Drilling tool jamming early warning method and system based on multiple attention mechanism recurrent neural network model

Abstract

The invention provides a drilling tool jamming early warning method and system based on a multi-attention mechanism recurrent neural network model. The method comprises the following steps: acquiring data values of a hook load sequence and a hook load association sequence; determining a hook load correlation sequence as a target input sequence, wherein the correlation between the hook load correlation sequence and the hook load sequence meets a set condition through a Pearson's correlation coefficient analysis method; establishing a recurrent neural network model with multiple attention modules by comprehensively considering the requirements of input information filtering, data space optimization, long-term dependency sequence processing and dynamic feature capture; and calculating a hook load predicted value by taking the hook load sequence and the hook load correlation sequence parameters as an input sequence of the model, and identifying a drilling tool jamming risk according to judgment rules corresponding to different working conditions and performing early warning. According to the scheme, the problems that in the prior art, interference influences are large, and calculation time and cost consumption are large can be effectively solved, key features are focused on the basis of the multiple attention architecture, time and space features are dynamically captured in a self-adaptive mode, an expert experience model is fused and applied, and drilling tool jamming early warning is efficiently and accurately achieved.

Description

Technical Field

[0001] This invention relates to the field of oil and gas development stuck drill prediction technology, and in particular to a stuck drill early warning method and system based on a recurrent neural network model with multiple attention mechanism. Background Technology

[0002] Stuck drill bit failure can damage drilling equipment, shorten its service life, increase the cost of maintenance and equipment replacement, and may also cause drilling operations to be interrupted, affecting the overall progress and prolonging the drilling cycle.

[0003] Traditional stuck pipe early warning methods mainly fall into two categories: one is prediction based on winch weight data; the other is based on parameter changes using expert experience models and rules to generate stuck pipe early warnings. Regarding the first method, by setting winch weight parameters, a PLC automatically records the weight at the winch's stopped state, and the weight increment is compared in real time with the set weight parameters. Early warnings are then generated based on the comparison results. However, due to the complex and variable conditions in actual drilling, setting only a weight parameter may lead to poor early warning accuracy under different formations and working conditions. Furthermore, this method only considers a single weight parameter without analyzing parameters related to stuck pipe, resulting in unobjective predictions that cannot reliably meet engineering requirements. Regarding the second method, expert experience models are more commonly used in engineering. However, due to the complex relationships and non-linear changes between various engineering parameters when stuck pipe occurs, this method has the following limitations: 1. Expert experience models typically rely on existing knowledge or experience to formulate corresponding judgment rules. This mechanism, based on static experience data, limits the model's ability to respond to new and abnormal situations. During drilling, conditions change rapidly, requiring the early warning system to respond in real time. 1. Rule-based expert models suffer from performance bottlenecks in real-time data processing and poor timeliness; 2. Expert experience models typically have extremely high data quality requirements, so noise and errors significantly impact the accuracy of their warnings; 3. Currently, expert models rely heavily on observing trends in drilling parameters to predict potential stuck pipe situations, which suffers from high false alarm rates and poor timeliness; 4. Developing and maintaining effective rules requires extensive expert knowledge and experience. With continuous technological and process advancements, rules need frequent updates, significantly increasing model maintenance costs.

[0004] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a stuck drill early warning method based on a recurrent neural network model with a multi-attention mechanism. This approach effectively overcomes the issues of significant interference, high computation time, and high cost associated with existing technologies. It focuses on key features based on a multi-attention architecture and adaptively captures temporal and spatial correlation features, integrating expert experience models to achieve efficient and accurate stuck drill early warning. The method acquires the data values ​​of the hook load sequence and its associated sequences; uses Pearson correlation coefficient analysis to determine the associated sequences that meet set conditions for correlation with the hook load sequence as target input sequences; comprehensively considers the needs of input information filtering, data space optimization, long-term dependent sequence processing, and dynamic feature capture to establish a recurrent neural network model with a multi-attention module; uses the parameters of the hook load sequence and its associated sequences as input sequences to calculate the hook load prediction value; and identifies stuck drill risks and provides early warnings based on the judgment rules corresponding to different working conditions. Preferably, in one embodiment, the method includes:

[0006] Step S100: Determine the relevant sequence of the hook load sequence based on the tool to be analyzed and the working condition, as the hook load associated sequence, obtain the historical value of the hook load sequence for a set time period, and obtain the historical value and current value of the hook load associated sequence for a set time period;

[0007] Step S200: Determine the hook-load correlation sequence that meets the set conditions with the hook-load sequence as the target input sequence by using the Pearson correlation coefficient analysis method;

[0008] Step S300: Taking into account the requirements for input information filtering, data space optimization, long-term dependent sequence processing, and dynamic feature capture, establish a recurrent neural network model with multiple attention modules;

[0009] Step S400: The hook load sequence and hook load associated sequence parameters are used as the input sequence of the recurrent neural network model to calculate the hook load prediction value. Based on this, the matching empirical identification model is used to identify the risk of stuck drill and realize early warning according to the identification rules corresponding to different working conditions.

[0010] In an optional embodiment, in step S100, the hook load associated sequence is determined to include a parameter sequence of torque, pump pressure, drilling fluid density, dogleg degree, total pool volume, and antisand amount; the parameter information of the hook load associated sequence is obtained from the logging tool operation data.

[0011] Furthermore, in one embodiment, the method further includes step S101: preprocessing the hook-load sequence parameters and hook-load associated sequence parameters, converting the sequence parameter file to CSV format, and then performing noise reduction, data cleaning, and outlier correction on the converted parameter file.

