Drill string vibration identification method during drilling process based on multi-scale temporal convolutional network
Through mutual information analysis and multi-scale time convolutional network model, the problem of identifying drill string vibration during drilling was solved, and the automatic identification and effective processing of drill string vibration during drilling was achieved, thereby improving drilling efficiency and safety.
Patent Information
- Application Number
- CN202311245618.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-25
- Publication Date
- 2025-09-16
- Estimated Expiration
- 2043-09-25
AI Technical Summary
Existing technologies have difficulty in effectively identifying and processing severe drill string vibration during drilling, resulting in reduced drilling efficiency and safety. In particular, it is difficult to automatically extract drill string vibration characteristics under complex formation conditions.
Mutual information analysis method is used to select process variables related to torque, and a multi-scale time convolutional network model is constructed. Automatic identification of drill string vibration is achieved through multi-scale feature extraction and recognition modules.
Effective identification of drill string vibration is achieved under complex formation conditions, which improves the accuracy and robustness of identification and ensures the efficiency and safety of the drilling process.
Smart Images
Figure CN117235625B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of geological drilling engineering, and in particular to a method for identifying drill string vibration during drilling based on a multi-scale time convolutional network. Background Art
[0002] Geological drilling is a complex industrial process that effectively explores deeply buried geological resources and energy, effectively ensuring my country's resource and energy needs. According to current mineral resource exploration statistics in my country, the average exploration depth for mineral resources is only 500 meters, while deep-seated oil resources account for approximately 40% of remaining oil resources. Oil and gas extraction depths average less than 4,500 meters, while deep-seated natural gas resources account for approximately 60% of remaining natural gas resources. Furthermore, preliminary estimates indicate that my country possesses substantial deep unconventional energy reserves. Recoverable shale gas reserves are estimated at 36.0825 trillion cubic meters, accounting for approximately 20% of the world's total proven reserves. Coalbed methane resources buried at depths exceeding 2,000 meters are estimated at 36.81 trillion cubic meters, representing approximately 13% of the world's total coalbed methane resources. Therefore, with the depletion of shallow resources and the discovery of numerous deep resources, deep geological exploration is necessary to alleviate my country's resource and energy constraints.
[0003] As deep geological exploration deepens, ensuring the efficiency and safety of the drilling process is crucial. During the actual drilling process, the drill rig transmits mechanical energy from top to bottom in the form of torque, driving the drill string and drill tools downhole for normal drilling. However, not all energy is consumed in the normal rotary drilling process. Some energy is lost in severe drill string vibration, which in turn seriously endangers drilling safety, inhibits drilling efficiency, and increases drilling costs. According to statistics, nearly 40% of drilling depths are affected by severe drill string vibration each year, and the negative impact becomes more severe with increasing drilling depth. If severe drill string vibration is not identified and addressed in a timely manner, it will result in huge economic losses. Therefore, effectively identifying severe drill string vibration is of great value and significance to ensuring the efficiency, safety, and economic benefits of the drilling process.
[0004] The drilling process involves numerous process parameters, primarily detected by surface sensors. These parameters reflect real-time drilling conditions, and their time series and distribution characteristics help monitor drill string vibration. As complex formations change during drilling, the time series and distribution characteristics of these process variables also change, posing a challenge to drill string vibration identification. Therefore, leveraging time series data from surface process variables to automatically extract drill string vibration characteristics in complex formations and effectively identify drill string vibration is crucial for state monitoring, efficiency optimization, and control of the drilling process. Summary of the Invention
[0005] Severe drill string vibration during drilling affects drilling efficiency and safety, and its severity and impact are directly affected by complex formation conditions. Numerous drilling process variables help reflect and monitor drill string vibration, but their timing and distribution characteristics vary with complex formations, making manual extraction of vibration features very difficult. To address these issues, the present invention provides a method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network. This method automatically extracts and effectively identifies drill string vibration features under complex formation conditions. The method primarily includes the following steps:
[0006] S1: The mutual information analysis method is used to calculate the mutual information value between the time series of each drilling process variable. The larger the mutual information value, the stronger the correlation between the process variables. Since torque reflects the vibration of the drill string, the process variable with a strong correlation with torque is selected as the characteristic variable, and the time series data S corresponding to the multivariate characteristic variable is obtained.
