Bilstm-based rapid inversion method and system for electromagnetic wave logging while drilling data

Abstract

Disclosed in the present invention are a BiLSTM-based rapid inversion method and system for electromagnetic wave logging while drilling data. By means of constructing a deep network structure consisting of four BiLSTM layers and one fully connected layer, efficient processing of logging data is achieved. In the present invention, ideal magnetic dipole responses are used as training data, thereby solving the problem that traditional data interpretation is liable to fall into local minima; recurrent dropout technology is used to reduce overfitting, thereby improving the generalization capability of models; the technology can not only accurately classify stratum models, but also predict various formation parameters, significantly improving the accuracy and processing speed of data interpretation, and providing strong technical support for geological exploration and formation evaluation.

Description

BiLSTM-based fast inversion method and system for electromagnetic wave logging data while drilling TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical exploration and information technology, in particular to a BiLSTM-based fast inversion method and system for electromagnetic wave logging data while drilling. BACKGROUND

[0002] In modern logging technology, electromagnetic wave logging while drilling is an important geologic exploration method. Electromagnetic wave logging data while drilling refers to the real-time acquisition of underground rock information and formation boundary detection data by electromagnetic wave instruments during drilling. These data are crucial for accurately evaluating formation properties, oil and gas reserves, and geological structure.

[0003] The principle is mainly to use the propagation and reflection characteristics of electromagnetic waves in the formation to detect formation information. The application scenarios mainly include the exploration and development of mineral resources such as oil and gas, providing important basis for geosteering and formation evaluation.

[0004] BiLSTM, i.e., bidirectional long short-term memory unit, is a recurrent neural network structure in deep learning models. It can handle long-term dependencies in sequence data by simultaneously capturing context information in sequences through forward and backward LSTM units, and is widely used in natural language processing, speech recognition, time series analysis, etc. In logging data inversion, BiLSTM can learn complex patterns in logging data and achieve efficient data interpretation.

[0005] Currently, the traditional deterministic or statistical methods are often used for data interpretation by electromagnetic wave forward-looking detection instruments while drilling. Deterministic methods such as Gauss-Newton method, etc., approximate the real formation model through iterative optimization, but are prone to local minimum, resulting in inaccurate interpretation results. Statistical methods can avoid local minimum to some extent, but have large computational complexity, making it difficult to meet the needs of real-time geosteering.

[0006] For example, in complex geological environments, the local minimum problem is particularly prominent. For example, in the exploration process of an oilfield, the interpretation results of the deterministic method are trapped in local minimum several times due to the multi-layer structure and anisotropy of the formation, resulting in incorrect drilling direction and causing waste of resources and time.

[0007] In order to avoid the problem of falling into local minimum and realize timely guidance of geosteering, various methods are often combined in the industry. For example, prior geological information is introduced to constrain the inversion process, or a more complex statistical model is used to improve the global search ability. However, these methods often increase the complexity of calculation, and even require high-performance computers and long-time operation, which is not practical in field operation.

[0008] For example, in a certain geological exploration, a complex statistical inversion method was used to obtain a more accurate stratigraphic model. Although satisfactory results were finally obtained, the inversion process took several days, far from meeting the real-time guidance drilling requirements.

[0009] Therefore, providing a while-drilling electromagnetic wave logging data inversion method that can both avoid falling into local minimum and meet real-time requirements has become a technical problem to be solved by the present application. SUMMARY

[0010] The technical problem solved by the present application is to provide a BiLSTM-based fast inversion method and system for while-drilling electromagnetic wave logging data to solve the problem that data interpretation is easy to fall into local minimum and the calculation amount is large, and cannot meet the real-time geosteering requirements in the background art.

