A method, device, computer equipment and storage medium for reducing noise in well logging data
Through the multi-head attention mechanism and soft threshold learning module combined with residual connection layer method, the problems of high redundancy and low signal-to-noise ratio of well logging data are solved, and the resistivity of denoising formations is accurately predicted, which improves the oil layer and water Accuracy of layer distribution judgment.
Patent Information
- Application Number
- CN202410676009.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-29
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-05-29
AI Technical Summary
The existing logging data is highly redundant and has a low signal-to-noise ratio, which cannot effectively reduce noise, resulting in difficulty in judging the distribution of oil and water layers.
The multi-head attention mechanism and soft threshold learning module combined with residual connection layer method is used to obtain the resistivity of the denoised formation through feature extraction, dimensionality reduction and denoising processing.
It effectively reduces data redundancy and noise interference, improves feature characterization capabilities, and improves the accuracy of distribution judgment of oil and water layers.
Smart Images

Figure CN118296307B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of well logging technology, and in particular to a well logging data noise reduction method, a management platform, a computer device and a storage medium. Background Art
[0002] With the continuous deepening of oil and gas exploration and development, downhole logging technology plays a key role in solving the definition of oil and water layers, perforation planning, etc. However, due to factors such as engineering progress, some wells did not conduct open hole resistivity logging before completion, which led to difficulties in accurately locating oil and water layers during the production stage. In the absence of open hole resistivity logging, many wells have deviations in perforation and cannot accurately determine the distribution of oil and water layers. To solve this problem, traditional technology can obtain resistivity information through transient electromagnetic logging to replace the missing true resistivity curve. However, transient electromagnetic logging is easily affected by a variety of interference factors, and the logging data is composed of a series of data. There are redundant parts between different data, which increases the complexity of noise processing and causes the effective signal to be masked or distorted. Summary of the invention
[0003] The purpose of the present application is to solve at least one of the above-mentioned technical defects, especially the problem that the logging data in the prior art has high redundancy and low signal-to-noise ratio and cannot be effectively noise-reduced.
[0004] In a first aspect, the present application provides a method for reducing noise in well logging data, comprising:
[0005] Obtain time series logging data; the time series logging data includes formation resistivity and multiple other logging data, and the dimensions of the time series logging data include time series length and feature quantity;
[0006] Inputting the time series logging data into the first linear layer for feature extraction to obtain first feature data;
[0007] The first feature data is input into the second linear layer to reduce the dimension of the feature quantity to obtain the second feature data;
[0008] The first feature data is input into a soft threshold learning module to obtain a denoising soft threshold; the soft threshold learning module identifies redundant components based on a multi-head attention mechanism and predicts a denoising soft threshold;
[0009] Process the second feature data according to a denoising soft threshold to obtain third feature data;
[0010] The time series logging data is passed through the residual connection layer and then summed with the third feature data. The summation result is synthesized by dimensionality reduction to obtain the denoised formation resistivity with a feature number of 1.
[0011] In one embodiment, the first feature data is input into a soft threshold learning module to obtain a denoising soft threshold, including:
[0012] The first feature data is absolute-valued and globally average-pooled to obtain non-time series feature data;
[0013] The non-time series feature data is input into the third linear layer, the feature quantity dimension is reduced to 1, and the fourth feature data is obtained;
[0014] Input the non-temporal feature data into the multi-head attention mechanism unit to obtain the multi-head attention splicing weight;
[0015] The multi-head attention splicing weights are processed in turn by the batch normalization layer, the ReLu layer, the second fully connected layer and the Sigmoid layer to obtain the fifth feature data, and the fifth feature data is multiplied by the fourth feature data to obtain the denoising soft threshold.
[0016] In one embodiment, non-temporal feature data is input into a multi-head attention mechanism unit to obtain a multi-head attention splicing weight, including:
[0017] The non-time series feature data are converted into corresponding calculation parameters through three linear layers respectively;
[0018] Input each calculation parameter into the linear layer corresponding to each attention head to obtain the multi-head attention features corresponding to each attention head;
[0019] After scaling the dot product and fusing the multi-head attention features, we get the multi-head attention concatenation weights.
[0020] In one embodiment, other logging data include sonic logging data, natural gamma logging data, induction logging data, natural potential logging data, and through-casing electromagnetic logging data obtained by multiple different current meter tests.
[0021] In one embodiment, the residual connection layer is a fourth linear layer.
[0022] In one embodiment, before inputting the time series logging data into the first linear layer for feature extraction to obtain the first feature data, the method further includes:
[0023] Preprocess the time series logging data.
[0024] In one embodiment, preprocessing the time series logging data includes any one or more of the following:
[0025] Correct outliers in time series logging data;
[0026] Fill missing values in time series well logging data.
[0027] In a second aspect, the present application provides a well logging data noise reduction device, comprising:
[0028] A data acquisition module is used to acquire time series logging data; the time series logging data includes formation resistivity and multiple other logging data, and the dimensions of the time series logging data include time series length and feature quantity;
[0029] A first processing module is used to input the time series logging data into a first linear layer for feature extraction to obtain first feature data;
[0030] A second processing module is used to input the first feature data into a second linear layer, reduce the dimension of the feature quantity, and obtain second feature data;
[0031] A third processing module is used to input the first feature data into a soft threshold learning module to obtain a denoising soft threshold; the soft threshold learning module identifies redundant components based on a multi-head attention mechanism and predicts a denoising soft threshold;
[0032] A fourth processing module, used for processing the second feature data according to a denoising soft threshold to obtain third feature data;
[0033] The fifth processing module is used to sum the time series logging data with the third feature data after passing through the residual connection layer, perform dimensionality reduction synthesis on the summation result, and obtain the denoised formation resistivity with a feature number of 1.
