time dimension is expanded by a factor of 2, and the output D is obtained after upsampling. (l-1)↑ Then, the encoder output feature E of the same layer is processed by skip connections. (l-1) With the decoder's output feature D (l-1)↑ Perform channel stitching and output a fused feature map. Among them, u 2× This indicates bilinear interpolation upsampling. This indicates concatenation of channel dimensions.

[0016] D(l-1)↑ = u 2× (D (l) )(17) In the convolutional block processing of the decoder, a convolution kernel with a size of 3*1 is used Sliding along the time dimension with a fixed step of 1, adopting a symmetric padding strategy, the padding size = 1; Ensure that the output time dimension length is unchanged, and cooperate with the batch normalization layer and the ReLU activation function to obtain the output D (l-1) .

[0017] Step four, loss function design and training optimization: In order to overcome the convergence deviation of the model caused by the energy intensity difference of longitudinal wave, transverse wave and Stoneley wave, a waveform intensity adaptive weighting mechanism is introduced, which respectively configures weight coefficients w P , w S , w ST , which are negatively correlated with the typical energy intensity of the corresponding waveform, and the loss function is defined as follows: Where, X is the target waveform, w X is the configurable weighting coefficient, w P +w S +w ST =1, N is the total number of single-channel signal sampling points, P X(t) is the Gaussian probability label, P output is the probability prediction value of the model output. In the training process, first, calculate the three-channel probability distribution sequence P output_P , P output_S , P output_ST output by the model through forward propagation, and then update the weight parameters of the network through the back propagation algorithm, and gradually minimize the value of the loss function; In this way, the first arrival time extraction of the acoustic logging signal is completed.

[0018] The advantages of the present application are: The U-shaped network first arrival time extraction method constructs a U-shaped encoding and decoding network, combines a bidirectional convolutional long short-term memory module (BiConvLSTM) to capture the forward and backward spatiotemporal features of the acoustic signal, and enhances the modeling ability of the first wave propagation law. In addition, a multi-scale attention module (ATT) is introduced in the U-shaped encoding and decoding network, which adaptively focuses on the key features of the first wave, effectively suppresses the interference of environmental noise, and solves the problems of large parameter setting error and poor noise immunity caused by traditional artificial experience. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1A schematic diagram of the network model of the present application; Figure 2 A true implementation of the monitoring record and first arrival time extraction result map; Figure 3 A loss curve of different optimizers changing with rounds for training; Figure 4 A verification accuracy curve of different optimizers changing with rounds; Figure 5 A real acoustic signal from an actual oil field; Figure 6 A first arrival time extraction result of the U-Net method; Figure 7 A first arrival time extraction result of the SegNet method; Figure 8 A first arrival time extraction result of the U-Net+ method; Figure 9 A first arrival time extraction result of the method in this paper; Figure 10 A synthetic acoustic signal after adding Gaussian white noise; Figure 11 A first arrival time extraction result of the U-Net method; Figure 12 A first arrival time extraction result of the SegNet method; Figure 13 A first arrival time extraction result of the U-Net+ method; Figure 14 A first arrival time extraction result of the method in this paper; Figure 15 The training loss of different methods on the training set; Figure 16 The training error of different methods on the training set; Figure 17 The verification loss of different methods on the verification set; Figure 18 The accuracy of different methods on the verification set for P-wave first arrival time extraction; Figure 19 The accuracy of different methods on the verification set for S-wave first arrival time extraction Figure 20 The accuracy of different methods on the verification set for ST-wave first arrival time extraction. DETAILED DESCRIPTION

