Method for recovering full wave train data damaged or missing in array sonic logging

Abstract

The application discloses a kind of array acoustic logging full wave train data damage or recovery method of loss, steps are: step 1: acquisition array acoustic logging data packet;Step 2: pretreatment full wave data;Step 3: construct multi-scale residual neural network model;Step 4: inversion waveform data;Step 5: output after recovery acoustic logging data packet.The application extracts depth prior information from effective array acoustic logging full wave train data channel using multi-scale residual convolutional neural network, recovers damaged or missing data channel in an unsupervised manner, avoids the cumbersome construction process of large-scale training data set, and provides complete measurement full wave data for subsequent acoustic property evaluation.The application is suitable for monopole, dipole, multipole open hole, casing and while-drilling acoustic logging.

Description

Technical Field

[0001] This invention belongs to the field of applied geophysical acoustic logging technology, specifically relating to a method for recovering damaged or missing array acoustic logging full-wave train data using depth prior information. Background Technology

[0002] Array acoustic logging is a method that uses acoustic signals to probe the formation outside the wellbore. It involves emitting acoustic waves from an internal sound source that propagate along the wellbore wall, which are then collected by an array of transducers above the instrument to obtain full-wave data at a single depth point. During actual logging, the instrument moves up and down within the well, acquiring full-wave data within the measurement depth range. After processing, continuous curves reflecting the acoustic properties of the surrounding medium are obtained, such as rock mechanics parameters, permeability, and porosity, providing high-resolution information support for oil and gas reservoir exploration and development.

[0003] In actual measurement situations, sonic logging instruments will face a variety of complex well conditions. In particular, in the current unconventional oil and gas exploration process, the harsh conditions of high temperature and high pressure downhole can interfere with the normal operation of sonic logging instruments, and in severe cases, cause instrument failure, resulting in damage or loss of measurement waveform data. Summary of the Invention

[0004] The purpose of this invention is to address the problems and shortcomings of the above-mentioned technologies by providing a method for recovering damaged or missing array acoustic logging full-wave train data.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A method for recovering damaged or missing array acoustic logging full-wave train data includes the following steps:

[0007] Step 1: Acquire full-wavelength data packets for array acoustic logging;

[0008] Step 2: Preprocess the full wave train data;

[0009] Step 3: Construct a multi-scale residual neural network model;

[0010] Step 4: Invert waveform data;

[0011] Step 5: Output the recovered acoustic logging data packet.

[0012] Preferably, in step 1, the specific method and steps for acquiring the full wave train data of the array acoustic logging are as follows:

[0013] Step 1-1: Lower the array acoustic logging instrument into the well to the starting measurement position;

[0014] Step 1-2: After the sound source is excited, the sound wave propagating along the well wall is collected using a receiving transducer array within a preset recording time to obtain the measured waveform data, gain and delay data at that depth point.

[0015] Steps 1-3: After moving the array acoustic logging instrument upwards according to the set measurement depth interval, repeat steps 1-2;

[0016] Steps 1-4: Repeat steps 1-3 until the measurement termination position is reached, obtain the array acoustic logging waveform data and gain and delay curves obtained from the measurement of the logging depth range, and store them in the array acoustic logging full wave train data package.

[0017] Preferably, in step 2, the specific method and steps for preprocessing the full-wavelength train data are as follows:

[0018] Step 2-1: Import the array acoustic logging full-wave train data package, obtain the array waveform data and corresponding gain and delay curves, perform gain and delay recovery on the array waveform data, and obtain the common source combined array waveform data. In the formula, nd is the number of measurement depth points, nr is the number of receiving transducers, and nt is the number of measurement time points, which is a combination of time-domain waveforms obtained by the receiving transducer array under the same sound source excitation condition at different depth points.

[0019] Step 2-2: Reconstruct the common source combination data X obtained in Step 2-1 for continuous depth points to obtain the common receiver combination data within the continuous depth profile. It is a combination of time-domain waveforms measured at different depth points by the same receiving transducer in the receiving transducer array.

[0020] Steps 2-3: Based on the waveform data D in the combined common receiver data, determine the measurement depth corresponding to the damaged waveform data of different common receiver profiles, and construct a sampling matrix. Fill all values ​​with 1, and set all measurement depth points corresponding to the damaged waveform data to 0.