[0034] In a third aspect, the present application provides a computer device comprising one or more processors and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the steps of the well logging data denoising method in any of the above-mentioned embodiments are executed.
[0035] In a fourth aspect, the present application provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the well logging data denoising method in any of the above-mentioned embodiments.
[0036] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:
[0037] First, the time series logging data containing multiple logging data is obtained as input, and then the input data is subjected to preliminary feature conversion through the first linear layer, and then the feature quantity dimension is reduced through the second linear layer to achieve feature compression and reduce redundancy. Then the output of the first linear layer is input into the soft threshold learning module, and the multi-head attention mechanism is used to capture the relevant features of the input data, and the denoising soft threshold is predicted. The predicted soft threshold is applied to the second feature data to complete the actual soft threshold denoising and redundancy removal processing. Finally, the original input information is combined with the denoised feature data through the residual connection, and the dimension is further reduced to synthesize the single-channel output, that is, the final denoised formation resistivity is obtained. This scheme combines the multi-head attention mechanism, the soft threshold attention mechanism and the residual network, which can not only ensure the effectiveness of training, but also effectively reduce data redundancy and noise interference through dimensionality reduction compression, making the feature distribution more compact and efficient, and reducing the computational complexity. It can also better capture the correlation and importance between features, thereby improving the representation ability of features. At the same time, the multi-head attention mechanism can better capture the redundant part and the random fluctuation part of the time series noise in the process of learning the soft threshold, so as to predict the corresponding denoising soft threshold, and finally the soft threshold processing unit realizes denoising, which has strong redundancy removal and noise reduction capabilities. The final output is a single-channel result, focusing on predicting accurate denoised formation resistivity curves to meet the core needs of well logging exploration. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0039] Figure 1 A schematic diagram of a process for reducing noise in well logging data provided by an embodiment of the present application;
[0040] Figure 2 A schematic diagram of the structure of well logging data noise reduction provided by one embodiment of the present application;
[0041] Figure 3 This is a schematic diagram of a process for predicting a denoising soft threshold in one embodiment of the present application;
[0042] Figure 4 A schematic diagram of a structure for noise reduction of well logging data provided by another embodiment of the present application;
[0043] Figure 5 A schematic diagram of a module of a well logging data noise reduction device provided in one embodiment of the present application;
[0044] Figure 6 An internal structure diagram of a computer device provided for one embodiment of the present application. DETAILED DESCRIPTION
[0045] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0046] In the field of well logging, different types of logging methods can obtain different measurement data. Combining these measurement data collected in time series constitutes a multi-dimensional time series logging data set. These data are easily affected by electromagnetic interference and electrical noise of the measuring equipment itself. And underground media, such as rocks, water and other materials, will also interfere with and attenuate the signal. More importantly, the test will also be affected by the casing. The casing is made of conductive materials, which will introduce additional interference signals into various response signals and cause distortion. Correcting the casing effect requires complex models and algorithms. The diversity of actual geological conditions will also increase the difficulty of correction. Many correction algorithms are only theoretical and cannot be applied to actual working conditions. Therefore, various types of transient electromagnetic logging curves through casing belong to multi-source data with low signal-to-noise ratio and high redundancy. In data with low signal-to-noise ratio, the impact of noise is more significant. Highly redundant data also increases the complexity of noise processing, because similar noise patterns may exist in redundant data, making noise removal more challenging. Therefore, in order to better utilize the logging data, this embodiment provides a logging data denoising method based on deep learning and neural network, including steps S102 to S112.
[0047] S102, acquiring time series logging data. The time series logging data includes formation resistivity and a plurality of other logging data, and the dimensions of the time series logging data include time series length and feature quantity.
[0048] It can be understood that time series logging data refers to a series of response data that changes with time and is obtained through a variety of different logging methods during the logging process. Time series logging data includes formation resistivity and multiple other logging data. Formation resistivity refers to the degree of resistance of the formation to the passage of electric current, and is a key parameter for describing the conductivity of the formation. Accurately obtaining formation resistivity is of great significance for identifying oil and gas reservoirs. Other logging data are data obtained by other logging methods in addition to formation resistivity, such as curve data obtained by various measurement methods such as sonic logging, density logging, natural gamma logging, and electrical logging. Regarding the dimension of time series logging data, the feature quantity dimension represents the number of types of measurement data in the time series logging data. For example, the time series logging data includes 40 types of data, including AC (acoustic logging curve), GR (natural gamma curve), CON (induction logging curve), SP (natural potential curve), through-casing electromagnetic logging curve STEMUB01~STEMUB35 (instrument ammeter readings 01 to 35), and RT (formation resistivity), so the number of features is 40. The length of the time series represents the number of different time points included in each measurement data. In order to improve processing efficiency, in the field of neural networks, all data will not be processed at once, but will be processed in batches. Therefore, the training batch dimension can also be increased based on these two dimensions. For example Figure 2 As shown, the data in this embodiment may include three dimensions: B, S, and F, where B represents the training batch, S represents the time series length, and F represents the number of features. The following text will take the training batch dimension as an example for explanation, but this dimension may not be included in actual use.