[0020] A U-shaped network first arrival time extraction method, comprising the following steps: Step one, acoustic data preprocessing and data enhancement: In actual array sonic logging work, due to the influence of environmental factors (such as formation complexity, wellbore conditions, etc.) and instrument factors (such as sensor accuracy, signal interference, etc.), the collected acoustic wave data often contains invalid data segments, abnormal fluctuations, noise interference and signal deviation. In order to ensure the reliability of subsequent analysis and the effectiveness of model training, the original data must be strictly preprocessed. This process specifically includes: (2) Filter processing: the first wave arrival time extraction is a key step in array sonic signal processing, and its extraction accuracy directly affects the calculation results of formation parameters such as porosity and anisotropy. Array sonic signals contain P waves, S waves and ST waves, and the frequency ranges of these waves are relatively close. When filtering data, the frequency band characteristics are required to be high, and the traditional filtering method is easy to cause the loss of signal details. The cosine bandpass filter avoids the problem of signal loss by constructing a smooth transition window function in the frequency domain, providing more complete information support for accurate first wave arrival time extraction, and its mathematical expression is: Where f represents the signal frequency, f1 represents the first cutoff frequency, signals below this frequency are completely filtered out, f2 represents the second cutoff frequency, f1 and f2 are connected by the cosine window function to realize low-pass filtering, f3 represents the third cutoff frequency, f2 and f3 form an ideal passband region, f4 represents the fourth cutoff frequency, f3 and f4 are connected by the cosine window function to realize high-pass filtering, and signals with frequency higher than f4 are completely filtered out.

[0021] (2) Data standardization: in order to eliminate the dimensional differences between different instruments and different channels, the acoustic wave signal is subjected to Z-score standardization processing, and the data distribution is adjusted to a normal distribution with a mean of 0 and a standard deviation of 1, and its mathematical expression is: s i represents the original acoustic wave data, s i ' represents the standardized data, N represents the total number of single-channel signal sampling points (1000 in the experiment), μ represents the signal mean, and σ represents the signal standard deviation.

[0022] (3) Data augmentation: In order to ensure that the deep learning model has stronger robustness and generalization ability in actual application, data augmentation strategy is further adopted to enrich the diversity of experimental samples. Specifically, on the basis of the preprocessed data set, a certain proportion of noise data simulating real scenes is artificially and controllably introduced. This strategy not only greatly enriches the diversity of training samples, but also enables the model to learn and adapt to noise patterns and interference characteristics in complex and variable environments during the training process.

[0023] Let the pre-processed normalized acoustic wave signal be z(t), and the signal data after data enhancement be z aug (t), the data enhancement operation can be expressed as: z aug (t) = z(t) + n g (t) (5) n g (t) ~ N(0, σ g 2 ), σ g = α·max(|z(t)|) (6) Wherein n g (t) is a Gaussian white noise, σ g is the noise standard deviation, and α is the noise intensity coefficient, which is in the range of 0.05≤α≤0.15.

[0024] Step two, acoustic wave data label processing: For single-channel waveform in array acoustic logging data, the traditional label generation method is to mark the first arrival time of artificial picked P wave, S wave and Stoneley wave (ST wave) as 1, and mark the rest time points as 0. However, the arrival point label obtained in this way is sparse relative to the non-arrival point label, which is not conducive to model learning of the first wave arrival feature, and at the same time, the first wave arrival time point picked artificially is influenced by subjective factors and signal quality, and there is a certain error. In order to overcome the above defects, the present application proposes a label generation method based on probability distribution: For P wave, S wave and Stoneley wave (ST wave), a Gaussian probability density function is constructed with the first wave arrival time picked artificially as the distribution center, and for time sampling point t, the probability P X(t) is defined as: Wherein, t X represents the first wave arrival time of the waveform X picked artificially (t P , t S , t ST ), σ X is the preset standard deviation parameter of the waveform X (σ P , σ S , σ ST ), which is used to control the width of the probability distribution. At the picked first wave arrival time point t X , the probability P X(t) reaches the maximum value, at the time point adjacent to t X , P X(t) smoothly decays with the increase of t X , and at the time point far away from t X , the probability tends to zero.

