Multiple longitudinal and transverse wave joint detection method based on seabed crawl device

Abstract

The invention provides a multiple longitudinal and transverse wave joint detection method based on a seabed crawler, which belongs to the technical field of geophysical exploration, and constructs an original longitudinal and transverse wave signal matrix by transmitting multi-frequency longitudinal and transverse wave signals and receiving stratum reflection signals. Fourier transform is utilized to extract frequency characteristics, physical analysis is carried out in combination with a wave equation, and a longitudinal and transverse wave interference matrix is constructed and weighted fusion is carried out. And constructing an error compensation matrix according to the crawl device attitude data, and performing interference compensation on the joint interference matrix. And performing time-frequency analysis through wavelet transform, extracting specific time-frequency features, inputting the specific time-frequency features into a pre-trained stratum classification prediction model, realizing anti-interference prediction on the type and thickness of the seabed stratum structure, finally drawing a seabed three-dimensional geological profile map according to a prediction result, and marking a key geological structure and a resource distribution region. The technical problem that accurate recognition and resource distribution positioning of the seabed stratum structure are affected due to much external interference in the seabed geological exploration process is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of geophysical detection, and in particular, relates to a multiple longitudinal and transverse wave joint detection method based on a seafloor crawler. Background Art

[0002] Seabed geological exploration is an important prerequisite for the development of marine resources and the construction of seabed engineering. Traditional seabed geological exploration mainly relies on ship-borne sonar, seismic waves or a single type of wave for detection. These methods usually use a single wave source with a fixed frequency to transmit detection signals from the water surface downward and analyze the seabed stratum structure through the echo signal. In practical applications, this type of technology has been widely used in offshore oil exploration, submarine cable laying route planning, and deep-sea mineral resource surveys.

[0003] However, traditional seabed exploration methods face serious interference problems. First, the signal emitted from the surface needs to pass through the entire water body to reach the seabed, and is interfered by factors such as waves, currents, and thermoclines during the process, resulting in severe signal attenuation and distortion; second, a single type of wave source (such as pure longitudinal waves or pure transverse waves) has limited ability to distinguish different geological structures, and the detection results of complex geological structures are easily interfered by the surrounding environmental noise; third, fixed-site or ship-borne detection methods are affected by factors such as ship shaking and positioning errors, making it difficult to obtain stable and continuous detection data.

[0004] These various interference factors seriously affect the accurate identification of the seabed stratigraphic structure and the accurate positioning of resource distribution, especially in areas with complex geological structures and changeable sea conditions. Traditional detection methods are difficult to eliminate the impact of various interferences on detection accuracy, thus limiting the effective exploration and development and utilization of seabed resources. In other words, there are many external interferences in the process of seabed geological exploration in the existing technology, which affects the technical problem of accurate identification of seabed stratigraphic structure and positioning of resource distribution. Summary of the invention

[0005] In view of this, the present invention provides a multiple longitudinal and transverse wave joint detection method based on a seabed crawler, which can solve the technical problems in the prior art of multiple external interferences in the process of seabed geological exploration, which affect the accurate identification of seabed stratum structure and resource distribution positioning.

[0006] The present invention is implemented as follows: The present invention provides a multiple P-wave and S-wave joint detection method based on a seabed crawler, including: deploying a multiple P-wave and S-wave joint detection device on the seabed crawler, so that the seabed crawler crawls along a preset detection path and emits multi-frequency P-wave and S-wave signals; receiving P-wave and S-wave signals reflected from the seabed strata, and constructing an original P-wave and S-wave signal matrix; extracting signal frequency characteristics, and constructing a P-wave and S-wave interference matrix; weightedly fusing the P-wave interference matrix and the S-wave interference matrix to generate a joint interference matrix; constructing an interference error compensation matrix based on real-time attitude data recorded by the seabed crawler attitude sensor, and compensating the joint interference matrix; performing time-frequency analysis through wavelet transform to extract specific time-frequency features; inputting the specific time-frequency features into a pre-trained stratum classification prediction model to realize classification of seabed stratum structure types and thickness prediction; drawing a three-dimensional seabed geological profile, and marking key geological structures and resource distribution areas.

[0007] Among them, the multiple longitudinal and transverse wave joint detection device is composed of a multi-frequency wave source generator, a longitudinal wave transmitting unit, a transverse wave transmitting unit, an omnidirectional receiving sensor array, a signal processing module and a data storage unit. It is installed at the bottom of the seabed crawler and can simultaneously transmit multi-frequency longitudinal waves and transverse waves, and receive reflected signals.

[0008] Among them, the P-wave interference matrix refers to the interference characteristic matrix generated by the propagation of P-waves in different geological structures, which is constructed by extracting the amplitude and phase of each frequency component after performing spectral analysis on the original P-wave signal matrix; the S-wave interference matrix refers to the interference characteristic matrix generated by the propagation of S-waves in different geological structures, which is constructed by extracting the amplitude and phase of each frequency component after performing spectral analysis on the original S-wave signal matrix; the joint interference matrix refers to the matrix formed by combining the P-wave interference matrix and the S-wave interference matrix through a weighted fusion algorithm, which comprehensively reflects the comprehensive interference characteristics of P-wave and S-wave propagation in seabed strata.

[0009] Among them, the interference error compensation matrix refers to the compensation matrix calculated based on the pitch angle, roll angle and depth data recorded by the seabed crawler attitude sensor, which is used to eliminate the interference caused by the seabed crawler attitude change on signal acquisition.

[0010] Among them, the specific time-frequency characteristics refer to the joint characteristics of time domain and frequency domain that can reflect the unique response characteristics of different geological structures, which are extracted after the time-frequency analysis of the joint interference matrix through wavelet transform.

[0011] Among them, the submarine stratum wave equation refers to the partial differential equation that describes the propagation characteristics of P- and S-waves in the submarine strata; the submarine stratum wave equation is used to calculate the propagation speed and attenuation characteristics of P- and S-waves in different geological structures; the input of the submarine stratum wave equation includes the stratum density obtained from the submarine crawler, the first Lamé parameter obtained from the rock sampling analysis of the submarine crawler, the second Lamé parameter obtained from the rock sampling analysis of the submarine crawler, and the displacement vector calculated from the original P-wave signal matrix and the original S-wave signal matrix; the output is the wave displacement field distribution and the wave energy attenuation law, which are used to construct the P-wave interference matrix and the S-wave interference matrix.

[0012] Among them, the layer classification prediction model refers to a neural network model used for stratigraphic structure recognition and thickness prediction; the specific structure of the stratigraphic classification prediction model is a U-shaped convolutional neural network based on the self-attention mechanism, which includes an encoder and a decoder; the encoder consists of five convolutional layers and four pooling layers, and the number of convolution kernels in each layer is 32, 64, 128, 256, and 512, respectively, and the convolution kernel size is 3×3; the decoder consists of five deconvolutional layers and four splicing layers, and the number of feature maps in each layer is the same as that of the corresponding layer of the encoder.

[0013] Among them, a self-attention mechanism is introduced between the encoder and the decoder to enhance the feature extraction capability. The sparse attention coefficient in the self-attention mechanism is dynamically adjusted according to three key parameters: the ratio of the P-wave and S-wave propagation velocities calculated from the P-wave interference matrix and the S-wave interference matrix, the stratum density obtained from the seabed crawler, and the seabed depth obtained from the seabed crawler.

[0014] Among them, the steps of establishing the training data set in the process of training the stratigraphic classification prediction model include collecting longitudinal and transverse wave reflection data of seabed strata of various geological structures from different sea areas, preprocessing the collected data to remove noise and outliers, using manual labeling to determine the true value of the stratigraphic type and thickness, and expanding the number of training samples through data enhancement technology.

[0015] The steps of training the formation classification prediction model include first initializing the formation classification prediction model using the preprocessed training data set, using the Adam optimizer for parameter optimization, setting the initial learning rate to 0.001, and decaying the learning rate to 0.1 times the original value every 50 training cycles. The loss function uses a weighted combination of the formation type classification cross entropy loss and the thickness prediction mean square error loss, with weights of 0.6 and 0.4 respectively.

[0016] Compared with the prior art, the present invention provides a method for joint detection of multiple P- and S-waves based on a seafloor crawler. The method proposed by the present invention realizes direct detection close to the seafloor by deploying a detection device that can simultaneously transmit and receive multi-frequency P- and S-waves on the seafloor crawler. The method uses a seafloor crawler to crawl along a preset path, continuously transmit multi-frequency P- and S-wave signals, and receive signals reflected from the formation to construct an original P- and S-wave signal matrix.

[0017] Frequency features were extracted through Fourier transform, and combined with the physical mechanism analysis of the seabed stratum wave equation, the longitudinal and transverse wave interference matrices were constructed and weighted fusion was performed. At the same time, the interference error compensation matrix was constructed based on the real-time attitude data recorded by the crawler attitude sensor, which effectively eliminated the signal interference introduced by the crawler attitude change. The compensated joint interference matrix was analyzed in time and frequency using wavelet transform to extract specific time and frequency features. Combined with the pre-trained stratum classification prediction model, the accurate prediction of the seabed stratum structure type and thickness was achieved.

