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

Abstract

The invention discloses a drill string vibration identification and regulation method based on multi-modal data fusion, and belongs to the crossing field of petroleum drilling and artificial intelligence. According to the method, multi-modal input fusing space and time-frequency information is constructed by using a three-axis vibration acceleration signal of an underground drill string and a three-channel time-frequency diagram corresponding to the three-axis vibration acceleration signal. Through a deep neural network in which BiGRU and ResNet18-SENet are combined, an attention mechanism is fused, and accurate recognition of complex vibration modes such as general vibration, stick-slip and vortex motion is realized. Further combining the recognition result, driving the self-adaptive regulation and control of drilling parameters, and realizing the closed-loop control of the vibration state of the drill string. The method has the advantages of being high in recognition precision, high in robustness, intelligent in regulation and control and the like, and is suitable for underground vibration monitoring and drilling parameter optimization in a complex stratum environment.

Description

Technical Field

[0001] The present invention belongs to the intersection field of petroleum drilling and artificial intelligence, and specifically relates to a drill string vibration identification and control method based on multimodal data fusion. Background Art

[0002] Drill string vibration is driven by the coupling of multiple factors, including the drill bit-rock interaction, formation heterogeneity, and wellbore constraints. This produces axial, lateral, and torsional vibrations, exhibiting strong nonlinear and nonstationary characteristics such as stick-slip and vortex. Excessive vibration can lead to drill tool failure, and different vibration modes correspond to different control strategies, making accurate vibration mode identification crucial.

[0003] In recent years, deep learning has shown great potential in the field of vibration analysis, but existing methods mostly rely on single time series signals or time-frequency graph modeling, failing to fully utilize the complementary advantages of multimodal data and cross-modal time series dependencies, and recognition accuracy and adaptability are still limited.

[0004] To this end, this paper proposes synchronously inputting vibration acceleration signals and their corresponding time-frequency plots into a dual-channel feature extraction network constructed by combining BiGRU and ResNet18-SENet, using an attention mechanism to fuse multidimensional information. This method, by simultaneously capturing the spatial coupling characteristics and time-frequency energy of vibration, accurately identifies various fault modes, including general vibration, stick-slip, and vortex. Based on these identification results, drilling parameters are automatically adjusted, forming a closed-loop control strategy. Summary of the Invention

[0005] To address the low accuracy and poor adaptability of traditional single-channel, single-modality drill string vibration identification methods, this paper proposes a drill string vibration identification and control method based on multimodal data fusion. This method integrates data perception, feature extraction, intelligent decision-making, and parameter control. It achieves multimodal synchronous acquisition through a combination of three-axis acceleration and three-channel time-frequency graphs. A dual-channel feature extraction network is constructed by combining BiGRU and ResNet18-SENet. An attention fusion mechanism is introduced to comprehensively model the temporal and spatial vibration characteristics of the drill string. Based on the fused features, this method identifies vibration types (e.g., general vibration, vortex, and stick-slip) and assesses their severity. Furthermore, a threshold strategy is used to generate corresponding speed and weight-on-bit control commands. Finally, this method can be integrated with downhole or surface control modules to form an identification-assessment-control closed-loop system, enabling vibration state identification and adaptive parameter optimization under complex operating conditions.

[0006] The technical solutions of the present invention are as follows: A drill string vibration identification and control method based on multimodal data fusion includes the following steps: Step 1: synchronously collect multi-source data to form an original three-dimensional time series signal, and divide it into equal-length segments; Step 2: Processing 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. Step 4: Concatenate the temporal modal features with the image modal features, and use the attention weights to adaptively weight the concatenated features to obtain weighted fusion features. Step 5: Identify the vibration type based on the weighted fusion features; Step 6: Based on the identified vibration type and real-time characteristic value, the risk level is determined and drilling parameter adjustment instructions are generated. The control instructions are sent via mud pulse or field bus to complete vibration control.

[0007] Furthermore, the specific process of step 1 is as follows: Step 1.1: Integrate a three-axis vibration sensor in the drill bit sub of the drill string to obtain the downhole X-axis, Y-axis, and Z-axis vibration acceleration in real time, and simultaneously record the speed and torque parameters to form the original three-dimensional time series signal. ; Step 1.2: Use the non-overlapping sliding window method to divide the original three-dimensional time series signal into equal length segments to obtain The original three-dimensional time series signal of the three-axis vibration signal of the drill string: ,in, is the segment length; the number 3 corresponds to triaxial vibration; For the Segment three-axis vibration signal, each segment of the three-axis vibration signal represents a length of Three-dimensional time series data.