[0012] In a preferred embodiment, in step S200, the covariance and standard deviation of each hook-load correlation sequence parameter and the hook-load sequence are calculated as the correlation evaluation index, and several hook-load correlation sequences that meet the set requirements of the correlation evaluation index are selected as target input sequences.

[0013] Optionally, in one embodiment, during step S300, in the process of establishing a recurrent neural network model with multiple attention modules, the input attention module is configured to selectively choose important information features in the input sequence by learning the weights of the input sequence, and suppress the influence of secondary information features.

[0014] In one embodiment, during step S300, in the process of establishing a recurrent neural network model with multiple attention modules, the multidimensional spatial attention module is set to selectively capture the dynamic correlation between different input sequences, so that the model focuses on the role of related features.

[0015] In one embodiment, during step S300, in the process of establishing a recurrent neural network model with multiple attention modules, the hidden state attention module is set to generate feature maps for each position of the input sequence by calculating attention weights, which is suitable for processing long-term dependent sequences.

[0016] In one embodiment, during step S300, in the process of establishing a recurrent neural network model with multiple attention modules, a time attention module is set to capture the dynamic temporal correlation between different time intervals in the parameter sequence, so that the model pays attention to the temporal correlation features.

[0017] Furthermore, in one embodiment, in the recurrent neural network model, the encoder and decoder employ a long short-term memory neural network. The encoder learns the implicit associations between input sequences and encodes the input sequences as contextual feature representations; the decoder decodes the input information encoded by the encoder.

[0018] In a preferred embodiment, in step S400, matching identification rules are set for the continuous tripping, the tripping process lowering, the continuous drilling, and the drilling process ascending conditions, respectively.

[0019] Based on other aspects of the methods described in any one or more of the foregoing embodiments, the present invention also provides a storage medium storing program code capable of implementing the methods as described in any one or more of the foregoing embodiments.

[0020] Based on the application aspects of the methods described in any one or more of the above embodiments, the present invention also provides a stuck drill early warning system based on a recurrent neural network model with a multiple attention mechanism, which performs the methods described in any one or more of the above embodiments.

[0021] Compared with the closest prior art, the present invention also has the following beneficial effects:

[0022] This invention provides a stuck drill bit early warning method and system based on a recurrent neural network model with a multiple attention mechanism. The method acquires the data values ​​of the hook load sequence and the hook load associated sequence; determines the hook load associated sequence that meets the set conditions with the hook load sequence through Pearson correlation coefficient analysis as the target input sequence; predicts the future value of the hook load based on the historical values ​​of various hook load associated parameter sequences, taking into account the nonlinear correlation between different parameters, and solving the problem that current methods only rely on single drilling parameter early warning.

[0023] Taking into account the requirements for input information filtering, data space optimization, long-term dependent sequence processing, and dynamic feature capture, a recurrent neural network model with multiple attention modules is established. Using an encoding-decoding-based multiple attention architecture, it can adaptively and dynamically capture the temporal and spatial correlations between input sequences, enabling the model to focus on features relevant to the prediction task, thereby improving the model's prediction accuracy and reducing the model's training cost.

[0024] The hook load sequence and its associated sequence parameters are used as input sequences to the recurrent neural network model to obtain the hook load prediction value. Based on this prediction, a matching empirical identification model is used to identify the risk of stuck drill bit and provide early warning according to the identification rules corresponding to different working conditions. The early warning identification result is based on a multi-attention architecture, focusing on key features and adaptively capturing temporal and spatial correlation features. It integrates and applies expert experience models to achieve efficient and accurate early warning of stuck drill bit.

[0025] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0027] Figure 1 This is a flowchart illustrating a stuck drill early warning method based on a recurrent neural network model with multiple attention mechanisms, provided in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the module connection of the neural network model of the stuck drill early warning method based on a recurrent neural network model with multiple attention mechanism provided in the embodiments of the present invention;

[0029] Figure 3 This is a schematic diagram of the processing mechanism of the HSAM module in the stuck drill early warning method based on a recurrent neural network model with multiple attention mechanism provided in the embodiments of the present invention;

[0030] Figure 4 This is a schematic diagram of a stuck drill early warning system based on a recurrent neural network model with a multiple attention mechanism, provided in another embodiment of the present invention. Detailed Implementation

[0031] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. Those skilled in the art will then fully understand how the present invention uses technical means to solve technical problems and achieve technical effects, and will be able to implement the present invention specifically based on the above-described implementation process. It should be noted that, as long as there is no conflict, the various embodiments and features of the present invention can be combined with each other, and the resulting technical solutions are all within the protection scope of the present invention.

[0032] Although the flowchart describes the operations as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. The order of the operations can be rearranged. A process can terminate when its operation is complete, but it may also have additional steps not included in the diagram. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0033] Computer equipment includes user equipment and network equipment. User equipment or clients include, but are not limited to, computers, smartphones, and PDAs (Personal Digital Assistants); network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Computer equipment can operate independently to implement this invention, or it can connect to a network and implement this invention through interaction with other computer equipment within the network. The network in which the computer equipment resides includes, but is not limited to, the Internet, wide area networks (WANs), metropolitan area networks (MANs), local area networks (LANs), and VPN networks.