[0049] S104, inputting the time series logging data into the first linear layer for feature extraction to obtain first feature data.
[0050] It can be understood that the method in this embodiment can be applied to Figure 2 The first linear layer is used to linearly transform the input and output through the weight matrix to achieve preliminary encoding and feature conversion, extract new feature representations that are more relevant to the formation conductivity from the original time series logging data, and prepare for subsequent deeper feature extraction and processing. The first linear layer will not affect the dimension of the time series logging data. The first feature data has the same dimension as the time series logging data, such as Figure 2 As shown, the dimension of the data input to the first linear layer is (B, S, F), and the dimension of the data output by the first linear layer is still (B, S, F).
[0051] S106, input the first feature data into the second linear layer, reduce the dimension of the feature quantity, and obtain second feature data.
[0052] It can be understood that the role of the second linear layer is to reduce the dimension of the first layer feature data in the dimension of feature quantity. Dimension reduction can reduce redundant information and make the feature distribution more compact. That is, the second linear layer only affects the first feature data in the dimension of feature quantity. The second feature data is the same as the first feature data in the dimensions of training batch and time series length, but the dimension of feature quantity is smaller than that of the first feature data. Figure 2 As shown, the dimension of the first feature data output by the first linear layer is (B, S, F), and the dimension of the second feature data output after the second linear layer is (B, S, H), where H is a positive integer less than F. Excessive dimensionality reduction in the second linear layer may lead to loss of effective information, so the ratio between H and F needs to be weighed through experiments. It can generally be set to 1 / 4-1 / 3. For example, the dimension of the first feature data is (32, 40, 42), and the dimension of the second feature data can be (32, 40, 12).
[0053] S108, inputting the first feature data into a soft threshold learning module to obtain a denoising soft threshold.
[0054] Understandably, Figure 2 The structure in is an improvement made on the basis of the traditional deep residual shrinkage network. The conventional deep residual shrinkage network is used in the image field, and this embodiment improves it to be suitable for time series data processing. And further combine the soft threshold processing mechanism with the soft threshold mechanism. The soft threshold learning module in this embodiment is used to identify redundant components in the input data based on the multi-head attention mechanism and predict the denoising soft threshold. The multi-head attention mechanism will focus on the part with a higher attention score in the input data, which is the redundant part in the multi-dimensional feature. The denoising soft threshold can be predicted based on the redundant part, and the redundant part can be removed later by the soft threshold mechanism. In addition, in the soft threshold learning module, the time series dimension can also be fused. By comparison, the random noise fluctuation of the time series dimension can be found. When predicting the soft threshold, the relevant information about the random noise fluctuation can be learned at the same time, so that the noise can be removed together during the subsequent soft threshold processing. For different time series logging data input, the soft threshold learning module can autonomously drive the learning of the optimal threshold according to the data distribution of the current input data, without manual setting. The multi-head attention mechanism can well capture the correlation and importance between features, thereby accurately predicting the soft threshold with good noise reduction and redundancy removal effects. The output of the soft threshold learning module is used for each group of data in the batch data separately. Therefore, the size of its feature quantity dimension and time series length dimension are both 1, such as Figure 2As shown, the dimension of the first feature data input to the soft threshold learning module is (B, S, F), and the dimension of the denoising soft threshold output by the soft threshold learning module is (B, 1, 1). Specifically, the dimension of the first feature data can be (32, 40, 42), and the dimension of the denoising soft threshold can be (32, 1, 1).
[0055] S110, processing the second feature data according to a denoising soft threshold to obtain third feature data.
[0056] It can be understood that the soft threshold mechanism refers to the use of flexible and variable thresholds to filter the input signal. The soft threshold output by the soft threshold learning module in this embodiment learns the distribution of the input data, that is, the soft threshold that varies according to the input data. Soft threshold processing is to set the features in the input data whose absolute value is lower than the denoising soft threshold to zero, and other features will also be adjusted to zero to achieve shrinkage. The expression of the soft threshold function and its solution are as follows:
[0057] ;
[0058] ;
[0059] In the above formula, x represents the second feature data, τ represents the denoising soft threshold, and y represents the denoised third feature data. This step applies the denoising soft threshold obtained in the previous step to the second feature data to complete the actual denoising process and effectively remove redundant features and noise interference. This step will not affect the dimension of the data, but only the specific value of the data. Therefore, the dimension of the second feature data is the same as that of the third feature data, such as Figure 2 As shown, the dimension of the second feature data output by the second linear layer is (B, S, H), and the dimension of the third feature data obtained after the soft threshold is also (B, S, H). For example, the dimension of the second feature data is (32, 40, 12), and the dimension of the third feature data can also be (32, 40, 12).
[0060] S112, summing the time series logging data with the third feature data after passing through the residual connection layer, performing dimension reduction synthesis on the summation result, and obtaining a denoised formation resistivity with a feature number of 1.