[0021] Preferably, in step 3, the specific method for constructing the multi-scale residual neural network model includes the following steps:

[0022] Step 3-1: Construct convolutional units, establish multi-scale feature splicing channels and residual channels to obtain a multi-resolution residual module, which includes the following steps:

[0023] Step 3-1-1: Construct a convolutional unit by connecting convolutional kernels, batch normalization layers, and the LeakyReLU activation function. Its size and stride correspond to the size and stride of the convolutional kernel.

[0024] Step 3-1-2: Stack three convolutional units of size 3×3 and stride 1×1 consecutively. After processing the input features, the output feature corresponding to each convolutional unit is {C}. 1 C 2 C 3}, where the number of channels in each convolutional unit is set to [W / 6, W / 3, W / 2], where W is the number of channels in the multi-resolution residual module;

[0025] Step 3-1-3: Transfer feature {C} 1 C 2 C 3 The multi-scale feature C is obtained by stitching along the channel direction. 4 ;

[0026] Step 3-1-4: Construct residual channels using a convolutional unit of size 1×1, stride 1×1, and number of channels N, and process the input features to obtain the output feature C. 5 ;

[0027] Step 3-1-5: Overlaying features at multiple scales C 4 and residual channel characteristics C 5 Step 3-2: Obtain the output features of the multi-resolution residual module; Step 3-2: Construct the path residual module by connecting convolutional units and residuals, which includes the following steps:

[0028] Step 3-2-1: Construct convolutional units with a size of 3×3 and a stride of 1×1 as described in Step 3-1-1 to form direct convolution channels. After processing the input features, feature C is obtained. 6 ;

[0029] Step 3-2-2: Construct convolutional units of size 1×1 and stride 1×1 as described in step 3-1-1 to form residual channels. After processing the input features, feature C is obtained. 7 ;

[0030] Step 3-2-3: Apply the features {C} obtained in steps 3-2-1 and 3-2-2 6 C 7 The output feature C of the residual path module is obtained by performing the addition. 8 ;

[0031] Step 3-3: Construct an encoder network module for processing input noise data using the multi-resolution residual module built in Step 3-1. This includes the following steps:

[0032] Step 3-3-1: Stack 5 multi-resolution residual modules {EM} consecutively. 1 EM 2 EM 3 EM4 EM 5}, where the number of channels corresponding to the modules are [16, 32, 64, 128, 256], and the multi-resolution residual module EM 1 The input is noise data;

[0033] Step 3-3-2: Using Step 3-1-1, construct convolutional units with a size of 3×3 and a stride of 2×2, and stack them on the multi-resolution residual module {EM}. 1 EM 2 EM 3 EM 4 Following this, its output features are downsampled to obtain the output features {E}. 1 E 2 E 3 E 4}, respectively serving as multi-resolution residual modules {EM 2 EM 3 EM 4 EM 5 Input features of};

[0034] Step 3-4: Construct a path residual network module using the path residual module built in Step 3-2 to process the output features of the encoder network built in Step 3-3. This includes the following steps:

[0035] Step 3-4-1: Based on step 3-2, build four path residual modules {RM} respectively. 1 RM 2 RM 3 RM 4};

[0036] Step 3-4-2: Connect the path residual module to the multi-resolution residual module {EM} in the encoder network built in Step 3-3. 1 EM 2 EM 3 EM 4} to make a connection;

[0037] Step 3-4-3: Path Residual Module {RM 1 RM 2 RM 3 RM 4} respectively for feature {E 1 E 2 E 3 E 4 The process is performed to obtain the output feature {R}. 1 R 2 R 3 R 4};

[0038] Step 3-5: Construct the decoder network module using the multi-resolution residual module built in Step 3-1, which includes the following steps:

[0039] Step 3-5-1: Stack four multi-resolution residual modules {DM} consecutively. 4 DM 3 DM 2 DM 1}, where the number of channels corresponding to the modules are [128, 64, 32, 16] respectively;

[0040] Step 3-5-2: Construct an upsampler of size 2×2 and stride 2×2 using linear interpolation, and cascade it with a batch normalization layer to build an upsampled module. These modules are then stacked on top of the multi-resolution residual module {EM}. 5 DM 4 DM 3 DM 2}after;