[0018] The present invention effectively overcomes various interference factors such as water propagation interference, environmental noise interference, and equipment attitude interference in traditional exploration through multiple longitudinal and transverse wave joint detection and multi-level interference compensation technology, and significantly improves the anti-interference ability and detection accuracy of seabed geological exploration. In particular, the stratum classification prediction model combined with the deep learning method can achieve stable and reliable stratum structure identification and resource distribution positioning in a complex and changeable seabed environment, solving the technical problem of many external interferences in the process of seabed geological exploration in the prior art, which affects the accurate identification of seabed stratum structure and resource distribution positioning, and provides more accurate technical support for the efficient development of seabed resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention.

[0020] Figure 2 This is a schematic diagram of the composition of the multiple longitudinal and shear wave joint detection device in Example 1 of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0022] like Figure 1 FIG. 1 is a flowchart of a method for joint detection of multiple longitudinal and transverse waves based on a seafloor crawler provided by the present invention. The method comprises the following steps:

[0023] S01. deploying a multiple P-wave and S-wave joint detection device on a seafloor crawler, causing the seafloor crawler to crawl along a preset detection path, and simultaneously starting the multiple P-wave and S-wave joint detection device to transmit multi-frequency P-wave and S-wave signals to the seafloor formation;

[0024] S02, using the multiple P-wave and S-wave joint detection device to receive P-wave and S-wave signals reflected from the seabed strata, and converting them into digital signals to construct an original P-wave signal matrix and an original S-wave signal matrix;

[0025] S03, extracting the frequency characteristics of the original P-wave signal matrix and the original S-wave signal matrix by Fourier transform, constructing the P-wave interference matrix and the S-wave interference matrix, and using the seabed formation wave equation to perform physical mechanism analysis on the P-wave and S-wave propagation;

[0026] S04, performing weighted fusion of the longitudinal wave interference matrix and the transverse wave interference matrix to generate a joint interference matrix, wherein the weight coefficient is dynamically adjusted based on seabed geological conditions;

[0027] S05, constructing an interference error compensation matrix according to the real-time attitude data recorded by the attitude sensor of the seabed crawler, and performing compensation processing on the joint interference matrix;

[0028] S06, performing time-frequency analysis on the compensated joint interference matrix through wavelet transform, extracting specific time-frequency features, and establishing a seabed stratum structure feature recognition model;

[0029] S07, inputting the specific time-frequency features into a pre-trained stratum classification prediction model to achieve classification of seabed stratum structure types and thickness prediction;

[0030] S08. Based on the stratigraphic structure type and the thickness prediction result, draw a three-dimensional geological profile of the seabed, and mark key geological structures and resource distribution areas.

[0031] Among them, the multiple longitudinal and transverse wave joint detection device is specifically composed of a multi-frequency wave source generator, a longitudinal wave transmitting unit, a transverse wave transmitting unit, an omnidirectional receiving sensor array, a signal processing module and a data storage unit. It is installed at the bottom of the seabed crawler and can simultaneously transmit multi-frequency longitudinal waves and transverse waves, and receive reflected signals.

[0032] The longitudinal wave interference matrix specifically refers to the interference characteristic matrix generated when the longitudinal wave propagates in different geological structures, which is constructed by extracting the amplitude and phase of each frequency component after performing spectral analysis on the original longitudinal wave signal matrix.

[0033] The shear wave interference matrix specifically refers to an interference characteristic matrix constructed by extracting the amplitude and phase of each frequency component after performing spectrum analysis on the original shear wave signal matrix, and characterizing the interference characteristic matrix generated when the shear wave propagates in different geological structures.

[0034] The joint interference matrix specifically refers to a matrix formed by combining the longitudinal wave interference matrix and the transverse wave interference matrix through a weighted fusion algorithm, which can comprehensively reflect the comprehensive interference characteristics of longitudinal and transverse waves propagating in the seabed strata.

[0035] Among them, the interference error compensation matrix specifically refers to a compensation matrix calculated based on the pitch angle, roll angle and depth data recorded by the seabed crawler attitude sensor, which is used to eliminate the interference caused by the seabed crawler attitude change on signal acquisition.

[0036] Among them, the specific time-frequency characteristics specifically refer to the joint characteristics of time domain and frequency domain that are extracted after the time-frequency analysis of the joint interference matrix through wavelet transform, and can reflect the unique response characteristics of different geological structures.

[0037] Among them, the seabed stratum wave equation specifically refers to the partial differential equation that describes the propagation characteristics of P- and S-waves in the seabed strata; the seabed stratum wave equation is used to calculate the propagation speed and attenuation characteristics of P- and S-waves in different geological structures, and the input includes the stratum density obtained from the seabed crawler, the first Lamé parameter obtained from the rock sampling analysis of the seabed crawler, the second Lamé parameter obtained from the rock sampling analysis of the seabed crawler, and the displacement vector calculated from the original P-wave signal matrix and the original S-wave signal matrix; the output is the wave displacement field distribution and the wave energy attenuation law, and the wave displacement field distribution and the wave energy attenuation law are used to construct the P-wave interference matrix and the S-wave interference matrix.

[0038] Among them, the stratum classification prediction model specifically refers to a neural network model used for stratum structure identification and thickness prediction; the specific structure of the stratum classification prediction model is a U-shaped convolutional neural network based on the self-attention mechanism, which includes an encoder and a decoder. The encoder consists of five convolutional layers and four pooling layers. The number of convolution kernels in each layer is 32, 64, 128, 256, and 512, respectively, and the convolution kernel size is 3×3. The decoder consists of five deconvolutional layers and four splicing layers. The number of feature maps in each layer is the same as that of the corresponding layer of the encoder, and a self-attention mechanism is introduced between the encoder and the decoder to enhance the feature extraction capability. The sparse attention coefficient in the self-attention mechanism is dynamically adjusted according to three key parameters: the ratio of the longitudinal and transverse wave propagation velocities calculated from the longitudinal wave interference matrix and the transverse wave interference matrix, the stratum density obtained from the seabed crawler, and the seabed depth obtained from the seabed crawler.

[0039] The steps of establishing the training data set in the training process of the stratigraphic classification prediction model specifically include collecting longitudinal and transverse wave reflection data of seabed strata of various geological structures from different sea areas, preprocessing the collected data to remove noise and outliers, using manual labeling to determine the true value of the stratigraphic type and thickness, and expanding the number of training samples through data enhancement technology, including adding different degrees of Gaussian noise, random cropping, and random rotation operations, and finally constructing a training set containing 100,000 sets of labeled data and a verification set containing 10,000 sets of labeled data.

[0040] The steps of training the formation classification prediction model specifically include first initializing the formation classification prediction model using the preprocessed training data set, using the Adam optimizer for parameter optimization, setting the initial learning rate to 0.001, and decaying the learning rate to 0.1 times the original value every 50 training cycles. The loss function adopts a weighted combination of the formation type classification cross entropy loss and the thickness prediction mean square error loss, with weights of 0.6 and 0.4 respectively. An early stopping mechanism is introduced during the training process to prevent overfitting. When the validation set loss does not decrease for 10 consecutive training cycles, the training is stopped, and finally the model parameters with the best performance on the validation set are selected as the final model.

[0041] The specific implementation of the above steps is described in detail below.

[0042] The specific implementation method of step S01 is to install a multiple longitudinal and transverse wave joint detection device at the bottom of the seabed crawler, and use a hydraulic drive mechanism to control the seabed crawler to crawl along a preset detection path. The preset detection path is pre-planned according to the geological characteristics of the detection target area, and generally a grid or spiral path is used to cover the target area, and the grid spacing is usually 5 to 10 meters. The multi-frequency wave source generator in the multiple longitudinal and transverse wave joint detection device generates an excitation signal in the range of 1 to 500 Hz, and the excitation signal is amplified by the power amplifier and transmitted to the longitudinal wave transmitting unit and the transverse wave transmitting unit respectively. The longitudinal wave transmitting unit is composed of a piezoelectric ceramic transducer array to convert electrical signals into mechanical vibration energy; the transverse wave transmitting unit is made of shear-type piezoelectric material, which can generate lateral vibrations perpendicular to the propagation direction. The two wave source units work simultaneously to transmit multi-frequency longitudinal and transverse wave signals to the seabed strata, and the penetration depth can reach 100 meters below the seabed. The purpose of this step is to achieve a comprehensive scanning detection of the seabed strata and provide original signal data for subsequent analysis.

[0043] The specific implementation method of step S02 is to use an omnidirectional receiving sensor array installed at the bottom of the seabed crawler to receive the longitudinal and transverse wave signals reflected from the seabed strata. The sensor array consists of 32 triaxial acceleration sensors arranged in a 4×8 matrix, which can simultaneously capture the longitudinal and transverse wave reflection signals in three-dimensional space. After the received analog signal is amplified by the preamplifier, it is converted into a digital signal by a 24-bit high-precision analog-to-digital converter at a sampling rate of 10 kHz. The digital signal is preliminarily filtered by the data processing module to filter out 50 Hz power frequency interference and high-frequency noise above 500 Hz. The processed data is rearranged according to the time series and sensor position information to construct the original longitudinal wave signal matrix and the original transverse wave signal matrix. The dimension of the original longitudinal wave signal matrix is ​​m×n, where m is the number of sensors and n is the number of time sampling points; the original transverse wave signal matrix is ​​composed of two m×n dimensional matrices, corresponding to the vibration components of the transverse wave in two orthogonal directions. The purpose of this step is to obtain high-quality original reflection signal data and lay the foundation for subsequent feature extraction.

