A drilling string vibration identification and regulation method based on multi-modal data fusion

Abstract

The application discloses a kind of drilling string vibration identification and regulation and control method based on multi-modal data fusion, belong to the intersection field of petroleum drilling and artificial intelligence.The method utilizes the triaxial vibration acceleration signal of downhole drill string and its corresponding three-channel time-frequency diagram, constructs the multi-modal input of fusion space and time-frequency information.Through the deep neural network of BiGRU and ResNet18-SENet combination, attention mechanism is fused, and the accurate identification of complex vibration modes such as general vibration, stick-slip, whirling, etc.is realized.Further combined with the identification result, the adaptive regulation and control of drilling parameters is driven, and the closed-loop control of drilling string vibration state is realized.The application has the advantages of high identification precision, strong robustness, intelligent regulation and control, etc., and is suitable for downhole vibration monitoring and drilling parameter optimization under complex formation environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of oil drilling and artificial intelligence, and particularly relates to a drilling string vibration identification and regulation method based on multi-modal data fusion. BACKGROUND

[0002] Drilling string vibration is coupled by multiple factors such as bit-rock interaction, formation heterogeneity and wellbore constraint, resulting in axial, lateral and torsional vibrations, and showing strong nonlinear and non-stationary characteristics such as stick-slip and whirling. Excessive vibration may lead to drilling tool failure, and different vibration modes correspond to different regulation strategies, so accurate identification of vibration mode is particularly important.

[0003] In recent years, deep learning has shown great potential in vibration analysis, but existing methods mostly rely on single time series signal or time-frequency diagram modeling, and cannot fully utilize the complementary advantages and cross-modal time series dependence of multi-modal data, so the recognition accuracy and adaptability are still limited.

[0004] Therefore, the application proposes to input vibration acceleration signals and their corresponding time-frequency diagrams into a double-channel feature extraction network constructed by joint BiGRU and ResNet18-SENet synchronously, and to fuse multi-dimensional information by using an attention mechanism. This method can accurately identify various fault modes such as general vibration, stick-slip and whirling by simultaneously capturing the spatial coupling characteristics and time-frequency energy of vibration, and automatically adjust drilling parameters based on the identification results to form a closed-loop regulation strategy. SUMMARY

[0005] In order to solve the problems of low recognition accuracy and poor working condition adaptability of the traditional "single-channel-single-modal" drilling string vibration identification method, the application proposes a drilling string vibration identification and regulation method based on multi-modal data fusion. This method integrates data perception, feature extraction, intelligent decision-making and parameter regulation, realizes multi-modal synchronous acquisition through "three-axis acceleration + three-channel time-frequency diagram", constructs a double-channel feature extraction network by joint BiGRU and ResNet18-SENet, and introduces an attention fusion mechanism to comprehensively model the time series and spatial vibration characteristics of the drilling string. This method realizes vibration type identification (such as general vibration, whirling and stick-slip) and severity evaluation according to the fused features, and generates corresponding speed and weight regulation instructions combined with threshold strategy. Finally, this method can be connected with downhole or ground control modules to construct an identification-evaluation-regulation closed-loop system, realize the identification of vibration state under complex working conditions and adaptive optimization of parameters.

[0006] The technical scheme of the application is as follows:

[0007] A drilling string vibration identification and regulation method based on multi-modal data fusion, comprising the following steps:

[0008] Step 1, synchronously collect multi-source data to form original three-dimensional time series signals, and segment them equally;

[0009] Step 2, data processing of original three-dimensional time series signals;

[0010] Step 3, dual-channel feature extraction; in the time series channel, a bidirectional gated recurrent unit network is used as a one-dimensional feature extractor for time series data to extract time series modal features; in the image channel, a ResNet18-SENet network is used as a two-dimensional feature extractor to extract image modal features;

[0011] Step 4, concatenate the time series modal features and the image modal features, and simultaneously use attention weights to adaptively weight the concatenated features to obtain weighted fusion features;

[0012] Step 5, identify the vibration type based on the weighted fusion features;

[0013] Step 6, according to the identified vibration type and real-time characteristic value, judge the risk level, and generate drilling parameter adjustment instructions, and the control instructions are issued through mud pulse or field bus to complete vibration control.

[0014] Further, the specific process of step 1 is:

[0015] Step 1.1, integrate a three-axis vibration sensor in the short section near the drill bit of the drill string, and real-time acquire X-axis, Y-axis and Z-axis three-axis vibration acceleration, and synchronously record the speed, torque working condition parameters to form original three-dimensional time series signals ;

[0016] Step 1.2, use a non-overlapping sliding window method to equally segment the original three-dimensional time series signals to obtain original three-dimensional time series signals containing three-axis vibration signals of the drill string: , wherein, is the segment length; the number 3 corresponds to three-axis vibration; is the three-axis vibration signal, and each three-axis vibration signal represents a three-dimensional time series data with a length of .

[0017] Further, the specific process of step 2 is:

[0018] Step 2.1, independently execute the noise reduction process for each-axis vibration signal;

[0019] Step 2.2, convert each segment of the denoised vibration signal into a two-dimensional time-frequency image based on short-time Fourier transform.

[0020] Further, the specific process of step 2.1 is:

[0021] Step 2.1.1, using a symmetric 5th-order wavelet basis to the current shaft vibration signal Seven-layer discrete wavelet decomposition is performed to obtain a wavelet coefficient set , wherein is the low-frequency approximation component corresponding to the first layer, ; is the high-frequency detail component corresponding to the first layer; the noise standard deviation is estimated using the high-frequency detail components of the first two layers , and the specific calculation is as follows:

[0022] (1);

[0023] , wherein denotes the median operation; , are the high-frequency detail components corresponding to the first layer and the second layer, respectively; 0.6745 is the correction factor under normal distribution;

[0024] Step 2.1.2, according to the high-frequency detail component corresponding to each layer, the adaptive threshold of each layer is calculated, and the soft threshold function is applied for compression processing;

[0025] The calculation formula of the adaptive threshold is as follows:

[0026] (2);

[0027] , wherein is the adaptive threshold corresponding to the first layer; is the length of the vibration signal;

[0028] The calculation formula of the soft threshold function compression processing is as follows:

[0029] (3);

[0030] , wherein is the soft threshold function, and the calculated value is the high-frequency detail component corresponding to the first layer after compression processing;

[0031] Step 2.1.3, inverse wavelet transform is performed on the wavelet coefficient set after compression processing to complete the reconstruction of the current shaft vibration signal, and obtain the one-dimensional vibration signal after noise reduction ;

[0032] Step 2.1.4, after the three-axis vibration signals are processed according to the noise reduction procedures of steps 2.1.1-2.1.3, they are recombined in the original axial order to obtain the three-dimensional time series signal after noise reduction .