[0008] Furthermore, the specific process of step 2 is as follows: Step 2.1: Perform the noise reduction process independently on each axis vibration signal; Step 2.2: Convert each vibration signal after noise reduction into a two-dimensional time-frequency image based on short-time Fourier transform.

[0009] Furthermore, the specific process of step 2.1 is as follows: Step 2.1.1: Use the symmetrical 5th-order wavelet basis to analyze the current shaft vibration signal Perform seven-layer discrete wavelet decomposition to obtain a set of wavelet coefficients ,in For the The low-frequency approximate component corresponding to the layer, ; For the The high-frequency detail components corresponding to the first two layers are used to estimate the noise standard deviation. , the specific calculation is: (1); in, Represents the median operation; 、 are the high-frequency detail components corresponding to the first and second layers respectively; 0.6745 is the correction factor under normal distribution; Step 2.1.2: Calculate the adaptive threshold of each layer based on the high-frequency detail components corresponding to each layer, and apply a soft threshold function for compression processing; The calculation formula of the adaptive threshold is: (2); in, For the The adaptive threshold corresponding to the layer; is the vibration signal length; The calculation formula for the soft threshold function compression processing is: (3); in, is the soft threshold function, and the calculated value is the first The high-frequency detail component corresponding to the layer; Step 2.1.3: Compress the wavelet coefficients set Perform inverse wavelet transform to complete the reconstruction of the current shaft vibration signal and obtain the one-dimensional vibration signal after noise reduction ; Step 2.1.4: After processing the three-axis vibration signals according to the noise reduction process of steps 2.1.1 to 2.1.3, reorganize them into the noise-reduced three-dimensional time series signals according to the original axial order. .

[0010] Furthermore, 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 transform respectively to obtain the one-dimensional vibration signal after noise reduction 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 processing flow is as follows: (4); in, is the time-frequency domain matrix Middle Frame No. Complex spectrum of frequency points; is the serial number of the one-dimensional vibration signal after denoising; is the window length; represents the window function; is the step length; Step 2.2.2: Take the logarithm of the time-frequency domain matrix amplitude modulus as the image intensity to form a two-dimensional time-frequency image: (5); in, is a two-dimensional time-frequency image; A small constant to prevent logarithmic divergence; Step 2.2.3: Process the three-axis vibration signals independently according to the process of steps 2.2.1-2.2.2 to obtain three time-frequency spectrum images. 、 and , corresponding to the two-dimensional time-frequency spectrum images in the X-axis, Y-axis and Z-axis directions respectively; Step 2.2.4: Use vertical splicing to combine the three time-frequency images in sequence to form a complete vibration time-frequency image. : (6); in, represents the length of the time dimension, is the length of the frequency dimension.

[0011] Furthermore, the specific process of step 3 is as follows: Step 3.1: Define the time step The corresponding multidimensional signal is ;in, 、 、 The three-axis vibration signals are respectively X-axis, Y-axis and Z-axis; the multidimensional signal corresponding to each time step is used as the input of the bidirectional gated recurrent unit network; The bidirectional gated recurrent unit network consists of two independent unidirectional gated recurrent units with no shared parameters; the forward gated recurrent unit arrive Sequential scanning, generating time steps The forward hidden state sequence ; Backward gated recurrent unit from arrive Scan in reverse order to generate time steps The backward hidden state sequence Each unidirectional gated recurrent unit controls the retention and forgetting of historical information through the update gate and reset gate. The unidirectional gated recurrent unit is The state update is calculated as follows: (7); (8); (9); (10); in, 、 and The time steps are Update door, reset door and hide door states; is the time step The hidden state of is the time step Candidate hidden states of ; For input To the update gate The weight matrix of For input To reset gate The weight matrix of For input To the candidate hidden state The weight matrix of Hidden To the update gate The weight matrix of Hidden To reset gate The weight matrix of Hidden go through Control to candidate hidden state The weight matrix of is the Sigmoid activation function; tanh is the hyperbolic tangent activation function; represents element-wise multiplication; Concatenate the forward hidden state and the backward hidden state of each time step to form a bidirectional hidden state vector, which is the time step Hidden door status , specifically: ,in, is the hidden layer dimension of the unidirectional gated recurrent unit; By applying a pooling operation to the bidirectional hidden state vector, the variable-length temporal feature sequence is compressed into a fixed-length feature vector to obtain the temporal modal feature ; Step 3.2: Use the ResNet18-SENet network to extract image modal features; the ResNet18-SENet network consists of two parts: the ResNet18 residual module and the SENet module.