[0044] The specific implementation method of step S03 is to perform Fourier transform processing on the original longitudinal wave signal matrix and the original shear wave signal matrix. First, perform fast Fourier transform on the time series signal recorded by each sensor, extract the frequency components in the range of 1 to 500 Hz, and obtain the spectrum amplitude and phase information. For the longitudinal wave signal, the propagation characteristics of the longitudinal wave at different frequencies are extracted to construct the longitudinal wave interference matrix; for the shear wave signal, the shear wave components in two orthogonal directions are subjected to spectrum analysis to construct the shear wave interference matrix. At the same time, the physical mechanism of the propagation of longitudinal and shear waves is analyzed based on the seabed stratum wave equation. The wave equation takes into account the physical properties of the stratum such as density, Lamé's first parameter, and Lamé's second parameter, and calculates the propagation speed and attenuation characteristics of longitudinal and shear waves in different geological structures. The longitudinal wave velocity v p and the shear wave velocity v s pass and It is calculated that λ is the first parameter of Lamé, μ is the second parameter of Lamé, and ρ is the formation density. The numerical solution of the wave equation adopts the finite difference time domain method, the grid spacing is one tenth of the propagation wavelength, and the time step is determined according to the Courant stability condition. The wave displacement field distribution and wave energy attenuation law are obtained by solving the wave equation, which are used to construct the longitudinal wave interference matrix and the transverse wave interference matrix. The purpose of this step is to extract the characteristic information of the propagation of longitudinal and transverse waves in the seabed formation through frequency domain analysis, so as to provide a basis for subsequent signal fusion.

[0045] The specific implementation method of step S04 is to perform weighted fusion of the longitudinal wave interference matrix and the shear wave interference matrix to generate a joint interference matrix. The weighted fusion adopts an adaptive weighted algorithm, and the weight coefficient is dynamically adjusted based on the seabed geological conditions. First, the signal-to-noise ratio of the longitudinal wave and the shear wave under the current seabed geological conditions is calculated, which are recorded as SNR longitudinal wave and SNR shear wave respectively, and then the initial weight coefficients α and β are determined according to the signal-to-noise ratio, satisfying α+β=1. When SNR longitudinal wave>SNR shear wave, increase the α value; otherwise, increase the β value. In practical applications, the α value is generally adjusted in the range of 0.3 to 0.7. In addition, the influence of the formation type on the weight is also considered: for hard formations such as basalt, the longitudinal wave has a strong penetration ability, and the α value can be set to 0.6 to 0.7; for soft formations such as muddy sediments, the shear wave is more sensitive to the interface response, and the β value can be set to 0.6 to 0.7. The calculation formula of the joint interference matrix is: joint interference matrix = α× longitudinal wave interference matrix + β× shear wave interference matrix. The purpose of this step is to comprehensively utilize the complementary advantages of longitudinal waves and shear waves to improve the accuracy of formation identification.

[0046] The specific implementation method of step S05 is to construct an interference error compensation matrix based on the real-time attitude data recorded by the attitude sensor of the submarine crawler, and compensate the joint interference matrix. The submarine crawler attitude sensor includes a high-precision gyroscope, an accelerometer and a depth sensor, which records the pitch angle, roll angle and depth data of the crawler in real time, and the sampling frequency is 100 Hz. First, a coordinate transformation model is established to convert the sensor measurement coordinate system to the global reference coordinate system to eliminate the signal distortion caused by the change of the crawler attitude. The coordinate transformation is implemented by a rotation matrix, and the three-dimensional rotation transformation matrix is ​​calculated according to the Euler angle. When the pitch angle or roll angle of the crawler exceeds 5 degrees, additional compensation processing is required. The dimension of the constructed interference error compensation matrix is ​​the same as that of the joint interference matrix, and its element value is nonlinearly calculated according to the attitude deviation angle. The compensation processing is implemented by matrix multiplication: the compensated joint interference matrix = interference error compensation matrix × joint interference matrix. The purpose of this step is to eliminate the influence of the attitude change of the submarine crawler on signal acquisition and improve data quality.

[0047] The specific implementation method of step S06 is to perform time-frequency analysis on the compensated joint interference matrix through wavelet transform to extract specific time-frequency features. First, a wavelet basis function suitable for stratigraphic analysis is selected, and db4 or db6 wavelets in the Daubechies wavelet family are usually used because they have good time-frequency localization characteristics. Multi-scale wavelet decomposition is performed on each row of data in the joint interference matrix, and the number of decomposition layers is 5 to obtain wavelet coefficients of different frequency bands. The energy distribution of each layer of wavelet coefficients is calculated, and a time-frequency energy map is constructed to reflect the energy distribution characteristics of the signal at different times and frequencies. According to the time-frequency energy map, the energy peak point and its distribution characteristics are extracted as specific time-frequency features. In addition, the statistical characteristics of the wavelet coefficients of each frequency band are calculated, including mean, variance, skewness, kurtosis, etc., as a supplement to the feature vector. According to geological prior knowledge, the feature vectors are screened and combined to retain the features with high contribution to stratigraphic structure identification and remove redundant features. The final feature vector dimension is 128, including time domain features, frequency domain features, and time-frequency joint features. The purpose of this step is to extract key features that can reflect the structural characteristics of the formation through time-frequency analysis, providing a basis for subsequent formation classification and thickness prediction.

[0048] The specific implementation method of step S07 is to input the specific time-frequency features into the pre-trained stratum classification prediction model to realize the classification and thickness prediction of the seabed stratum structure type. The stratum classification prediction model is based on a U-shaped convolutional neural network with a self-attention mechanism, which includes an encoder and a decoder. The encoder reshapes the 128-dimensional feature vector into a 16×8 feature map, and then extracts features through five convolutional layers and four pooling layers. The convolutional layer uses a 3×3 convolution kernel, and the number of convolution kernels is 32, 64, 128, 256, and 512 respectively. Each convolution layer is followed by batch normalization and ReLU activation function, and the pooling layer uses 2×2 maximum pooling. The decoder realizes feature reconstruction through five deconvolution layers and four splicing layers. The deconvolution layer uses a 3×3 convolution kernel, and the number of feature maps is the same as that of the corresponding layer of the encoder. A self-attention mechanism is introduced between the encoder and the decoder to enhance the feature extraction capability. The sparse attention coefficient in the self-attention mechanism is dynamically adjusted according to three key parameters: the ratio of the longitudinal and transverse wave propagation speeds, the stratum density, and the seabed depth. The model output includes the probability distribution of formation types and thickness prediction values. The formation types are divided into 8 categories: sandy sedimentary layer, muddy sedimentary layer, carbonate rock layer, volcanic rock layer, metamorphic rock layer, gas permeable zone, liquid permeable zone and solid mineral enrichment zone. The thickness prediction accuracy can reach more than 90% of the actual thickness. The purpose of this step is to use deep learning technology to realize the automatic identification and parameter prediction of the submarine formation structure.

[0049] The specific implementation method of step S08 is to draw a three-dimensional geological profile of the seabed based on the prediction results of the stratigraphic structure type and thickness, and mark the key geological structures and resource distribution areas. First, the prediction results are reconstructed into three-dimensional spatial distribution data according to the crawler path information, and the undetected area is estimated by the Kriging interpolation method to generate a continuous three-dimensional stratigraphic distribution model. Then, different colors and texture identifiers are set according to different stratigraphic types, and the OpenGL graphics library is used to realize three-dimensional visual rendering. Key geological structures, including faults, folds, intrusive bodies, etc., are marked on the three-dimensional profile, and resource distribution areas are predicted according to stratigraphic characteristics, mainly including oil and gas resources, natural gas hydrates, and seabed hydrothermal deposits. The marking process adopts a combination of automatic recognition and manual verification to automatically mark areas with a prediction probability of more than 85%, and areas with a probability between 60% and 85% need to be manually confirmed before marking. The drawn three-dimensional geological profile is displayed in an interactive form, supporting operations such as rotation, scaling, and slicing, which is convenient for intuitive analysis of geological structures. The purpose of this step is to display the detection results in an intuitive form to provide a decision-making basis for resource exploration and development.