[0033] Further, the specific process of step 2.2 is as follows:

[0034] Step 2.2.1, for each segment of the denoised one-dimensional vibration signal respectively, a short-time Fourier transform is performed to obtain the denoised one-dimensional vibration signal The corresponding time-frequency domain matrix The time-frequency domain matrix is a complex matrix composed of all frames and frequency points; the short-time Fourier transform process is as follows:

[0035] (4) ;

[0036] wherein, is the complex spectrum of the time-frequency domain matrix frame frequency point; is the serial number of the denoised one-dimensional vibration signal; is the window length; indicates the window function; is the step length;

[0037] Step 2.2.2, taking the logarithm of the amplitude modulus length of the time-frequency domain matrix as the image intensity to form a two-dimensional time-frequency image:

[0038] (5) ;

[0039] wherein, is the two-dimensional time-frequency image; is a small constant to prevent the divergence of the logarithm;

[0040] Step 2.2.3, independently process the three-axis vibration signals according to the process of step 2.2.1-step 2.2.2 to obtain three time-frequency spectrum images , and , respectively corresponding to the two-dimensional time-frequency spectrum images of X-axis, Y-axis and Z-axis;

[0041] Step 2.2.4, vertically splice the three time-frequency spectrum images in order to form a complete vibration time-frequency image :

[0042] (6) ;

[0043] wherein, indicates the length of the time dimension, is the length of the frequency dimension.

[0044] Further, the specific process of step 3 is as follows:​

[0045] Step 3.1, defining time steps The corresponding multi-dimensional signal is ; wherein, , , are three-axis vibration signals in X-axis, Y-axis and Z-axis respectively; the multi-dimensional signal corresponding to each time step is taken as the input of the bidirectional gated recurrent unit network;

[0046] The bidirectional gated recurrent unit network is composed of two independent unidirectional gated recurrent units which do not share parameters; the forward gated recurrent unit sequentially scans from to to generate the forward hidden state sequence of time step ; The backward gated recurrent unit inversely scans from to to generate the backward hidden state sequence of time step ; Each unidirectional gated recurrent unit controls the retention and forgetting of historical information through update gate and reset gate; the state update calculation of the unidirectional gated recurrent unit at time step is as follows:

[0047] (7);

[0048] (8);

[0049] (9);

[0050] (10);

[0051] wherein, , and are the update gate, reset gate and hidden gate states of time step ; is the hidden state of time step ; is the candidate hidden state of time step ; is the weight matrix of the input to the update gate ; is the weight matrix of the input to the reset gate ; is the weight matrix of the input to the candidate hidden state ; is the hidden state To the update gate The weight matrix; Hidden state To reset door The weight matrix; Hidden state go through After control, proceed to the candidate hidden state The weight matrix; It is the Sigmoid activation function; tanh is the hyperbolic tangent activation function. This represents element-wise multiplication;

[0052] The forward hidden state and backward hidden state of each time step are concatenated to form a bidirectional hidden state vector, which is the time step. Hidden door status Specifically: ,in, The hidden layer dimension of the unidirectional gated loop unit;

[0053] By applying pooling operations to the bidirectional hidden state vector, the variable-length temporal feature sequence is compressed into a fixed-length feature vector, thus obtaining the temporal modal features. ;

[0054] Step 3.2: Extract image modal features using the ResNet18-SENet network; the ResNet18-SENet network consists of two parts: the ResNet18 residual module and the SENet module.

[0055] Furthermore, the specific process of step 3.2 is as follows:

[0056] Step 3.2.1: Obtain the vibration time-frequency image The input is fed into the ResNet18-SENet network; firstly, a 7×7 convolutional layer is used for the first stage of convolution output, with the following formula:

[0057] (11);

[0058] in, This represents the output of the first stage of convolution; For the convolution output at the th The sample, the first Each channel, location The value; For convolution kernel; The convolution operation outputs the position index in the matrix. Vertical direction In the horizontal direction; and These are the height and width of the convolution kernel, respectively; and These are the indices of the convolution kernel in the horizontal and vertical directions; For the first The bias term for each channel; This is a convolution operation;

[0059] The first-stage convolution output is batch normalized, and the calculation formula for each channel is as follows:

[0060] (12);

[0061] (13);

[0062] (14);

[0063] in, and The first Mean and variance of each channel; and All are the first Learnable parameters for each channel; It is a small constant that prevents division by zero; Batch size; This is the normalized and linearly transformed value of the first-stage convolution output;

[0064] Using the ReLU activation function on the normalized output enhances its nonlinear expressiveness:

[0065] (15);

[0066] in, This is the result of the first stage convolution output after ReLU activation;

[0067] Then, max pooling is used to reduce the feature map space size:

[0068] (16);

[0069] in, This is the feature map after pooling; This is the first phase of the max pooling operation; For pooled windows;

[0070] Step 3.2.2: Pool the feature map The first residual block includes two first residual units connected in series.

[0071] The first residual unit and the second residual unit both adopt a standard unit structure, sequentially performing 3x3 convolution, first batch normalization, a ReLU activation function, 3x3 convolution, and second batch normalization to form a convolution mapping output.

[0072] (17);

[0073] wherein, is the convolution mapping output. is the input of the residual unit, the input of the first residual unit in the first residual block is the pooled feature map , and the input of other residual units is the output of the previous residual unit. is the second batch normalization operation. is the 3x3 convolution. is the first batch normalization operation.

[0074] The skip connection of the first residual unit is an identity mapping.

[0075] (18);

[0076] wherein, is the skip connection mapping.

[0077] The second residual unit adds an additional 1x1 convolution in the skip connection for matching the number of channels or implementing down-sampling.

[0078] (19);

[0079] wherein, is the 1x1 convolution.

[0080] The output of each residual unit is

[0081] (20);

[0082] The output of each residual unit is subjected to channel attention re-labeling via an SENet module. The SENet module includes two stages: compression and excitation.