[0034] The terms “first,” “second,” etc., may be used herein to describe various units, but these units should not be limited by these terms; they are used merely to distinguish one unit from another. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items. When a unit is referred to as “connected” or “coupled” to another unit, it may be directly connected or coupled to said other unit, or there may be intermediate units present.

[0035] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a” and “an” as used herein are also intended to include the plural. It should also be understood that the terms “comprising” and / or “including” as used herein specify the presence of the stated features, integers, steps, operations, units, and / or components, without excluding the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0036] Stuck drill bit failures not only damage drilling equipment, shorten its service life, and increase the cost of maintenance and equipment replacement, but also cause drilling operations to be interrupted, affecting the overall progress and prolonging the drilling cycle.

[0037] Traditional stuck pipe early warning methods mainly fall into two categories. One is prediction based on winch weight data. This involves setting winch weight parameters and automatically recording the weight when the winch is stopped using a PLC. The weight increment is compared with the set weight parameters in real time, and an early warning is issued based on the comparison results. However, due to the complexity and variability of actual drilling conditions, setting only a weight parameter may lead to poor early warning accuracy under different formations and working conditions. Furthermore, this method only considers a single weight parameter without analyzing parameters related to stuck pipe, resulting in unobjective predictions that cannot reliably meet engineering requirements.

[0038] Secondly, stuck pipe warnings are based on expert experience models using rules to predict drilling failures, a method widely used in current engineering projects. However, due to the complex relationships and non-linear changes among various engineering parameters when stuck pipe occurs, expert experience models have the following limitations: 1. Expert experience models typically rely on existing knowledge or experience to formulate corresponding judgment rules. This mechanism, based on static experience data, limits the model's ability to respond to new and abnormal situations. During drilling, conditions change rapidly, requiring real-time responses from the warning system. Rule-based expert models suffer from performance bottlenecks in real-time data processing, resulting in poor timeliness. 2. Expert experience models typically have extremely high data quality requirements; therefore, noise and errors significantly impact the accuracy of their warnings. 3. Currently, expert models rely heavily on observing trends in drilling engineering parameters to predict potential stuck pipe situations. This method suffers from high false alarm rates and poor timeliness. 4. Developing and maintaining effective rules requires extensive expert knowledge and experience. With continuous technological and process advancements, rules need frequent updates, significantly increasing model maintenance costs.

[0039] Stuck drill pipe failures are influenced by various factors, including reservoir conditions, geology, and engineering aspects. Changes in hook load are one of the key signs of stuck drill pipe. Therefore, data-driven research into hook load prediction models and their applications plays a crucial role in early warning of stuck drill pipe.

[0040] To overcome the above-mentioned technical problems, the present invention proposes a...

[0041] The aim is to predict the future value of hook load, an important indicator affecting stuck drill bit warning. By using the changing trend of historical hook load data and a set calculation method, the future hook load value is reliably predicted, and the predicted value is used as an important basis for judging stuck drill bit warning. A recurrent neural network model based on multiple attention (MA-RNN) is proposed, which takes multiple types of drilling engineering parameters related to stuck drill bit, including hook load, as input to the model.

[0042] The Multi-Attention Recurrent Neural Network (MA-RNN) model follows an "encoder-decoder" architecture, dividing different attention mechanisms into four stages: The first stage uses an input attention module to weight the input data, highlighting information useful for the current task. This input attention mechanism can learn to weight different input features, thereby enhancing attention to relevant information and reducing attention to irrelevant information. The second stage uses a multidimensional attention module to selectively capture the dynamic correlations between the multi-channel ocean observation data output by the input attention module. The third stage uses a hidden state attention module (HSAM) to generate feature maps for each position in the input sequence by calculating attention weights. These feature maps can be used as the model's representation, making the model more effective in handling long-term dependent sequences. The fourth stage uses a temporal attention module to adaptively capture the dynamic temporal correlations between different time intervals of the multi-channel ocean observation data output by the hidden state attention module. This invention enables the fusion modeling of multiple drilling engineering parameters affecting stuck pipe, thus avoiding the problem of poor prediction accuracy.

[0043] Rule-based expert models typically rely on predefined rules and lack self-learning capabilities, limiting their ability to adapt to changes and optimize early warning systems. The solution proposed in this invention overcomes the problems of traditional expert experience models, such as high susceptibility to external conditions, poor timeliness, and high maintenance costs. The researchers considered the numerous variables involved in drilling processes and the complex nonlinear relationships between them. Rule-based expert experience models struggle to effectively capture these complex relationships, impacting the accuracy and real-time performance of early warnings. To address the shortcomings of expert models, this invention employs a MA-RNN model that utilizes historical values ​​of the hook load sequence and the current and historical values ​​of other related sequences to effectively predict the current value of the hook load sequence. During the calculation process, this method adaptively and dynamically updates model parameters based on different formations and working conditions.

[0044] The following describes the detailed flow of the method according to an embodiment of the present invention with reference to the accompanying drawings, the steps of which can be executed in a computer system containing, for example, a set of computer-executable instructions. Although the logical order of the steps is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0045] Example 1

[0046] Figure 1 This diagram illustrates a flowchart of a stuck drill early warning method based on a recurrent neural network model with a multiple attention mechanism, as provided in Embodiment 1 of the present invention. Figure 1 As can be seen, the method includes the following steps.

[0047] Step S100: Determine the relevant sequence of the hook load sequence based on the tool to be analyzed and the working condition, as the hook load associated sequence, obtain the historical value of the hook load sequence for a set time period, and obtain the historical value and current value of the hook load associated sequence for a set time period;

[0048] Step S200: Determine the hook-load correlation sequence that meets the set conditions with the hook-load sequence as the target input sequence by using the Pearson correlation coefficient analysis method.