[0050] Furthermore, the detailed structure of the stratigraphic classification prediction model and the specific implementation methods of the training data set construction steps are as follows: The stratigraphic classification prediction model adopts a U-shaped convolutional neural network structure based on a self-attention mechanism, which can effectively process the multi-scale information of stratigraphic characteristics. The encoder part consists of five convolutional layers and four pooling layers. The first convolution layer uses 32 3×3 convolution kernels with a step size of 1, inputs a feature map of 16×8×1, and outputs a feature map of 16×8×32; the second convolution layer uses 64 3×3 convolution kernels, inputs a feature map of 8×4×32, and outputs a feature map of 8×4×64; the third convolution layer uses 128 3×3 convolution kernels, inputs a feature map of 4×2×64, and outputs a feature map of 4×2×128; the fourth convolution layer uses 256 3×3 convolution kernels, inputs a feature map of 2×1×128, and outputs a feature map of 2×1×256; the fifth convolution layer uses 512 3×3 convolution kernels, inputs a feature map of 1×1×256, and outputs a feature map of 1×1×512. Each convolution layer is connected to a batch normalization layer and a ReLU activation function to enhance the network's expressiveness and prevent gradient disappearance. The pooling layer uses 2×2 maximum pooling with a step size of 2 to gradually reduce the spatial dimension of the feature map. The decoder part consists of five deconvolution layers and four splicing layers. The deconvolution layer uses transposed convolution to achieve feature map upsampling. The number of feature maps is the same as that of the corresponding layer of the encoder. The splicing layer splices the feature map of the corresponding layer of the encoder with the feature map of the decoder through jump connections to retain high-resolution detail information. The self-attention mechanism is located between the encoder and the decoder. By calculating the correlation between different positions of the feature map, the model's ability to capture long-distance dependencies is enhanced. The attention matrix calculation formula is: Attention=softmax(Q×K / √d)×V, where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, and d is the feature dimension. The sparse attention coefficient is dynamically adjusted according to three key parameters: when the ratio of the longitudinal and transverse wave propagation speeds is greater than 1.9, the attention coefficient is inclined to the longitudinal wave feature; when the formation density exceeds 2.5 g / cm3, the attention coefficient is inclined to the high-frequency feature; when the seabed depth exceeds 3000 meters, the attention coefficient is inclined to the low-frequency feature. The model output layer is divided into two parts: the classification head outputs the probability distribution of 8 categories, using the softmax activation function; the regression head outputs the predicted value of the stratum thickness, using the linear activation function. The process of establishing the training data set involves four stages: data collection, preprocessing, annotation, and enhancement. In the data collection stage, longitudinal and transverse wave reflection data of submarine strata of various geological structures were collected from different sea areas such as the Beibu Gulf, the South China Sea, and the East China Sea. The collection point spacing was 5 meters, covering an area of ​​more than 500 square kilometers, including various typical submarine geological structures. In the preprocessing stage, the collected data was bandpass filtered to filter out noise signals above 500 Hz and below 1 Hz, and then outliers were removed by median filtering. Finally, the data was normalized and the signal amplitude was scaled to the range of [-1, 1].In the annotation stage, a team of more than 5 experts with geological backgrounds, combined with drilling verification data, manually annotated the formation type and thickness. The annotation consistency requires the Inter-rater agreement coefficient to be no less than 0.85. In the data enhancement stage, the noise resistance of the model is increased by adding different degrees of Gaussian noise, and the noise level is set to 5% to 20% of the signal standard deviation; random cropping expands the model's adaptability to incomplete data; random rotation operations enhance the model's robustness to directional changes, and the rotation angle range is [-10°, 10°]. Finally, a training set containing 100,000 sets of annotated data and a validation set of 10,000 sets of annotated data are constructed. The ratio of training set to validation set is 10:1. The number of samples of different formation types is basically balanced, and the ratio of the number of samples of the largest category to the smallest category does not exceed 3:1. The model training adopts the small batch gradient descent method, with a batch size of 64, an initial learning rate of 0.001, and a cosine annealing strategy. The learning rate decays to 0.1 times the original every 50 training cycles. The loss function adopts a weighted combination of the cross entropy loss for formation type classification and the mean square error loss for thickness prediction, with weights of 0.6 and 0.4 respectively. The total loss function is expressed as: Loss = 0.6 × CrossEntropy + 0.4 × MSE. An early stopping mechanism is introduced during the training process. When the loss of the validation set does not decrease for 10 consecutive training cycles, the training is stopped to avoid overfitting. In addition, regularization techniques such as weight decay and Dropout are used to improve the generalization ability of the model. Finally, the model parameters with the best performance on the validation set are selected as the final model, and convergence is usually achieved after 200 to 300 training cycles.

[0051] The mathematical model or calculation process involved in the present invention is described in detail below.

[0052] In step S03, when the physical mechanism of the propagation of longitudinal and transverse waves is analyzed by wave equations, the wave equations of the seabed formation are specifically expressed as follows:

[0053]

[0054] Where ρ is the seabed stratum density, in kilograms per cubic meter; is the displacement vector, in meters; t is the time variable, in seconds; λ is the Lamé first parameter, in Pascal; μ is the Lamé second parameter, in Pascal; is the gradient operator; is the Laplace operator; is the external force, measured in Newtons per cubic meter.

[0055] The parameter acquisition method is as follows: ρ is directly measured by the density measuring instrument carried by the seafloor crawler, and the measurement range is 1500-3500 kg / m3; λ and μ are obtained by static mechanical testing of rock samples collected by the seafloor crawler in the laboratory, and the typical ranges are 10 9～10 11 Pascal and 5×10 8～5×10 10 Pascal; It is provided by the acoustic wave signal emitted by the multiple longitudinal and transverse wave joint detection device, and its size and direction are dynamically adjusted according to the frequency and amplitude of the excitation wave.

[0056] In actual numerical calculations, the finite difference time domain method is used to solve the wave equation, discretizing the continuous space and time to form a difference equation:

[0057]

[0058] In the formula, It represents the displacement vector when the position coordinates are (i, j, k) and the time step is n; Δt is the time step in seconds; and They are the discrete gradient operator and the discrete Laplace operator respectively.

[0059] The time step Δt is determined according to the Courant stability condition:

[0060]

[0061] Where Δh is the spatial grid spacing, in meters; d is the spatial dimension, which is 3; v max is the maximum wave speed in the medium, in meters per second, and the calculation formula is where ρ min is the minimum density value.

[0062] P-wave velocity v p and the shear wave velocity v s The calculation formula is as follows:

[0063]

[0064] The wave energy density E can be calculated from the displacement field obtained by solving the wave equation:

[0065]

[0066] In the formula, u i Represents the displacement vector The i-th component of j Represents the j-th component of the spatial coordinate.

[0067] In step S02, the original longitudinal wave signal matrix P and the original shear wave signal matrix S are constructed.x , S y The expression is as follows:

[0068]

[0069]

[0070] In the formula, p i，j represents the amplitude of the longitudinal wave signal received by the i-th sensor at the j-th time sampling point; and They represent the amplitudes of the shear wave signals in two orthogonal directions received by the ith sensor at the jth time sampling point; m is the total number of sensors, which is 32; n is the number of time sampling points, which depends on the sampling rate and sampling duration, and its typical value is 10000 to 50000.

[0071] In step S03, the frequency characteristics are extracted by Fourier transform to construct the longitudinal wave interference matrix P F and the shear wave interference matrix S F The expression is as follows:

[0072]

[0073] In the formula, represents the amplitude of the longitudinal wave signal received by the i-th sensor at the k-th frequency component; Indicates the corresponding phase; and They represent the amplitudes of the two orthogonal shear wave signals received by the i-th sensor at the k-th frequency component; θ i，k It represents the phase of the synthetic shear wave; j is the imaginary unit; q is the number of frequency components, usually 500.

[0074] The above amplitude and phase are calculated by fast Fourier transform (FFT):

[0075]

[0076] Among them, P i (f k ), and is a complex spectrum, which can be expressed as:

[0077]

[0078] The phase θ of the composite shear wave i，k The calculation formula is:

[0079]

[0080] In step S04, the longitudinal wave interference matrix and the transverse wave interference matrix are weightedly fused to generate a joint interference matrix C:

[0081] C=αP F +βS F ;

[0082] Wherein, α and β are weight coefficients, satisfying α+β=1, 0.3≤α≤0.7.

[0083] The weight coefficient is dynamically adjusted based on the signal-to-noise ratio of longitudinal and shear waves, and the calculation formula is:

[0084]

[0085] β = 1-α;

[0086] In the formula, SNR p is the signal-to-noise ratio of the longitudinal wave signal; SNR s is the signal-to-noise ratio of the shear wave signal; γ is the formation type adjustment factor, ranging from -0.2 to 0.2. When the formation is a hard formation (such as basalt), γ takes a positive value and increases α; when the formation is a soft formation (such as mud sedimentary layer), γ takes a negative value and decreases α.

[0087] The signal-to-noise ratio calculation formula is:

[0088]

[0089] In the formula, and are the noise amplitudes of longitudinal waves and shear waves in two directions, respectively, which are obtained by Fourier analysis of the signal-free interval.

[0090] In step S05, the interference error compensation matrix E is constructed according to the real-time attitude data recorded by the attitude sensor of the seabed crawler:

[0091]

[0092] In the formula, e i，k is the compensation coefficient, and the calculation formula is:

[0093] e i，k =1+a 1 (θ p -θ p0 ) 2 +a 2 (θ r -θ r0 ) 2 +a 3 (dd 0 );

[0094] In the formula, θ pis the pitch angle, in degrees; θ r is the roll angle, in degrees; d is the seabed depth, in meters; θ p0 ,θ r0 and d 0 are the pitch angle, roll angle and depth in the reference state respectively; a 1 、a 2 and a 3 is the compensation coefficient, which is obtained through experimental calibration, and the typical values ​​are 0.01, 0.01 and 0.001 respectively.

[0095] The calculation formula of the compensated joint interference matrix C′ is:

[0096]

[0097] In the formula, Represents the element-wise product (Hadamard product) of matrices.

[0098] In step S06, the compensated joint interference matrix is ​​subjected to time-frequency analysis by wavelet transform. In specific implementation, each row of the matrix C′ is subjected to wavelet transform, which can be expressed as:

[0099]

[0100] In the formula, c′ i represents the i-th row of the matrix C′, i.e., the signal of the i-th sensor; ψ k，l (t) indicates a scale of 2 k , wavelet basis function with a translation of l; W i，k，l represents the wavelet coefficients.

[0101] The db4 wavelet in the Daubechies wavelet family is used. Its mother wavelet function ψ(t) has no analytical expression and is obtained by numerical calculation through an iterative method.

[0102] Wavelet coefficient energy distribution E i，k，l The calculation formula is:

[0103] E i，k，l =|W i，k，l | 2 ;

[0104] Time-Frequency Energy (TFE) i (f, t) is obtained by wavelet coefficient energy reconstruction:

[0105]

[0106] In the formula, f k For scale 2 k The corresponding frequency; t lis the time corresponding to the translation l; δ is the Dirac function.