[0007] S2: Build a drill string vibration recognition model, which includes a multi-scale feature extraction module and a drill string vibration recognition module. Time series data S is used as a sample and input into the model for training.
[0008] The multi-scale feature extraction module uses a multi-scale temporal convolutional network with different dilated convolution factors (d1, d2, d3) as a feature extractor. The corresponding temporal convolutional network under a single dilated convolution factor extracts different single-scale vibration features, which are finally fused into multi-scale features, and vibration features are obtained based on the multi-scale features.
[0009] In the drill string vibration recognition module, a one-dimensional global maximum pooling layer (GMP) is first used to reduce the dimensionality of the vibration features to obtain high-level features. Then, a fully connected layer (FC) is used to learn and combine the high-level features and map them to the sample label space. Finally, a softmax layer is used to calculate and output the probability that the sample corresponds to different categories. The probability classification layer then identifies and outputs the corresponding vibration category.
[0010] S3: Use the trained drill string vibration recognition model to identify the drill string vibration during the actual drilling process.
[0011] Furthermore, the specific process of determining the input characteristic variables of the drill string vibration identification model is as follows:
[0012] (1-1): V={X i |i=1,...,M} represents the set of drilling process variables, where X i represents the time series of the i-th process variable, M represents the total number of process variable types; the mutual information value calculation formula shown below is used to perform correlation analysis on the drilling process variable V:
[0013]
[0014] Where X and Y represent the time series of any two variables in V, p(x,y) represents the joint probability distribution of variables X and Y, and p(x) and p(y) represent the marginal probability distribution of variables X and Y, respectively.
[0015] (1-2): Sort the mutual information values calculated in step (1-1) in descending order, and select the process variables with higher mutual information values with torque as the preliminary input feature variables of the drill string vibration identification model;
[0016] (1-3): Analyze the mutual information value between the characteristic variables in step (1-2), eliminate the redundant characteristic variables with strong correlation, and determine the time series data S = {X i |i=1,...,N}, where N represents the number of types of feature variables.
[0017] Furthermore, the specific process of extracting the multi-scale features of drill string vibration and obtaining the vibration characteristics is as follows:
[0018] (2-1): The feature extractor is constructed by a multi-scale temporal convolutional network, which consists of two layers with different dilation factors d j Atrous Causal Convolutional Layer (DCC) j , batch normalization layer BN, rectified linear unit ReLU, dropout layer and additional one-dimensional convolution layer constitute residual connection, j = 1, 2, 3, capture and extract multivariate feature variable time series data S = {X i |i=1,...,N}, and obtain the vibration characteristics of the drill string vibration; the operation output of each network layer in the multi-scale time convolution network is expressed as follows:
[0019] Atrous causal convolution layer operation:
[0020]
[0021] Among them, F(s) represents the dilated convolution operation on element s in the time series data S, f(i) represents the i-th convolution filter, k is the filter size, d is the dilation factor, and f d represents a convolution filter with a dilation factor of d, S s-d.i Represents the sequence of operations of time series data S on the i-th convolution filter, with different hole factors d in the same layer j Atrous Causal Convolutional Layer (DCC) j The output can be expressed as follows:
[0022] Y 1j =DCC j (S,W j )
[0023] Among them, Y 1j Indicates DCC j The intermediate features obtained by the layer, W j Indicates DCC j Layer trainable parameters;
[0024] Batch normalization layer BN output:
[0025] Y 2j =BN(Y 1j )
[0026] Among them, Y 2j Represents the intermediate features obtained by the BN layer;
[0027] Rectified linear unit layer ReLU output:
[0028] Y 3j =ReLU(Y 2j )=max(0,W BN Y 2j +b BN )
[0029] Among them, Y 3j Represents the intermediate features obtained by the ReLU layer, W BN and b BN Represents the trainable parameters of the ReLU layer;
[0030] (2-2): Based on different void factors d j Atrous Causal Convolutional Layer (DCC) j The extracted single-scale features and the fused multi-scale features are expressed as follows:
[0031]
[0032] Among them, Y4 represents multi-scale features, Y 3j represents the void factor d j The single-scale features obtained by the second ReLU layer, j = 1, 2, 3, Represents feature fusion operation;
[0033] (2-3): The residual connection operation is expressed as follows:
[0034] Y=Y4+WS
[0035] Among them, Y represents the vibration feature, W represents an additional one-dimensional convolution operation, and “+” represents the feature element addition operation.