[0049] Step S300: Taking into account the requirements for input information filtering, data space optimization, long-term dependent sequence processing, and dynamic feature capture, establish a recurrent neural network model with multiple attention modules;

[0050] Step S400: The hook load sequence and hook load associated sequence parameters are used as the input sequence of the recurrent neural network model to calculate the hook load prediction value. Based on this, the matching empirical identification model is used to identify the risk of stuck drill and realize early warning according to the identification rules corresponding to different working conditions.

[0051] The stuck drill bit early warning method provided in this invention predicts the future value of the hook load based on the associated parameter values ​​of the hook load. It takes into account the nonlinear correlation between different parameters, solving the problem that current methods only rely on single drilling parameter early warning. It uses a multi-attention architecture based on encoding and decoding, which can adaptively and dynamically capture the temporal and spatial correlation between input sequences, enabling the model to focus on features related to the prediction task, improving the prediction accuracy of the model, reducing the training cost of the model, and integrating the output data of the prediction model and the expert experience model in the identification process, thus truly realizing stuck drill bit early warning. It solves the problems of pure expert experience model methods, such as being greatly affected by external conditions, poor timeliness, and high maintenance costs. It is computationally efficient and highly reliable.

[0052] First, in this embodiment of the invention, the relevant sequence of the hook load sequence is determined based on the tool to be analyzed and the working condition, which is used as the hook load associated sequence. The historical value of the hook load sequence for a set time period is obtained, as well as the historical value and current value of the hook load associated sequence for a set time period.

[0053] In an optional embodiment, in step S100, the hook load associated sequence is determined to include a parameter sequence of torque, pump pressure, drilling fluid density, dogleg degree, total pool volume, and antisand amount.

[0054] In practical applications, raw logging parameter datasets including hook load, torque, pump pressure, drilling fluid density, dogleg density, total pool volume, and reverse sand volume are obtained from the integrated logging instrument.

[0055] The embodiments of the present invention predict the future value of hook load based on the historical values ​​of sequence parameters such as hook load, torque, pump pressure, drilling fluid density, dogleg degree, total pool volume, and anti-sand amount. It takes into account the nonlinear correlation between different parameters and solves the problem that the current method only relies on a single drilling parameter for early warning.

[0056] Furthermore, in one embodiment, the method further includes:

[0057] Step S101: Preprocess the hook load sequence parameters and hook load associated sequence parameters. After converting the sequence parameter file to CSV format, perform noise reduction, data cleaning, and outlier correction on the converted parameter file.

[0058] Drilling engineering parameters related to stuck pipe, such as hook load, torque, pump pressure, drilling fluid density, dogleg density, total pool volume, and sand removal volume, collected by the integrated logging tool, are used as the original input sequence for the model. Data preprocessing is performed on the original input sequence. The preprocessing stage includes: converting the original JSON format data file to a CSV format file; performing noise reduction, data cleaning, and correcting outliers on the file.

[0059] The set length sequence parameters of torque, pump pressure, drilling fluid density, and total pool volume are selected as the target input sequence, which, together with the hook load sequence, form the comprehensive input sequence of the recurrent neural network model.

[0060] The present invention aims to obtain hook load prediction data based on the sequence parameters related to the hook load using a prediction model; therefore, step S200 is executed next, and the hook load associated sequence that meets the set conditions with the correlation with the hook load sequence is determined as the target input sequence by Pearson correlation coefficient analysis.

[0061] Specifically, in an optional embodiment, in step S200, the covariance and standard deviation of each hook-load correlation sequence parameter and the hook-load sequence are calculated as the correlation evaluation index, and several hook-load correlation sequences that meet the set requirements of the correlation evaluation index are selected as target input sequences.

[0062] In practical applications, the hook load sequence is used as the target sequence, and the Pearson correlation coefficient is used to measure its correlation with other logging sequences such as torque, pump pressure, drilling fluid density, dogleg degree, total pool volume, and antisand amount. By calculating the quotient of the covariance and standard deviation between the target sequence and the other sequences, the six sequences of torque, pump pressure, drilling fluid density, dogleg degree, total pool volume, and antisand amount are selected to form the input sequence of the model along with the hook load sequence.

[0063] Next, step S300 is executed, which comprehensively considers the requirements for input information filtering, data space optimization, long-term dependent sequence processing, and dynamic feature capture to establish a recurrent neural network model with multiple attention modules as a prediction model.

[0064] In a preferred embodiment, in step S300, during the process of establishing a recurrent neural network model with multiple attention modules, the input attention module is configured to selectively choose important information features in the input sequence by learning the weights of the input sequence, and suppress the influence of secondary information features.

[0065] In the process of building a recurrent neural network model with multiple attention modules, a multidimensional spatial attention module is set to selectively capture the dynamic correlation between different input sequences, so that the model focuses on the role of relevant features.

[0066] In the process of building a recurrent neural network model with multiple attention modules, the hidden state attention module is set to generate feature maps for each position of the input sequence by calculating attention weights, which is suitable for processing long-term dependent sequences.

[0067] In the process of building a recurrent neural network model with multiple attention modules, a temporal attention module is set up to capture the dynamic temporal correlation between different time intervals in the parameter sequence, so that the model pays attention to the temporal correlation features.