[0107] Specific time-frequency feature extraction includes the following aspects:

[0108] 1. Energy peak characteristics:

[0109] EP i =max f，t TFE i (f, t);

[0110] FP i = argmax f ∑ t TFE i (f, t);

[0111] TP i = argmax t ∑ f TFE i (f, t);

[0112] 2. Frequency band energy distribution:

[0113] BE i，k =∑ l E i，k，l ;

[0114] 3. Statistical characteristics:

[0115] Mean:

[0116] variance:

[0117] Skewness:

[0118] Kurtosis:

[0119] Where, L k is the total number of wavelet coefficients at scale k.

[0120] The final specific time-frequency feature vector F i Combination of the above features:

[0121]

[0122] By processing the 32 sensor data collected by the seafloor crawler, the feature matrix F of the final input formation classification prediction model is obtained:

[0123] F=[F 1 , F 2 , …, F 32 ] T ;

[0124] The principle and significance of selecting these functional relationships in the above equations are: the wave equation is the basic physical equation that describes the propagation of waves in elastic media. The partial differential equation form can accurately reflect the propagation, reflection and attenuation characteristics of waves; the Fourier transform can convert the time domain signal into the frequency domain representation, which is convenient for extracting the response characteristics of the formation to different frequency components; the adaptive weighted fusion algorithm based on the signal-to-noise ratio can dynamically adjust the weight of the P- and S-wave information according to the signal quality to improve the fusion effect; the attitude compensation matrix adopts the quadratic function form, which can better fit the nonlinear influence of attitude changes on the signal; the wavelet transform has good time-frequency localization characteristics, can simultaneously obtain the time domain and frequency domain information of the signal, and is particularly suitable for analyzing the non-stationary characteristics of the submarine formation; the statistical characteristics (mean, variance, skewness, kurtosis, etc.) can comprehensively describe the distribution characteristics of the wavelet coefficients, which is helpful to identify the differences in different formation structures. These equations comprehensively consider the physical principles, signal processing technology and feature extraction methods of submarine formation detection, forming a complete theoretical system of multiple P- and S-wave joint detection methods, which improves the accuracy and resolution of submarine formation structure identification compared with the existing technology.

[0125] Specifically, the principle of the present invention is: the core principle of the technical solution of the present invention is to overcome various interference factors in seabed geological exploration through the synergistic effect of multiple anti-interference technologies, and realize accurate identification of the stratum structure. First, the seabed crawler is used to directly detect close to the seabed, avoiding the attenuation and interference problems caused by the signal passing through the entire water body, and reducing the adverse effects of water body factors such as water flow and thermocline. Secondly, two different types of elastic waves, longitudinal waves and shear waves, are used for detection at the same time, and the complementarity of their propagation characteristics in the stratum is utilized to enhance the anti-interference ability and resolution ability of the detection signal.

[0126] The multi-frequency longitudinal and transverse wave signal design makes the detection system more adaptable. Waves of different frequencies can maintain the effectiveness of some frequency bands under different interference environments. High-frequency waves are suitable for detecting shallow details, while low-frequency waves can penetrate deeper strata and have stronger anti-interference capabilities. The original signal is analyzed in the frequency domain by Fourier transform, and the physical mechanism analysis is performed in combination with the seabed stratum wave equation. The characteristic differences between effective signals and interference signals can be distinguished, and the longitudinal and transverse wave interference matrix that characterizes the characteristics of different geological structures can be constructed.

[0127] The key innovation of the present invention is the introduction of an interference error compensation mechanism based on the attitude data of the seabed crawler. The attitude changes (pitch angle, roll angle, etc.) of the crawler when moving on an uneven seabed will cause obvious interference to signal acquisition. By constructing an interference error compensation matrix through real-time attitude data, the signal fluctuations and distortion caused by these attitude changes can be effectively eliminated. The time-frequency analysis capability of wavelet transform enables the system to simultaneously identify and filter interference signals in the time domain and frequency domain, and extract time-frequency joint features that can reflect the unique response characteristics of different geological structures.

[0128] The U-shaped convolutional neural network based on the self-attention mechanism is used as a stratum classification prediction model. Through training with a large amount of labeled data, it learns the signal characteristic patterns under different interference environments, can adaptively identify and suppress various interferences, and improves the recognition ability of complex geological structures. The model dynamically adjusts the attention coefficient according to key parameters such as stratum density, depth, and the ratio of longitudinal and transverse wave propagation velocities, further enhancing the system's adaptability and recognition accuracy under various interference conditions.

[0129] A specific embodiment 1 of the present invention is provided below, and the specific implementation method of each step in this embodiment 1 is described in detail as follows.

[0130] The specific implementation of step S01 is the same as above and will not be repeated here.

[0131] The specific implementation method of step S02 is to use an omnidirectional receiving sensor array installed at the bottom of the seabed crawler to receive the longitudinal and transverse wave signals reflected from the seabed strata. The sensor array consists of 32 triaxial acceleration sensors arranged in a 4×8 matrix, which can simultaneously capture the longitudinal and transverse wave reflection signals in three-dimensional space. After the received analog signal is amplified by the preamplifier, it is converted into a digital signal by a 24-bit high-precision analog-to-digital converter at a sampling rate of 10 kHz. The digital signal is preliminarily filtered by the data processing module to filter out 50 Hz power frequency interference and high-frequency noise above 500 Hz. The processed data is rearranged according to the time series and sensor position information to construct the original longitudinal wave signal matrix and the original transverse wave signal matrix. The dimension of the original longitudinal wave signal matrix is ​​m×n, where m is the number of sensors and n is the number of time sampling points; the original transverse wave signal matrix is ​​composed of two m×n dimensional matrices, corresponding to the vibration components of the transverse wave in two orthogonal directions. The purpose of this step is to obtain high-quality original reflection signal data and lay the foundation for subsequent feature extraction. The original longitudinal wave signal matrix P and the original shear wave signal matrix S x , S y The expression is as follows:

[0132]

[0133] In the formula, p i，jrepresents the amplitude of the longitudinal wave signal received by the i-th sensor at the j-th time sampling point; and They represent the amplitudes of the shear wave signals in two orthogonal directions received by the ith sensor at the jth time sampling point; m is the total number of sensors, which is 32; n is the number of time sampling points, which depends on the sampling rate and sampling duration, and its typical value is 10000 to 50000.

[0134] The specific implementation method of step S03 is to perform Fourier transform processing on the original longitudinal wave signal matrix and the original shear wave signal matrix. First, perform fast Fourier transform on the time series signal recorded by each sensor, extract the frequency components in the range of 1 to 500 Hz, and obtain the spectrum amplitude and phase information. For the longitudinal wave signal, the propagation characteristics of the longitudinal wave at different frequencies are extracted to construct the longitudinal wave interference matrix; for the shear wave signal, the shear wave components in two orthogonal directions are subjected to spectrum analysis to construct the shear wave interference matrix. At the same time, the physical mechanism of the propagation of longitudinal and shear waves is analyzed based on the seabed stratum wave equation. The wave equation takes into account the physical properties of the stratum such as density, Lamé's first parameter, and Lamé's second parameter, and calculates the propagation speed and attenuation characteristics of longitudinal and shear waves in different geological structures. The longitudinal wave velocity v p and the shear wave velocity v s pass and It is calculated that λ is the first parameter of Lamé, μ is the second parameter of Lamé, and ρ is the formation density. The numerical solution of the wave equation adopts the finite difference time domain method, the grid spacing is one tenth of the propagation wavelength, and the time step is determined according to the Courant stability condition. The wave displacement field distribution and wave energy attenuation law are obtained by solving the wave equation, which are used to construct the longitudinal wave interference matrix and the transverse wave interference matrix. The purpose of this step is to extract the characteristic information of the propagation of longitudinal and transverse waves in the seabed formation through frequency domain analysis, so as to provide a basis for subsequent signal fusion.

[0135] The wave equation of the seabed formation is specifically expressed as follows:

[0136]

[0137] Where ρ is the seabed stratum density, in kilograms per cubic meter; is the displacement vector, in meters; t is the time variable, in seconds; λ is the Lamé first parameter, in Pascal; μ is the Lamé second parameter, in Pascal; is the gradient operator; is the Laplace operator; is the external force, measured in Newton / cubic meter.

[0138] The parameter acquisition method is as follows: ρ is directly measured by the density measuring instrument carried by the seafloor crawler, and the measurement range is 1500-3500 kg / m3; λ and μ are obtained by static mechanical testing of rock samples collected by the seafloor crawler in the laboratory, and the typical ranges are 10 9 ~10 11 Pascal and 5×10 8 ~5×10 10 Pascal; It is provided by the acoustic wave signal emitted by the multiple longitudinal and transverse wave joint detection device, and its size and direction are dynamically adjusted according to the frequency and amplitude of the excitation wave.

[0139] In actual numerical calculations, the finite difference time domain method is used to solve the wave equation, discretizing the continuous space and time to form a difference equation:

[0140]

[0141] In the formula, It represents the displacement vector when the position coordinates are (i, j, k) and the time step is n; Δt is the time step in seconds; and They are the discrete gradient operator and the discrete Laplace operator respectively.

[0142] The time step Δt is determined according to the Courant stability condition:

[0143]

[0144] Where Δh is the spatial grid spacing, in meters; d is the spatial dimension, which is 3; v max is the maximum wave speed in the medium, in meters per second, and the calculation formula is where ρ min is the minimum density value.