[0011] To solve the above technical problems, the technical solutions adopted by the present application are as follows:

[0012] A BiLSTM-based fast inversion method for while-drilling electromagnetic wave logging data, comprising the following steps:

[0013] Step 1, obtaining the electromagnetic wave signal received by the antenna and pre-processing to obtain logging data;

[0014] Step 2, inputting the pre-processed logging data as input into the BiLSTM layer and neural network, wherein the BiLSTM layer contains a correction linear unit ReLU as an activation function, and recurrent dropout technology is used in the BiLSTM layer to reduce overfitting phenomenon;

[0015] Step 3, the neural network performs feature extraction, convolution and fusion on the input logging data, and outputs classification results and prediction results, wherein the classification results are any one of the pre-set models, and the prediction results are a plurality of stratigraphic parameters corresponding to the model categories;

[0016] Step 4, according to the classification results and prediction results, geosteering and stratigraphic evaluation are performed;

[0017] As a further scheme of the present application, in step 1, before obtaining the electromagnetic wave signals received by the antenna, a propagation matrix method is first used as a forward algorithm to calculate the response of a magnetic dipole in a one-dimensional layered medium, in which the influence of the wellbore environment is ignored, and the response of an ideal magnetic dipole in a layered medium is used to simulate the instrument response to generate a training data set and a verification data set, the training data set will be used for subsequent training of the neural network weights, and the verification data set is used to verify the training results during the training process, after the preparation of the training data set and the verification data set is completed, the electromagnetic wave signals received by the antenna are obtained and preprocessed to obtain the logging data used for neural network input.

[0018] As a further scheme of the present application, the neural network of step 3 uses a recurrent neural network, the input information of the hidden layer neurons of which is not only derived from the output of the input layer neurons, but also derived from the output of the hidden layer neurons at the current time; and each layer of the recurrent neural network shares parameters U, V and W to reduce the number of parameters and improve the network training speed.

[0019] As a further scheme of the present application, the output value st of the hidden layer neurons of the recurrent neural network at time t is calculated according to the input information It at the current time and the output value st-1 of the hidden layer neurons at the previous time through an activation function and shared parameters U, V and W; wherein the input information It includes the current logging data input and possible feedback of the output layer neurons at the previous time; the output ot of the output layer neurons at time t is calculated according to the output value st of the hidden layer neurons at time t through a connection weight and an activation function.

[0020] As a further scheme of the present application, the electromagnetic wave signals received by the antenna in step 1 are specifically the electromagnetic wave response signals from the formation received by the antenna on the logging instrument during the process of logging while drilling, which include but are not limited to formation resistivity, boundary position, anisotropy coefficient and formation dip, and these signals are preprocessed to be converted into logging data that can be used for neural network input.

[0021] As a further scheme of the present application, the neural network in step 2 uses an RMSprop optimizer to train the logging data with a loss function defined as mean square error.

[0022] As a further scheme of the present application, the output classification result in step 3 is any one of four preset formation models, and the four formation models correspond to different formation characteristics respectively; the prediction result is parameters of a plurality of specific formation characteristics corresponding to the classified formation model, wherein the parameters of formation characteristics include but are not limited to formation thickness, resistivity, porosity, permeability and water saturation, which are obtained through the feature extraction, convolution and fusion process of the neural network to provide detailed data support for geosteering and formation evaluation.

[0023] A drilling-while-circulating electromagnetic wavefront-viewing logging data inversion system based on a recurrent neural network, comprising:

[0024] A data preprocessing module is configured to acquire electromagnetic wave signals received by three antennas and perform preprocessing to obtain logging data.

[0025] A neural network module is configured to receive the preprocessed logging data as input, perform feature extraction, convolution and fusion through a neural network composed of four bidirectional long short-term memory units (BiLSTM) layers and one fully connected layer, and output a classification result and a prediction result.

[0026] A result output module is configured to perform geosteering and formation evaluation according to the output of the neural network.

[0027] As a further scheme of the present application, the neural network module uses an RMSprop optimizer to train the logging data with a loss function defined as mean square error.