[0041] Step 3-5-3: Use the upsampling module from step 3-5-2 to process the multi-resolution residual module EM 5 Output characteristics E 5 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 4 The data is stitched together to form a multi-resolution residual module (DM). 4 The input features are used to obtain the output features D. 4 ;

[0042] Step 3-5-4: Use the upsampling module in step 3-5-2 to process the output feature D obtained in step 3-5-3. 4 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 3 The data is stitched together to form a multi-resolution residual module (DM). 3 The input features are used to obtain the output features D. 3 ;

[0043] Step 3-5-5: Use the upsampling module in step 3-5-2 to process the output feature D obtained in step 3-5-4. 3 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 2 The data is stitched together to form a multi-resolution residual module (DM). 2 The input features are used to obtain the output features D. 2 ;

[0044] Step 3-5-6: Use the upsampling module in step 3-5-2 to process the output feature D obtained in step 3-5-5. 2 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3.1 The data is stitched together to form a multi-resolution residual module (DM). 1 The input features are used to obtain the predicted co-receiver combined waveform data;

[0045] Preferably, in step 4, the specific method and steps for retrieving the inverted waveform data are as follows:

[0046] Step 4-1: Obtain the multi-size residual neural network model from Step 3 through slicing and preprocessing. This model is used to extract reference waveform data for deep prior information. The steps include:

[0047] Step 4-1-1: Process the combined waveform data of the common receiver obtained in Step 2-2. Slicing along the first channel yields waveform data corresponding to a specific receiving transducer.

[0048] Step 4-1-2: Process waveform data d obs Preprocessing is performed using the following formula:

[0049]

[0050] Where σ represents the measured waveform d obs The standard deviation of the waveform data is α, which is the target standard deviation. Adjusting the standard deviation of the waveform data to the target standard deviation helps the neural network extract deep prior information more effectively.

[0051] Step 4-2: Construct the input noise data for the multi-scale residual neural network model in Step 3 using Gaussian noise, which includes the following steps:

[0052] Step 4-2-1: Determine the probability distribution of Gaussian noise, the calculation formula is as follows:

[0053]

[0054] In the formula, μ and σ are the mean and standard deviation, respectively, and are set to 0 and 1.

[0055] Step 4-2-2: Construct two-dimensional Gaussian noise based on the probability distribution formula in Step 4-2-1. Its size is the same as the waveform data in step 4-2-1.

[0056] Step 4-3: Optimize the neural network and gradually invert the waveform data. The specific steps are as follows:

[0057] Step 4-3-1: Input the two-dimensional Gaussian noise constructed in Step 4-2 into the multi-scale residual neural network constructed in Step 3 to obtain the predicted waveform data f. θ (z);

[0058] Step 4-3-2: Combine predicted and measured waveform data to calculate the loss function:

[0059] E(f θ (z), d obs )=|Sf θ (z)-Sd obs | (3)

[0060] In the formula, E(·) represents the l1 distance, and S is the sampling matrix generated in steps 2-4. Since the corrupted data are all set to 0 in the sampling matrix S, the loss function only uses the waveform data of the effective measurements during the calculation process, enabling the neural network to learn rich prior information from it.

[0061] Step 4-3-3: Backpropagate the loss function calculated in step 4-3-2 using the gradient descent algorithm to update the neural network parameters and iteratively invert the waveform data;

[0062] Step 4-3-4: Repeat steps 4-3-1, 4-3-2, and 4-3-3 to continuously optimize the network until the loss function no longer decreases within the set number of iterations, and finally obtain the recovered complete waveform data.

[0063] Step 4-4: Repeat steps 4-1 and 4-3 to process the waveform data corresponding to different receiving transducers and evaluate to obtain the recovered waveforms of all measured waveforms.

[0064] Preferably, in step 5, the specific method and steps for outputting the recovered acoustic logging data packet are as follows:

[0065] Step 5-1: Reverse process the combined data of each co-receiver group according to equation (1) to obtain the original data distribution;

[0066] Step 5-2: Reconstruct the common receiver combination to obtain the common sound source combination data at each depth point and save it.

[0067] The beneficial effects of this invention are:

[0068] (1) Multiscale residual neural networks can effectively extract deep prior information from measured waveform data, and then accurately convert random noise into complete waveform data. No training dataset needs to be established, no complex parameter settings are required, no expert knowledge is needed, and missing or damaged waveform data can be automatically restored, providing complete waveform data for subsequent acoustic parameter evaluation.