[0012] Furthermore, the specific process of step 3.2 is as follows: Step 3.2.1, the vibration time-frequency image It is fed into the ResNet18-SENet network as input; first, a 7×7 convolutional layer is used to perform the first stage of convolution output, and the formula is: (11); in, represents the convolution output of the first stage; The convolution output is Samples, Channels, positions The value of is the convolution kernel; is the position index in the convolution operation output matrix, is the vertical direction, is the horizontal direction; and are the height and width of the convolution kernel respectively; and is the index of the convolution kernel in the horizontal and vertical directions; For the The bias term of each channel; is the convolution operation; The convolution output of the first stage is batch normalized, and the calculation formula on each channel is: (12); (13); (14); in, and Respectively The mean and variance of each channel; and All are Learnable parameters of channels; is a small constant that prevents division by zero; is the batch size; The convolution output of the first stage is normalized and linearly transformed. Use the ReLU activation function on the normalized output to enhance the nonlinear expression ability: (15); in, The convolution output of the first stage is the result after ReLU activation; Then use maximum pooling to reduce the spatial size of the feature map: (16); in, is the feature map after pooling; It is the maximum pooling operation in the first stage; is the pooling window; Step 3.2.2, the pooled feature map The network is fed into the ResNet18 residual module for deep feature extraction. The 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 first residual units connected in series. The second, third, and fourth residual blocks have the same structure, and all include a second residual unit and a first residual unit connected in series. The first residual unit and the second residual unit both use the standard unit structure, and perform 3×3 convolution, the first batch normalization, the ReLU activation function, the 3×3 convolution, and the second batch normalization in sequence to form the convolution map output: (17); in, is the convolution map output; The input of the residual unit is the pooled feature map. , the input of 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 normalization operation; The skip connection of the first residual unit is the identity mapping: (18); in, Mapping for skip connections; The second residual unit adds one more 1×1 convolution in the jump connection to match the number of channels or achieve downsampling: (19); in, It is a 1×1 convolution; Each residual unit outputs for: (20); The output of each residual unit is recalibrated through the SENet module for channel attention; the SENet module consists of two stages: compression and excitation; The compression stage performs global average pooling on each channel: (twenty one); in, To indicate the The global average pooling value of the channels, , is the number of channels in the convolutional network; is the global average pooling operation; is the output of the residual unit Feature map corresponding to each channel; Then the vector obtained in the compression phase is Input two fully connected layers for excitation, adopt a bottleneck structure, first reduce the dimension and then increase the dimension, and generate channel weights : (twenty two); in, and are the learnable weight matrices of the two fully connected layers; and is the bias vector of the two fully connected layers; Finally, the channel weighting operation is used to multiply the channel weights obtained in the excitation stage back to the original features channel by channel to achieve adaptive recalibration of the feature map: (twenty three); in, is the first Feature maps of channels; Weighting operation for channels; For the The channel weight of each channel; After four residual blocks, the spatial dimensions of the feature map are compressed into tensor values. ;in, 、 are all spatial dimensions, is the number of channels; Step 3.2.3. Next, use the global average pooling operation to compress each channel: (twenty four); in, For the Channels at position The tensor value of ; For the The eigenvalues ​​after channel compression; The eigenvalues ​​after compression of all channels are spliced ​​together to obtain the global average pooling feature : (25); Step 3.2.4: Project the fully connected layer to obtain the final image modality features : (26); in, is a fully connected layer.

[0013] Furthermore, the specific process of step 4 is as follows: The temporal modal features and image modality features Perform splicing operations to generate fusion features : (27); Introduce the self-attention mechanism to calculate the attention weight: (28); (29); in, is the feature vector after attention mapping; is the ReLU activation function; 、 All are learnable parameter matrices; 、 is the learnable bias vector; Represents the attention weights of fusion features in different dimensions; The attention weight Multiply with the fusion feature point to complete the weighted fusion and obtain the final weighted fusion feature : (30).

[0014] Furthermore, in step 5, the final weighted fusion feature is sent to the fully connected layer, and the Softmax activation function is used for nonlinear mapping to output the corresponding vibration state classification result: (31); in, 、 All are hierarchical parameters; Indicates belonging The probability distribution of the vibration state class is calculated; the class corresponding to the highest probability is taken as the drill string vibration type under the current drilling state. The vibration types include general vibration, vortex vibration, and stick-slip vibration.