[0028] As a further scheme of the present application, the formation parameters include horizontal resistivity of each layer, distance to the boundary, anisotropy coefficient of the layer where the instrument is located, and formation dip angle.

[0029] Compared with the prior art, the present application has the beneficial effects that: by constructing a deep network structure composed of four BiLSTM layers and one fully connected layer, efficient processing of logging data is achieved. The present application uses ideal magnetic dipole response as training data, which solves the problem of easy falling into local minimum in traditional data interpretation. The recurrent dropout technology is used to reduce overfitting and improve the generalization ability of the model. This technology not only accurately classifies the formation model, but also predicts a plurality of formation parameters, significantly improving the accuracy and processing speed of data interpretation, and providing strong technical support for geological exploration and formation evaluation.

[0030] Additional aspects and advantages of the present application will be given in part in the following description, will become apparent in part from the following description, or will be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and all other embodiments obtained by those of ordinary skill in the art without creative effort based on these accompanying drawings also belong to the protection scope of the present application.

[0032] Fig. 1 is a schematic diagram of the principle of a recurrent neural network.

[0033] Fig. 2 is a schematic diagram of the principle of an LSTM.

[0034] Fig. 3 is a neural network structure for inversion of data of a while-drilling azimuthal electromagnetic wave logging.

[0035] Fig. 4 is a schematic diagram of instrument detection.

[0036] Fig. 5 is a diagram for setting a formation model.

[0037] Fig. 6 is a flowchart of deep learning inversion. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort also belong to the protection scope of the present application.

[0039] Please refer to Figs. 1-6. In the embodiments of the present application, a while-drilling electromagnetic wave logging data fast inversion method based on BiLSTM has the following steps.

[0040] Step 1: Obtain electromagnetic wave signals received by, for example, three antennas, and perform preprocessing to obtain logging data.

[0041] In step 1, the propagation matrix method is used as a forward algorithm for calculating the response of a magnetic dipole in a one-dimensional layered medium, the influence of the wellbore environment is ignored, the response of an ideal magnetic dipole in a layered medium is used to simulate the instrument response, and the training data set and the verification data set are generated, wherein the training data set is used to train the weight of the neural network, and the verification data set is used to verify the training result in the training process; specifically, in step 1, before obtaining the electromagnetic wave signal received by the antenna, the propagation matrix method is used as a forward algorithm to calculate the response of a magnetic dipole in a one-dimensional layered medium, for the purpose of simplifying the model and improving the calculation efficiency, the influence of the wellbore environment is ignored, and the response of an ideal magnetic dipole in a layered medium is used to simulate the instrument response. In this way, the training data set and the verification data set are generated. The training data set will be used for subsequent training of the neural network weight, and the verification data set is used to verify the training result in the training process. After the preparation of the data set is completed, the electromagnetic wave signal received by the antenna is obtained and preprocessed to obtain the logging data used as the input of the neural network.

[0042] In step 2, the preprocessed logging data is input into a neural network composed of four bidirectional long short-term memory units BiLSTM layers and a fully connected layer, wherein each BiLSTM layer includes a rectified linear unit ReLU as an activation function, and a recurrent dropout technique is used in the BiLSTM layer to reduce the overfitting phenomenon.

[0043] In step 3, the neural network performs feature extraction, convolution and fusion on the input logging data, and outputs a classification result and a prediction result, wherein the classification result is any one of the four set models, and the prediction result is a plurality of formation parameters corresponding to the model category.

[0044] In step 4, according to the classification result and the prediction result, the geosteering and formation evaluation are performed.

[0045] In step 3, the neural network uses a recurrent neural network, the input information of the hidden layer neurons comes not only from the output of the input layer neurons, but also from the output of the current time hidden layer neurons; and each layer of the recurrent neural network shares parameters U, V and W to reduce the parameter amount and improve the network training speed.

[0046] Wherein, "t time" refers to a specific time point when the recurrent neural network processes sequence data. In the recurrent neural network, data is processed in time sequence, and each time point corresponds to an input data and a state update. Therefore, "t time" is a specific time point in this sequence.