[0069] (2) This method can be used in naked-eye or drilling conditions, and can also be used in various sound source conditions such as monopole, dipole, and multipole.

[0070] (3) This method can also be used to interpolate the acquired full-wave data in the depth domain, improve the vertical resolution of the full-wave data, and thus provide high-resolution acoustic parameters to meet the current needs of fine exploration and development. Attached Figure Description

[0071] Figure 1 This is a flowchart of the method of the present invention.

[0072] Figure 2 This is a schematic diagram of array acoustic logging data combination.

[0073] Figure 3 This is a schematic diagram of a multi-resolution residual module.

[0074] Figure 4 This is a schematic diagram of the path residual module.

[0075] Figure 5 This is a schematic diagram of a multi-scale residual neural network structure.

[0076] Figure 6 The recovered waveforms are shown at different iteration times. Detailed Implementation

[0077] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0078] A method for recovering damaged or missing array acoustic logging full-wave train data specifically includes the following steps:

[0079] Step 1: Acquire array acoustic logging data packets:

[0080] Step 1-1: As Figure 2 As shown, the cable array acoustic logging instrument is placed in the well, consisting of a lower acoustic source and an upper array of receiving transducers. At the initial measurement position, after the acoustic source is excited, the energy propagates to each receiving transducer and is recorded as full-wavelength data;

[0081] Step 1-2: After moving the logging instrument up according to the set measurement depth interval, repeat step 1-1;

[0082] Steps 1-3: Repeat steps 1-2 until the termination measurement position is reached to obtain the array acoustic logging data packet for the logging depth range;

[0083] Step 2: Preprocess the full wave train data;

[0084] Step 2-1: Import the array acoustic logging data package, obtain the array waveform data and corresponding gain and delay curves, perform gain and delay recovery on the array waveform data, and obtain the common source combined array waveform data. In the formula, nd represents the number of depth points measured, nr represents the number of receiving transducers, and nt represents the number of time points measured. Figure 2 The vertical box in the middle shows a combination of time-domain waveforms measured by a receiving transducer array under the same sound source excitation condition at different depth points.

[0085] Step 2-2: Reconstruct the common source combination data X obtained in Step 2-1 for continuous depth points to obtain the common receiver combination data within the continuous depth profile. like Figure 2 The oblique box in the middle shows a combination of time-domain waveforms measured at different depth points from the same receiving transducer in the receiving transducer array.

[0086] Steps 2-3: Based on the waveform data D in the combined common receiver data, determine the measurement depth corresponding to the damaged waveform data of different common receiver profiles, and construct a sampling matrix. Fill all values ​​with 1, and set all measurement depth points corresponding to the damaged waveform data to 0;

[0087] Step 3: Construct a multi-scale residual neural network model;

[0088] Step 3-1: As Figure 3 As shown, convolutional units are constructed, multi-scale feature concatenation channels and residual channels are established to obtain multi-resolution residual modules, which extract rich semantic information. The process includes the following steps:

[0089] Step 3-1-1: Construct a convolutional unit by connecting convolutional kernels, batch normalization layers, and the LeakyReLU activation function. Its size and stride correspond to the size and stride of the convolutional kernel.

[0090] Step 3-1-2: Stack three convolutional units of size 3×3 and stride 1×1 consecutively. After processing the input features, the output feature corresponding to each convolutional unit is (C 1 C 2 C 3}, where the number of channels in each convolutional unit is set to [W / 6, W / 3, W / 2], where W is the number of channels in the multi-resolution residual module;

[0091] Step 3-1-3: Feature {C 1 C 2 C 3 The} represent features extracted from different receptive fields. Therefore, by concatenating them along the channel direction, we can obtain the multi-scale feature C. 4 ;

[0092] Step 3-1-4: Construct residual channels using a convolutional unit of size 1×1, stride 1×1, and number of channels N, and process the input features to obtain the output feature C. 5 ;

[0093] Step 3-1-5: Overlaying features at multiple scales C 4 and residual channel characteristics C 5 The output features of the multi-resolution residual module are obtained. By adding residual branches, the learning ability of the model can be improved without increasing the number of parameters.