[0036] Furthermore, the process of identifying drill string vibration is as follows:
[0037] (3-1): Use the one-dimensional global maximum pooling layer to perform feature dimensionality reduction on the vibration features. The output of the one-dimensional global maximum pooling layer GMP is as follows:
[0038] Y GMP =GMP(Y)
[0039] Among them, Y GMP Represents the output of the GMP layer, i.e., high-level features, and GMP() represents a one-dimensional global maximum pooling layer;
[0040] (3-2): The output of the fully connected layer FC is expressed as follows:
[0041] Y FC =FC(Y GMP )=W FC Y GMP +b FC
[0042] Among them, Y FC represents the FC layer output, FC() represents the fully connected layer, W FC and b FC Represent the weight and bias parameters of the FC layer respectively;
[0043] (3-3): The output of the Softmax layer is expressed as follows:
[0044]
[0045] Among them, P i Indicates the output probability that the sample belongs to the i-th category, satisfying n represents the number of drill string vibration categories; softmax() represents the Softmax layer; ψ i Represents the output layer parameters.
[0046] A storage device stores instructions and data for implementing the method for identifying drill string vibration during drilling based on a multi-scale time convolutional network.
[0047] A device for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network, comprising: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network.
[0048] The beneficial effects of the technical solution provided by the present invention are as follows: the present invention adopts the mutual information analysis method to calculate the mutual information value between the time series of each drilling process variable, selects the process variable with strong correlation with the torque as the feature variable, obtains the time series data S corresponding to the multivariate feature variable, constructs a drill string vibration recognition model, uses the time series data S as a sample, and inputs it into the model for model training; uses a multi-scale time convolution network with different void convolution factors as a feature extractor to extract effective multi-scale features and vibration features of the drill string vibration, adopts a one-dimensional global maximum pooling layer to perform feature dimensionality reduction processing on the vibration features to obtain high-level features; then, adopts A fully connected layer learns and combines these high-level features, mapping them to the sample label space. Finally, a Softmax layer calculates and outputs the probability of samples corresponding to different categories. The probability classification layer then identifies and outputs the corresponding vibration category. This trained model can thus determine whether severe drill string vibration is occurring during drilling using only surface drilling data. This technology enables end-to-end drill string vibration identification using only surface drilling data. It also effectively adapts to complex formation variations during actual drilling, effectively identifying severe drill string vibration, and improving the accuracy and robustness of drill string vibration identification, thereby ensuring drilling efficiency and safety. This technology lays a foundation for research in drilling process control, condition monitoring, and efficiency optimization, demonstrating its practicality and applicability. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0050] Figure 1 This is a flow chart of a method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network in an embodiment of the present invention;
[0051] Figure 2 is a structural diagram of a multi-scale temporal convolutional network in an embodiment of the present invention;
[0052] Figure 3 This is a visualization result diagram of FC layer feature activation during model training in an embodiment of the present invention;
[0053] Figure 4 This is a visualization result diagram of the Softmax layer feature activation during the model training process in an embodiment of the present invention;
[0054] Figure 5 This is a visualization result diagram of FC layer feature activation during the model verification process in an embodiment of the present invention;
[0055] Figure 6 This is a visualization result diagram of the Softmax layer feature activation during the model verification process in an embodiment of the present invention. DETAILED DESCRIPTION
[0056] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0057] The embodiment of the present invention provides a method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network. Figure 1 As shown, the following steps are included:
[0058] S1: The mutual information analysis method is used to calculate the mutual information value between the time series of each drilling process variable, and the process variable that is strongly correlated with the torque is selected as the characteristic variable to obtain the time series data S corresponding to the multivariate characteristic variable;
[0059] S2: Construct a drill string vibration recognition model, which includes a multi-scale feature extraction module and a drill string vibration recognition module. Time series data S is used as a sample and input into the model for model training.