[0061] It can be understood that the main branch of data processing is the time series feature data passing through the first linear layer, the second linear layer and the soft threshold processing unit in sequence. However, as the neural network deepens, it will also bring about problems such as gradient vanishing and gradient explosion, which will cause many troubles for model training. In order to solve this problem, the original input, that is, the time series logging data, is summed with the output of the main branch through the residual connection layer to form a residual network structure. This structure can solve the problem of gradient dissipation very well, and can also enhance the expression effect of the model and reduce the difficulty of model training. The residual connection layer will adjust the dimension of the time series logging data to be the same as the third feature data to facilitate better addition with the third feature data. Figure 2 As shown, the dimension of the time series logging data is (B, S, F), and the dimension of the third feature data obtained after soft thresholding is (B, S, H). The residual connection layer will reduce the dimension of the time series logging data to (B, S, H).
[0062] For example, if the dimension of the time series logging data is (32, 40, 42), and the dimension of the third feature data is (32, 40, 12), then the dimension of the data output by the residual connection layer is also (32, 40, 12). Finally, the denoised multi-channel data needs to be fused to reduce the feature quantity dimension to 1, that is, only the denoised formation resistivity of a single channel is output, and only the single-channel result is produced, focusing on the main task of the model in predicting formation resistivity. The dimensionality reduction synthesis here is generally achieved using a fully connected layer. The function of this fully connected layer is to sum the multi-channel data of the feature quantity dimension according to the weight obtained by training to become the final single-channel data. The denoised formation resistivity obtained in the end is still time series data. That is, the training batch dimension and the time series length dimension are the same between the summation result and the result output by the fully connected layer, but the fusion is performed on the feature quantity dimension. Specifically, if the dimension of the summation result is (B, S, H), the dimension of the result after the fully connected layer becomes (B, S, 1). For example, the dimension of the summation result is (32, 40, 12), and the dimension of the final denoised resistivity is (32, 40, 1).
[0063] First, the time series logging data containing multiple logging data is obtained as input, and then the input data is subjected to preliminary feature conversion through the first linear layer, and then the feature quantity dimension is reduced through the second linear layer to achieve feature compression and reduce redundancy. Then the output of the first linear layer is input into the soft threshold learning module, and the multi-head attention mechanism is used to capture the relevant features of the input data, and the denoising soft threshold is predicted. The predicted soft threshold is applied to the second feature data to complete the actual soft threshold denoising and redundancy removal processing. Finally, the original input information is combined with the denoised feature data through the residual connection, and the dimension is further reduced to synthesize the single-channel output, that is, the final denoised formation resistivity is obtained. This scheme combines the multi-head attention mechanism, the soft threshold attention mechanism and the residual network, which can not only ensure the effectiveness of training, but also effectively reduce data redundancy and noise interference through dimensionality reduction compression, making the feature distribution more compact and efficient, and reducing the computational complexity. It can also better capture the correlation and importance between features, thereby improving the representation ability of features. At the same time, the multi-head attention mechanism can better capture the redundant part and the random fluctuation part of the time series noise in the process of learning the soft threshold, so as to predict the corresponding denoising soft threshold, and finally the soft threshold processing unit realizes denoising, which has strong redundancy removal and noise reduction capabilities. The final output is a single-channel result, focusing on predicting accurate denoised formation resistivity curves to meet the core needs of well logging exploration.
[0064] The above embodiments need to rely on Figure 2The well logging data denoising model in is implemented, and the model includes a first linear layer, a second linear layer, a soft threshold learning module, a soft threshold processing unit, a residual connection layer, a summation unit and a fully connected layer. Among them, the first linear layer is used to perform preliminary feature extraction on the time series well logging data, realize preliminary encoding and feature conversion, and extract new feature representations that are more relevant to the formation conductivity from the original time series well logging data. The first linear layer is connected to the second linear layer and the soft threshold learning module respectively. The second linear layer is used to reduce the dimension of the first feature data, and the dimension reduction can reduce redundant information and make the feature distribution more compact. The soft threshold learning module is used to identify redundant components in the input data based on the multi-head attention mechanism and predict the denoising soft threshold. The soft threshold learning module and the second linear layer are both connected to the soft threshold processing unit. The soft threshold processing unit will soft-threshold the second feature data output by the second linear layer according to the denoising soft threshold output by the soft threshold learning module to complete the actual denoising and de-redundancy process. The soft threshold processing unit is connected to the summation unit, and the input of the entire model is also connected to the summation unit through the residual connection layer to form a residual network structure. The third feature data after soft thresholding and the time series logging data after the residual connection layer are input to the summation unit for summation. The summation unit is connected to the fully connected layer. The fully connected layer fuses and reduces the dimension of the summation result in the feature quantity dimension, and finally fuses the multi-channel data in the feature quantity dimension into a single-channel denoised formation resistivity.
[0065] The training of this model is similar to the training methods of traditional deep residual shrinkage networks and multi-head attention mechanisms. Collect multiple different time-series logging data as training sets, and use the corresponding standard formation resistivity as annotations. Since the goal is to predict accurate denoised formation resistivity, regression loss functions can be used, such as mean square error (MSE) and smooth L1 loss. The denoised resistivity predicted by the model is compared with the annotated standard formation resistivity to calculate the loss value. After initializing all weights in the model, a small batch stochastic gradient descent optimization algorithm is used to forward calculate the denoised resistivity output by the model, calculate the loss with the annotated value, and back propagate according to the loss value to update the parameters of each layer. Strategies such as learning rate decay and momentum can also be used to accelerate convergence. The soft threshold learning module is a submodule that needs to be trained end-to-end with the entire network. Finally, by adjusting the model weights, optimizer parameters, number of layers, etc., a model with performance indicators that meet the requirements is obtained.