[0047] "Output value st" represents the output state of the recurrent neural network hidden layer neuron at time t. This state is calculated based on the current input and the state at the previous time step, and it contains all relevant information in the sequence up to the current time point.

[0048] "Output value st-1" represents the output state of the recurrent neural network hidden layer neuron at time t-1, i.e., the previous time point. This state is passed to the next time point (t) as part of the calculation of the new state.

[0049] "U" is the connection weight matrix from the input layer neurons to the hidden layer neurons. In recurrent neural networks, this matrix is used to transform and pass the input layer data to the hidden layer, and is an important parameter for the neural network to learn the features of the input data.

[0050] "V" is the connection weight matrix between the hidden layer neurons and the output layer. It is responsible for transforming the state of the hidden layer into the output of the output layer. During training, this matrix is adjusted to minimize the prediction error of the network.

[0051] "W" is the connection weight matrix from the hidden layer neurons at the previous time step to the hidden layer neurons at the current time step. In recurrent neural networks, this matrix is used to save and pass the timing information, so that the network can remember the previous state and affect the current and future output. This is also the key to the ability of recurrent neural networks to process sequence data.

[0052] The output value st of the hidden layer neuron of the recurrent neural network at time t is calculated based on the input information It at the current time step and the output value st-1 of the hidden layer neuron at the previous time step, through the activation function and shared parameters U, V, W; where the input information It includes the current time step logging data input and possibly the feedback of the output layer neuron at the previous time step; the output ot of the output layer neuron at time t is calculated based on the output value st of the hidden layer neuron at time t through the connection weight and the activation function.

[0053] The electromagnetic wave signal received by the antenna in step 1 refers to the electromagnetic wave response signal from the formation received by the antenna on the logging instrument during the process of logging while drilling. These signals include but are not limited to formation resistivity, boundary position, anisotropy coefficient and formation dip. These signals are preprocessed and converted into logging data that can be used as input for the neural network.

[0054] In step 2, the neural network uses the RMSprop optimizer to train the logging data with the loss function defined as mean square error.

[0055] The output classification result in step 3 is any one of four preset stratigraphic models, which correspond to different stratigraphic characteristics; the prediction result is the parameters of various specific stratigraphic characteristics corresponding to the classified stratigraphic model.

[0056] The parameters of formation characteristics include, but are not limited to, formation thickness, resistivity, porosity, permeability, and water saturation. These parameters are obtained through feature extraction, convolution, and fusion processes of neural networks to provide detailed data support for geological guidance and formation evaluation.

[0057] A drilling electromagnetic wave forward-looking logging data inversion system based on a recurrent neural network includes:

[0058] The data preprocessing module is used to acquire the electromagnetic wave signals received by the three antennas and perform preprocessing to obtain well logging data.

[0059] The neural network module receives preprocessed logging data as input, performs feature extraction, convolution, and fusion through a neural network consisting of four bidirectional long short-term memory (BiLSTM) layers and a fully connected layer, and outputs classification and prediction results.

[0060] The results output module is used to perform geological guidance and stratigraphic evaluation based on the output of the neural network.

[0061] The neural network module uses the RMSprop optimizer to train logging data with a loss function defined as mean square error. Formation parameters include the horizontal resistivity of each layer, distance to the boundary, anisotropy coefficient of the formation where the instrument is located, and formation dip angle.

[0062] This invention achieves efficient processing of well logging data by constructing a deep network structure consisting of four BiLSTM layers and one fully connected layer. It uses ideal magnetic dipole responses as training data, solving the problem of getting trapped in local minima in traditional data interpretation. Recurrent dropout technology is used to reduce overfitting and improve the model's generalization ability. This technique can not only accurately classify formation models but also predict various formation parameters, significantly improving the accuracy and processing speed of data interpretation, and providing strong technical support for geological exploration and formation evaluation.