[0060] The multi-scale feature extraction module uses a multi-scale temporal convolutional network with different dilated convolution factors as a feature extractor. The corresponding temporal convolutional network under a single dilated convolution factor extracts different single-scale vibration features respectively, which are finally fused into multi-scale features, and vibration features are obtained based on the multi-scale features.
[0061] In the drill string vibration recognition module, a one-dimensional global maximum pooling layer is first used to reduce the dimensionality of the vibration features to obtain high-level features. Then, a fully connected layer is used to learn and combine the high-level features and map them to the sample label space. Finally, a softmax layer is used to calculate and output the probability that the sample corresponds to different categories. The corresponding vibration category is identified and output based on the probability classification layer.
[0062] S3: Use the trained drill string vibration recognition model to identify the drill string vibration during the actual drilling process.
[0063] The mutual information analysis method is used to perform correlation analysis on drilling process variables and determine the input characteristic variables of the drill string vibration identification model. The specific process is as follows:
[0064] (1-1): V={X i |i=1,...,M} represents the set of drilling process variables, where X i represents the time series of the i-th process variable, M represents the total number of process variable types; the mutual information analysis method is used to perform correlation analysis on the drilling process variable V, and the calculation formula of the mutual information value is as follows:
[0065]
[0066] Where X and Y represent the time series of any two variables in V, p(x,y) represents the joint probability distribution of variables X and Y, and p(x) and p(y) represent the marginal probability distribution of variables X and Y, respectively.
[0067] (1-2): Sort the mutual information values in descending order, and select the process variable with the higher mutual information value with the torque as the preliminary input feature variable of the vibration recognition model. In this embodiment, the mutual information value greater than or equal to 0.5 is considered to be the higher mutual information value;
[0068] (1-3): Analyze the mutual information value between the preliminary input feature variables, eliminate the redundant feature variables with strong correlation, and determine the time series data S = {X i |i=1,...,N}, where N represents the number of types of input feature variables.
[0069] According to Figure 2 The time series data S for the multivariate feature variables shown in the figure has different representations of the drill string vibration characteristics under different formation conditions, and the single-scale time convolutional network has limited ability to capture and characterize vibration characteristics. Therefore, a multi-scale time convolutional network with different void factors is constructed to capture and extract the multi-scale features of the time series of the multivariate feature variables to characterize the drill string vibration, so as to adapt to complex formation changes and provide advanced feature representation for the subsequent effective identification of drill string vibration.
[0070] The specific process of constructing a multi-scale time convolutional network to extract multi-scale features of the multivariate feature variable time series to characterize the drill string vibration is as follows:
[0071] (2-1): In the actual drilling process, the drilling process variables are directly affected by the formation environment. The time series characteristics of the drilling process variables under different formation conditions are different, resulting in the limited ability of the single-scale time series characteristics to characterize the drill string vibration. In order to enhance the characterization ability of the drill string vibration characteristics under complex formation conditions, a time series with different void factors d is constructed. j (j=1,2,3) multi-scale temporal convolutional network to capture and extract multivariate feature variables S={X i |i=1,...,N}, and obtain the vibration characteristics of the drill string vibration; the operations and outputs of each network layer of the constructed multi-scale temporal convolutional network are expressed as follows:
[0072] Atrous causal convolution layer operation:
[0073]
[0074] Among them, F(s) represents the dilated convolution operation on element s in sequence S, f represents the convolution filter, f(i) represents the i-th convolution filter, k is the filter size, d is the dilation factor, and f d represents a convolution filter with a dilation factor of d, S s-d.i Represents the sequence of operations of sequence S on the i-th convolution filter, sd·i represents the past direction. The same layer has different hole factors d j Atrous Causal Convolutional Layer (DCC) j The output of (j=1,2,3) can be expressed as follows:
[0075] Y 1j =DCC j (S,W j )
[0076] Among them, Y 1j Indicates DCC j The intermediate features obtained by the layer, W j Indicates DCC j Layer trainable parameters.