[0015] Furthermore, in step 6, the real-time characteristic values ​​include acceleration mean, peak difference, and stick-slip ratio; Normal vibration is harmless and does not require any treatment. Just continue drilling and monitor the vibration status of the drill string. For vortex vibration, by analyzing the three-dimensional acceleration signal, the acceleration mean range and peak difference within a specific time window are extracted as the judgment basis; when the acceleration mean is between 1-2g and the peak difference is between 1-5g, it is judged as mild vortex; when the acceleration mean is between 3-5g and the peak difference is between 5-10g, it is judged as moderate vortex; and when the acceleration mean is greater than 5g and the peak difference is greater than 10g, it is judged as severe vortex. 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 considered low-level stick-slip; between 0.5 and 1.0, it is considered medium-level stick-slip; between 1.0 and 1.5, it is considered high-level stick-slip; and when it exceeds 1.5, it is classified as severe stick-slip. Based on the evaluation results, control recommendations are given: in case of moderate or severe vortex, it is recommended to reduce the rotation speed and increase the drilling pressure. When the severe vortex exceeds the limit, drilling should be stopped. In case of advanced or severe stick-slip, it is recommended to increase the rotation speed and reduce the drilling pressure. When the severe stick-slip exceeds the limit, drilling should be stopped.

[0016] The beneficial technical effects brought about by the present invention are as follows: the present invention introduces multimodal data fusion and a dual-channel deep network structure, fully combines the advantages of time series and time-frequency characteristics, realizes high-precision identification and severity level assessment of drill string vibration modes, and forms a closed-loop control strategy on this basis, which can automatically optimize key parameters such as rotation speed and drilling pressure, not only effectively improving the accuracy and adaptability of vibration monitoring to working conditions, but also reducing the risk of drill tool failure and improving drilling efficiency, and has good engineering application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention is a flowchart of a drill string vibration identification and control method based on multimodal data fusion.

[0018] Figure 2 Schematic diagram of the drill string vibration time series feature extractor in the present invention.

[0019] Figure 3 Schematic diagram of the drill string vibration time-frequency feature extractor in the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: like Figure 1As shown, to achieve accurate identification and real-time control of drill string vibration, the present invention proposes a multimodal data fusion method of "three axes - three channels - dual-channel network - closed-loop decision-making". This method mainly 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 recognition is performed through the fusion model, and vibration control is performed based on the recognition results. The specific steps are as follows: Step 1: At the data input layer, synchronously collect multi-source data to form the original three-dimensional time series signal, and divide it into equal-length segments; the specific process is as follows: Step 1.1: Integrate a triaxial vibration sensor in the drill bit sub of the drill string to obtain real-time downhole vibration acceleration along the X-axis (transverse 1, one of the vibration components perpendicular to the wellbore direction), the Y-axis (transverse 2, another vibration component orthogonal to the X-axis), and the Z-axis (axial, the vibration component in the wellbore direction). Simultaneously record operating parameters such as speed and torque to form a raw three-dimensional time series signal. The vibration sensor may specifically be a MEMS accelerometer, which is an acceleration sensor manufactured using Micro-Electro-Mechanical Systems (MEMS) technology.

[0021] Step 1.2: Use the non-overlapping sliding window method to divide the original three-dimensional time series signal into equal length segments to obtain The original three-dimensional time series signal of the three-axis vibration signal of the drill string: , set the segment length to , that is, the length of each triaxial vibration signal is ; For example, under the condition of sampling rate of 120Hz, It can be taken as 2400 (corresponding to a 20-second signal). For the Segment three-axis vibration signal, each segment of the three-axis vibration signal represents a length of Three-dimensional time series data; the number 3 corresponds to triaxial vibration.