[0063] In this model, the ideal magnetic dipole refers to an object or system with a magnetic moment that produces a specific response under the influence of an external magnetic field. In this model, the specific shape and size of the magnetic dipole are usually ignored, focusing only on the magnitude and direction of its magnetic moment. This simplifies the problem and makes the mathematical model easier to handle.

[0064] The generation of response data is achieved by calculating the response of an ideal magnetic dipole under different external magnetic field conditions, which can generate a set of response data. These data describe the behavior of the magnetic dipole in different magnetic field environments. The response data may include the magnetic field strength, direction, and other physical quantities generated by the magnetic dipole, which are closely related to the characteristics of the external magnetic field.

[0065] In addition, the application of the ideal magnetic dipole model as training data, such as electromagnetic wave logging or magnetic target detection and positioning, requires fast and accurate processing of data related to the magnetic dipole. To achieve this, a machine learning-based method such as a neural network can be used. When training these machine learning models, "ideal magnetic dipole response" is used as training data. Through these data, the model can learn the response rules of the magnetic dipole under different magnetic field conditions. After training, the model can quickly and accurately predict the state or related parameters of the magnetic dipole according to the actual measurement data, thereby achieving fast detection and positioning of magnetic targets.

[0066] Embodiment 1:

[0067] This embodiment demonstrates the application of the BiLSTM-based fast inversion method for electromagnetic wave logging while drilling in actual oil exploration.

[0068] During the exploration of a certain oil field, due to the complex structure of the formation, the existence of multi-layer structure and anisotropy influence, the traditional deterministic method such as Gauss-Newton method often falls into local minimum value during data interpretation, resulting in drilling direction judgment error, causing waste of resources and time. In order to solve this problem, the BiLSTM-based fast inversion method for electromagnetic wave logging while drilling is used.

[0069] First, the electromagnetic wave response signals from the formation are received by the three antennas on the logging while drilling instrument, which contain information such as formation resistivity, boundary position, anisotropy coefficient, and formation dip angle. After preprocessing, these signals are converted into logging data that can be used as input for the neural network.

[0070] Next, the preprocessed logging data is input into a neural network composed of four bidirectional long short-term memory units (BiLSTMs) and a fully connected layer. Each BiLSTM layer includes a rectified linear unit (ReLU) as the activation function, and recurrent dropout techniques are used in the BiLSTM layers to reduce overfitting. The neural network outputs classification results and prediction results through feature extraction, convolution, and fusion. The classification results are any of the four set models, which correspond to different formation characteristics. The prediction results are parameters of various specific formation characteristics corresponding to the classified formation model, including formation thickness, resistivity, porosity, permeability, and water saturation.

[0071] The RMSprop optimizer is used to train the logging data with a loss function defined as the mean square error to improve the prediction accuracy of the neural network. Through efficient interpretation of the neural network, the problem of falling into a local minimum is successfully avoided, and due to the fast processing capability of the neural network, this method meets the needs of real-time geosteering.

[0072] During exploration, real-time geosteering and formation evaluation data are used to guide the selection of drilling direction. Compared with traditional deterministic methods, the method proposed in the present invention significantly improves the success rate of drilling and reduces the waste of resources and time.

[0073] In summary, the present embodiment demonstrates the successful application of the BiLSTM-based fast inversion method for electromagnetic logging data while drilling in oil exploration. This method not only avoids falling into a local minimum, but also meets the real-time requirements, providing accurate and efficient data support for geosteering and formation evaluation.

[0074] Example 2:

[0075] In this embodiment, we will elaborate on how the BiLSTM-based fast inversion method for electromagnetic logging data while drilling combines the recurrent neural network structure shown in FIG. 1 to achieve efficient and accurate geosteering and formation evaluation.