[0066] In one embodiment, see Figure 3 , input the first feature data into the soft threshold learning module to obtain the denoising soft threshold, including steps S302 to S308. In addition, please refer to Figure 4 ,The soft threshold learning module includes layers for absolute value and global mean pooling, multi-head attention mechanism units, batch normalization layers, ReLu layers, the second fully connected layer and the Sigmoid layer.
[0067] S302: Absolutely value the first feature data and globally average pool the first feature data to obtain non-time series feature data.
[0068] It can be understood that the operation of taking the absolute value of the data is to cancel the positive and negative signs of the value. Global mean pooling performs an average pooling operation on the entire feature map, compressing a certain dimension of the entire feature into a single-channel scalar value. Specifically in this embodiment, global mean pooling is to reduce the multi-channel data of the time series length dimension of the first feature data into single-channel data, so that it changes from the original time series data to non-time series data, extracts the global features of the first feature data, and provides a feature basis for the subsequent multi-head attention mechanism unit to compare multi-dimensional redundant components. Assume that the dimension of the first feature data is (B, S, F), and the dimension of the non-time series feature data output by the absolute value and global mean pooling layer is (B, 1, F). Specifically, the dimension of the first feature data can be (32, 40, 42), and the dimension of the denoising soft threshold can be (32, 1, 42).
[0069] S304, input the non-time series feature data into the third linear layer, reduce the feature quantity dimension to 1, and obtain fourth feature data.
[0070] It can be understood that the third linear layer further compresses the non-time series feature data so that it has only one unit value in both the number of features and the length of the time series. This is equivalent to concentrating the entire feature on a scalar. Through this extreme dimensionality reduction operation, the model can be forced to learn the statistical distribution characteristics of the input data as a whole. Assuming that the dimension of the non-time series feature data output by the absolute value and global mean pooling layers is (B, 1, F), the dimension of the fourth feature data output by the third linear layer is (B, 1, 1). Specifically, the dimension of the first feature data can be (32, 1, 42), and the dimension of the denoising soft threshold can be (32, 1, 1).
[0071] S306, input the non-time series feature data into the multi-head attention mechanism unit to obtain the multi-head attention splicing weight.
[0072] It can be understood that the multi-head attention mechanism unit can divide attention into multiple subspaces, and perform attention calculation in each subspace, so that the correlation between inputs and inputs can be captured from different representation subspaces. For each subspace, the attention weight reflects the correlation between the current position feature and all other position features. By splicing these subspace weights, an attention representation that fully describes the entire input distribution can be obtained. Specifically, the multi-head attention mechanism unit includes multiple linear layers, and the non-time series feature data will pass through the linear layers corresponding to Query, Key and Value respectively, and the non-time series feature data will be converted into three calculation parameters: Query, Key and Value. These three calculation parameters are respectively input into the corresponding linear layers of multiple attention heads, and the weights in the linear layers corresponding to each attention head are different, so as to obtain multiple groups of multi-head attention features corresponding to each attention head but with different outputs. After each group of multi-head attention features is processed by scaling dot product and fusion (concat), the multi-head attention splicing weight can be obtained. Since there are three calculation parameters, the feature number dimension of the multi-head attention splicing weight obtained after fusion processing is three times that of the input. The dimension of the non-time series feature data is (B, 1, F), and the dimension of the multi-head attention splicing weight is (B, 1, 3F). Specifically, the dimension of the first feature data can be (32, 1, 42), and the dimension of the denoising soft threshold can be (32, 1, 126).
[0073] S308, the multi-head attention splicing weights are processed in sequence by a batch normalization layer, a ReLu layer, a second fully connected layer, and a Sigmoid layer to obtain fifth feature data, and the fifth feature data is multiplied by the fourth feature data to obtain a denoising soft threshold.
[0074] It can be understood that the batch normalization layer is used to normalize batch data, which can speed up the convergence of the model and alleviate the gradient diffusion problem in the deep network to a certain extent. The ReLu layer is a layer that uses the ReLu activation function for nonlinear mapping. The expression of the ReLu activation function is max(0, I), and I represents the input of the ReLu layer. The ReLu layer can better produce feature selection. The second fully connected layer synthesizes the features selected and activated by the ReLu layer, and merges the multi-channel features in the feature quantity dimension into one channel. Finally, after the Sigmoid activation function layer, the mapping of f(x)=1 / (1+e^-x) is realized to obtain the fifth feature data. The fifth feature data is multiplied by the fourth feature data to obtain the denoising soft threshold. The denoising soft threshold is not only a positive number, but also not too large, that is, it will not make all the features after soft thresholding zero.
[0075] In one embodiment, the residual connection layer is a fourth linear layer. Since the dimensions of the third feature data and the well logging time series feature data are different after the dimensionality reduction process, the identity mapping cannot be used directly. In order to construct the residual network, this embodiment uses the fourth linear layer to reduce the dimension of the time series well logging data so that its dimension remains the same as the third feature data, and can be directly summed in the summation unit.
[0076] In one embodiment, before inputting the time series logging data into the first linear layer for feature extraction to obtain the first feature data, the method further includes: preprocessing the time series logging data.