[0025] Step three, network model: The present application takes U-Net network as the basic architecture (Ronneberger et al. (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation), and for the first wave arrival time extraction task of sound wave signal, the core innovation is realized by embedding BiConvLSTM module and multi-scale attention module (ATT). The innovation points are as follows: the output of the first layer of the encoder and the output of the last layer of the decoder are fused through BiConvLSTM, the directional dependence of sound wave propagation is captured by using the bidirectional space-time modeling capability, and the environmental noise interference is effectively suppressed; further, the multi-scale attention module (ATT) focuses on the short-time high-frequency feature (the first wave take-off point) and suppresses the subsequent noise, which significantly improves the signal-to-noise ratio of the key signal area, so as to optimize the positioning accuracy and robustness of the first wave arrival time, and the specific processing process is as follows: Let the preprocessed sound wave signal be X ∈ R B×C×T , B, C, T represent batch size (positive integer, the experimental value is 32), channel number (the number of receivers, the experimental value is 8), and time sequence length (the experimental value is 1000). The encoder takes the sound wave signal X ∈ R B×C×T as the input, and extracts multi-scale features through multi-level convolution block processing and down-sampling layer alternation operation.

[0026] In the processing process of the convolution block, the one-dimensional convolution kernel slides along the time dimension to extract waveform features (kernel size is fixed as 3x1, step is 1, and padding is 1), and the output is sequentially subjected to batch normalization layer (BN) and ReLU activation function. The channel number of the convolution block is configured according to the hierarchical progression rule, and is increased to 64, 128, 256, 512, and 1024 channels in turn, forming a multi-scale feature extraction structure. In the down-sampling process, a maximum pooling down-sampling layer MaxPool 1d is connected after each convolution block (kernel size is fixed as 2x1, step is 2), which makes the output time dimension constant as half of the input (i.e. T out = T in / 2). The above "convolution block→down-sampling" combination is repeatedly executed 6 times to form a cyclic iteration processing structure, and finally realizes a 6-level compression path of the time dimension: T→T / 2→T / 4→T / 8→T / 16→T / 32→T / 64.

[0027] E (l)↓ = MaxPool 1d (E (l) )(9) In Equations 8 and 9, E (l-1) E represents the feature map input to the (l-1)th convolutional block. (l) E represents the feature map output by the l-th convolutional block. (l)↓ This represents the feature map output by the l-th layer after downsampling. represents a one-dimensional convolution kernel, * represents a one-dimensional convolution operation, BN represents batch normalization, and ReLU is the modified linear unit activation function.

[0028] The first convolutional block output feature E (1) With the decoder's last layer upsampled output D last After channel concatenation, the data are input into the BiConvLSTM module, where forward and backward propagation operations are performed simultaneously to form a new feature map with a complete temporal context.

[0029] F = W f *H f +W b *H b +b(12) E (1) D last These are the outputs of the first layer of the encoder and the last layer of the decoder, respectively. Represents feature concatenation, ConvLSTM f and ConvLSTM b H represents the forward convolution LSTM unit and the backward convolution LSTM unit, respectively. f H b These represent the forward hidden state sequence and the reverse hidden state sequence, respectively, used to record the forward / reverse propagation characteristics of sound waves. W f W b represents the forward weight kernel and the reverse weight kernel, respectively; * represents the convolution operation; b is the bias term; and F represents the final output feature.

[0030] The feature maps F output by the second and third convolutional blocks (2) F (3) The input is directly fed into the multi-scale attention module (ATT) with the feature map F. (3) Taking the processing procedure as an example, F (2) Similarly: by using 3×1 convolution kernels Conv1D in parallel 3×1 and a 5×1 convolution kernel Conv1D 5×1 The high-frequency feature F3 and waveform envelope feature F5 of the first wave transition are captured respectively, and finally the two features are adaptively fused for output, where α∈[0,1].

[0031] F3 = Conv1D 3×1 (F (3) )(13) F5 = Conv1D 5×1 (F (3) )(14) F ATT = a · F3 + (1 - a) · F5 (15) In the deep feature initialization stage of the decoder, the deep features E (l) output by the encoder are input into the decoder, first using a transpose convolution kernel with a size of 1x1 to compress the number of channels to half, then normalizing the feature distribution through a batch normalization layer (BN), and introducing non-linear changes through a ReLU activation function to obtain the output D (l) . Wherein, represents the transpose convolution kernel weight of the decoder l layer.