[0022] Step 2: In the data processing layer, the original three-dimensional time series signal is processed: sym5 is independently executed on each axis to perform seven-layer discrete wavelet decomposition and soft threshold noise reduction. Finally, the denoised signal is reconstructed through the inverse wavelet transform. The denoised three-axis signal is then converted into a time-frequency diagram through the short-time Fourier transform (STFT) to achieve time series-image modality alignment. The specific process is as follows: Step 2.1: To ensure that the three-way information does not interfere with each other, the noise reduction process is performed independently on each axis vibration signal. The specific process is as follows: Step 2.1.1. First, use the Symlet 5 wavelet basis (symmetric 5th order wavelet basis, referred to as sym5 wavelet basis) to analyze the vibration signal of each axis. Perform seven-layer discrete wavelet decomposition to obtain a set of wavelet coefficients ,in For the The low-frequency approximate component corresponding to the layer, ; For the The high-frequency detail components corresponding to the first two layers contain high-frequency fluctuations and potential noise in the signal. In order to estimate the noise level, the high-frequency detail components of the first two layers are used to estimate the noise standard deviation. , the specific calculation is: (1); in, Represents the median operation; 、 are the high-frequency detail components corresponding to the first and second layers respectively; 0.6745 is the correction factor under normal distribution.

[0023] Step 2.1.2: After obtaining the noise estimate, calculate the adaptive threshold for each layer based on the high-frequency detail components corresponding to each layer, and apply a soft threshold function for compression. This method can maximize the preservation of effective signal details while suppressing high-frequency noise. The low-frequency approximate components are not processed, preserving the main energy and overall change trend of the signal.

[0024] The calculation formula of the adaptive threshold is: (2); in, For the The adaptive threshold corresponding to the layer; is the vibration signal length; The calculation formula for the soft threshold function compression processing is: (3); in, is the soft threshold function, and the calculated value is the first The high-frequency detail component corresponding to the layer; Step 2.1.3: Then, the compressed wavelet coefficients are set Input inverse wavelet transform to complete the reconstruction of the shaft vibration signal and obtain the one-dimensional vibration signal after noise reduction .

[0025] Step 2.1.4: After the three-axis vibration signals have completed the above noise reduction process, they are reorganized into the noise-reduced three-dimensional time series signal according to the original axial order. .

[0026] Step 2.2: To extract richer 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: Step 2.2.1: For each segment of the denoised one-dimensional vibration signal Perform short-time Fourier transform (STFT) to obtain the one-dimensional vibration signal after noise reduction The corresponding time-frequency domain matrix , the time-frequency domain matrix is ​​a complex matrix composed of all frames and frequency points. The STFT processing flow is as follows: (4); in, is the time-frequency domain matrix Middle Frame No. Complex spectrum of frequency points; Indicates the frame index, which controls the starting position of the window; It is the frequency index, which determines the spectrum resolution; is the serial number of the one-dimensional vibration signal after denoising; is the window length; represents the one-dimensional vibration signal after noise reduction, For the A one-dimensional vibration signal after noise reduction; Represents the window function, where Hamming window is used; is the step length, the formula is , is the overlap rate.

[0027] Step 2.2.2: After calculating the time-frequency domain matrix representation, take the logarithm of its amplitude modulus as the image intensity to form a two-dimensional time-frequency image: (5); in, is a two-dimensional time-frequency image; A small constant to prevent logarithmic divergence.

[0028] Step 2.2.3: Process the three-axis vibration signals independently according to the process of steps 2.2.1-2.2.2 to obtain three time-frequency spectrum images. 、 and , corresponding to the two-dimensional time-frequency spectrum images of the X-axis, Y-axis and Z-axis directions respectively. Each image is a two-dimensional array, and each size is A two-dimensional image of represents the length of the time dimension, is the length of the frequency dimension.

[0029] Step 2.2.4: To integrate the three-axis information into a unified input form, vertical stacking is used to combine the three time-frequency images in sequence to form a complete vibration time-frequency image. : (6); After signal noise reduction and time-frequency conversion, the original triaxial vibration data is simultaneously represented as two complementary data modes: three-dimensional vibration time series signal (time series data) The corresponding vibration time-frequency image .

[0030] Step 3: In the feature extraction layer, dual-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 (time series modal features) of the vibration signal; in the image channel, a ResNet18-SENet network is used as a two-dimensional feature extractor to extract the time series features (image modal features) of the time-frequency image. The specific process is as follows: Step 3.1, such as Figure 2 As shown, in the modeling process, firstly, a multidimensional signal corresponding to each time step is formed as the input of BiGRU, and the time step is defined The corresponding multi-dimensional signal (three-axis acceleration) is ;in, 、 、 They are the three-axis vibration signals of the X-axis, Y-axis and Z-axis.