[0076] First, we clarify the meanings of the symbols in FIG. 1: x represents the input of the hidden layer neurons, o represents the output of the network, s represents the output of the hidden layer neurons, and U, V, and W represent different connection weights. These connection weights play a crucial role in the recurrent neural network, as they are responsible for converting input information into hidden layer states and then converting hidden layer states into final outputs.

[0077] In the process of inverting electromagnetic logging data while drilling, we first preprocess the electromagnetic signals received by the three antennas, converting them into an input format acceptable to the neural network. This preprocessed logging data is then input into a neural network consisting of four BiLSTM layers and one fully connected layer.

[0078] Each BiLSTM layer contains a ReLU activation function to enhance the network's non-linear expressive power. Additionally, we employ recurrent dropout in the BiLSTM layers to reduce overfitting and improve the model's generalization ability.

[0079] The hidden layers of the neural network employ a recurrent neural network structure. A key feature of this structure is that the input information to the hidden layer neurons comes not only from the outputs of the input layer neurons but also from the outputs of the hidden layer neurons at the current time step. This structure enables the neural network to capture long-term dependencies in sequential data, thereby providing a better understanding of the inherent patterns in well logging data.

[0080] Furthermore, each layer of a recurrent neural network shares parameters U, V, and W, which significantly reduces the number of parameters and improves network training speed. Let st be the output value of a hidden layer neuron at time t, which is calculated based on the current input information It and the output value st-1 of the hidden layer neuron at the previous time step. This information transmission method allows the neural network to remember previous states and influence current and future outputs, thus better handling sequential data.

[0081] In this embodiment, we use a neural network to extract features, perform convolution, and fuse the input well logging data, ultimately outputting classification and prediction results. The classification result is any one of four preset models, and the prediction result is various formation parameters corresponding to the model category. These outputs provide accurate and efficient data support for geological steering and formation evaluation.

[0082] Compared to traditional deterministic methods, the method proposed in this invention combines the advantages of recurrent neural networks, enabling the learning of complex patterns in well logging data and achieving efficient data interpretation. Furthermore, by employing a BiLSTM structure, our method avoids getting trapped in local minima while meeting real-time requirements. This has significant practical implications for field operations, significantly improving drilling success rates and reducing resource and time waste.

[0083] In summary, this embodiment demonstrates the successful application of a BiLSTM-based rapid inversion method for electromagnetic logging-while-drilling data combined with a recurrent neural network structure in geological exploration. This method provides strong technical support for geological steering and formation evaluation through efficient data processing and accurate prediction results.

[0084] Example 3:

[0085] To address the problem that gradient vanishing and gradient exploding limit the network's ability to utilize long-term historical information and thus affect training performance when training recurrent neural networks, Long Short-Term Memory (LSTM) units are employed.

[0086] LSTM is an improvement on traditional recurrent neural networks. It carries and spans information across multiple time steps by adding an extra data stream (c_t), also known as the cell state. This design allows past information to continuously flow through the network and re-enter computation when needed, thus effectively solving the gradient vanishing problem, as shown in Figure 2.

[0087] To further improve the performance of LSTM, we adopted a bidirectional LSTM (BiLSTM) structure. Bidirectional LSTM consists of two standard LSTM networks that process the input sequence in forward and reverse order, respectively. By merging the representations in these two directions, bidirectional LSTM can capture complex patterns and contextual information that unidirectional LSTM might miss.

[0088] In our implementation, a bidirectional LSTM was applied to process time-series data, such as electromagnetic logging-while-drilling data. Through training, the network learned to extract meaningful features from the data and generate accurate predictions and classifications. These results have significant guiding implications for geological exploration and formation evaluation.

[0089] Example 4:

[0090] This invention employs a specific neural network structure to process azimuth electromagnetic logging data while drilling, enabling rapid data inversion. The following is a detailed description of the neural network structure used in this invention and its workflow, along with in-depth analysis of Figures 3, 4, 5, and 6.