[0083] The compression stage performs global average pooling on each channel. ​

[0084] (twenty one);

[0085] in, To indicate the first The global average pooling value for each channel. , This represents the number of channels in the convolutional network. This is a global average pooling operation; The output of the residual unit is the first Feature maps corresponding to each channel;

[0086] The vector obtained during the compression stage will then be... Two fully connected layers are input for excitation. A bottleneck structure is adopted, first reducing the dimensionality and then increasing it, and channel weights are generated. :

[0087] (twenty two);

[0088] in, and Here are the learnable weight matrices for the two fully connected layers; and These are the bias vectors for the two fully connected layers;

[0089] Finally, channel weighting is used to multiply the channel weights obtained in the activation phase back into the original features channel by channel, thereby achieving adaptive recalibration of the feature map:

[0090] (twenty three);

[0091] in, For the first channel attention recalibration Feature maps of each channel; Weighted operation for the channel; For the first Channel weights for each channel;

[0092] After four residual blocks, the spatial dimensions of the feature map are compressed into tensor values. ;in, , All are spatial dimensions. Number of channels;

[0093] Step 3.2.3: Next, global average pooling is used to compress each channel:

[0094] (twenty four);

[0095] in, For the first Each channel is located in The tensor value; For the first The compressed feature values ​​of each channel;

[0096] The compressed feature values ​​from all channels are concatenated to obtain the global average pooling feature. :

[0097] (25);

[0098] Step 3.2.4: Obtain the final image modal features through projection onto the fully connected layer. :

[0099] (26);

[0100] in, It is a fully connected layer.

[0101] Furthermore, the specific process of step 4 is as follows:

[0102] Temporal modal features Image modal features Perform a stitching operation to generate fused features. :

[0103] (27);

[0104] Introducing a self-attention mechanism to calculate attention weights:

[0105] (28);

[0106] (29);

[0107] in, The feature vector after attention mapping; It is the ReLU activation function; , All are learnable parameter matrices; , It is a learnable bias vector; This represents the attention weights of the fused features across different dimensions;

[0108] Attention weight Multiply the feature by the fusion feature to complete the weighted fusion, and obtain the final weighted fused feature. :

[0109] (30).

[0110] Further, in step 5, the final weighted fusion features are sent to a fully connected layer, and a Softmax activation function is used for nonlinear mapping, and the corresponding vibration state classification result is output:

[0111] (31);

[0112] wherein, , are hierarchical parameters; represent the probability distribution belonging to the vibration state class; the class corresponding to the highest probability is taken as the vibration type of the drill string under the current drilling state, and the vibration type includes general vibration, whirling vibration and stick-slip vibration.

[0113] Further, in step 6, the real-time feature values include acceleration mean value, peak difference and stick-slip ratio;

[0114] The general vibration is a vibration without risk, and no treatment is needed, and continuous drilling and monitoring of the drill string vibration state can be performed;

[0115] For the whirling vibration, the acceleration mean value range and the peak difference of the three-dimensional acceleration signal in a specific time window are extracted as the judgment basis; when the acceleration mean value is 1-2g and the peak difference is 1-5g, it is determined as mild whirling; when the acceleration mean value is 3-5g and the peak difference is 5-10g, it is determined as moderate whirling; and when the acceleration mean value is greater than 5g and the peak difference is greater than 10g, it is determined as severe whirling;

[0116] For the stick-slip vibration, the stick-slip ratio is used as the risk judgment index; when the stick-slip ratio is less than 0.5, it is determined as low-level stick-slip; between 0.5 and 1.0 is medium-level stick-slip; 1.0 to 1.5 is high-level stick-slip; and more than 1.5 is severe stick-slip;

[0117] According to the evaluation result, a regulation suggestion is given: when the moderate or severe whirling is suggested to reduce the rotating speed and increase the drilling pressure, when the severe whirling exceeds the limit, the drilling is stopped; when the high-level or severe stick-slip is suggested to increase the rotating speed and reduce the drilling pressure, when the severe stick-slip exceeds the limit, the drilling is stopped.

[0118] The beneficial technical effects brought by the present application: the present application introduces multi-modal data fusion and double-channel deep network structure, fully combines the advantages of time sequence and time-frequency characteristics, realizes high-precision identification and severe grade evaluation of the drill string vibration mode, and forms a closed-loop regulation strategy on this basis, which can automatically optimize the key parameters such as rotating speed and drilling pressure, not only effectively improves the accuracy and working condition adaptability of vibration monitoring, but also reduces the failure risk of drilling tools and improves the drilling efficiency, has good engineering application value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0119] Figure 1 This is a flowchart of the drill string vibration identification and control method based on multimodal data fusion according to the present invention.

[0120] Figure 2 This is a schematic diagram of the drill string vibration timing feature extractor in this invention.

[0121] Figure 3 This is a schematic diagram of the drill string vibration time-frequency feature extractor in this invention. Detailed Implementation

[0122] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0123] like Figure 1 As shown, to achieve accurate identification and real-time control of drill string vibration, this invention proposes a multimodal data fusion method based on a "three-axis, three-channel, dual-channel network, and closed-loop decision-making" architecture. This method primarily constructs a fusion model, ResNet18SENet-BiGRUG, and proposes a vibration control strategy. The fusion model includes a data input layer, a data processing layer, a feature extraction layer, a feature fusion layer, and a classification output layer. Intelligent vibration identification is performed through the fusion model, and vibration control is implemented based on the identification results. The specific steps are as follows:

[0124] Step 1: In the data input layer, multi-source data are synchronously acquired to form the original three-dimensional time-series signal, and then segmented into equal-length segments; the specific process is as follows:

[0125] Step 1.1: Integrate a triaxial vibration sensor within the drill string near the drill bit section to acquire real-time downhole vibration accelerations in three directions: X-axis (lateral 1, one of the vibration components perpendicular to the wellbore direction), Y-axis (lateral 2, another vibration component orthogonal to the X-axis), and Z-axis (axial, the vibration component in the wellbore direction). Simultaneously record operating parameters such as rotational speed and torque to form the original three-dimensional time-series signal. Specifically, the vibration sensor can be a MEMS accelerometer, which is an accelerometer manufactured using Micro-Electro-Mechanical Systems (MEMS) technology.