[0068] In the recurrent neural network model, a four-fold matching attention module is set up around each of the four stages. The attention modules for the four stages are as follows:

[0069] (i) Input Attention Module: The input attention module filters out unimportant information from the input sequence while highlighting information useful for the current task. It learns different weights in the input sequence to selectively focus on important features (high-value features) and suppress unimportant features. It can be used to process high-dimensional and complex data, helping the model extract useful features.

[0070] (ii) Multidimensional spatial attention module: The multidimensional spatial attention module concentrates the relevant information between the query vector and the input vector into a local region related to the query vector, thereby improving the model's attention to and accuracy of relevant information.

[0071] (iii) Hidden State Attention Module (HSAM): The Hidden State Attention Module can generate feature maps for each position of the input sequence by calculating attention weights. These feature maps can be used as the representation of the model. After this process, the model is more effective in processing long-term dependent sequences.

[0072] (iv) Temporal attention module: The temporal attention module enables the model to adaptively capture the dynamic temporal correlation between different time intervals in the input sequence.

[0073] In the recurrent neural network model, the encoder and decoder employ a long short-term memory neural network. The encoder learns the implicit associations between input sequences and encodes the input sequences into contextual feature representations. The decoder decodes the input information encoded by the encoder.

[0074] like Figure 2 As shown, Input attention refers to the input attention module, which filters out unimportant information in the input sequence while highlighting information useful for the current task. It can selectively focus on important features and suppress unimportant features by learning different weights in the input sequence.

[0075] Multidimensional attention refers to a multidimensional spatial attention module, which can selectively capture the dynamic correlation between different input sequences;

[0076] LSTM stands for Long Short-Term Memory Neural Network. It can learn the hidden layer representation of the input sequence and update the hidden state at the current time step based on the input sequence and its previous hidden state and memory units, so as to better capture long-term dependencies in the sequence.

[0077] HSAM stands for Hidden State Attention Module, which generates feature maps for each position of the input sequence by computing attention weights. These feature maps can be used as the representation of the model, and the model is more effective in processing long-term dependent sequences after this process.

[0078] Temporal attention refers to a time attention module that can adaptively capture the dynamic temporal correlation between different time intervals in a sequence.

[0079] In the model, both the encoder and decoder are Long Short-Term Memory (LSTM) neural networks, which can learn the implicit associations between input sequences and encode the input sequences into contextual feature representations.

[0080] Based on the constructed prediction model, this embodiment of the invention realizes the construction of prediction and the early warning of stuck drill through step S400; the hook load sequence and hook load associated sequence parameters are used as the input sequence of the recurrent neural network model to calculate the hook load prediction value, and based on it, the matching experience judgment model is used to identify the risk of stuck drill according to the judgment rules corresponding to different working conditions and realize the early warning.

[0081] Combining the above-mentioned recurrent neural network model, this invention implements a hook load value prediction model based on a multi-attention recurrent neural network combined with an expert experience model when predicting hook load data. The overall stuck drill risk warning process includes the following steps:

[0082] (1) Select the four drilling engineering parameter sequences of torque, pump pressure, drilling fluid density, and total tank volume, and combine them with the hook load sequence to form the input sequence. X=(x 1 ,x 2 ,x 3 ,...,x n ) T =(x1,x2,x3,...,x T )∈R n×T , where n represents the number of different types of sequences, T represents the length of the input sequence, and X represents the dataset consisting of five sequences: hook load, torque, pump pressure, drilling fluid density, and total tank volume.

[0083] (2) Decompose the input sequence after (2) into a hook sequence Y = (y1, y2, ..., y2) T-1 The multi-feature time series X = (x1, x2, ..., x3) consists of four sequences: torque, pump pressure, drilling fluid density, and total pool volume. T ).

[0084] (3) In the encoding stage, the encoder is actually a recurrent neural network (RNN). The encoder learns the implicit relationships between the input sequences and encodes the input sequences into contextual feature representations. Given an input sequence X = (x1, x2, ..., x... T The encoder can learn the input sequence x at time t. t to hidden state h t Mapping:

[0085] h t =f(h) t-1 ,x t )

[0086] Here, f(·) is a non-linear activation function, which is set here as a Long Short-Term Memory (LSTM) neural network unit to capture long-term dependencies in the sequence.

[0087] Each LSTM has memory cells. Let the encoder's memory cell state vector at time t be s_e. t Three sigmoid gates control the memory units: the forget gate (f...). t Input gate i t Output gate o t The LSTM process for processing input information is shown in the following formula:

[0088] f t =σ(W f [h t-1 ;x t ]+b f )

[0089] i t =σ(W i [h t-1 ;x t ]+b i )

[0090] o t =σ(W o [h t-1 ;x t ]+b o )

[0091]

[0092]

[0093] Where [;] represents a join operation, h t f is the hidden state of the encoder at time t. t i t o t These represent the outputs of the forget gate, input gate, and output gate, respectively.t W represents the input at time t. f W i W o W represents the weight vector matrices of the forget gate, input gate, and output gate, respectively. s Let b represent the trainable parameter matrix. f b i b o b s These represent the biases of the forget gate, input gate, and output gate, respectively. σ(·) represents the sigmoid activation function. This indicates element-wise multiplication.