[0077] It is understandable that in the actual logging process, the time series logging data may have inconsistent formats, missing data, abnormal data, and other situations that do not meet the model input conditions. Therefore, before the time series logging data is input into the first linear layer, it can be preprocessed so that the time series logging data meets the input requirements. Specifically, the preprocessing of the time series logging data includes any one or more of the following: correcting abnormal values in the time series logging data. Filling missing values in the time series logging data.
[0078] In a specific embodiment, since the time series logging data is the logging data collected continuously at a certain time interval, the data at adjacent moments have a strong correlation. Therefore, the missing values can be filled by using interpolation methods, such as linear interpolation, spline interpolation, etc., to estimate the missing values according to the values of adjacent data points. In addition, when interpolating, it is possible to consider weighting the values of adjacent data points, and the adjacent data points closer to the missing data will be given a greater weight. It is also possible to use the statistical characteristics of the logging curve, such as the mean, median, etc., to replace the missing values with statistical values. In addition, a regression model can be established based on the values of other related logging curves to estimate the missing values. Finally, if the amount of data is sufficient, machine learning methods, such as K nearest neighbors, decision trees, etc., can also be used to predict missing values based on known data.
[0079] In a specific embodiment, outliers can be removed based on empirical rules, setting a threshold interval, and values outside the interval are considered outliers. The three-sigma principle can also be used to consider outliers that exceed 3 times the standard deviation of the mean as outliers. In addition, based on box plot analysis, values that deviate too much from the upper and lower quartiles can be considered outliers.
[0080] This application provides a well logging data noise reduction device, please refer to Figure 5 , including a data acquisition module 510, a first processing module 520, a second processing module 530, a third processing module 540, a fourth processing module 550 and a fifth processing module 560.
[0081] The data acquisition module 510 is used to acquire time series logging data. The time series logging data includes formation resistivity and multiple other logging data. The dimensions of the time series logging data include time series length and feature quantity.
[0082] The first processing module 520 is used to input the time series logging data into the first linear layer for feature extraction to obtain first feature data.
[0083] The second processing module 530 is used to input the first feature data into the second linear layer, reduce the dimension of the feature quantity, and obtain second feature data.
[0084] The third processing module 540 is used to input the first feature data into the soft threshold learning module to obtain the denoising soft threshold. The soft threshold learning module identifies redundant components based on the multi-head attention mechanism and predicts the denoising soft threshold.
[0085] The fourth processing module 550 is used to process the second feature data according to the denoising soft threshold to obtain the third feature data.
[0086] The fifth processing module 560 is used to sum the time series logging data with the third feature data after passing through the residual connection layer, and perform dimensionality reduction synthesis on the summation result to obtain the denoised formation resistivity with a feature number of 1.
[0087] In one embodiment, the third processing module 540 is used to absolute value and globally mean pool the first feature data to obtain non-time series feature data; input the non-time series feature data into the third linear layer, reduce the feature quantity dimension to 1, and obtain fourth feature data; input the non-time series feature data into the multi-head attention mechanism unit to obtain the multi-head attention splicing weight; process the multi-head attention splicing weight in turn through the batch normalization layer, ReLu layer, second fully connected layer and Sigmoid layer to obtain the fifth feature data, and multiply the fifth feature data by the fourth feature data to obtain the denoising soft threshold.
[0088] In one embodiment, the third processing module 540 is used to convert non-time series feature data into corresponding calculation parameters through three linear layers respectively; input each calculation parameter into the linear layer corresponding to each attention head respectively to obtain the multi-head attention feature corresponding to each attention head; and obtain the multi-head attention splicing weight after scaling dot product processing and fusion processing of each multi-head attention feature.
[0089] In one embodiment, other logging data include sonic logging data, natural gamma logging data, induction logging data, natural potential logging data, and through-casing electromagnetic logging data obtained by multiple different current meter tests.
[0090] In one embodiment, the residual connection layer is a fourth linear layer.
[0091] In one embodiment, the well logging data noise reduction module further includes a preprocessing module. The preprocessing module is used to preprocess the time series well logging data.
[0092] In one embodiment, the preprocessing module is used to perform preprocessing in any one or more of the following ways: correcting outliers in the time series logging data; filling missing values in the time series logging data.
[0093] For the specific definition of the logging data denoising module, please refer to the definition of the logging data denoising method above, which will not be repeated here. Each module in the above-mentioned logging data denoising device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation.
[0094] The present application provides a computer device, comprising one or more processors and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the following are executed: obtaining time series logging data; the time series logging data comprises formation resistivity and a plurality of other logging data, and the dimension of the time series logging data comprises the length of the time series and the number of features; inputting the time series logging data into a first linear layer for feature extraction to obtain first feature data; inputting the first feature data into a second linear layer to reduce the dimension of the feature number to obtain second feature data; inputting the first feature data into a soft threshold learning module to obtain a denoising soft threshold; the soft threshold learning module identifies redundant components based on a multi-head attention mechanism and predicts a denoising soft threshold; processing the second feature data according to the denoising soft threshold to obtain a third feature data; summing the time series logging data with the third feature data after passing through a residual connection layer, and performing dimensionality reduction synthesis on the sum result to obtain a denoised formation resistivity with a feature number of 1.