[0126] Step 1.2: The original three-dimensional time series signal is segmented into equal-length segments using a non-overlapping sliding window method to obtain the signal containing... The original three-dimensional time-series signal of the triaxial vibration signal of the drill string segment: Set the segment length to That is, the length of each segment of triaxial vibration signal is For example, under a sampling rate of 120Hz, It can be 2400 (i.e. corresponding to 20 seconds signal). Among them, is the first segment three-axis vibration signal, each segment three-axis vibration signal represents a segment of three-dimensional time series data with a length of ; the number 3 corresponds to three-axis vibration.

[0127] Step 2, in the data processing layer, the original three-dimensional time series signal is processed: sym5 is independently executed for each axis to perform seven-layer discrete wavelet decomposition and soft threshold denoising, and finally the denoised signal is reconstructed through inverse wavelet transform, and then the denoised three-axis signal is converted into a time-frequency graph through short-time Fourier transform (STFT), realizing time series-image modal alignment; the specific process is as follows:

[0128] Step 2.1, in order to ensure that the three-way information does not interfere with each other, the denoising process is independently executed for each axis vibration signal. The specific process is as follows:

[0129] Step 2.1.1, first, the Symlet 5 wavelet basis (sym5 wavelet basis) is used to perform seven-layer discrete wavelet decomposition on each axis vibration signal , to obtain a set of wavelet coefficients , wherein is the low-frequency approximate component corresponding to the first layer, ; is the high-frequency detail component corresponding to the first layer, which contains high-frequency fluctuations and potential noise in the signal. In order to estimate the noise level, the noise standard deviation is estimated using the high-frequency detail components of the first two layers, and the specific calculation is as follows:

[0130] (1);

[0131] wherein, indicates the median operation; , are the high-frequency detail components corresponding to the first layer and the second layer respectively; 0.6745 is the correction factor under normal distribution.

[0132] Step 2.1.2, after obtaining the noise estimate, the adaptive threshold of each layer is calculated according to the high-frequency detail component corresponding to each layer, and the soft threshold function is applied for compression processing; this method can maximize the preservation of the effective details of the signal while suppressing high-frequency noise. The low-frequency approximate component is not processed, so as to preserve the main energy and overall trend of the signal.

[0133] The calculation formula of the adaptive threshold is as follows:

[0134] (2);

[0135] wherein, is the adaptive threshold value corresponding to the vibration signal length;

[0136] The calculation formula of the soft threshold function compression processing is:

[0137] (3);

[0138] wherein, is the soft threshold function, and the calculated value is the high-frequency detail component corresponding to the compressed wavelet coefficient set;

[0139] Step 2.1.3, then, the compressed wavelet coefficient set is input into the inverse wavelet transform to complete the reconstruction of the axial vibration signal, and a one-dimensional vibration signal after noise reduction is obtained.

[0140] Step 2.1.4, after the three-axis vibration signals complete the above noise reduction process in turn, they are reorganized into a three-dimensional time sequence signal after noise reduction in the original axial order. .

[0141] Step 2.2, in order to extract more abundant vibration information and support subsequent image-based deep feature modeling, this step converts each vibration signal after noise reduction into a two-dimensional time-frequency image. The specific process is as follows:

[0142] Step 2.2.1, the one-dimensional vibration signal after noise reduction is respectively subjected to short-time Fourier transform (STFT) to obtain the time-frequency domain matrix corresponding to the one-dimensional vibration signal after noise reduction . The time-frequency domain matrix is a complex matrix composed of all frames and frequency points. The STFT processing procedure is as follows:

[0143] (4);

[0144] wherein, is the complex spectrum of the frame and the frequency point in the time-frequency domain matrix ; represents the frame index, which controls the starting position of the window; is the frequency point index, which determines the frequency spectrum resolution; is the serial number of the one-dimensional vibration signal after noise reduction; is the window length; represents the one-dimensional vibration signal after noise reduction, is the a one-dimensional vibration signal after noise reduction; represents a window function, and a Hamming window is used here; is a step size, and the formula is , is an overlap rate.

[0145] Step 2.2.2, after obtaining the time-frequency domain matrix representation, the logarithm of the amplitude modulus length is taken as the image intensity to form a two-dimensional time-frequency image:

[0146] (5) ;

[0147] wherein, is a two-dimensional time-frequency image; is a small constant to prevent the divergence of the logarithm.

[0148] Step 2.2.3, according to the process of step 2.2.1-step 2.2.2, the three-axis vibration signals are independently processed to obtain three time-frequency spectrum images , and , corresponding to the two-dimensional time-frequency spectrum images of X-axis, Y-axis and Z-axis respectively. Each image is a two-dimensional array, and each image has a size of two-dimensional image, wherein represents the length of the time dimension, is the length of the frequency dimension.

[0149] Step 2.2.4, in order to integrate the three-axis information into a unified input form, the three time-frequency spectrum images are combined in order by vertical stacking to form a complete vibration time-frequency image :

[0150] (6) ;

[0151] After signal noise reduction and time-frequency conversion processing, the original three-axis vibration data is represented as two complementary data modalities: three-dimensional vibration time series signal (time series data) and its corresponding vibration time-frequency image .

[0152] Step 3, in the feature extraction layer, double-channel feature extraction is performed: in the time series channel, a bidirectional gated recurrent unit network (BiGRU) is used as a one-dimensional feature extractor for time series data to extract the time series features of the vibration signal (time series modal features); in the image channel, a ResNet18-SENet network is used as a two-dimensional feature extractor to extract the time series features of the time-frequency image (image modal features). The specific process is as follows:

[0153] Step 3.1, as Figure 2As shown, in the modeling process, the multidimensional signal corresponding to each time step is first formed as the input of the BiGRU, and the time step is defined. The corresponding multidimensional signal (triaxial acceleration) is ;in, , , The signals are triaxial vibration signals in the X, Y, and Z axes, respectively.