[0145] The frequency characteristics are extracted by Fourier transform to construct the longitudinal wave interference matrix P F and the shear wave interference matrix S F The expression is as follows:

[0146]

[0147] In the formula, represents the amplitude of the longitudinal wave signal received by the i-th sensor at the k-th frequency component; Indicates the corresponding phase; and They represent the amplitudes of the two orthogonal shear wave signals received by the i-th sensor at the k-th frequency component; θ i，kIt represents the phase of the synthetic shear wave; j is the imaginary unit; q is the number of frequency components, usually 500.

[0148] The above amplitude and phase are calculated by fast Fourier transform (FFT):

[0149]

[0150] Among them, P i (f k ), and is a complex spectrum, which can be expressed as:

[0151]

[0152] The phase θ of the composite shear wave i，k The calculation formula is:

[0153]

[0154] The specific implementation method of step S04 is to perform weighted fusion of the longitudinal wave interference matrix and the shear wave interference matrix to generate a joint interference matrix. The weighted fusion adopts an adaptive weighted algorithm, and the weight coefficient is dynamically adjusted based on the seabed geological conditions. First, the signal-to-noise ratio of the longitudinal wave and the shear wave under the current seabed geological conditions is calculated, which are recorded as SNR longitudinal wave and SNR shear wave respectively, and then the initial weight coefficients α and β are determined according to the signal-to-noise ratio, satisfying α+β=1. When SNR longitudinal wave>SNR shear wave, increase the α value; otherwise, increase the β value. In practical applications, the α value is generally adjusted in the range of 0.3 to 0.7. In addition, the influence of the formation type on the weight is also considered: for hard formations such as basalt, the longitudinal wave has a strong penetration ability, and the α value can be set to 0.6 to 0.7; for soft formations such as muddy sediments, the shear wave is more sensitive to the interface response, and the β value can be set to 0.6 to 0.7. The calculation formula of the joint interference matrix is: joint interference matrix = α× longitudinal wave interference matrix + β× shear wave interference matrix. The purpose of this step is to comprehensively utilize the complementary advantages of longitudinal waves and shear waves to improve the accuracy of formation identification.

[0155] The longitudinal wave interference matrix and the transverse wave interference matrix are weightedly fused to generate the joint interference matrix C:

[0156] C=αP F +βS F ;

[0157] Wherein, α and β are weight coefficients, satisfying α+β=1, 0.3≤α≤0.7.

[0158] The weight coefficient is dynamically adjusted based on the signal-to-noise ratio of longitudinal and shear waves, and the calculation formula is:

[0159]

[0160] In the formula, SNR p is the signal-to-noise ratio of the longitudinal wave signal; SNR s is the signal-to-noise ratio of the shear wave signal; γ is the formation type adjustment factor, ranging from -0.2 to 0.2. When the formation is a hard formation (such as basalt), γ takes a positive value and increases α; when the formation is a soft formation (such as mud sedimentary layer), γ takes a negative value and decreases α.

[0161] The signal-to-noise ratio calculation formula is:

[0162]

[0163] In the formula, and are the noise amplitudes of longitudinal waves and shear waves in two directions, respectively, which are obtained by Fourier analysis of the signal-free interval.

[0164] The specific implementation method of step S05 is to construct an interference error compensation matrix based on the real-time attitude data recorded by the attitude sensor of the submarine crawler, and compensate the joint interference matrix. The submarine crawler attitude sensor includes a high-precision gyroscope, an accelerometer and a depth sensor, which records the pitch angle, roll angle and depth data of the crawler in real time, and the sampling frequency is 100 Hz. First, a coordinate transformation model is established to convert the sensor measurement coordinate system to the global reference coordinate system to eliminate the signal distortion caused by the change of the crawler attitude. The coordinate transformation is implemented by a rotation matrix, and the three-dimensional rotation transformation matrix is ​​calculated according to the Euler angle. When the pitch angle or roll angle of the crawler exceeds 5 degrees, additional compensation processing is required. The dimension of the constructed interference error compensation matrix is ​​the same as that of the joint interference matrix, and its element value is nonlinearly calculated according to the attitude deviation angle. The compensation processing is implemented by matrix multiplication: the compensated joint interference matrix = interference error compensation matrix × joint interference matrix. The purpose of this step is to eliminate the influence of the attitude change of the submarine crawler on signal acquisition and improve data quality.

[0165] According to the real-time attitude data recorded by the attitude sensor of the seabed crawler, the interference error compensation matrix E is constructed:

[0166]

[0167] In the formula, e i，k is the compensation coefficient, and the calculation formula is:

[0168] e i，k =1+a 1 (θ p -θ p0 ) 2 +a 2 (θ r -θ r0 ) 2+a 3 (dd 0 );

[0169] In the formula, θ p is the pitch angle, in degrees; θ r is the roll angle, in degrees; d is the seabed depth, in meters; θ p0 ,θ r0 and d 0 are the pitch angle, roll angle and depth in the reference state respectively; a 1 、a 2 and a 3 is the compensation coefficient, which is obtained through experimental calibration, and the typical values ​​are 0.01, 0.01 and 0.001 respectively.

[0170] The calculation formula of the compensated joint interference matrix C′ is:

[0171] C′=E°C;

[0172] Where ° represents the element-by-element product of matrices (Hadamard product).

[0173] The specific implementation method of step S06 is to perform time-frequency analysis on the compensated joint interference matrix through wavelet transform to extract specific time-frequency features. First, a wavelet basis function suitable for stratigraphic analysis is selected, and db4 or db6 wavelets in the Daubechies wavelet family are usually used because they have good time-frequency localization characteristics. Multi-scale wavelet decomposition is performed on each row of data in the joint interference matrix, and the number of decomposition layers is 5 to obtain wavelet coefficients of different frequency bands. The energy distribution of each layer of wavelet coefficients is calculated, and a time-frequency energy map is constructed to reflect the energy distribution characteristics of the signal at different times and frequencies. According to the time-frequency energy map, the energy peak point and its distribution characteristics are extracted as specific time-frequency features. In addition, the statistical characteristics of the wavelet coefficients of each frequency band are calculated, including mean, variance, skewness, kurtosis, etc., as a supplement to the feature vector. According to geological prior knowledge, the feature vectors are screened and combined to retain the features with high contribution to stratigraphic structure identification and remove redundant features. The final feature vector dimension is 128, including time domain features, frequency domain features, and time-frequency joint features. The purpose of this step is to extract key features that can reflect the structural characteristics of the formation through time-frequency analysis, providing a basis for subsequent formation classification and thickness prediction.

[0174] The time-frequency analysis of the compensated joint interference matrix is ​​performed by wavelet transform. Each row of the matrix C′ is subjected to wavelet transform, which can be expressed as:

[0175]

[0176] In the formula, c′ i represents the i-th row of the matrix C′, i.e., the signal of the i-th sensor; ψk，l (t) indicates a scale of 2 k , wavelet basis function with a translation of l; W i，k，l represents the wavelet coefficients.

[0177] Wavelet coefficient energy distribution E i，k，l The calculation formula is:

[0178] E i，k，l =|W i，k，l | 2 ;

[0179] Time-Frequency Energy (TFE) i (f, t) is obtained by wavelet coefficient energy reconstruction:

[0180]

[0181] In the formula, f k For scale 2 k The corresponding frequency; t l is the time corresponding to the translation l; δ is the Dirac function.

[0182] Specific time-frequency feature extraction includes the following aspects:

[0183] 1. Energy peak characteristics:

[0184] EP i =max f， tTFEi (f, t);

[0185] FP i = argmax f ∑ t TFE i (f, t);

[0186] TP i = argmax t ∑ f TFE i (f, t);

[0187] 2. Frequency band energy distribution:

[0188] BE i，k =∑ l E i，k，l ;

[0189] 3. Statistical characteristics:

[0190] Mean:

[0191] variance:

[0192] Skewness:

[0193] Kurtosis:

[0194] Where, L k is the total number of wavelet coefficients at scale k.

[0195] The final specific time-frequency feature vector F i Combination of the above features:

[0196]

[0197] By processing the 32 sensor data collected by the seafloor crawler, the feature matrix F of the final input formation classification prediction model is obtained:

[0198] F=[F 1 , F 2 , …, F 32 ] T ;

[0199] The specific implementation of steps S07-S08 is the same as above and will not be repeated here.

[0200] Optional, such as Figure 2 As shown in the figure, the multiple longitudinal and transverse wave joint detection device is the core component of the seabed crawler detection system. It is composed of a multi-frequency wave source generator, a longitudinal wave transmitting unit, a transverse wave transmitting unit, an omnidirectional receiving sensor array, a signal processing module and a data storage unit. The whole is packaged in a titanium alloy shell and has a pressure resistance of 20 MPa, which is suitable for deep-sea environments within 6000 meters.

[0201] The multi-frequency wave source generator adopts digital signal synthesis technology and can generate continuously adjustable frequency signals in the range of 1 to 500 Hz with a frequency resolution of 0.1 Hz. The generator integrates a high-precision digital frequency synthesis chip DDS (direct digital frequency synthesizer) and is controlled by a 32-bit microprocessor. It can simultaneously generate up to 16 excitation signals of different frequencies. These signals are waveform shaped by digital filters to ensure high signal purity and total harmonic distortion of less than 0.01%. The multi-frequency wave source generator also has an adaptive frequency selection function, which intelligently adjusts the excitation signal frequency combination according to the characteristics of different seabed geological environments to obtain the best detection effect. In practical applications, logarithmically spaced frequency combinations are usually used, such as 1, 2, 5, 10, 20, 50, 100, 200, and 500 Hz, covering a wide frequency band to meet the detection needs of different depths and different physical properties of strata.