[0095] In one embodiment, when the computer-readable instructions are executed by one or more processors, the following are performed: absolute value and global mean pooling of the first feature data to obtain non-time series feature data; inputting the non-time series feature data into the third linear layer, reducing the feature quantity dimension to 1, and obtaining fourth feature data; inputting the non-time series feature data into the multi-head attention mechanism unit to obtain the multi-head attention splicing weight; processing the multi-head attention splicing weight in turn through the batch normalization layer, the ReLu layer, the second fully connected layer and the Sigmoid layer to obtain the fifth feature data, and multiplying the fifth feature data by the fourth feature data to obtain the denoising soft threshold.
[0096] In one embodiment, when the computer-readable instructions are executed by one or more processors, the following are performed: converting the non-time series feature data into corresponding calculation parameters through three linear layers respectively; inputting each calculation parameter into the linear layer corresponding to each attention head respectively to obtain the multi-head attention feature corresponding to each attention head; and obtaining the multi-head attention splicing weight after scaling dot product processing and fusion processing of each multi-head attention feature.
[0097] In one embodiment, other logging data include sonic logging data, natural gamma logging data, induction logging data, natural potential logging data, and through-casing electromagnetic logging data obtained by multiple different current meter tests.
[0098] In one embodiment, the residual connection layer is a fourth linear layer.
[0099] In one embodiment, when the computer readable instructions are executed by one or more processors, the computer readable instructions perform: pre-processing the time series logging data.
[0100] In one embodiment, when the computer readable instructions are executed by one or more processors, the computer readable instructions perform: correcting outliers in the time series logging data; and / or filling missing values in the time series logging data.
[0101] Indicatively, if Figure 6 As shown, Figure 6 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. Figure 6 The computer device 600 includes a processing component 602, which further includes one or more processors, and a memory resource represented by a memory 601, for storing instructions executable by the processing component 602, such as an application. The application stored in the memory 601 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 602 is configured to execute instructions to perform the steps of the method for reducing noise in well logging data in any of the above embodiments.
[0102] The computer device 600 may further include a power supply component 603 configured to perform power management of the computer device 600 , a wired or wireless network interface 604 configured to connect the computer device 600 to a network, and an input / output (I / O) interface 605 .
[0103] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0104] The present application provides a storage medium, in which computer-readable instructions are stored. When the computer-readable instructions are executed by one or more processors, the one or more processors execute: obtaining time series logging data; the time series logging data includes formation resistivity and multiple other logging data, and the dimensions of the time series logging data include time series length and feature quantity; inputting the time series logging data into a first linear layer for feature extraction to obtain first feature data; inputting the first feature data into a second linear layer, reducing the dimension of the feature quantity to obtain second feature data; inputting the first feature data into a soft threshold learning module to obtain a denoising soft threshold; the soft threshold learning module identifies redundant components based on a multi-head attention mechanism and predicts a denoising soft threshold; processing the second feature data according to the denoising soft threshold soft threshold to obtain a third feature data; summing the time series logging data with the third feature data after passing through a residual connection layer, and performing dimensionality reduction synthesis on the sum result to obtain a denoised formation resistivity with a feature quantity of 1.
[0105] In one embodiment, when the computer-readable instructions are executed by one or more processors, the one or more processors execute: absolute value and global mean pooling of the first feature data to obtain non-time series feature data; inputting the non-time series feature data into the third linear layer, reducing the feature quantity dimension to 1, and obtaining fourth feature data; inputting the non-time series feature data into the multi-head attention mechanism unit to obtain the multi-head attention splicing weight; processing the multi-head attention splicing weight in turn through the batch normalization layer, the ReLu layer, the second fully connected layer and the Sigmoid layer to obtain the fifth feature data, and multiplying the fifth feature data by the fourth feature data to obtain the denoising soft threshold.
[0106] In one embodiment, when the computer-readable instructions are executed by one or more processors, the one or more processors execute: converting the non-time series feature data into corresponding calculation parameters through three linear layers respectively; inputting each calculation parameter into the linear layer corresponding to each attention head respectively to obtain the multi-head attention feature corresponding to each attention head; and obtaining the multi-head attention splicing weight after scaling dot product processing and fusion processing of each multi-head attention feature.
[0107] In one embodiment, other logging data include sonic logging data, natural gamma logging data, induction logging data, natural potential logging data, and through-casing electromagnetic logging data obtained by multiple different current meter tests.
[0108] In one embodiment, the residual connection layer is a fourth linear layer.
[0109] In one embodiment, when the computer readable instructions are executed by one or more processors, the one or more processors are caused to perform: pre-processing the time series logging data.
[0110] In one embodiment, when the computer readable instructions are executed by one or more processors, the one or more processors are caused to perform: correcting outliers in the time series logging data; and / or filling missing values in the time series logging data.
[0111] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.