[0031] BiGRU consists of two independent unidirectional GRUs (Gated Recurrent Units) with no shared parameters; in the forward layer, the forward GRU arrive Sequential scanning, generating time steps The forward hidden state sequence ; In the reverse layer, the backward GRU is arrive Scan in reverse order to generate time steps The backward hidden state sequence Each unidirectional GRU controls the retention and forgetting of historical information through the update gate and reset gate. The state update is calculated as follows: (7); (8); (9); (10); where, , and are the update gate, reset gate and hidden gate states at time step ; is the hidden state at time step , also referred to as the hidden state of the previous time step; is the candidate hidden state at time step ; is the weight matrix of input to the update gate ; is the weight matrix of input to the reset gate ; is the weight matrix of input to the candidate hidden state ; is the weight matrix of hidden state to the update gate ; is the weight matrix of hidden state to the reset gate ; is the weight matrix of hidden state , after being controlled by , to the candidate hidden state ; is the Sigmoid activation function; tanh is the hyperbolic tangent activation function; denotes element-wise multiplication.

[0032] Figure 2 The "1-" in corresponds to the "1-" in the of formula (10).

[0033] At the output layer, the forward hidden state and the backward hidden state of each time step are spliced to form a bidirectional hidden state vector, which is the hidden gate state at time step , specifically: where, is the hidden layer dimension of the unidirectional GRU.

[0034] In summary, by applying the pooling operation on the hidden state of the sequence, the variable-length time sequence feature sequence is compressed into a fixed-length feature vector, i.e., the time sequence modal feature , as the time domain input for subsequent feature fusion.

[0035] ​Step 3.2: In order to achieve multimodal feature fusion, in addition to temporal features, the ResNet18-SENet network is also required to extract image modality information. The ResNet18-SENet network mainly consists of two parts: ResNet18 and SENet. The ResNet18 basic network architecture solves the gradient vanishing problem in deep network training through residual learning; SENet explicitly models the dependency between feature channels through feature recalibration strategies, improving the model's ability to recognize important features. Figure 3 As shown in the figure, the data processing process of the ResNet18-SENet network is: Step 3.2.1, and three-axis acceleration Corresponding vibration time-frequency image The three-channel time-frequency image is fed into the ResNet18-SENet network. First, the first-stage convolution output is passed through a standard 7×7 convolution layer to extract the initial local features, whose mathematical form is: (11); in, represents the convolution output of the first stage; The convolution output is Samples, Channels, positions The value of is the convolution kernel; is the vertical direction (row coordinate); is the horizontal direction (column coordinate); The position index in the output matrix of the convolution operation; and are the height and width of the convolution kernel respectively; and is the index of the convolution kernel in the horizontal and vertical directions; For the The bias term of each channel; is the convolution operation.

[0036] In order to speed up training and stabilize the network, the convolution output of the first stage is batch normalized (BN), and the calculation formula on each channel is: (12); (13); (14); in, and Respectively The mean and variance of each channel; and All are Learnable parameters of channels; is a small constant that prevents division by zero; is the batch size; The convolution output of the first stage is normalized and linearly transformed.

[0037] The ReLU activation function is used on the normalized output to enhance the nonlinear expression ability and improve the model's ability to express complex vibration modes: (15); in, This is the result of the convolution output of the first stage after ReLU activation.

[0038] Then use maximum pooling to reduce the spatial size of the feature map and extract the most significant local features: (16); in, is the feature map after pooling; It is the maximum pooling operation in the first stage; is the pooling window; Step 3.2.2, then the pooled feature map The data is fed into the ResNet18 residual module for deep feature extraction. This residual module comprises four residual blocks: the first residual block, the second residual block, the third residual block, and the fourth residual block, which are connected in sequence. Each residual block is composed of several residual units. In the present invention, each residual block consists of two residual blocks. The first residual block includes two first residual units connected in series. The second, third, and fourth residual blocks have the same structure, each consisting of a second residual unit and a first residual unit connected in series. Each residual unit uses a combination of convolution, batch normalization, ReLU, and skip connections to form a residual network to alleviate the vanishing gradient problem.

[0039] The first residual unit adopts the standard unit structure, which performs 3×3 convolution, the first batch normalization, ReLU activation function, 3×3 convolution, and the second batch normalization in sequence to form the convolution map output: (17); Its skip connection is the identity mapping: (18); in, is the convolution map output; The input of the residual unit is the first residual unit in the first residual block, and the input of the first residual unit is the pooled feature map , the input of 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 normalization operation; The second residual unit (with 1×1 convolution unit) also uses the standard unit basis, but it adds an additional 1×1 convolution in the jump connection to match the number of channels or achieve downsampling: (19); in, Mapping for skip connections; It is a 1×1 convolution; Each residual unit outputs for: (20); The output of each residual unit is recalibrated through the SENet module for channel attention, which includes two stages: compression and excitation.