[0202] The longitudinal wave transmitting unit adopts a piezoelectric ceramic transducer array structure, which consists of 64 high-performance piezoelectric ceramic elements arranged in an 8×8 matrix with a total area of ​​400 square centimeters. Each piezoelectric element is made of modified PZT-5H material, which has high motor conversion efficiency and good temperature stability, and the operating temperature range is -10 to 60 degrees Celsius. The piezoelectric element is connected to the titanium alloy shell through a special backing material. The backing material adopts a composite formula of polymer and tungsten powder, which has high damping characteristics, effectively suppresses the backward radiation of the transducer, and improves the forward sound energy output. The working principle of the longitudinal wave transmitting unit is to use piezoelectric materials to generate mechanical vibrations in the thickness direction under the action of an electric field, and excite longitudinal sound waves parallel to the propagation direction. The overall sound power of the unit can reach 2 kilowatts, and the sound intensity can reach up to 5 watts / square centimeter, which can generate enough energy to penetrate complex stratum structures. To prevent overheating, the longitudinal wave transmitting unit has a built-in water circulation cooling system to ensure long-term stable operation.

[0203] The shear wave transmitting unit adopts a shear-type piezoelectric ceramic design, which consists of 32 pairs of orthogonally arranged shear mode piezoelectric elements, and can generate shear waves in two orthogonal directions. Each pair of shear elements is 30×30×15 mm in size and works in d15 mode. The electric field applied on the electrode surface generates shear vibrations perpendicular to the polarization direction. The shear wave transmitting unit is arranged in a 4×8 array with a total area of ​​300 square centimeters. The shear elements are fixed on a specially designed flexible coupling layer, which is made of silicone rubber material filled with metal particles and has an acoustic impedance of 5×10 6 Pa·s / m, which is close to the acoustic impedance of seabed sediments, achieving efficient energy transfer. The operating frequency range of the shear wave transmitting unit is 5 to 200 Hz, which is lower than that of the longitudinal wave transmitting unit. This is because the propagation speed of shear waves in the medium is usually lower than that of longitudinal waves. The acoustic power of the shear wave unit is 1.5 kilowatts, and the sound intensity can reach up to 3 watts / square centimeter. The shear wave signal can provide complementary formation information to the longitudinal wave, and is particularly suitable for identifying fluid filling structures and interface characteristics.

[0204] The omnidirectional receiving sensor array consists of 32 triaxial acceleration sensors arranged in a 4×8 rectangular array, covering an area of ​​600 square centimeters. Each triaxial sensor contains three high-sensitivity MEMS accelerometers in orthogonal directions, with a sensitivity of 100 millivolts per gram, a frequency response of 0.1 to 1000 Hz, and a dynamic range of ±50 grams. The sensor adopts a low-noise design, with an equivalent noise level of less than 1 microgram per √Hz and a signal-to-noise ratio of more than 80 decibels. The triaxial sensor can simultaneously receive longitudinal and transverse wave signals and distinguish wave components in different vibration directions. The receiving array adopts an elastic suspension installation method, which maintains good contact with the bottom of the crawler while reducing the interference of the crawler body vibration on the receiving signal through the microspring isolation system. An acoustic barrier ring is set around the array to reduce lateral noise interference. The sampling rate of the omnidirectional receiving sensor array is 10 kHz, which meets the requirements of the Nyquist sampling theorem and avoids spectrum aliasing.

[0205] The signal processing module adopts a high-performance digital signal processor architecture. The main chip is a dual-core DSP with a main frequency of 3GHz, which meets the needs of real-time signal processing. The module contains a multi-channel preamplifier with an adjustable gain range of 20 to 60 decibels and an input impedance of 10 kilo-ohms. The signal conditioning circuit adopts a differential input design with a high common mode rejection ratio (>80 decibels) to effectively suppress environmental electromagnetic interference. The analog-to-digital conversion uses a 24-bit Sigma-Delta ADC with a sampling rate of 10 kHz and a signal-to-noise ratio greater than 110 decibels. Digital filtering uses a programmable filter group, including high-pass, low-pass, band-pass and band-stop filters, and the cut-off frequency can be dynamically adjusted according to actual needs. The signal processing algorithm includes adaptive noise elimination, waveform feature extraction, spectrum analysis, etc., which is realized through FPGA hardware acceleration, and the processing delay is less than 10 milliseconds. The module also has an automatic gain control function, which automatically adjusts the gain according to the received signal strength to maintain the optimal signal level and avoid saturation distortion or noise drowning.

[0206] The data storage unit adopts a solid-state hard disk array design and adopts a RAID5 redundant storage strategy to improve data reliability. The unit is also equipped with a low-power standby mode, which reduces power consumption by more than 90% when not in operation, extending battery life. The data storage unit is connected to the signal processing module through an optical fiber interface, with a data transmission rate of 10 Gbit / s to avoid electromagnetic interference. The storage unit is also equipped with a data backup module, which automatically backs up key data to an independent storage area every hour to prevent accidental data loss.

[0207] The power supply system of the multiple longitudinal and transverse wave joint detection device uses a lithium-ion battery pack with a capacity of 100 kWh, which supports continuous operation for 12 hours. The power management system realizes intelligent power distribution, dynamically adjusts the power supply according to the working status of each module, and maximizes energy utilization efficiency. The device shell is made of high-strength titanium alloy with a strength grade of HY-100, with a pressure-resistant depth of 6,000 meters. The outer surface is coated with anti-corrosion polyurethane material, and the seawater corrosion resistance life is more than 5 years. The entire device weighs 120 kilograms and has a volume of 0.6 cubic meters. It adopts a modular design for easy maintenance and upgrading.

[0208] The working process of the multiple P- and S-wave joint detection device is as follows: the multi-frequency wave source generator generates an excitation signal with a preset frequency combination, which drives the P-wave transmitting unit and the S-wave transmitting unit to transmit detection signals to the seabed strata respectively; the omnidirectional receiving sensor array receives the P- and S-wave signals reflected from different strata interfaces; the received signals are amplified, filtered and preliminarily feature extracted by the signal processing module; the processed data is saved to the data storage unit together with the timestamp and location information to provide raw materials for subsequent data analysis. The device significantly improves the accuracy and resolution of seabed strata structure identification through complementary P- and S-wave detection, and is particularly suitable for the detection and analysis of complex geological structures and multiphase media.

[0209] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: Researchers conducted a multiple P- and S-wave joint detection experiment based on a seafloor crawler in a hydrothermal activity area in a trough. The water depth in this area is about 1400 to 1600 meters, and it is a known hydrothermal activity area rich in sulfide deposits. The seafloor crawler used in the experiment is equipped with a multiple P- and S-wave joint detection device to detect the seafloor stratum structure and resource distribution.

[0210] The multiple longitudinal and transverse wave joint detection device consists of a multi-frequency wave source generator, a longitudinal wave transmitting unit, a transverse wave transmitting unit, an omnidirectional receiving sensor array, a signal processing module and a data storage unit. The operating frequency range of the wave source generator is 1 to 500Hz, and the transmitting power is 2500W. The longitudinal wave transmitting unit is composed of 24 piezoelectric ceramic transducers, the transverse wave transmitting unit is made of 16 shear-type piezoelectric materials, and the omnidirectional receiving sensor array is composed of 32 triaxial acceleration sensors with a signal sampling rate of 10kHz. The crawler is equipped with a high-precision attitude sensor system, including a gyroscope, an accelerometer and a depth sensor, with a sampling frequency of 100Hz.

[0211] The researchers planned a detection path with a total length of about 3.6 km, which was laid out in a grid with a grid spacing of 8 meters. The seafloor crawler crawled along the preset path at an average speed of 0.3 m / s, and simultaneously activated the multiple P- and S-wave joint detection device to transmit multi-frequency P- and S-wave signals in the range of 1 to 500 Hz to the seafloor formation. The density measuring instrument carried by the crawler measured the density of the seafloor formation in real time, and collected rock samples for subsequent laboratory analysis to obtain the Lamé parameters.

[0212] Through the reflection signals received by the multiple P-wave and S-wave joint detection device, the researchers constructed the original P-wave signal matrix and the original S-wave signal matrix. These matrices were processed by Fourier transform, the frequency characteristics were extracted, and the P-wave interference matrix and S-wave interference matrix were constructed. The typical formation physical parameters obtained in the experiment are shown in Table 1:

[0213] Table 1 Measurement results of physical parameters of different formations in the Okinawa Trough hydrothermal area

[0214]

[0215] Based on these physical parameters, the researchers used the wave equations of the seafloor to analyze the physical mechanism of the propagation of longitudinal and transverse waves. The numerical simulation of the wave equations uses the finite difference time domain method, and the spatial grid spacing is set to one tenth of the propagation wavelength, specifically 5 cm. The time step is determined to be 5×10 based on the Courant stability condition. -6 s. By solving the wave equation, the wave displacement field distribution and wave energy attenuation law are obtained, and the longitudinal wave interference matrix and the transverse wave interference matrix are further improved.