[0094] (4) Input the hook sequence X into the input attention module. The input attention module is essentially constructed from a multilayer perceptron. It uses the input sequence X and the hidden state h of the encoder at time t-1 to achieve the desired effect. t-1 and memory cell state s_e t-1 Calculate attention weights:

[0095]

[0096]

[0097]

[0098] Among them, V e W e U e These are all trainable parameter matrices, representing weight vector matrices at different stages. This is the attention weight, which measures the importance of the nth input sequence at time t. The attention weights are normalized using the Softmax function to obtain... Ensure the sum of the attention weights is 1. After obtaining the attention weights, adaptively extract features from the input sequence, weight the input sequence, reconstruct the input sequence, and obtain a new sequence x'. t .

[0099] The new sequence x' after processing by the input attention module t The input sequence serves as the input sequence for the multidimensional spatial attention module. The multidimensional spatial attention module calculates the correlation between the input sequence and the query sequence, adaptively capturing the dynamic spatial relationships that exist between the input sequences.

[0100] Calculated using the multidimensional spatial attention module The encoder uses the LSTM unit to calculate the hidden state unit h from the previous time step (t-1). t-1 Compared with the input sequence at the current time The hidden state unit h at the current time (time t) is calculated using the activation function. t ;Right now:

[0101]

[0102]

[0103]

[0104] Wherein, the query vector q T This is a parameter vector learned by the model from the input sequence, used to calculate the correlation between the input sequence and the query vector, and to aggregate key information in the sequence through weighted summation. Its feature dimensions are the same as those of the input sequence. The attention weights at time t are calculated. It represents the query vector q. T The correlation between x and the position (i,j,...,k,l) ​​in the input vector. i,j,...,k,l Let represent the feature vector at position (i,j,...,k,l) ​​in the input vector, and exp(·) denote the exponential function. The denominator of the above equation is the sum of the relevances between all positions in the input sequence and the query vector, ensuring that the sum of all attention weights is 1. After calculating the attention weights, a weighted summation is performed to obtain the result.

[0105] (5) In the decoding stage, in order to predict the final output value Another LSTM-based recurrent neural network is used to decode the input information encoded by the encoder. The hidden state attention module (HSAM) enhances the model's representational ability by refining the input hidden state units and mapping the hidden states to H. T Update the hidden state feature map h' by performing element-wise multiplication with all hidden states h of the encoder.

[0106] like Figure 3 As shown, the hidden state attention module extracts feature context descriptors by average pooling and max pooling along the hidden state feature axis of the encoder, respectively. and These are fed into a shared multilayer perceptron (MLP) layer. Then, the output feature map is obtained through two nonlinear transformations performed by the MLP. and They are concatenated and used as input to the convolutional layer to obtain the hidden state map H. T ∈R 1×T The details are as follows:

[0107]

[0108] in, and This indicates that HSAM extracts feature context descriptors generated by average pooling and max pooling along the hidden state feature axis of the encoder, respectively. r and W s The weights of the ReLU and Sigmoid activation functions in the fully connected layer of the MLP are shown below, and h represents all the hidden states of the encoder.

[0109] Map the hidden state H T Update the hidden state feature map h' by performing element-wise multiplication with all hidden states h of the encoder:

[0110]

[0111] The time attention mechanism, by weighting information at different time steps, enables the model to pay more attention to information relevant to the current time step. It can effectively handle the relationship between different time steps and improve the model's prediction accuracy and generalization ability.

[0112] When the encoder obtains the new hidden state feature map h at time t... t Then, it is compared with the hidden state d of the decoder at time t-1. t-1 and memory cell state s_d t-1 The inputs are combined and fed into the time-attention module to calculate the attention weights of the decoder's hidden state at time t. The calculated attention weights are processed using the Softmax function. Normalization yields Temporal attention mechanism calculates context vector c t As a weighted sum of the hidden new states of all decoders:

[0113]

[0114]

[0115]

[0116] Among them, W m U m V m These represent the weight vector matrices for three different computational stages. This indicates the importance of the hidden state of the i-th decoder to the prediction.

[0117] In actual calculations, W mU m V m These represent three different weight vector matrices, indicating that in the first calculation stage, the weight vector matrix W... m The second computation stage involves multiplying the vector matrix U, which is concatenated with the hidden state and the memory unit. m In the third stage, the vector matrix is ​​multiplied by the vector representing the hidden state feature mapping. Then, the tanh activation function is used to calculate the weight vector matrix V. m Multiply by the vector matrix calculated using the tanh activation function.

[0118] (6) The decoder will input the sequence y at time t-1. t-1 With context vector c t-1 The hidden state of the decoder at time t is updated by combining the two elements. Finally, the decoder concatenates the context vector vT at time T with the hidden state unit dT to calculate the final prediction result.

[0119] The decoder obtains the context vector c at time t-1. t-1 Combine it with the decoder input y at time t-1 t-1 Combined to update the decoder input sequence y t ' -1 Then, the decoder updates the input sequence y. t ' -1 and the decoder hidden state d at time t-1 t-1 Used to update the decoder hidden state d at time t t :

[0120] y' t-1 =ω T [y t-1 ;c t-1 ]+b

[0121] d t =f(d t-1 ;y t ' -1 )

[0122] Where, ω T b is the parameter matrix that maps the connection to the decoder input, and ω T is the parameter matrix, and b is the bias term.

[0123] (7) The context vector c at time T T and the hidden state d of the decoder T Combined to generate the final predicted value at time T.

[0124]

[0125] Where [·:·] represents a join operation, [d T ;c T The symbol ] represents the concatenation operation between the decoder's hidden state and the context vector at time T, using weights V. y The bias term is b v The final prediction result is obtained from the linear function.