[0154] BiGRU consists of two independent unidirectional GRUs (gated recurrent units) with non-shared parameters; in the forward layer, the forward GRU starts from... arrive Sequential scanning to generate time steps Forward hidden state sequence In the reverse layer, the backward GRU from arrive Reverse scan to generate time steps Backward hidden state sequence Each one-way GRU internally regulates the retention and forgetting of historical information through update and reset gates. The one-way GRU operates at time steps... The state update is calculated as follows:

[0155] (7);

[0156] (8);

[0157] (9);

[0158] (10);

[0159] in, , and They are time steps Update the status of doors, reset doors, and hidden doors; It is a time step The hidden state, also known as the hidden state of the previous time step; For time step The candidate hidden state; For input To the update gate The weight matrix; For input To reset door The weight matrix; For input To candidate hidden state The weight matrix; Hidden state To the update gate The weight matrix; Hidden state To reset door The weight matrix; Hidden state go through After control, proceed to the candidate hidden state The weight matrix; It is the Sigmoid activation function; tanh is the hyperbolic tangent activation function. This indicates element-wise multiplication.

[0160] Figure 2 The “1-” in formula (10) corresponds to the “ "1-".

[0161] In the output layer, the forward hidden state and the backward hidden state of each time step are concatenated to form a bidirectional hidden state vector, which is the time step's hidden state vector. Hidden door status Specifically: ,in, The hidden layer dimension of a unidirectional GRU.

[0162] In summary, by applying pooling operations to the hidden states of this sequence, the variable-length temporal feature sequence is compressed into a fixed-length feature vector, i.e., temporal modal features. This serves as the temporal input for subsequent feature fusion.

[0163] Step 3.2: To achieve multimodal feature fusion, in addition to temporal features, the ResNet18-SENet network is also needed to extract image modality information. The ResNet18-SENet network mainly consists of two parts: ResNet18 and SENet. ResNet18 is the basic network architecture, which solves the gradient vanishing problem in deep network training through residual learning; SENet explicitly models the dependencies between feature channels through a feature recalibration strategy, improving the model's ability to identify important features. Figure 3 As shown, the data processing procedure of the ResNet18-SENet network is as follows:

[0164] Step 3.2.1, and triaxial acceleration Corresponding vibration time-frequency image The three-channel time-frequency image is fed into the ResNet18-SENet network. First, a standard 7×7 convolutional layer is used for the first stage of convolution output to extract initial local features, the mathematical form of which is:

[0165] (11);

[0166] in, This represents the output of the first stage of convolution; For the convolution output at the th The sample, the first Each channel, location The value; For convolution kernel; Vertical direction (row coordinates); Horizontal direction (column coordinates); The convolution operation outputs the position index in the matrix; and These are the height and width of the convolution kernel, respectively; and These are the indices of the convolution kernel in the horizontal and vertical directions; For the first The bias term for each channel; This is a convolution operation.

[0167] To accelerate training and stabilize the network, the first-stage convolutional output is batch normalized (BN), and the calculation formula for each channel is as follows:

[0168] (12);

[0169] (13);

[0170] (14);

[0171] in, and The first Mean and variance of each channel; and All are the first Learnable parameters for each channel; It is a small constant that prevents division by zero; This refers to the batch size. This is the normalized and linearly transformed value of the first-stage convolution output.

[0172] Applying the ReLU activation function to the normalized output enhances its nonlinear expressiveness and improves the model's ability to represent complex vibration modes.

[0173] (15);

[0174] in, This is the result of the first stage convolution output after ReLU activation.

[0175] Then, max pooling is used to reduce the feature map space size and extract the most salient local features:

[0176] (16);

[0177] in, This is the feature map after pooling; This is the first phase of the max pooling operation; For pooled windows;

[0178] Step 3.2.2: Next, the pooled feature map... The data is fed into a ResNet18 residual module for deep feature extraction. This residual module consists of four residual blocks: a first residual block, a second residual block, a third residual block, and a fourth residual block, which are connected sequentially. Each residual block is composed of several residual units. In this invention, each residual block consists of two residual blocks. The first residual block includes two cascaded first residual units. The second, third, and fourth residual blocks have the same structure, each including cascaded second residual units and first residual units. Each residual unit uses a combination of convolution, batch normalization (BN), ReLU, and skip connections to form a residual network to alleviate the gradient vanishing problem.

[0179] The first residual unit adopts a standard unit structure, and sequentially performs 3×3 convolution, first batch normalization, ReLU activation function, 3×3 convolution, and second batch normalization to form a convolutional mapping output:

[0180] (17);

[0181] Its jump connection is an identity mapping:

[0182] (18);

[0183] in, The output is a convolutional mapping; The input to the residual unit is the pooled feature map. The input to other residual units is the output of the previous residual unit; This is the second batch normalization operation; It is a 3×3 convolution; This is the first batch of normalization operations;

[0184] The second residual unit (with a 1×1 convolutional unit) also uses the standard unit base, but it adds an extra 1×1 convolution in the skip connections to match the number of channels or to achieve downsampling:

[0185] (19);

[0186] in, Mapping for skip connections; It is a 1×1 convolution;

[0187] Each residual unit outputs for:

[0188] (20);

[0189] The output of each residual unit is recalibrated for channel attention via the SENet module, which includes two stages: squeezing and excitation.

[0190] The Squeeze phase performs global average pooling on each channel:

[0191] (twenty one);

[0192] in, To indicate the first The global average pooling value for each channel. , This represents the number of channels in the convolutional network. This is a global average pooling operation; The output of the residual unit is the first Feature maps corresponding to each channel.

[0193] The vector obtained in the Squeeze phase is then... Excitation is performed on two fully connected layers using a bottleneck structure, first reducing dimensionality and then increasing it, and channel weights are generated. :

[0194] (twenty two);

[0195] in, and Here are the learnable weight matrices for two fully connected layers. , , This refers to the compression ratio; and These are the bias vectors for the two fully connected layers. , .

[0196] Finally, the Scale (channel weighting) operation is used to multiply the channel weights obtained in the activation phase back into the original features channel by channel, thereby achieving adaptive recalibration of the feature map:

[0197] (twenty three);

[0198] in, For the first channel attention recalibration Feature maps of each channel; Weighted operation for the channel; For the first Channel weights for each channel;

[0199] After four residual blocks, the spatial dimensions of the feature map are compressed into tensor values. ;in, , All are spatial dimensions. In this invention, ResNet18 corresponds to... , Both are 7; The number of channels depends on the number of output channels of the last residual. In this invention, the number of channels corresponding to ResNet18 is... It is 512.

[0200] Step 3.2.3: Next, use Global Average Pooling (GAP) to compress each channel:

[0201] (twenty four);

[0202] in, For the first Each channel is located in The tensor value; For the first The feature values ​​after compression of each channel.