[0077] Batch normalization layer BN output:
[0078] Y 2j =BN(Y 1j )
[0079] Among them, Y 2j Represents the intermediate features obtained by the BN layer.
[0080] Rectified linear unit layer ReLU output:
[0081] Y 3j =ReLU(Y 2j )=max(0,W BN Y 2j +b BN )
[0082] Among them, Y 3j Represents the intermediate features obtained by the ReLU layer, W BN and b BN Represents the trainable parameters of the ReLU layer.
[0083] (2-2): Based on different void factors d j Atrous Causal Convolutional Layer (DCC) j The single-scale features extracted by (j=1,2,3) and the fused multi-scale features are expressed as follows:
[0084]
[0085] Among them, Y4 represents multi-scale features, Y 3j(j=1,2,3) represents the void factor d j The single-scale feature obtained by the second ReLU layer, Represents the feature fusion operation.
[0086] (2-3): The residual connection operation is expressed as follows:
[0087] Y=Y4+WS
[0088] Among them, Y represents the final high-level features, W represents an additional one-dimensional convolution operation, and "+" represents the feature element addition operation.
[0089] The drill string vibration recognition module consists of a one-dimensional global maximum pooling layer (GMP), a fully connected layer (FC), a Softmax layer, and a classification layer to effectively identify mild and severe drill string vibrations. The GMP layer performs dimensionality reduction on the multi-scale vibration features captured and extracted by the multi-scale temporal convolutional network. The FC layer learns and combines the high-level features after dimensionality reduction processing by the GMP layer and maps them to the sample label space. The Softmax layer calculates the probability that a sample belongs to different categories based on the output vector of the FC layer. The classification layer outputs the corresponding vibration category based on the statistical probability.
[0090] The operations and outputs of each network layer in the vibration recognition process are shown as follows:
[0091] (3-1): The output of the one-dimensional global maximum pooling layer GMP is as follows:
[0092] Y GMP =GMP(Y)
[0093] Among them, Y GMP Indicates GMP layer output.
[0094] (3-2): The output of the fully connected layer FC is expressed as follows:
[0095] Y FC =FC(Y GMP )=W FC Y GMP +b FC
[0096] Among them, Y FC represents the FC layer output, W FC and b FC Represent the weight and bias parameters of the FC layer respectively.
[0097] (3-3): The output of the Softmax layer is expressed as follows:
[0098]
[0099] Among them, P iIndicates the output probability that the sample belongs to the i-th category, satisfying n represents the number of drill string vibration categories; ψ i Represents the output layer parameters.