[0032] In the upsampling process of the decoder, the time dimension is expanded by 2, and after upsampling, the output D (l-1)↑ is obtained, followed by channel splicing of the encoder output features E (l-1) and the decoder output features D (l-1)↑ in the same layer through a jump connection, outputting the fused feature map Wherein, u 2× represents bilinear interpolation upsampling, represents channel dimension splicing.

[0033] D (l-1)↑ = u 2× (D (l) )(17) In the convolution block processing of the decoder, a convolution kernel with a size of 3x1 is used to slide along the time dimension with a fixed step size of 1, and a symmetric padding strategy (padding size = 1) is used to ensure that the output time dimension length remains unchanged, and a batch normalization layer (BN) and a ReLU activation function are used to obtain the output D (l-1) .

[0034] Step four, loss function design and training optimization: To overcome the convergence deviation of the model caused by the energy intensity difference of P wave, S wave and ST wave, a waveform intensity adaptive weighting mechanism is introduced, which respectively configures weight coefficients w P , w S , w ST for each waveform, which are negatively correlated with the typical energy intensity of the corresponding waveform, and the loss function is defined as follows: wherein X is a target waveform, w X is a configurable weighting coefficient, satisfying w P +w S +w ST =1, N is the total number of single-channel signal sampling points, P X(t) is a Gaussian probability label, P output is a probability prediction value output by the model. In the training process, firstly, a three-channel probability distribution sequence P output_P , P output_S , P output_ST output by the model is calculated through forward propagation, and then the weight parameters of the network are updated through the back propagation algorithm (https: / / blog.csdn.net / leonardotu / article / details / 136509444?ops_request_misc=%257B%2522request%255Fid%2522%253A%25220196716baaa875772097477344480af7%2522%252C%2522scm%2522%253A%252220140713.130102334..%2522%257D&request_id=0196716baaa875772097477344480af7&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~all~top_positive~default-1-136509444-null-null.142^v102^control&utm_term=%E5%8F%8D%E5%90%91%E4%BC%A0%E6%92%AD%E7%AE%97%E6%B3%95&spm=1018.2226.3001.4187) to gradually minimize the value of the loss function; Thus, the first arrival time extraction of the acoustic logging signal is completed.

[0035] Compared with the existing manual picking method and the conventional automatic picking algorithm, the application can effectively extract the time sequence features of the sound wave signal by combining the time sequence feature extraction ability of the bidirectional convolution long short-term memory module (BiConvLSTM) and the feature refinement ability of the multi-scale attention module (ATT), and maintain high extraction accuracy in a low signal-to-noise ratio environment. Experiments show that the accuracy of the method in the present application in the first wave arrival time extraction of P wave, S wave and ST wave reaches 95.202%, 98.259% and 99.193% respectively, which is significantly better than the existing STA / LTA algorithm and the convolution neural network-based segmentation network.

[0036] In addition, the application automatically extracts the first wave arrival time of the sound wave signal through the deep learning model, avoids the time-consuming and subjectivity of the traditional manual picking method, greatly improves the processing efficiency. At the same time, the model can automatically adapt to the sound wave signal processing under different signal-to-noise ratios and complex geological conditions, reducing the need for artificial intervention and parameter optimization, and providing the possibility for real-time processing of acoustic logging data.

[0037] In order to verify the feasibility of the present application, the applicant has carried out the following experiments: In the model training stage, in order to determine the optimal parameter update strategy, the applicant designed and carried out a strict optimizer comparison experiment. Under the condition of the same model architecture, data set (containing fixed training set and validation set), batch size (64), loss function, four mainstream optimizers (Adam, RMSprop, SGD, Adagrad) were used for model training.

[0038] All experiments were carried out on the same hardware platform (NVIDIA A100 GPU) and the same deep learning framework (PyTorch2.0.1) environment, and each experiment was independently run to the preset maximum training round (100 rounds). The model performance was evaluated by the training set loss and the validation set accuracy, and finally, by comparing the performance indicators of each optimizer under the same conditions, the Adam optimizer was selected for parameter optimization. The training loss and validation accuracy results using different optimizers (see the accompanying drawings Figure 3 and 4 ).