[0203] By concatenating the compressed feature values ​​from all channels, we can obtain the global average pooling feature. :

[0204] (25);

[0205] Step 3.2.4: Subsequently, the final image modal features are obtained through projection onto the fully connected layer. :

[0206] (26);

[0207] in, It is a fully connected layer.

[0208] Step 4: In the feature fusion layer, feature fusion is performed. To fuse data representations from multiple modalities, temporal modal features and image modal features are concatenated. Simultaneously, attention weights are used to adaptively weight the concatenated features, resulting in weighted fused features, achieving spatial-frequency-condition complementarity. The specific process is as follows:

[0209] Temporal modal features Image modal features Perform a concatenation operation to construct a joint representation and generate fused features. :

[0210] (27);

[0211] To guide the model to focus more on important features, a self-attention mechanism is further introduced to calculate attention weights:

[0212] (28);

[0213] (29);

[0214] in, The feature vector after attention mapping; It is the ReLU activation function; , All are learnable parameter matrices; , It is a learnable bias vector; This represents the attention weights of the fused features across different dimensions. Attention weights Then, multiply the weighted feature by the fusion feature to complete the weighted fusion, resulting in the final weighted fused feature. :

[0215] (30);

[0216] Step 5: In the classification output layer, perform vibration classification; feed the final weighted fused features into a fully connected layer (or multilayer perceptron), and use the Softmax activation function for nonlinear mapping to output the corresponding vibration state classification result:

[0217] (31);

[0218] in, , All are layered parameters; Indicates belonging to Probability distribution of vibration-like states.

[0219] The category with the highest probability is taken as the drill string vibration type under the current drilling condition. Vibration types include general vibration, vortex, and stick-slip.

[0220] Step 6: Vibration Control: Based on the intelligently identified vibration type and real-time characteristic values ​​(such as mean acceleration, peak difference, and SSR ratio), the risk level is determined, and drilling parameter adjustment commands are generated (whirling vibration: reduce rotational speed by 8%, increase drilling pressure by 6%; stick-slip vibration: increase rotational speed by 10%, reduce drilling pressure by 5%). These control commands are sent via mud pulses or the fieldbus, forming a closed loop of "identification-assessment-control". The specific process is as follows:

[0221] Normal vibration is a risk-free vibration and does not require treatment. Just continue drilling and monitor the vibration status of the drill string.

[0222] For lateral vibration (eddying), the mean acceleration range and peak value difference within a specific time window are extracted by analyzing the triaxial acceleration signals as the basis for judgment. Mild eddying is characterized by a mean acceleration of 1–2g and a peak value difference of 1–5g; moderate eddying corresponds to a mean acceleration of 3–5g and a peak value difference of 5–10g; and when the mean acceleration exceeds 5g and the peak value difference exceeds 10g, it is judged as severe eddying.

[0223] For stick-slip vibration, the stick-slip ratio is used as the core indicator:

[0224] (32);

[0225] in, Viscosity-slip ratio; This represents the maximum rotational speed of the drill bit during the sampling period. This represents the minimum rotational speed of the drill bit during the sampling period. This represents the average rotational speed of the drill bit during the sampling period.

[0226] Based on actual measurements and experience, it is divided into four levels: when the stickiness ratio is less than 0.5, it is considered low stickiness; between 0.5 and 1.0, it is medium stickiness; between 1.0 and 1.5, it is high stickiness; and when it exceeds 1.5, it is classified as severe stickiness.

[0227] Based on the assessment results, the system provides control recommendations: for moderate / severe whirl, it is recommended to reduce the drilling speed and increase the drilling pressure; when severe whirl exceeds the limit, drilling should be stopped. For advanced / severe stick-slip, it is recommended to increase the drilling speed and reduce the drilling pressure; when severe stick-slip exceeds the limit, drilling should be stopped. The control strategies can be embedded in the rule base and linked with the control system to form a closed loop of identification-assessment-control, improving drilling stability and efficiency.

[0228] To demonstrate the feasibility and superiority of this invention, the following comparative experiments are presented.

[0229] ResNet18, ResNet18-SENet, and BiGRU were selected as comparison models for the fused ResNet18SENet-BiGRU model of this invention. The ResNet18 model mainly includes residual blocks (skip connections) to solve the gradient vanishing problem and optimize deep networks; the ResNet18-SENet model mainly includes SE modules (channel attention mechanism) to dynamically adjust feature channel weights; the BiGRU model mainly captures temporal dependencies bidirectionally and simplifies the gating mechanism (update gate + reset gate); the ResNet18SENet-BiGRU model of this invention is a dual-channel feature extraction network constructed by jointly using BiGRU and ResNet18-SENet. This model jointly models image modalities and temporal modalities and performs attention-weighted fusion. The training time and various evaluation metrics (accuracy, precision, recall, F1 score) of different models on the validation set were compared and analyzed. The architecture depth and evaluation metric results of each model are shown in Table 1. The ResNet18SENet-BiGRU model of this invention achieves an accuracy of 0.9906 on the test set, which is 1.95% higher than that of single-channel ResNet18SENet and 1.45% higher than that of BiGRU, demonstrating the significant advantages of the fusion model of this invention in vibration pattern recognition.

[0230] Table 1. Models with different architectures

[0231] .