[0126] The above embodiments of the present invention use a multi-attention architecture based on encoding and decoding, which can adaptively and dynamically capture the temporal and spatial correlations between input sequences, enabling the model to focus on features relevant to the prediction task, improving the prediction accuracy of the model, and reducing the training cost of the model.

[0127] The prediction results are input into the expert experience model, which then issues a warning about stuck drills through rule-based judgment. After obtaining the prediction results, the expert experience model issues a warning about stuck drills through rule-based judgment, threshold setting, and other methods.

[0128] In step S400, matching identification rules are set for the continuous tripping, tripping and lowering, continuous descent, and descent and ascent conditions, respectively; the specific identification rules are as follows:

[0129] 1. During the continuous ascent of the large hook, it did not descend in the middle when it reached the highest point (continuous drilling):

[0130] (1) (Current hook load value + 1 drill rod weight) - previous hook load maximum value > 50KN, alarm.

[0131] 2. When the large hook is continuously rising but has not reached the highest point, and there is a situation where the movable drill string is lowered in the middle (the movable drill string is raised and lowered during the tripping process; rule (1) applies to the raising situation, and rule (2) applies to the lowering situation):

[0132] (2) (Moving drill string, lowering action), the difference between the current hook load fluctuation and the first hook load fluctuation of this string is >= 50KN, alarm.

[0133] 3. During the continuous descent of the large hook, it did not rise in the middle when it reached the lowest point (continuous drilling):

[0134] (3) Rule: (Current hook load - weight of 1 drill rod) - minimum value of previous hook load >= 50KN, alarm.

[0135] 4. When the large hook continues to descend but has not reached the lowest point, and there is a situation where the moving drill string is raised or lowered in the middle (the moving drill string is raised or lowered during the descent; rule (3) applies to the lowering situation, and rule (4) applies to the raising situation):

[0136] (4) Rule: (Moving drill bit, lifting action), if the difference between the current hook load fluctuation and the first hook load fluctuation of this drill bit is >= 50KN, alarm will be triggered.

[0137] It is evident that this invention integrates predictive models and expert experience models, truly achieving drill jam warning and solving the problems of high susceptibility to external conditions, poor timeliness, and high maintenance costs that currently exist when relying solely on expert experience models.

[0138] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0139] It should be noted that, in other embodiments of the present invention, the method can also combine one or more of the above embodiments to obtain a new stuck drill early warning method based on a recurrent neural network model with a multiple attention mechanism, so as to efficiently and accurately realize the prediction and analysis of stuck drills in engineering, and provide a direct and reliable basis for processing decisions.

[0140] Example 2

[0141] It should be noted that, based on the methods in any one or more embodiments of the present invention described above, the present invention also provides a storage medium storing program code that can implement the methods described in any one or more embodiments. When the program code is executed by the operating system, it can implement the stuck drill warning method based on the recurrent neural network model with multiple attention mechanism as described above.

[0142] Example 3

[0143] The methods described in detail in the above-disclosed embodiments of the present invention can be implemented using various forms of devices or systems. Therefore, based on other aspects of the methods described in any one or more of the above embodiments, the present invention also provides a stuck drill warning system based on a recurrent neural network model with a multiple attention mechanism. This system is used to execute the stuck drill warning method based on a recurrent neural network model with a multiple attention mechanism described in any one or more of the above embodiments. Specific embodiments are given below for detailed description.

[0144] Specifically, Figure 4 The diagram shows a schematic representation of a stuck drill early warning system based on a recurrent neural network model with a multiple attention mechanism, as provided in an embodiment of the present invention. Figure 4 As shown, the system includes:

[0145] The parameter sequence acquisition module is configured to determine the relevant sequence of the hook load sequence around the tool to be analyzed and the working condition, as the hook load associated sequence, acquire the historical value of the hook load sequence for a set time period, and acquire the historical value and current value of the hook load associated sequence for a set time period;

[0146] The input sequence determination module is configured to use Pearson correlation coefficient analysis to determine the hook-load correlation sequence that meets the set conditions with the hook-load sequence as the target input sequence.

[0147] The prediction model building module is configured to build a recurrent neural network model with multiple attention modules, taking into account input information filtering requirements, data space optimization requirements, long-term dependent sequence processing requirements, and dynamic feature capture requirements.

[0148] The calculation and early warning module is configured to take the hook load sequence and hook load associated sequence parameters as the input sequence of the recurrent neural network model, calculate the hook load prediction value, and use the matching empirical judgment model to identify the risk of stuck drill and realize early warning according to the judgment rules corresponding to different working conditions.

[0149] In an optional embodiment, the parameter sequence acquisition module is configured to: determine the hook load associated sequence including a parameter sequence of torque, pump pressure, drilling fluid density, dogleg degree, total pool volume, and anti-sand amount; the parameter information of the hook load associated sequence is obtained from the logging tool operation data.

[0150] Furthermore, in one embodiment, the system further includes a parameter processing module configured to preprocess the hook-load sequence parameters and hook-load associated sequence parameters, convert the sequence parameter file to CSV format, and then perform noise reduction, data cleaning, and outlier correction on the converted parameter file.

[0151] In a preferred embodiment, the input sequence determination module uses the quotient of the covariance and standard deviation of each hook-load correlation sequence parameter and the hook-load sequence as a correlation evaluation index, and selects several hook-load correlation sequences that meet the set requirements as target input sequences.