[0091] First, the neural network structure used in this invention is shown in Figure 3. This is a deep network consisting of four BiLSTM layers and one fully connected layer. Each BiLSTM layer is equipped with a Rectified Linear Unit (ReLU) as an activation function to increase the network's non-linearity. The BiLSTM layer is characterized by its ability to process both forward and reverse sequence data simultaneously, thereby capturing information that unidirectional LSTMs might miss. Furthermore, to reduce overfitting, this invention implements recurrent dropout technology in the BiLSTM layers. This technique randomly discards some connections in the network, thereby improving the model's generalization ability.

[0092] Figure 3 illustrates the structure of this neural network in detail, with each layer clearly visible, including the input layer, four BiLSTM layers, a fully connected layer, and the output layer. Data enters the network from the input layer, is processed by the BiLSTM layers, then integrated by the fully connected layers, and finally the result is obtained from the output layer.

[0093] In terms of data preparation, this invention uses the response of an ideal magnetic dipole in a layered medium as the instrument response and employs the propagation matrix method to generate training and validation datasets. These datasets will be used for training and validating the neural network.

[0094] Figure 4 shows a schematic diagram of the instrument used in this invention. The instrument has a maximum detection depth of 30m; therefore, this invention considers the scenario where at most one interface is detected above or below the instrument when designing the model.

[0095] Based on this consideration, this invention establishes four different models, as shown in Figure 5. These models represent different stratigraphic structures and exploration conditions. The inversion targets for stratigraphic parameters include the horizontal resistivity of each layer, the distance to the boundary, the anisotropy coefficient of the stratum where the instrument is located, and the dip angle of the stratum. These parameters are of great significance for geological exploration and stratigraphic evaluation.

[0096] During training, this invention employs the RMSprop optimizer and defines the loss function as mean squared error. This setup helps the network learn and optimize model parameters more effectively. Simultaneously, to prevent overfitting and improve training efficiency, this invention also uses the EarlyStopping callback method and sets an appropriate tolerance.

[0097] Once the network is trained, it can be used to invert actual well logging data. The inversion process is shown in Figure 6. First, the preprocessed well logging data is input into the trained neural network; then, the network performs operations such as feature extraction, convolution, and fusion on this data; finally, the network outputs classification and prediction results. The classification result indicates which of the four defined models the data belongs to, while the prediction result includes various formation parameters corresponding to the model category.

[0098] In summary, this invention achieves rapid inversion of azimuth electromagnetic logging data during drilling by employing a specific neural network structure and training strategy. This method not only improves the accuracy of data interpretation but also significantly increases processing speed, providing strong technical support for geological exploration and formation evaluation. Furthermore, this invention offers flexibility and scalability, allowing for adjustments and optimizations based on actual needs.

[0099] In this application, unless otherwise clearly indicated and limited, the terms "mounting", "setting", "connecting", "fixing", "screwing" and the like should be interpreted broadly, for example, can be fixed connection, can also be detachable connection, or integrated; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be the internal communication of two elements or the interaction relationship of two elements, unless otherwise clearly limited, the above terms in this application can be understood according to the specific meaning of the above terms in this application for those skilled in the art.

[0100] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application.

Claims

1. A fast inversion method for drilling electromagnetic logging data based on BiLSTM, characterized in that, Includes the following steps: Step 1: Acquire the electromagnetic wave signal received by the antenna and perform preprocessing to obtain well logging data; Step 2: The preprocessed logging data is used as input to the BiLSTM layer and the neural network. The BiLSTM layer contains a ReLU calibrated linear unit as the activation function, and the recurrent dropout technique is used in the BiLSTM layer to reduce overfitting. Step 3: The neural network performs feature extraction, convolution and fusion on the input well logging data, and outputs classification results and prediction results. The classification result is any one of the pre-defined models, and the prediction result is a variety of formation parameters corresponding to the model category. Step 4: Based on the classification and prediction results, conduct geological guidance and stratigraphic evaluation.