[0232] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for identifying and controlling drill string vibration based on multimodal data fusion, characterized in that, Includes the following steps: Step 1: Synchronously acquire multi-source data to form the original three-dimensional time-series signal, and then divide it into equal-length segments; Step 2: Process the original three-dimensional time-series signal; Step 3: Perform dual-channel feature extraction; in the time-series channel, a bidirectional gated recurrent unit network is used as a one-dimensional feature extractor for time-series data to extract time-series modal features; in the image channel, a ResNet18-SENet network is used as a two-dimensional feature extractor to extract image modal features. The specific process is as follows: Step 3.1: Define the time step The corresponding multidimensional signal is ;in, , , The signals are triaxial vibration signals in the X, Y, and Z axes, respectively. This is the denoised three-dimensional time-series signal. The segment length is used; the multidimensional signal corresponding to each time step is used as the input to the bidirectional gated recurrent unit network. Step 3.2: Extract image modal features using the ResNet18-SENet network; the ResNet18-SENet network consists of two parts: a ResNet18 residual module and an SENet module; the specific process is as follows: Step 3.2.1: Obtain the vibration time-frequency image The input is fed into the ResNet18-SENet network; firstly, a 7×7 convolutional layer is used for the first stage of convolution output; the first stage of convolution output is batch normalized; the normalized output is then subjected to the ReLU activation function to enhance the non-linear expressive power; subsequently, max pooling is used to reduce the feature map space size; Step 3.2.2: Pool the feature map The data is fed into a ResNet18 residual module for deep feature extraction. This residual module includes a first residual block, a second residual block, a third residual block, and a fourth residual block, which are connected in sequence. Each residual block consists of two residual units. The first residual block includes two cascaded first residual units. The second, third, and fourth residual blocks have the same structure, each including a cascaded second residual unit and a cascaded first residual unit. Both the first and second residual units adopt the standard unit structure, and are sequentially subjected to 3×3 convolution, first batch normalization, ReLU activation function, 3×3 convolution, and second batch normalization to form a convolutional mapping output; After four residual blocks, the spatial dimensions of the feature map are compressed into tensor values. ;in, , All are spatial dimensions. Number of channels; Step 3.2.3: Next, global average pooling is used to compress each channel; the compressed feature values ​​of all channels are concatenated to obtain the global average pooling feature. ; Step 3.2.4: Obtain the final image modal features through projection onto the fully connected layer. ; Step 4: Concatenate the temporal modality features with the image modality features, and simultaneously use attention weights to adaptively weight the concatenated features to obtain weighted fused features; Step 5: Identify vibration types based on weighted fusion features; Step 6: Based on the identified vibration type and real-time characteristic value, determine the risk level and generate drilling parameter adjustment instructions. The control instructions are sent through mud pulses or fieldbus to complete vibration control.

2. The drill string vibration identification and control method based on multimodal data fusion according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Integrate a triaxial vibration sensor within the drill string near the drill bit section to acquire downhole X-axis, Y-axis, and Z-axis vibration accelerations in real time, and simultaneously record rotational speed and torque parameters to generate the original three-dimensional time-series signal. ; Step 1.2: The original three-dimensional time series signal is segmented into equal-length segments using a non-overlapping sliding window method to obtain the signal containing... The original three-dimensional time-series signal of the triaxial vibration signal of the drill string segment: ,in, This represents the segment length; the number 3 corresponds to triaxial vibration. For the first Each segment of the triaxial vibration signal represents a segment of length [missing information]. Three-dimensional time series data.

3. The drill string vibration identification and control method based on multimodal data fusion according to claim 2, characterized in that, The specific process of step 2 is as follows: Step 2.1: Perform noise reduction independently on each axis vibration signal; Step 2.2: Based on the short-time Fourier transform, convert each segment of the denoised vibration signal into a two-dimensional time-frequency image.

4. The drill string vibration identification and control method based on multimodal data fusion according to claim 3, characterized in that, The specific process of step 2.1 is as follows: Step 2.1.1: Apply a symmetric 5th-order wavelet basis to the current axis vibration signal. Perform a seven-level discrete wavelet decomposition to obtain the set of wavelet coefficients. ,in For the first The low-frequency approximation component corresponding to the layer, ; For the first The high-frequency detail components corresponding to the first two layers are used to estimate the noise standard deviation. The specific calculation is as follows: （1）； in, This indicates median operation; , These represent the high-frequency detail components corresponding to the first and second layers, respectively; 0.6745 is the correction factor under the normal distribution. Step 2.1.2: Calculate the adaptive threshold for each layer based on the high-frequency detail components corresponding to each layer, and apply a soft thresholding function for compression processing; The formula for calculating the adaptive threshold is: （2）； in, For the first The adaptive threshold corresponding to the layer; The length of the vibration signal; The calculation formula for soft threshold function compression is as follows: （3）； in, This is a soft thresholding function; the calculated value is the [number]th [value] after compression. High-frequency detail components corresponding to the layer; Step 2.1.3: Process the compressed wavelet coefficient set. Perform inverse wavelet transform to reconstruct the vibration signal of the current axis, obtaining the denoised one-dimensional vibration signal. ; Step 2.1.4: After processing the triaxial vibration signals according to the noise reduction process in steps 2.1.1-2.1.3, reassemble them into the noise-reduced three-dimensional time-series signals in their original axial order. .

5. The drill string vibration identification and control method based on multimodal data fusion according to claim 4, characterized in that, The specific process of step 2.2 is as follows: Step 2.2.1: For each segment of the denoised one-dimensional vibration signal Perform short-time Fourier transforms on each signal to obtain the denoised one-dimensional vibration signal. Corresponding time-frequency domain matrix The time-frequency domain matrix is ​​a complex matrix composed of all frames and frequency points; the short-time Fourier transform processing flow is as follows: （4）； in, Time-frequency domain matrix The Middle Frame number Complex spectrum at frequency points; The sequence number of the one-dimensional vibration signal after noise reduction; The length of the window; Indicates the window function; Step size; Step 2.2.2: Take the logarithm of the magnitude of the time-frequency domain matrix as the image intensity to form a two-dimensional time-frequency image: （5）； in, It is a two-dimensional time-frequency image; A small constant to prevent logarithmic divergence; Step 2.2.3: Process the triaxial vibration signal independently according to the procedures in steps 2.2.1-2.2.2 to obtain three time-spectrum images. , and These correspond to two-dimensional time-frequency images along the X, Y, and Z axes, respectively. Step 2.2.4: Combine the three time-frequency images sequentially using a vertical stitching method to form a complete vibration time-frequency image. : （6）； in, Indicates the length of the time dimension. is the length of the frequency dimension.