[0152] Optionally, in one embodiment, during the process of the prediction model building module building a recurrent neural network model with multiple attention modules, the input attention module is configured to selectively select important information features in the input sequence by learning the weights of the input sequence, and suppress the influence of secondary information features.

[0153] In one embodiment, during the process of building a recurrent neural network model with multiple attention modules, the prediction model building module sets a multidimensional spatial attention module to selectively capture the dynamic correlation between different input sequences, so that the model pays attention to the role of relevant features.

[0154] In one embodiment, during the process of building a recurrent neural network model with multiple attention modules, the prediction model building module sets a hidden state attention module to generate feature maps for each position of the input sequence by calculating attention weights, which is suitable for processing long-term dependent sequences.

[0155] In one embodiment, during the process of building a recurrent neural network model with multiple attention modules, the prediction model building module sets a time attention module to capture the dynamic temporal correlation between different time intervals in the parameter sequence, so that the model pays attention to the temporal correlation features.

[0156] Furthermore, in one embodiment, in the recurrent neural network model established by the prediction model building module, the encoder and decoder employ a long short-term memory neural network. The encoder learns the implicit correlations between input sequences and encodes the input sequences into contextual feature representations; the decoder decodes the input information encoded by the encoder.

[0157] In a preferred embodiment, the calculation and early warning module is configured to set matching identification rules for the continuous tripping drilling condition, the tripping drilling process lowering condition, the continuous drilling process lowering condition, and the drilling process ascending condition, respectively.

[0158] In the stuck drill early warning system based on a recurrent neural network model with multiple attention mechanism provided in this invention, each module or unit structure can operate independently or in combination according to the actual parameter correlation analysis requirements and model operation requirements to achieve the corresponding technical effects.

[0159] It should be understood that the embodiments disclosed herein are not limited to the specific structures, processing steps, or materials disclosed herein, but should be extended to equivalent substitutions of these features as understood by those skilled in the art. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting.

[0160] The phrase "an embodiment" in the specification means that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0161] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and variations in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection for this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A stuck drill warning method based on a multiple attention mechanism recurrent neural network model, characterized in that, The method comprises: Step S100: determining the relevant sequence of the load hook sequence around the tool and the working condition to be analyzed, obtaining the historical value of the load hook sequence for a set time period as the load hook correlation sequence, and obtaining the historical value and the current value of the load hook correlation sequence for a set time period; Step S200: determining the load hook correlation sequence meeting the set condition in terms of the correlation with the load hook sequence as the target input sequence by the Pearson correlation coefficient analysis method; Step S300: comprehensively considering the input information filtering requirement, the data space optimization requirement, the long-term dependent sequence processing requirement, and the dynamic feature capturing requirement to establish a recurrent neural network model with multiple attention modules; Step S400: taking the load hook sequence and the load hook correlation sequence parameters as the input sequence of the recurrent neural network model, calculating the load hook prediction value, and identifying and warning the drill pipe sticking risk according to the identification rules corresponding to different working conditions based on the matched experience identification model.

2. The method according to claim 1, characterized in that, In step S100, the load hook correlation sequence includes the parameter sequence of torque, pump pressure, drilling fluid density, dogleg, total tank volume, and sand amount; and the parameter information of the load hook correlation sequence is obtained from the logging instrument operation data.

3. The method of claim 1, wherein, The method further comprises step S101: preprocessing the load hook sequence parameters and the load hook correlation sequence parameters, converting the sequence parameter file into a csv format, and performing noise reduction processing, data cleaning, and abnormal value correction on the converted parameter file.

4. The method according to claim 1, characterized by, In step S200, the correlation evaluation index is calculated by taking the quotient of the covariance and the standard deviation of each load hook correlation sequence parameter and the load hook sequence as the correlation evaluation index, and a plurality of load hook correlation sequences meeting the set requirement in terms of the correlation evaluation index are selected as the target input sequence.

5. The method according to claim 1, characterized by, In step S300, in the process of establishing the recurrent neural network model with multiple attention modules, the input attention module is set to learn the weight of the input sequence, selectively select important information features in the input sequence, and suppress the influence of secondary information features.

6. The method according to claim 1, characterized by, In step S300, in the process of establishing the recurrent neural network model with multiple attention modules, the multi-dimensional space attention module is set to selectively capture the dynamic correlation between different input sequences, so that the model focuses on the role of related features.

7. The method according to claim 1, characterized by, In step S300, in the process of establishing the recurrent neural network model with multiple attention modules, the hidden state attention module is set to generate feature mapping for each position of the input sequence by calculating the attention weight, which is suitable for processing long-term dependent sequences.

8. The method according to claim 1, characterized by, In step S300, in the process of establishing the recurrent neural network model with multiple attention modules, the time attention module is set to capture the dynamic time correlation between different time intervals in the parameter sequence, so that the model focuses on the time correlation features.

9. The method according to claim 1, characterized by, In the recurrent neural network model, the encoder and the decoder adopt a long short-term memory neural network, the encoder learns the implicit correlation between the input sequences and encodes the input sequences into context feature representations, and the decoder decodes the input information encoded by the encoder.

10. The method according to claim 1, characterized by, In step S400, the matching identification rules are respectively set around the continuous tripping condition, the tripping process lowering condition, the continuous drilling condition and the drilling process rising condition.

11. A storage medium, characterized by, The storage medium has program codes stored thereon, which can implement the method in any one of claims 1-10.

12. A stuck drill alert system based on a multiple attention mechanism recurrent neural network model, characterized in that, The system performs the method in any one of claims 1-10.

Images