[0112] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for reducing noise in well logging data, characterized in that: include: Acquire time series logging data; the time series logging data includes formation resistivity and multiple other logging data, the dimensions of the time series logging data include time series length and feature quantity, and the feature quantity dimension represents the number of types of measurement data in the time series logging data; Preprocessing the time series logging data; the preprocessing of the time series logging data includes any one or more of the following: correcting abnormal values in the time series logging data; filling missing values in the time series logging data; Inputting the time series logging data into a first linear layer for feature extraction to obtain first feature data, wherein the first feature data has the same dimension as the time series logging data and has a higher degree of correlation with formation conductivity; Input the first feature data into a second linear layer, reduce the dimension of the feature quantity, and obtain second feature data; the second feature data has the same time series length dimension as the first feature data; Input the first feature data into a soft threshold learning module to obtain a denoising soft threshold; the inputting the first feature data into the soft threshold learning module to obtain the denoising soft threshold includes: absolute value and global mean pooling of the first feature data to reduce the time series length dimension of the first feature data into single-channel data to obtain non-time series feature data; the feature quantity dimension of the non-time series feature data is the same as the first feature data; input the non-time series feature data into a third linear layer to reduce the feature quantity dimension to 1 to obtain fourth feature data; input the non-time series feature data into a multi-head attention mechanism unit to obtain a multi-head attention splicing weight; the multi-head attention splicing weight is processed by a batch normalization layer, a ReLu layer, a second fully connected layer and a Sigmoid layer in sequence to obtain a fifth feature data, and the fifth feature data is multiplied by the fourth feature data to obtain the denoising soft threshold; The step of inputting the non-time series feature data into a multi-head attention mechanism unit to obtain a multi-head attention splicing weight comprises: converting the non-time series feature data into corresponding calculation parameters through three linear layers respectively; inputting each of the calculation parameters into a linear layer corresponding to each attention head respectively to obtain a multi-head attention feature corresponding to each attention head; and performing scaling dot product processing and fusion processing on each of the multi-head attention features to obtain the multi-head attention splicing weight. Processing the second feature data according to the denoising soft threshold to obtain third feature data; the second feature data has the same dimension as the third feature data; The time series logging data is passed through a residual connection layer and then summed with the third feature data, and the summation result is subjected to dimensionality reduction synthesis only in the feature quantity dimension to obtain a denoised formation resistivity with the feature quantity being 1; the residual connection layer adjusts the dimension of the time series logging data to be the same as that of the third feature data, and the denoised formation resistivity is the same as the time series length dimension of the time series logging data; wherein the residual connection layer is a fourth linear layer.
2. The method for reducing noise in well logging data according to claim 1, characterized in that: The other logging data include sonic logging data, natural gamma logging data, induction logging data, natural potential logging data and through-casing electromagnetic logging data obtained by testing with multiple different current meters.
3. A logging data noise reduction device, characterized in that: include: A data acquisition module, used to acquire time series logging data; the time series logging data includes formation resistivity and a plurality of other logging data, the dimensions of the time series logging data include time series length and feature quantity, and the feature quantity dimension represents the number of types of measurement data in the time series logging data; A preprocessing module is used to preprocess the time series logging data; the preprocessing of the time series logging data includes any one or more of the following: correcting abnormal values in the time series logging data; filling missing values in the time series logging data; A first processing module, configured to input the time series logging data into a first linear layer for feature extraction to obtain first feature data, wherein the first feature data has the same dimension as the time series logging data and has a higher correlation with formation conductivity; A second processing module is used to input the first feature data into a second linear layer, reduce the dimension of the feature quantity, and obtain second feature data; the second feature data has the same time series length dimension as the first feature data; A third processing module, configured to input the first feature data into a soft threshold learning module to obtain a denoising soft threshold; The soft threshold learning module identifies redundant components based on a multi-head attention mechanism, and predicts the denoising soft threshold according to the redundant components; The step of inputting the first feature data into a soft threshold learning module to obtain a denoising soft threshold comprises: performing absolute value and global mean pooling on the first feature data to reduce the time series length dimension of the first feature data into single-channel data to obtain non-time series feature data; the feature quantity dimension of the non-time series feature data is the same as that of the first feature data; inputting the non-time series feature data into a third linear layer to reduce the feature quantity dimension to 1 to obtain fourth feature data; inputting the non-time series feature data into a multi-head attention mechanism unit to obtain a multi-head attention splicing weight; and sequentially passing the multi-head attention splicing weight through a batch normalization layer, ReLu layer, second fully connected layer and Sigmoid layer are processed to obtain fifth feature data, and the fifth feature data is multiplied by the fourth feature data to obtain the denoising soft threshold; wherein, the non-time series feature data is input into the multi-head attention mechanism unit to obtain the multi-head attention splicing weight, including: converting the non-time series feature data into corresponding calculation parameters through three linear layers respectively; inputting each of the calculation parameters into the linear layer corresponding to each attention head respectively to obtain the multi-head attention feature corresponding to each attention head; after scaling dot product processing and fusion processing of each of the multi-head attention features, the multi-head attention splicing weight is obtained; A fourth processing module, configured to process the second feature data according to the denoising soft threshold to obtain third feature data; the second feature data has the same dimension as the third feature data; The fifth processing module is used to sum the time series logging data with the third feature data after passing through the residual connection layer, and perform dimensionality reduction synthesis on the summation result only in the feature quantity dimension to obtain the denoised formation resistivity with the feature quantity of 1; the residual connection layer adjusts the dimension of the time series logging data to be the same as the third feature data, and the denoised formation resistivity is the same as the time series length dimension of the time series logging data; wherein the residual connection layer is the fourth linear layer.
4. A computer device, characterized in that: The method comprises one or more processors and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the method for reducing noise in well logging data as described in any one of claims 1 or 2 are executed.
5. A storage medium, characterized in that: The storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the well logging data denoising method as described in any one of claims 1 or 2.