[0040] The Squeeze stage performs global average pooling on each channel: (twenty one); in, To indicate the The global average pooling value of the channels, , is the number of channels in the convolutional network; is the global average pooling operation; is the output of the residual unit The feature map corresponding to each channel.

[0041] Then the vector obtained in the Squeeze stage is Input two fully connected layers for Excitation, using a bottleneck structure, first reducing the dimension and then increasing the dimension, and generate channel weights : (twenty two); in, and are the learnable weight matrices of the two fully connected layers, , , is the compression ratio; and are the bias vectors of the two fully connected layers, , .

[0042] Finally, the Scale (channel weighting) operation is used to multiply the channel weights obtained in the excitation stage back to the original features channel by channel to achieve adaptive recalibration of the feature map: (twenty three); in, is the first Feature maps of channels; Weighting operation for channels; For the The channel weight of each channel; After four residual blocks, the spatial dimensions of the feature map are compressed into tensor values. ;in, 、 Both 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.

[0043] Step 3.2.3. Next, use the global average pooling (GAP) operation to compress each channel: (twenty four); in, For the Channels at position The tensor value of ; For the The eigenvalues ​​after channel compression.

[0044] The global average pooling feature can be obtained by splicing the compressed feature values ​​of all channels : (25); Step 3.2.4: Then, the final image modality features are obtained through full connection layer projection : (26); in, is a fully connected layer.

[0045] Step 4: In the feature fusion layer, feature fusion is performed. To fuse the data representations from multiple modalities, the temporal modal features and the image modal features are spliced ​​together. At the same time, the spliced ​​features are adaptively weighted using attention weights to obtain weighted fusion features, thus achieving space-frequency-condition complementarity. The specific process is as follows: The temporal modal features and image modality features Perform splicing operations, construct joint representations, and generate fusion features : (27); In order to guide the model to pay more attention to important features, the self-attention mechanism is further introduced to calculate the attention weight: (28); (29); in, is the feature vector after attention mapping; is the ReLU activation function; 、 Both are learnable parameter matrices; 、 is the learnable bias vector; Indicates the attention weight of the fusion feature in different dimensions. Attention weight Then multiply it with the fusion feature to complete the weighted fusion and obtain the final weighted fusion feature. : (30); Step 5: Perform vibration classification in the classification output layer; send the final weighted fusion features to the fully connected layer (or multi-layer perceptron), and use the Softmax activation function for nonlinear mapping to output the corresponding vibration state classification results: (31); in, 、 All are hierarchical parameters; Indicates belonging Probability distribution of vibration-like states.

[0046] The category with the highest probability is taken as the drill string vibration type under the current drilling state. The vibration types include general vibration, vortex, and stick-slip.

[0047] Step 6: Vibration control: Based on the intelligently identified vibration type and real-time characteristic values ​​(such as acceleration mean, peak difference, and SSR ratio), the risk level is determined and drilling parameter adjustment instructions are generated (vortex speed reduction 8%, increase drilling pressure 6%, stick-slip speed increase 10%, reduce drilling pressure 5%). The control instructions are issued via mud pulses or fieldbus, forming an "identification-assessment-control" closed loop. The specific process is as follows: General vibration is a vibration without risk and does not require any treatment. Just continue drilling and monitor the vibration status of the drill string. For lateral vibration (eddy), the three-dimensional acceleration signals are analyzed, and the range of mean acceleration and peak-to-peak differences within a specific time window are extracted as the basis for judgment. Mild eddy is characterized by a mean acceleration of 1–2g and a peak-to-peak difference of 1–5g; moderate eddy corresponds to a mean acceleration of 3–5g and a peak-to-peak difference of 5–10g; and severe eddy is determined when the mean acceleration exceeds 5g and the peak-to-peak difference exceeds 10g.

[0048] For stick-slip vibration, the stick-slip ratio is used as the core indicator: (32); in, is the stick-slip ratio; is the maximum speed of the drill bit during the sampling period; is the minimum speed of the drill bit during the sampling period; is the average speed of the drill bit during the sampling period; Based on measured and empirical data, it is divided into four levels: when the stick-slip ratio is less than 0.5, it is considered 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 when it exceeds 1.5, it is classified as a severe stick-slip state.