[0039] On the basis of parameter optimization of the Adam optimizer, the key parameters of the optimizer are set as follows through experimental tuning: initial learning rate 0.001, weight decay (L2 regularization coefficient) 0.00001, exponential decay rate of first moment estimation (β1) 0.9, and exponential decay rate of second moment estimation (β2) 0.99. Considering the GPU memory capacity and input image size, the batch size is set to 64. To accelerate the training convergence and optimize the performance, a step learning rate decay strategy (StepLR) is adopted: the learning rate is multiplied by the decay factor 0.1 to be adjusted down every 10 epochs.

[0040] To comprehensively evaluate the effectiveness and generalization ability of the proposed method (U-shaped first wave arrival time extraction network), the experimental data set is composed of measured data and synthetic data (see the data enhancement part of step 1 in the technical solution for the synthesis method), totaling 10,000 sample data. Specifically, it contains 8,000 measured acoustic signals and 2,000 artificially synthesized acoustic signals. The mixed data set is randomly divided into a training set (6,000), a validation set (2,000), and a test set (2,000) in a ratio of 6:2:2. The real acoustic signals of the actual oil field and the synthetic signals after adding Gaussian white noise (see the attached drawings Figure 5 and 10 ) in the specification.

[0041] In the comparative verification aspect, the mainstream convolutional neural network architecture is selected as the benchmark model, including U-Net, SegNet, and U-Net+. All comparative models are trained and evaluated under the same training rounds, the same training set, and the same validation set. From the loss results, the loss of the model on the training set is 0.88% (U-Net), 0.82% (Segnet), 0.81% (U-Net+), and 0.72% (the method of this paper), respectively. The loss of the model on the validation set is 0.89% (U-Net), 0.87% (Segnet), 0.82% (U-Net+), and 0.78% (the method of this paper), respectively. The extraction of the three waves (P wave, S wave, and ST wave) by the four models is as follows: the first wave arrival time extraction accuracy of the P wave is 75% (U-Net), 86% (Segnet), 90% (U-Net+), and 95% (the method of this paper), respectively; the first wave arrival time extraction accuracy of the S wave is 90% (U-Net), 91% (Segnet), 96% (U-Net+), and 98% (the method of this paper), respectively; and the first wave arrival time extraction accuracy of the ST wave is 96% (U-Net), 97.5% (Segnet), 97.8% (U-Net+), and 99% (the method of this paper), respectively. The performance results of different methods in the training set and the validation set (see the attached drawings Figure 15 、 16, 17, 18, 19, 20), the results of P-wave, S-wave, and ST-wave first arrival extraction in actual oilfield data by different methods (see the description of the accompanying drawings Figure 6 、 7 , 8, 9), the results of P-wave, S-wave, and ST-wave first arrival extraction in synthetic data by different methods (see the description of the accompanying drawings Figure 11 、 12 , 13, 14).

[0042] The experimental results show that, under the same training conditions: The U-shaped first arrival extraction network proposed in this paper achieves the highest overall accuracy (i.e., the lowest validation loss of 0.78%) on the validation set, significantly outperforming the comparative U-Net, SegNet, and U-Net+ models.

[0043] In the key P-wave, S-wave, and ST-wave first arrival extraction task, the method proposed in this paper exhibits the optimal accuracy performance (P-wave: 95.202%, S-wave: 98.259%, ST-wave: 99.193%), significantly surpassing all the comparative models.

[0044] The U-shaped network first arrival extraction method constructs a U-shaped encoding-decoding network, combines a bidirectional convolutional long short-term memory module to capture the forward and backward spatiotemporal features of the sound wave signal, enhances the modeling ability of the first wave propagation law, and, in addition, introduces a multi-scale attention module in the U-shaped encoding-decoding network to adaptively focus on the key features of the first wave, effectively suppresses the interference of environmental noise, and solves the problem of large parameter setting error and poor noise resistance caused by traditional manual experience.