[0094] Step 3-2: As Figure 4 As shown, a path residual module is constructed by using convolutional units and residual connections to extract shallow features from the encoder module and further preserve spatial details in the waveform data, reducing the semantic gap between the encoder and decoder. This includes the following steps:

[0095] Step 3-2-1: Construct convolutional units with a size of 3×3 and a stride of 1×1 as described in Step 3-1-1 to form direct convolution channels. After processing the input features, feature C is obtained. 6 ;

[0096] Step 3-2-2: Construct convolutional units of size 1×1 and stride 1×1 as described in step 3-1-1 to form residual channels. After processing the input features, feature C is obtained. 7 ;

[0097] Step 3-2-3: Apply the features {C} obtained in steps 3-2-1 and 3-2-2 6 C 7 The output feature C of the residual path module is obtained by performing the addition. 8 ;

[0098] Step 3-3: As Figure 5 As shown in the middle left section, an encoder network module is built using the multi-resolution residual module constructed in step 3-1 to process input noise data. This includes the following steps:

[0099] Step 3-3-1: Stack 5 multi-resolution residual modules {EM} consecutively. 1 EM 2 EM 3 EM 4 EM 5}, where the number of channels corresponding to the modules are [16, 32, 64, 128, 256], and the multi-resolution residual module EM 1 The input is noise data;

[0100] Step 3-3-2: Using Step 3-1-1, construct convolutional units with a size of 3×3 and a stride of 2×2, and stack them on the multi-resolution residual module {EM}. 1 EM 2 EM 3 EM 4 Following this, its output features are downsampled to obtain the output features {E}. 1 E 2 E 3 E 4}, respectively serving as multi-resolution residual modules {EM 2 EM 3 EM 4 EM 5 Input features of};

[0101] Steps 3-4: As Figure 5 The path residual network module, constructed using the path residual module built in step 3-2, is shown in the middle section. This module is used to process the output features of the encoder network built in step 3-3, and includes the following steps:

[0102] Step 3-4-1: Based on step 3-2, build four path residual modules {RM} respectively. 1 RM 2 RM 3 RM 4};

[0103] Step 3-4-2: Connect the path residual module to the multi-resolution residual module {EM} in the encoder network built in Step 3-3. 1 EM 2 EM 3 EM 4} to make a connection;

[0104] Step 3-4-3: Path Residual Module {RM 1 RM 2 RM 3 RM 4} respectively for feature {E 1 E 2 E 3 E 4 The process is performed to obtain the output features {R1, R}. 2 R 3 R 4};

[0105] Steps 3-5: (e.g.) Figure 5 As shown on the right, the decoder network module is built using the multi-resolution residual module constructed in step 3-1, which includes the following steps:

[0106] Step 3-5-1: Stack four multi-resolution residual modules {DM} consecutively. 4 DM 3 DM 2 DM 1}, where the number of channels corresponding to the modules are [128, 64, 32, 16] respectively;

[0107] Step 3-5-2: Construct an upsampler of size 2×2 and stride 2×2 using linear interpolation, and cascade it with a batch normalization layer to build an upsampled module. These modules are then stacked on top of the multi-resolution residual module {EM}. 5 DM 4 DM 3 DM 2}after;

[0108] Step 3-5-3: Use the upsampling module from step 3-5-2 to process the multi-resolution residual module EM 5 Output characteristics E 5 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 4 The data is stitched together to form a multi-resolution residual module (DM). 4 The input features are used to obtain the output features D. 4 ;

[0109] Step 3-5-4: Use the upsampling module in step 3-5-2 to process the output feature D obtained in step 3-5-3. 4 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 3 The data is stitched together to form a multi-resolution residual module (DM). 3 The input features are used to obtain the output features D. 3 ;

[0110] Step 3-5-5: Use the upsampling module in step 3-5-2 to process the output feature D obtained in step 3-5-4. 3 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 2 The data is stitched together to form a multi-resolution residual module (DM). 2 The input features are used to obtain the output features D. 2 ;

[0111] Step 3-5-6: Use the upsampling module in step 3-5-2 to process the output feature D obtained in step 3-5-5. 2 After upsampling, the output feature R of step 3-4-3 is compared with that of step 4-3. 1 The data is stitched together to form a multi-resolution residual module (DM). 1 The input features are used to obtain the predicted co-receiver combined waveform data;