[0100] In this implementation, the process variables selected in the drilling process are: well depth, drilling pressure, hook load, drilling speed, rotation speed, pump volume, riser pressure, torque, and drill pipe weight. The actual operation data of a certain engineering well is used as a specific object. The time series data of the process variables every three minutes constitute an experimental sample, and 213 mild vibration samples and 102 severe vibration samples are obtained. On this basis, the mutual information analysis method is first used to perform correlation analysis on the drilling process variables to determine the input feature variables of the drill string vibration identification model; then a multi-scale time convolution network is constructed as a feature extractor to capture and extract the multi-scale features of the time series of the determined input feature variables, and obtain a high-level feature representation of the drill string vibration that adapts to changes in complex formation conditions; finally, a drill string vibration identification module is constructed, and the high-level feature map is combined into the sample label space, the probability of the vibration category to which the sample belongs is calculated, and the vibration category corresponding to the maximum probability is output. The specific steps are as follows:
[0101] (1) Determination of model input characteristic variables
[0102] Due to the strong nonlinearity and strong coupling between drilling process variables, input feature redundancy occurs when all process variables are input into the model, which in turn reduces model recognition accuracy. Therefore, it is necessary to analyze the correlation between process variables and determine the appropriate process variables for input into the model. First, mutual information analysis was used to evaluate the degree of correlation between process variables. The results are shown in Table 1. Then, the mutual information values between torque and other process variables were sorted in descending order: well depth, drill pipe weight, hook load, bit pressure, drilling rate, rotational speed, standpipe pressure, and pump volume. Since bit pressure, rotational speed, and pump volume are drilling operational variables, and drilling rate is strongly correlated with torque and reflects formation conditions, they should be selected as input features. Finally, the mutual information values between well depth, drill pipe weight, hook load, and standpipe pressure were compared and analyzed, and hook load was selected as the input feature. Therefore, the final input feature variables for the vibration recognition model were determined to be torque, bit pressure, rotational speed, pump volume, drilling rate, and hook load.
[0103] Table 1 Normalized mutual information values
[0104]
[0105]
[0106] (2) Construct a multi-scale temporal convolutional network to extract multi-scale features of multivariate process variable time series
[0107] Based on the determination of the model input feature variables, a multi-scale temporal convolutional network (MTCN) was constructed as a feature extractor to capture and extract multi-scale features of the time series of multivariate process variables, thereby obtaining a high-level representation of drill string vibration. The constructed MTCN consists of two layers of dilated causal convolutional layers (DCC), a batch normalization layer (BN), a rectified linear unit (ReLU), a spatial dropout layer, and a residual connection formed by an additional one-dimensional convolutional layer. Within the same network layer, DCC layers with dilation factors of 1, 2, and 4 were designed to process the temporal input in parallel to extract temporal features at different scales, ultimately fusing them into multi-scale features. To improve model performance and robustness, a five-fold cross-validation method was used to optimize the parameters of the MTCN based on the F1 score, a performance evaluation metric. The resulting parameter settings are as follows: convolution filter size k is 8, the number of filters is 32, the minimum batch size is 8, and the spatial dropout factor is 0.09.
[0108] (3) Construct a vibration recognition module to classify multi-scale features and identify vibration categories
[0109] Based on the multi-scale vibration features extracted by a multi-scale temporal convolutional network, a vibration recognition module is constructed to perform drill string vibration classification and identification. The vibration recognition module consists of a one-dimensional global maximum pooling layer (GMP), a fully connected layer (FC), a softmax layer, and a classification layer. The GMP layer performs dimensionality reduction on the multi-scale vibration features. The FC layer combines the high-level features output by the GMP layer and maps them to the sample label space. The softmax layer calculates the probability of a sample belonging to different vibration categories based on the FC layer output vector. The classification layer classifies and identifies the corresponding drill string vibration category based on the statistical probability output by the softmax layer.
[0110] In order to visualize the feature activation of each network layer during model training and verification, the t-SNE visualization method is used to reduce the high-dimensional vibration features to two-dimensional features and perform sample-label space mapping on the sample distribution. Figures 3 to 6 As shown. Figure 3 and Figure 4 It can be seen that during the model training process, based on the extracted multi-scale features, the FC layer can well combine these features and map them to the sample label space, and the Softmax layer can effectively divide the corresponding vibration categories. Figure 5 and Figure 6 It can be seen that during the model verification process, each network layer obtains an increasingly accurate sample-to-label space mapping relationship by continuously learning and characterizing the drill string vibration characteristics. Finally, the Softmax layer realizes the effective classification and recognition of vibration categories.