Claims

1. A method for extracting the first arrival time of a U-shaped network, characterized in that: The first wave arrival time extraction method of the U-shaped network comprises the following steps: Step one, sound wave data preprocessing and data enhancement: Step two, sound wave data label processing: For the P-wave, S-wave and ST-wave, the Gaussian probability density function is constructed with the manually picked first arrival time as the distribution center, for the time sampling point t, which belongs to the wave type X, X can be P, S, ST; the probability P of the first arrival event X(t) is defined as: wherein t X represents the first arrival time (t P , t S , t ST ) of the first wave of the waveform X picked up by the human being; σ X is a standard deviation parameter (σ P , σ S , σ ST ) preset for the waveform X and used to control the width of the probability distribution; at the picked-up first arrival time point t X , the probability P X(t) reaches a maximum value, at the time points adjacent to t X , P X(t) smoothly decays with the increase of the distance of t X , and at the time points far away from t X , the probability approaches zero; Step three, network model: The application takes U-Net network as the basic framework, and aims at the first wave arrival time extraction task of sound wave signal. The core innovation is realized by embedding BiConvLSTM module and multi-scale attention module, and the innovation points are as follows: the output of the first layer of the encoder and the output of the last layer of the decoder are fused through BiConvLSTM, the directional dependence of sound wave propagation is captured by using the bidirectional space-time modeling capability, and the environmental noise interference is effectively suppressed; further, the multi-scale attention module focuses on short-time high-frequency features and suppresses subsequent noise, significantly improves the signal-to-noise ratio of the key signal area, and optimizes the positioning accuracy and robustness of the first wave arrival time; Step four, loss function design and training optimization: In order to overcome the model convergence deviation caused by the energy intensity difference of longitudinal waves, transverse waves and Stoneley waves, the waveform intensity self-adaptive weighting mechanism is introduced, the mechanism respectively configures weight coefficients w P 、w S 、w ST for each waveform, the value of which is negatively correlated with the typical energy intensity of the corresponding waveform, and the loss function is defined as follows: wherein X is a target waveform, w X is a configurable weighting coefficient, satisfying w P +w S +w ST =1, N is the total number of single-channel signal sampling points, P X(t) is a Gaussian probability label, P output is a probability prediction value output by the model; in the training process, firstly, a three-channel probability distribution sequence P output_P , P output_S , P output_ST output by the model is calculated through forward propagation, and then the weight parameters of the network are updated through a back propagation algorithm, and the value of the loss function is gradually minimized. In this way, the first wave arrival time extraction of the sound wave logging signal is completed.

2. The method of claim 1, wherein: The sound wave data preprocessing and data enhancement in step one comprises the following steps: (1) filtering processing: to provide more complete information support for the accurate extraction of the first wave arrival time, the mathematical expression is: Wherein, f represents the signal frequency, f1 represents the first cutoff frequency, the signal below this frequency is completely filtered out, f2 represents the second cutoff frequency, f1 and f2 realize low-pass filtering through the cosine window function, f3 represents the third cutoff frequency, f2 and f3 form an ideal passband region, f4 represents the fourth cutoff frequency, f3 and f4 realize high-pass filtering through the cosine window function, and the signal with frequency higher than f4 is completely filtered out; (2) data standardization: the sound wave signal is processed by Z-score standardization, the data distribution is adjusted to normal distribution with mean value of 0 and standard deviation of 1, and the mathematical expression is: s i s represents the original acoustic wave data i N represents the total number of single-channel signal sampling points (1000 in the experiment), μ represents the signal mean value, and σ represents the signal standard deviation (3) data enhancement.