[0112] Step 4: Invert waveform data:

[0113] Step 4-1: Obtain the reference waveform data used by the multi-size residual neural network model in Step 3 to extract depth prior information through slicing and preprocessing. This includes the following steps:

[0114] Step 4-1-1: Process the combined waveform data of the common receiver obtained in Step 2-2. Slicing along the first channel yields waveform data corresponding to a specific receiving transducer. like Figure 6 As shown on the right side of the middle;

[0115] Step 4-1-2: Process waveform data d obs Preprocessing is performed using the following formula:

[0116]

[0117] Where σ represents the measured waveform d obs The standard deviation of the waveform data is α, which is the target standard deviation. Adjusting the standard deviation of the waveform data to the target standard deviation helps the neural network extract deep prior information more effectively.

[0118] Step 4-2: Construct the input noise data for the multi-scale residual neural network model in Step 3 using Gaussian noise, which includes the following steps:

[0119] Step 4-1-1: Determine the probability distribution of Gaussian noise, the calculation formula is as follows:

[0120]

[0121] In the formula, μ and σ are the mean and standard deviation, respectively, and are set to 0 and 1.

[0122] Step 4-1-2: Construct two-dimensional Gaussian noise based on the probability distribution formula in Step 4-1-1. Its size is the same as the waveform data in step 4-1-1, such as Figure 6 As shown on the left side of the middle section.

[0123] Step 4-3: Optimize the neural network and gradually invert the waveform data. The specific steps are as follows:

[0124] Step 4-3-1: Input the two-dimensional Gaussian noise constructed in Step 4-2 into the multi-scale residual neural network constructed in Step 3 to obtain the predicted waveform data f. θ (z);

[0125] Step 4-3-2: Combine predicted and measured waveform data to calculate the loss function:

[0126] E(f θ (z), dobs )=|Sf θ (z)-Sd obs | (3)

[0127] In the formula, E(·) represents the l1 distance, and S is the sampling matrix generated in steps 2-4. Since the corrupted data are all set to 0 in the sampling matrix S, the loss function only uses the waveform data of the effective measurements during the calculation process, enabling the neural network to learn rich prior information from it.

[0128] Step 4-3-3: Backpropagate the loss function calculated in step 4-3-2 using the gradient descent algorithm to update the neural network parameters and iteratively invert the waveform data;

[0129] Step 4-3-4: Repeat steps 4-3-1, 4-3-2, and 4-3-3 to continuously optimize the network until the loss function no longer decreases within the set number of iterations, obtaining the final recovered complete waveform data. For example... Figure 6 As shown, the waveform data generated by the network when the number of iterations is 0, 40, 80, and 1000 are given respectively. It can be clearly seen that the waveform at the effective measurement position is completely restored, and the waveform data of the missing and damaged channels are also effectively generated.

[0130] Step 4-4: Repeat steps 4-1 and 4-3 to process the waveform data corresponding to different receiving transducers and evaluate to obtain the recovered waveforms of all measured waveforms.

[0131] Step 5: Output the recovered acoustic logging data packet:

[0132] Step 5-1: Reverse process the combined data of each co-receiver group according to equation (1) to obtain the original data distribution;

[0133] Step 5-2: Reconstruct the common receiver combination to obtain the common sound source combination data at each depth point and save it.

[0134] Working principle:

[0135] In acoustic logging, energy generated by a sound source propagates along the wellbore and is collected by an array of receiver transducers at the top. Through continuous sliding measurements within the well section, array acoustic logging waveform data packets at that depth are obtained. Due to the complex working environment of high temperature and high pressure downhole, some waveform data may be damaged or missing. To recover these signals, the waveform data packets are first decoded and reconstructed to generate common receiver combined data and a corresponding sampling matrix. Then, a multi-scale residual neural network is built to learn the depth prior information of the valid data channels, converting random noise into waveform data within the valid data channels. Simultaneously, the waveform data of missing or damaged channels is recovered.