[0111] Therefore, the constructed multiscale temporal convolutional network effectively extracts multiscale features of drill string vibration based on the time series of multivariate process variables, providing a reasonable and accurate high-level feature representation for subsequent drill string vibration identification. The constructed vibration identification model can adapt to the complex formation changes during actual drilling, effectively identify drill string vibration, and has good performance indicators, as shown in Table 2.
[0112] Table 2 Model evaluation indicators
[0113] method Accuracy Accuracy Recall F1 score Multi-scale Temporal Convolutional Networks 96.19％ 96.18％ 92.1％ 93.98％
[0114] A storage device stores instructions and data for implementing the method for identifying drill string vibration during drilling based on a multi-scale time convolutional network.
[0115] A device for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network, comprising: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network.
[0116] The beneficial effects of the present invention are as follows: the present invention adopts the mutual information analysis method to calculate the mutual information value between the time series of each drilling process variable, selects the process variable with strong correlation with the torque as the characteristic variable, obtains the time series data S corresponding to the multivariate characteristic variable, constructs the drill string vibration recognition model, takes the time series data S as the sample, inputs it into the model for model training; uses the multi-scale time convolution network with different void convolution factors as the feature extractor, extracts the effective multi-scale features and vibration features of the drill string vibration, adopts the one-dimensional global maximum pooling layer, performs feature dimensionality reduction processing on the vibration features, and obtains high-level features; then, adopts the fully connected layer, learns and combines the high-level features and maps them to the samples Label space; finally, the Softmax layer is used to calculate and output the probability of samples corresponding to different categories, and the corresponding vibration category is identified and output according to the probability classification layer. Therefore, the trained model can be used to determine whether severe drill string vibration occurs during the drilling process. This technology can achieve "end-to-end" drill string vibration recognition using only on-well drilling data. At the same time, it can effectively adapt to the complex formation changes in the actual drilling process, realize the effective identification of severe drill string vibration during the drilling process, improve the accuracy and robustness of drill string vibration recognition, thereby ensuring the efficiency and safety of the drilling process, and laying the foundation for research on drilling process control, condition monitoring and efficiency optimization. It is practical and applicable.
[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network, characterized by: The following steps are involved: S1: The mutual information analysis method is used to calculate the mutual information value between the time series of each drilling process variable, and the process variable that is strongly correlated with the torque is selected as the characteristic variable to obtain the time series data S corresponding to the multivariate characteristic variable; S2: Construct a drill string vibration recognition model, which includes a multi-scale feature extraction module and a drill string vibration recognition module. Time series data S is used as a sample and input into the model for model training. The specific process of extracting the multi-scale features of drill string vibration and obtaining vibration characteristics is as follows: (2-1): The feature extractor is constructed by a multi-scale temporal convolutional network, which consists of two layers with different dilation factors d j Atrous Causal Convolutional Layer (DCC) j , batch normalization layer BN, rectified linear unit ReLU, dropout layer and additional one-dimensional convolution layer constitute residual connection, j = 1, 2, 3, capture and extract multivariate feature variable time series data S = {X i |i=1,...,N}, and obtain the vibration characteristics of the drill string vibration; the operation output of each network layer in the multi-scale time convolution network is expressed as follows: Atrous causal convolution layer operation: Among them, F(s) represents the dilated convolution operation on element s in the time series data S, f(i) represents the i-th convolution filter, k is the filter size, d is the dilation factor, and f d represents a convolution filter with a dilation factor of d, S s-d . i Represents the sequence of operations of time series data S on the i-th convolution filter, with different hole factors d in the same layer j Atrous Causal Convolutional Layer (DCC) j The output can be expressed as follows: Y 1j =DCC j (S,W j ) Among them, Y 1j Indicates DCC j The intermediate features obtained by the layer, W j Indicates DCC j Layer trainable