3. The method of claim 2, wherein: The steps of data enhancement are as follows: Let the pre-processed normalized acoustic signal be z(t), and the signal data after data enhancement be z aug (t), and the data enhancement operation can be expressed as: z aug (t) = z(t) + n g (t) (5) In the formula, n g (t) is a Gaussian white noise, σ g is a noise standard deviation, and a is a noise intensity coefficient, which is in a range of 0.05≤α≤0.15。 4. The method of claim 1, wherein: The specific processing process of step three is as follows: Let the pre-processed acoustic signal be X ∈ R B×C×T , B, C, T represent batch size, channel number, and time sequence length, respectively; the encoder takes the acoustic signal X ∈ R B×C×T as input, extracts multi-scale features through multi-level convolution block processing and alternating operation with down-sampling layers; In the processing of the convolution block, a one-dimensional convolution kernel slides along the time dimension to extract waveform features, the kernel size is fixed at 3x1, the step is 1, and the padding is 1; the output successively passes through a batch normalization layer and a ReLU activation function, the channel number of the convolution block is configured according to a hierarchical progression rule, and is sequentially increased to 64, 128, 256, 512, and 1024 channels, forming a multi-scale feature extraction structure; in the down-sampling process, a maximum pooling down-sampling layer MaxPool 1d is connected after each convolution block, the kernel size is fixed at 2x1, and the step is 2; this operation makes the output time dimension constant at half of the input, i.e., T out =T in / 2; the above-mentioned "convolution block -> down-sampling" combination is repeatedly executed 6 times, forming a cyclic iteration processing structure, and finally realizing a 6-level compression path of the time dimension: T -> T / 2 -> T / 4 -> T / 8 -> T / 16 -> T / 32 -> T / 64; E (l)↓ = MaxPool 1d (E (l) )(9) In formula 8 and formula 9, E (l-1) represents the feature map input by the (l-1)th convolutional block, E (l) represents the feature map output by the lth convolutional block, E (l)↓ represents the feature map output by the lth convolutional block after down-sampling, represents a one-dimensional convolution kernel, * represents a one-dimensional convolution operation, BN represents a batch normalization operation, and ReLU is a rectified linear unit activation function. The first convolutional block outputs a feature E (1) The output D of the last layer of the decoder is upsampled last After channel splicing, the common input is input into the BiConvLSTM module, and forward propagation and backward propagation operations are performed to form a new feature map with complete timing context. F = W f *H f +W b *H b +b(12) E (1) , D last are the output of the first layer of the encoder and the output of the last layer of the decoder, respectively, denotes feature concatenation, ConvLSTM f and ConvLSTM b denote forward and backward convolutional LSTM units, respectively, H f , H b denote forward and backward hidden state sequences, respectively, used to record the forward / backward propagating features, W f , W b denote forward and backward weight kernels, respectively, * denotes convolution operation, b is the bias term, and F denotes the final output feature; The feature map F output by the second layer and third layer convolutional blocks (2) , F (3) is directly input to a multi-scale attention module (ATT) to obtain a feature map F (3) . The processing process of F (2) is as follows: by using a 3x1 convolution kernel Conv1D 3×1 and a 5x1 convolution kernel Conv1D 5×1 in parallel, high-frequency features F3 of the first wave jump and waveform envelope features F5 are captured respectively, and finally the features of the two are adaptively fused for output, where α∈[0,1]. F3 = Conv1D 3×1 (F (3) )(13) F5 = Conv1D 5×1 (F (3) )(14) F ATT = a - F3+ (1 - a) - F5(15) In the deep feature initialization stage of the decoder, the deep feature E (l) output by the encoder is input into the decoder, first compressed to half of the channel number using a transpose convolution kernel with a size of 1x1, then standardized by a batch normalization layer, and introduced into a nonlinear change by a ReLU activation function to obtain the output D (l) ; wherein, represents the transpose convolution kernel weight of the decoder l layer. In the up-sampling process of the decoder, the time dimension is expanded by 2 times, and the output D is obtained after up-sampling (l-1)↑ Then, the output features E (l-1) of the same layer of the encoder are connected in a skip connection manner, and the output features D (l-1)↑ of the decoder are channel spliced to output the fused feature map Wherein, u 2× represents bilinear interpolation up-sampling, represents channel dimension splicing; In the convolution block processing of the decoder, a convolution kernel with a size of 3x1 is used Sliding along the time dimension with a fixed step of 1, using a symmetric padding strategy, and the padding size is 1; ensure that the output time dimension length is unchanged, and cooperate with the batch normalization layer and the ReLU activation function to obtain the output D (l-1) ;

Images