[0136] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for recovering damaged or missing array acoustic logging full-wave train data, characterized in that: Includes the following steps: Step 1: Acquire full wave train data from array acoustic logging. The specific methods and steps are as follows: Step 1-1: Lower the array acoustic logging instrument into the well to the starting measurement position; Step 1-2: After the sound source is excited, the sound waves propagating along the well wall are collected using a receiving transducer array within a preset recording time to obtain the current depth point measurement waveform data, gain and delay data. Steps 1-3: After moving the array acoustic logging instrument upwards according to the set measurement depth interval, repeat steps 1-2; Steps 1-4: Repeat steps 1-3 until the measurement termination position is reached, obtain the array acoustic logging waveform data and gain and delay curves obtained from the current logging depth range, and store them in the array acoustic logging full wave train data package; Step 2, preprocess the full-wavelength data. The specific methods and steps are as follows: Step 2-1: Import the array acoustic logging full-wave train data package, obtain the array waveform data and corresponding gain and delay curves, perform gain and delay recovery on the array waveform data, and obtain the common source combined array waveform data. In the formula, nd is the number of measurement depth points, nr is the number of receiving transducers, and nt is the number of measurement time points. It is a combination of time-domain waveforms obtained by the receiving transducer array under the same sound source excitation condition at different depth points. Step 2-2: Reconstruct the common source combination data X obtained in Step 2-1 for continuous depth points to obtain the common receiver combination data within the continuous depth profile. It is a combination of time-domain waveforms measured at different depth points by the same receiving transducer in the receiving transducer array. Steps 2-3: Based on the waveform data D in the combined common receiver data, determine the measurement depth corresponding to the damaged waveform data of different common receiver profiles, and construct a sampling matrix. Fill all values ​​with 1, and set all measurement depth points corresponding to the damaged waveform data to 0. Step 3: Construct a multi-scale residual neural network model. The specific method includes the following steps: Step 3-1: Construct convolutional units, establish multi-scale feature splicing channels and residual channels to obtain multi-resolution residual modules. Step 3-2: Construct a path residual module using convolutional units and residual connections. Step 3-3: Construct an encoder network module for processing input noise data using the multi-resolution residual module built in Step 3-1. Step 3-4: Using the path residual module built in Step 3-2, construct a path residual network module to process the output features of the encoder network built in Step 3-3. Step 3-5: Construct the decoder network module using the multi-resolution residual module built in Step 3-1. Step 4, invert waveform data. The specific methods and steps are as follows: Step 4-1: Obtain the multi-size residual neural network model from Step 3 through slicing and preprocessing, which is used to extract reference waveform data for depth prior information. Step 4-2: Construct the input noise data for the multi-scale residual neural network model in Step 3 using Gaussian noise. Step 4-3: Optimize the neural network and gradually invert the waveform data. Step 4-4: Repeat steps 4-1 and 4-3 to process the waveform data corresponding to different receiving transducers and evaluate to obtain the recovered waveforms of all measured waveforms; Step 5: Output the recovered acoustic logging data packet. The specific methods and steps are as follows: Step 5-1: Reverse process the combined data of each co-receiver group according to equation (1) to obtain the original data distribution; Step 5-2: Reconstruct the common receiver combination to obtain the common sound source combination data at each depth point and save it.

2. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of step 3-1 described in step 3 are as follows: Step 3-1-1: Construct a convolutional unit by concatenating convolutional kernels, batch normalization layers, and the LeakyReLU activation function. Its size and stride correspond to the size and stride of the convolutional kernel. Step 3-1-2: Stack three convolutional units of size 3×3 with a stride of 1×1 consecutively. After processing the input features, the output feature corresponding to each convolutional unit is... The number of channels in each convolutional unit is set to , where W is the number of channels in the multi-resolution residual module; Step 3-1-3: Features By stitching along the channel direction, multi-scale features are obtained. ; Step 3-1-4: Construct residual channels using a convolutional unit of size 1×1, stride 1×1, and number of channels N, and process the input features to obtain the output features. ; Step 3-1-5: Overlaying features at multiple scales and residual channel characteristics The output characteristics of the multi-resolution residual module are obtained.

3. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of step 3-2 described in step 3 are as follows: Step 3-2-1: Construct the structure with the dimensions specified in Step 3-1-1. Step size is The convolutional units form direct convolutional channels, which process the input features to obtain the features. ; Step 3-2-2: Construct the structure with the dimensions specified in Step 3-1-1. Step size is The convolutional units form residual channels, which process the input features to obtain the feature vectors. ; Step 3-2-3: Analyze the features obtained in steps 3-2-1 and 3-2-2. The output features of the residual path module are obtained by performing addition. .

4. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of step 3-3 described in step 3 are as follows: Step 3-3-1: Stack 5 multi-resolution residual modules consecutively. }, where the number of channels corresponding to the modules are respectively Among them, the multi-resolution residual module The input is noise data; Step 3-3-2: Using step 3-1-1, construct a structure with the following dimensions. Step size is The convolutional units are stacked on top of the multi-resolution residual module. Following this, its output features are downsampled to obtain the output features { }, respectively serving as multi-resolution residual modules { The input features of}.

5. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of steps 3-4 described in step 3 are as follows: Step 3-4-1: Based on step 3-2, build four path residual modules respectively. }; Step 3-4-2: Connect the path residual module to the multi-resolution residual module in the encoder network built in Step 3-3. } to make a connection; Step 3-4-3: Path Residual Module { } respectively for features { } is processed to obtain the output features { } 6. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of steps 3-5 in step 3 are as follows: Step 3-5-1: Stack four multi-resolution residual modules consecutively. }, where the number of channels corresponding to the modules are [128, 64, 32, 16] respectively; Step 3-5-2: Construct the dimensions using linear interpolation. Step size is An upsampler is constructed by concatenating it with a batch normalization layer to build an upsampling module, which is then stacked on top of the multi-resolution residual module. }after; Step 3-5-3: Use the upsampling module from step 3-5-2 to process the multi-resolution residual module. Output characteristics After upsampling, the output features are compared with those in step 3-4-3. The data is stitched together to form a multi-resolution residual module. Input features, output features ; Step 3-5-4: Use the upsampling module in step 3-5-2 to process the output features obtained in step 3-5-3. After upsampling, the output features are compared with those in step 3-4-3. The data is stitched together to form a multi-resolution residual module. Input features, output features ; Step 3-5-5: Use the upsampling module in step 3-5-2 to process the output features obtained in step 3-5-4. After upsampling, the output features are compared with those in step 3-4-3. The data is stitched together to form a multi-resolution residual module. Input features, output features ; Step 3-5-6: Use the upsampling module in step 3-5-2 to process the output features obtained in step 3-5-5. After upsampling, the output features are compared with those in step 3-4-3. The data is stitched together to form a multi-resolution residual module. The input characteristics are used to obtain the predicted combined waveform data of the common receiver.

7. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of step 4-1 described in step 4 are as follows: Step 4-1-1: Process the combined waveform data of the common receiver obtained in Step 2-2. Slicing along the first channel yields waveform data corresponding to a specific receiving transducer. ; Step 4-1-2: Process waveform data Preprocessing is performed using the following formula: （1）； in Indicates the measured waveform standard deviation By adjusting the standard deviation of the waveform data to the target standard deviation, the neural network can extract deep prior information more effectively.

8. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of step 4-2 described in step 4 are as follows: Step 4-2-1: Determine the probability distribution of Gaussian noise, the calculation formula is as follows: （2）； In the formula, and These represent the mean and standard deviation, set to 0 and 1 respectively. Step 4-2-2: Construct two-dimensional Gaussian noise based on the probability distribution formula in Step 4-2-1. Its size is the same as the waveform data in step 4-2-1.

9. The method for recovering damaged or missing array acoustic logging full-wave train data according to claim 1, characterized in that: The specific content, methods, and steps of step 4-3 described in step 4 are as follows: Step 4-3-1: Input the two-dimensional Gaussian noise constructed in Step 4-2 into the multi-scale residual neural network constructed in Step 3 to obtain the predicted waveform data. ; Step 4-3-2: Combine predicted and measured waveform data to calculate the loss function: （3）； In the formula, express Distance, S is the sampling matrix generated in steps 2-4. Since the damaged data are all set to 0 in the sampling matrix S, the loss function only uses the waveform data of the effective measurement during the calculation process, so that the neural network can learn rich prior information from it. Step 4-3-3: Backpropagate the loss function calculated in step 4-3-2 using the gradient descent algorithm to update the neural network parameters and iteratively invert the waveform data; Step 4-3-4: Repeat steps 4-3-1, 4-3-2, and 4-3-3 to continuously optimize the network until the loss function no longer decreases within the set number of iterations, and finally obtain the recovered complete waveform data.

Images