parameters; Batch normalization layer BN output: AND 2j =BN(Y 1j ) Among them, Y 2j Represents the intermediate features obtained by the BN layer; Rectified linear unit layer ReLU output: Y 3j =ReLU(Y 2j )=max(0,W BN Y 2j +b BN ) Among them, Y 3j Represents the intermediate features obtained by the ReLU layer, W BN and b BN Represents the trainable parameters of the ReLU layer; (2-2): Based on different void factors d j Atrous Causal Convolutional Layer (DCC) j The extracted single-scale features and the fused multi-scale features are expressed as follows: Among them, Y4 represents multi-scale features, Y 3j represents the void factor d j The single-scale features obtained by the second ReLU layer, j = 1, 2, 3, Represents feature fusion operation; (2-3): The residual connection operation is expressed as follows: Y=Y4+WS Where Y represents the vibration feature, W represents an additional one-dimensional convolution operation, and "+" represents the addition operation of feature elements; The multi-scale feature extraction module uses a multi-scale temporal convolutional network with different dilated convolution factors as a feature extractor. The corresponding temporal convolutional network under a single dilated convolution factor extracts different single-scale vibration features respectively, which are finally fused into multi-scale features, and vibration features are obtained based on the multi-scale features. In the drill string vibration recognition module, a one-dimensional global maximum pooling layer is first used to reduce the dimensionality of the vibration features to obtain high-level features. Then, a fully connected layer is used to learn and combine these high-level features and map them to the sample label space. Finally, a softmax layer is used to calculate and output the probability that the sample corresponds to different categories. The probability classification layer then identifies and outputs the corresponding vibration category. S3: Use the trained drill string vibration recognition model to identify the drill string vibration during the actual drilling process.
2. The method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network according to claim 1, wherein: In step S1, the specific process of determining the input characteristic variables of the drill string vibration identification model is as follows: (1-1): V={X i |i=1,...,M} represents the set of drilling process variables, where X i represents the time series of the i-th process variable, M represents the total number of process variable types; the mutual information value calculation formula shown below is used to perform correlation analysis on the drilling process variable V: Where X and Y represent the time series of any two variables in V, p(x,y) represents the joint probability distribution of variables X and Y, and p(x) and p(y) represent the marginal probability distribution of variables X and Y, respectively. (1-2): Sort the mutual information values calculated in step (1-1) in descending order, and select the process variables with higher mutual information values with torque as the preliminary input feature variables of the drill string vibration identification model; (1-3): Analyze the mutual information value between the characteristic variables in step (1-2), eliminate the redundant characteristic variables with strong correlation, and determine the time series data S = {X i |i=1,...,N}, where N represents the number of types of feature variables.
3. The method for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network according to claim 1, wherein: In step S3, the process of identifying drill string vibration is as follows: (3-1): Use the one-dimensional global maximum pooling layer to perform feature dimensionality reduction on the vibration features. The output of the one-dimensional global maximum pooling layer GMP is as follows: AND GMP =GMP(Y) Among them, Y GMP Represents the output of the GMP layer, i.e., high-level features, and GMP() represents a one-dimensional global maximum pooling layer; (3-2): The output of the fully connected layer FC is expressed as follows: Y FC =FC(Y GMP )=W FC Y GMP +b FC Among them, Y FC represents the FC layer output, FC() represents the fully connected layer, W FC and b FC Represent the weight and bias parameters of the FC layer respectively; (3-3): The output of the Softmax layer is expressed as follows: Among them, P i Indicates the output probability that the sample belongs to the i-th category, satisfying n represents the number of drill string vibration categories; softmax() represents the Softmax layer; ψ i Represents the output layer parameters.
4. A storage device, characterized in that: The storage device stores instructions and data for implementing the method for identifying drill string vibration during drilling based on a multi-scale time convolutional network as described in any one of claims 1 to 3.
5. A device for identifying drill string vibration during drilling based on a multi-scale temporal convolutional network, characterized by: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the method for identifying drill string vibration during drilling based on a multi-scale time convolutional network as described in any one of claims 1 to 3.