2. The method for fast inversion of drilling electromagnetic logging data based on BiLSTM according to claim 1, characterized in that, In step 1, before acquiring the electromagnetic wave signal received by the antenna, the propagation matrix method is first used as the forward modeling algorithm to calculate the response of the magnetic dipole in the one-dimensional layered medium. During this process, the influence of the wellbore environment is ignored, and the response of the ideal magnetic dipole in the layered medium is used to simulate the instrument response to generate training and validation datasets. The training dataset will be used for the subsequent training of neural network weights, while the validation dataset is used to verify the training results during the training process. After the preparation of the training and validation datasets is completed, the electromagnetic wave signal received by the antenna is acquired and preprocessed to obtain the logging data used as input to the neural network.

3. The method for fast inversion of drilling electromagnetic logging data based on BiLSTM according to claim 1, characterized in that, The neural network in step 3 is a recurrent neural network, in which the input information of the hidden layer neurons comes not only from the output of the input layer neurons, but also from the output of the hidden layer neurons at the current moment; and each layer of the recurrent neural network shares parameters U, V and W to reduce the number of parameters and improve the network training speed.

4. The method for rapid inversion of drilling electromagnetic logging data based on BiLSTM according to claim 2, characterized in that, The output value st of the hidden layer neuron of the recurrent neural network at time t is calculated based on the input information It at the current time and the output value st-1 of the hidden layer neuron at the previous time, through the activation function and shared parameters U, V, and W. The input information It includes the logging data input at the current time and possible feedback from the output layer neuron at the previous time. The output ot of the output layer neuron at time t is calculated based on the output value st of the hidden layer neuron at time t through the connection weights and the activation function.

5. The method for fast inversion of drilling electromagnetic logging data based on BiLSTM according to claim 1, characterized in that, The electromagnetic wave signals received by the antenna in step 1 specifically refer to the electromagnetic wave response signals from the formation received by the antenna on the logging instrument during the logging-while-drilling process. These signals include, but are not limited to, formation resistivity, boundary location, anisotropy coefficient, and formation dip angle. These signals are preprocessed and converted into logging data that can be used as input to the neural network.

6. The fast inversion method for drilling electromagnetic logging data based on BiLSTM according to claim 1, characterized in that, In step 2, the neural network is trained using the RMSprop optimizer on well logging data with a loss function defined as mean square error.

7. The method for rapid inversion of drilling electromagnetic logging data based on BiLSTM according to claim 1, characterized in that, The output classification result mentioned in step 3 is any one of four preset stratigraphic models, which correspond to different stratigraphic characteristics. The prediction result is the parameters of various specific stratigraphic characteristics corresponding to the classified stratigraphic model. The parameters of the stratigraphic characteristics include, but are not limited to, stratigraphic thickness, resistivity, porosity, permeability and water saturation. These parameters are obtained through feature extraction, convolution and fusion processes of neural networks to provide detailed data support for geological guidance and stratigraphic evaluation.

8. A drilling electromagnetic wave forward-looking logging data inversion system based on a recurrent neural network, characterized in that, include: The data preprocessing module is used to acquire the electromagnetic wave signals received by the three antennas and perform preprocessing to obtain well logging data. The neural network module receives preprocessed logging data as input, performs feature extraction, convolution, and fusion through a neural network consisting of four bidirectional long short-term memory (BiLSTM) layers and a fully connected layer, and outputs classification and prediction results. The results output module is used to perform geological guidance and stratigraphic evaluation based on the output of the neural network.

9. A rapid inversion system for drilling electromagnetic logging data based on BiLSTM according to claim 8, characterized in that, The neural network module uses the RMSprop optimizer to train the logging data with a loss function defined as mean square error.

10. A rapid inversion system for drilling electromagnetic logging data based on BiLSTM according to claim 8, characterized in that, The formation parameters include the horizontal resistivity of each layer, the distance to the boundary, the anisotropy coefficient of the formation where the instrument is located, and the dip angle of the formation.

Images