[0216] The signal-to-noise ratio evaluation results before the fusion of the longitudinal wave interference matrix and the shear wave interference matrix are shown in Table 2:

[0217] Table 2 Evaluation results of the signal-to-noise ratio of P- and S-wave signals in different formation types

[0218]

[0219] According to the signal-to-noise ratio and formation type, the researchers determined the weighted fusion weight coefficient suitable for each formation type, and weightedly fused the longitudinal wave interference matrix with the transverse wave interference matrix to generate a joint interference matrix. During the detection process, the attitude data recorded by the seabed crawler showed that the maximum change in pitch angle was ±8.3° and the maximum change in roll angle was ±6.7°. These changes interfered with signal acquisition. The researchers constructed an interference error compensation matrix based on the attitude sensor data, and the compensation coefficient a 1 、a 2 and a 3 The values ​​are 0.012, 0.011 and 0.0008 respectively.

[0220] The time-frequency analysis of the compensated joint interference matrix was performed through wavelet transform. The researchers selected the db4 wavelet in the Daubechies wavelet family as the basis function, performed a 5-layer wavelet decomposition, and extracted specific time-frequency features. These features were input into the pre-trained stratigraphic classification prediction model, which is based on a U-shaped convolutional neural network with a self-attention mechanism and consists of an encoder and a decoder.

[0221] The model training used a training set containing 105,000 sets of annotated data and a validation set containing 10,000 sets of annotated data. The training used the Adam optimizer, with an initial learning rate of 0.001, and the learning rate decayed to 0.1 times the original value every 50 training cycles. The loss function used a weighted combination of the cross entropy loss for formation type classification and the mean square error loss for thickness prediction, with weights of 0.6 and 0.4 respectively. After 178 training cycles, the model achieved the best performance on the validation set.

[0222] The final formation classification and thickness prediction accuracy are shown in Table 3:

[0223] Table 3. Strata classification and thickness prediction results accuracy evaluation

[0224] Strata type Classification accuracy (%) Thickness prediction relative error (%) Sample size Sandy sediments 94.6 6.8 1263 Mud deposits 95.2 5.4 1587 Hydrothermal deposits 93.8 7.2 826 Sulfide ore bodies 96.3 4.5 342 Basalt basement 97.5 3.8 982

[0225] Based on the results of stratigraphic classification and thickness prediction, the researchers drew a three-dimensional geological profile of the hydrothermal area of ​​the Okinawa Trough, marking the key geological structures and resource distribution areas. Through this process, three sulfide ore-rich areas and two potential distribution areas of natural gas hydrates were successfully identified. Their specific locations and estimated resource quantities are shown in Table 4:

[0226] Table 4 Location of identified resource-rich areas and estimated resource quantities

[0227]

[0228] Traditional seabed stratum detection mainly relies on a single acoustic wave detection method, such as single-frequency longitudinal wave detection or side-scan sonar. These methods have problems such as limited detection depth and insufficient ability to distinguish different geological structures. Especially in complex seabed hydrothermal areas, traditional methods are difficult to accurately identify ore bodies and resource distribution. Traditional methods can generally only achieve a stratum classification accuracy of 70% to 80%, and the relative error of thickness prediction is usually between 15% and 20%. The multiple longitudinal and transverse wave joint detection method adopted by the present invention significantly improves the detection accuracy by simultaneously utilizing the complementary characteristics of longitudinal and transverse waves and combining the advantages of the seabed crawler platform. It can be seen from the experimental data that the stratum classification accuracy of this method reaches 93.8% to 97.5%, and the relative error of thickness prediction is reduced to 3.8% to 7.2%. In addition, traditional methods often require the use of multiple devices and are complicated to operate. The present invention integrates multiple detection functions into one system, simplifies the operation process, and improves work efficiency. Most importantly, the present invention can more accurately identify the distribution of seabed resources by integrating longitudinal and transverse wave information and adopting advanced machine learning algorithms, providing more reliable technical support for seabed resource exploration.

[0229] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5 and 6 below.

[0230] Table 5 Variable explanation table (Part I)

[0231]

[0232]

[0233] Table 6 Variable explanation table (Part II)

[0234]

[0235] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for joint detection of multiple longitudinal and transverse waves based on a seafloor crawler, characterized in that: include: Deploy a multiple P- and S-wave joint detection device on a seafloor crawler, so that the seafloor crawler crawls along a preset detection path and emits multi-frequency P- and S-wave signals; Receive the P- and S-wave signals reflected from the seabed strata and construct the original P- and S-wave signal matrix; Extract signal frequency characteristics and construct the longitudinal and transverse wave interference matrix; The longitudinal wave interference matrix and the shear wave interference matrix are weightedly fused to generate a joint interference matrix; based on the real-time attitude data recorded by the seabed crawler attitude sensor, an interference error compensation matrix is ​​constructed to compensate the joint interference matrix; time-frequency analysis is performed through wavelet transform to extract specific time-frequency features; the specific time-frequency features are input into a pre-trained formation classification prediction model to achieve classification of seabed formation structure types and thickness prediction; Draw a three-dimensional geological profile of the seabed and mark key geological structures and resource distribution areas.

2. The multiple longitudinal and shear wave joint detection method according to claim 1 is characterized in that: The multiple longitudinal and transverse wave joint detection device consists of a multi-frequency wave source generator, a longitudinal wave transmitting unit, a transverse wave transmitting unit, an omnidirectional receiving sensor array, a signal processing module and a data storage unit. It is installed at the bottom of the seabed crawler and is used to simultaneously transmit multi-frequency longitudinal waves and transverse waves and receive reflected signals.

3. The multiple longitudinal and shear wave joint detection method according to claim 2 is characterized in that: The P-wave interference matrix refers to the interference characteristic matrix generated when the P-wave propagates in different geological structures, which is constructed by extracting the amplitude and phase of each frequency component after performing spectrum analysis on the original P-wave signal matrix. The shear wave interference matrix refers to the interference characteristic matrix generated when the shear wave propagates in different geological structures, which is constructed by extracting the amplitude and phase of each frequency component after performing spectral analysis on the original shear wave signal matrix; the joint interference matrix refers to the matrix formed by combining the longitudinal wave interference matrix and the shear wave interference matrix through a weighted fusion algorithm, which comprehensively reflects the comprehensive interference characteristics of longitudinal and shear waves propagating in seabed strata.

4. The multiple longitudinal and shear wave joint detection method according to claim 3 is characterized in that: The interference error compensation matrix refers to the compensation matrix calculated based on the pitch angle, roll angle and depth data recorded by the seabed crawler attitude sensor, which is used to eliminate the interference caused by the seabed crawler attitude change on signal acquisition.

5. The multiple longitudinal and shear wave joint detection method according to claim 4 is characterized in that: Specific time-frequency features refer to the joint features of time domain and frequency domain extracted after time-frequency analysis of the joint interference matrix through wavelet transform, which are used to reflect the unique response characteristics of different geological structures.

6. The multiple longitudinal and shear wave joint detection method according to claim 5, characterized in that: The submarine stratum wave equation refers to the partial differential equation that describes the propagation characteristics of P- and S-waves in the submarine strata; the submarine stratum wave equation is used to calculate the propagation speed and attenuation characteristics of P- and S-waves in different geological structures; the input of the submarine stratum wave equation includes the stratum density obtained from the submarine crawler, the first Lame parameter obtained from the rock sampling analysis of the submarine crawler, the second Lame parameter obtained from the rock sampling analysis of the submarine crawler, and the displacement vector calculated from the original P-wave signal matrix and the original S-wave signal matrix; the output is the wave displacement field distribution and the wave energy attenuation law, which are used to construct the P-wave interference matrix and the S-wave interference matrix.

7. The multiple longitudinal and shear wave joint detection method according to claim 6, characterized in that: The layer classification prediction model refers to a neural network model used for stratum structure recognition and thickness prediction; the specific structure of the stratum classification prediction model is a U-shaped convolutional neural network based on the self-attention mechanism, which includes an encoder and a decoder; the encoder consists of five convolutional layers and four pooling layers, and the number of convolution kernels in each layer is 32, 64, 128, 256, and 512, respectively, and the convolution kernel size is 3×3; the decoder consists of five deconvolutional layers and four splicing layers, and the number of feature maps in each layer is the same as that of the corresponding layer of the encoder.

8. The multiple longitudinal and shear wave joint detection method according to claim 7, characterized in that: A self-attention mechanism is introduced between the encoder and the decoder to enhance the feature extraction capability. The sparse attention coefficient in the self-attention mechanism is dynamically adjusted according to three key parameters: the ratio of the P-wave and S-wave propagation velocities calculated from the P-wave interference matrix and the S-wave interference matrix, the stratum density obtained from the seabed crawler, and the seabed depth obtained from the seabed crawler.

9. The multiple longitudinal and shear wave joint detection method according to claim 8, characterized in that: The steps of establishing the training data set in the process of training the stratigraphic classification prediction model include collecting longitudinal and transverse wave reflection data of seabed strata of various geological structures from different sea areas, preprocessing the collected data to remove noise and outliers, using manual labeling to determine the true value of the stratigraphic type and thickness, and expanding the number of training samples through data enhancement technology.

10. The multiple longitudinal and shear wave joint detection method according to claim 9, characterized in that: The steps of training the stratigraphic classification prediction model include first initializing the stratigraphic classification prediction model using the preprocessed training data set, using the Adam optimizer for parameter optimization, setting the initial learning rate to 0.001, and decaying the learning rate to 0.1 times the original value every 50 training cycles. The loss function uses a weighted combination of the cross entropy loss for stratigraphic type classification and the mean square error loss for thickness prediction, with weights of 0.6 and 0.4 respectively.

Images