6. The drill string vibration identification and control method based on multimodal data fusion according to claim 5, characterized in that, In step 3.1, the bidirectional gated loop unit network consists of two independent unidirectional gated loop units that do not share parameters; the forward gated loop unit starts from... arrive Sequential scanning to generate time steps Forward hidden state sequence ; Backward gated loop unit from arrive Reverse scan to generate time steps Backward hidden state sequence Each one-way gated loop unit regulates the retention and forgetting of historical information through update and reset gates. The one-way gated loop unit operates at time steps... The state update is calculated as follows: （7）； （8）； （9）； （10）； in, , and They are time steps Update the status of doors, reset doors, and hidden doors; It is a time step The hidden state; For time step The candidate hidden state; For input To the update gate The weight matrix; For input To reset door The weight matrix; For input To candidate hidden state The weight matrix; Hidden state To the update gate The weight matrix; Hidden state To reset door The weight matrix; Hidden state go through After control, proceed to the candidate hidden state The weight matrix; It is the Sigmoid activation function; tanh is the hyperbolic tangent activation function. This represents element-wise multiplication; The forward hidden state and backward hidden state of each time step are concatenated to form a bidirectional hidden state vector, which is the time step. Hidden door status Specifically: ,in, The hidden layer dimension of the unidirectional gated loop unit; By applying pooling operations to the bidirectional hidden state vector, the variable-length temporal feature sequence is compressed into a fixed-length feature vector, thus obtaining the temporal modal features. .

7. The drill string vibration identification and control method based on multimodal data fusion according to claim 6, characterized in that, In step 3.2.1, the formula for calculating the convolution output in the first stage is: （11）； in, This represents the output of the first stage of convolution; For the convolution output at the th The sample, the first Each channel, location The value; For convolution kernel; The convolution operation outputs the position index in the matrix. Vertical direction In the horizontal direction; and These are the height and width of the convolution kernel, respectively; and These are the indices of the convolution kernel in the horizontal and vertical directions; For the first The bias term for each channel; This is a convolution operation; The calculation formula for batch normalization on each channel is: （12）； （13）； （14）； in, and The first Mean and variance of each channel; and All are the first Learnable parameters for each channel; It is a small constant that prevents division by zero; Batch size; This is the normalized and linearly transformed value of the first-stage convolution output; The formula for calculating the ReLU activation function is: （15）； in, This is the result of the first stage convolution output after ReLU activation; The formula for calculating max pooling is: （16）； in, This is the feature map after pooling; This is the first phase of the max pooling operation; For pooled windows; In step 3.2.2, the formula for calculating the convolutional mapping output is as follows: （17）； in, The output is a convolutional mapping; The input to the residual unit is the pooled feature map. The input to other residual units is the output of the previous residual unit; This is the second batch normalization operation; It is a 3×3 convolution; This is the first batch of normalization operations; The skip connection of the first residual unit is an identity mapping: （18）； in, Mapping for skip connections; The second residual unit adds an extra 1×1 convolution in the skip connection to match the number of channels or to achieve downsampling: （19）； in, It is a 1×1 convolution; Each residual unit outputs for: （20）； The output of each residual unit undergoes channel attention recalibration via the SENet module; the SENet module consists of two stages: compression and excitation. During the compression phase, global average pooling is performed on each channel: （21）； in, To indicate the first The global average pooling value for each channel. , This represents the number of channels in the convolutional network. This is a global average pooling operation; The output of the residual unit is the first Feature maps corresponding to each channel; The vector obtained during the compression stage will then be... Two fully connected layers are input for excitation. A bottleneck structure is adopted, first reducing the dimensionality and then increasing it, and channel weights are generated. : （22）； in, and Here are the learnable weight matrices for the two fully connected layers; and These are the bias vectors for the two fully connected layers; Finally, channel weighting is used to multiply the channel weights obtained in the activation phase back into the original features channel by channel, thereby achieving adaptive recalibration of the feature map: （23）； in, For the first channel attention recalibration Feature maps of each channel; Weighted operation for the channel; For the first Channel weights for each channel; In step 3.2.3, the calculation formula for the global average pooling operation is as follows: （24）； in, For the first Each channel is located in The tensor value; For the first The compressed feature values ​​of each channel; Global average pooling features The calculation formula is: （25）； In step 3.2.4, the final image modal features The calculation formula is: （26）； in, It is a fully connected layer.

8. The drill string vibration identification and control method based on multimodal data fusion according to claim 7, characterized in that, The specific process of step 4 is as follows: Temporal modal features Image modal features Perform a stitching operation to generate fused features. : （27）； Introducing a self-attention mechanism to calculate attention weights: （28）； （29）； in, The feature vector after attention mapping; It is the ReLU activation function; , All are learnable parameter matrices; , It is a learnable bias vector; This represents the attention weights of the fused features across different dimensions; Attention weight Multiply the feature by the fusion feature to complete the weighted fusion, and obtain the final weighted fused feature. : （30）。 9. The drill string vibration identification and control method based on multimodal data fusion according to claim 8, characterized in that, In step 5, the final weighted fused features are fed into a fully connected layer, and a Softmax activation function is used for nonlinear mapping to output the corresponding vibration state classification result. （31）； in, , All are layered parameters; Indicates belonging to The probability distribution of vibration states; the category with the highest probability is taken as the drill string vibration type under the current drilling state. Vibration types include general vibration, eddy vibration, and stick-slip vibration.

10. The drill string vibration identification and control method based on multimodal data fusion according to claim 9, characterized in that, In step 6, the real-time feature values ​​include the mean acceleration, the peak difference, and the stick-slip ratio; Normal vibration is risk-free and requires no treatment; simply continue drilling and monitor the vibration status of the drill string. For eddy vibration, the mean and peak value range of acceleration within a specific time window are extracted by analyzing the triaxial acceleration signal as the basis for judgment. When the mean acceleration is between 1 and 2g and the peak value difference is between 1 and 5g, it is judged as mild eddy vibration; when the mean acceleration is between 3 and 5g and the peak value difference is between 5 and 10g, it is judged as moderate eddy vibration; and when the mean acceleration is greater than 5g and the peak value difference is greater than 10g, it is judged as severe eddy vibration. For stick-slip vibration, the stick-slip ratio is used as a risk assessment indicator; when the stick-slip ratio is less than 0.5, it is judged as low-level stick-slip; and between 0.5 and 1.0, it is judged as medium-level stick-slip. A viscosity rating of 1.0 to 1.5 indicates high stickiness; a value exceeding 1.5 indicates severe stickiness. Based on the assessment results, the following control recommendations are given: For moderate or severe whirl, it is recommended to reduce the rotation speed and increase the drilling pressure. When severe whirl exceeds the limit, drilling should be stopped. For advanced or severe stick-slip, it is recommended to increase the rotation speed and reduce the drilling pressure. When severe stick-slip exceeds the limit, drilling should be stopped.

Images