[0049] Based on the assessment results, the system provides control recommendations: For moderate / severe turbulence, it recommends reducing the rotational speed and increasing the drilling pressure. Drilling should be stopped if severe turbulence exceeds the limit. For advanced / severe stick-slip, it recommends increasing the rotational speed and reducing the drilling pressure. Drilling should be stopped if severe stick-slip exceeds the limit. Control strategies can be embedded in a rule base and integrated with the control system, forming a closed loop of identification, assessment, and control, improving drilling stability and efficiency.

[0050] In order to demonstrate the feasibility and superiority of the present invention, the following comparative experiments are given.

[0051] The ResNet18, ResNet18-SENet, and BiGRU models were selected as comparison models for the fused ResNet18SENet-BiGRU model of this invention. The ResNet18 model primarily includes residual blocks (skip connections) to address the vanishing gradient problem and optimize deep networks. The ResNet18-SENet model primarily includes the SE module (channel attention mechanism) to dynamically adjust feature channel weights. The BiGRU model primarily 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 combining BiGRU and ResNet18-SENet. This model jointly models the image modality and the temporal modality, performing attention-weighted fusion. The training time and various evaluation metrics (accuracy, precision, recall, and F1 score) of the different models on the validation set were compared and analyzed. The results of the model architecture depth and evaluation metrics are shown in Table 1. The ResNet18SENet-BiGRU model of the present invention achieved an accuracy of 0.9906 on the test set, an increase of 1.95% compared to the single-channel ResNet18SENet and 1.45% compared to BiGRU, demonstrating the significant advantages of the fusion model of the present invention in vibration pattern recognition.

[0052] Table 1 Models with different architectures .

[0053] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A drill string vibration identification and control method based on multimodal data fusion, characterized in that: The steps include: Step 1: synchronously collect multi-source data to form an original three-dimensional time series signal, and divide it into equal-length segments; Step 2: Processing 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, the ResNet18-SENet network is used as a two-dimensional feature extractor to extract image modality features; Step 4: Concatenate the temporal modal features with the image modal features, and use the attention weights to adaptively weight the concatenated features to obtain weighted fusion features. Step 5: Identify the vibration type based on the weighted fusion features; Step 6: Based on the identified vibration type and real-time characteristic value, the risk level is determined and drilling parameter adjustment instructions are generated. The control instructions are sent via mud pulse or field bus 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: Step 1.1: Integrate a three-axis vibration sensor in the drill bit sub of the drill string to obtain the downhole X-axis, Y-axis, and Z-axis vibration acceleration in real time, and simultaneously record the speed and torque parameters to form the original three-dimensional time series signal. ; Step 1.2: Use the non-overlapping sliding window method to divide the original three-dimensional time series signal into equal length segments to obtain The original three-dimensional time series signal of the three-axis vibration signal of the drill string: ,in, is the segment length; the number 3 corresponds to triaxial vibration; For the Segment three-axis vibration signal, each segment of the three-axis vibration signal represents a length of 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: Step 2.1: Perform the noise reduction process independently on each axis vibration signal; Step 2.2: Convert each vibration signal after noise reduction into a two-dimensional time-frequency image based on short-time Fourier transform.

4. The drill string vibration identification and control method based on multimodal data fusion according to claim 3 is characterized in that: The specific process of step 2.1 is as follows: Step 2.1.1: Use the symmetrical 5th-order wavelet basis to analyze the current shaft vibration signal Perform seven-layer discrete wavelet decomposition to obtain a set of wavelet coefficients ,in For the The low-frequency approximate component corresponding to the layer, ; For the The high-frequency detail components corresponding to the first two layers are used to estimate the noise standard deviation. , the specific calculation is: （1）； in, Represents the median operation; 、 are the high-frequency detail components corresponding to the first and second layers respectively; 0.6745 is the correction factor under normal distribution; Step 2.1.2: Calculate the adaptive threshold of each layer based on the high-frequency detail components corresponding to each layer, and apply a soft threshold function for compression processing; The calculation formula of the adaptive threshold is: （2）； in, For the The adaptive threshold corresponding to the layer; is the vibration signal length; The calculation formula for the soft threshold function compression processing is: （3）； in, is the soft threshold function, and the calculated value is the first The high-frequency detail component corresponding to the layer; Step 2.1.3: Compress the wavelet coefficients set Perform inverse wavelet transform to complete the reconstruction of the current shaft vibration signal and obtain the one-dimensional vibration signal after noise reduction ; Step 2.1.4: After processing the three-axis vibration signals according to the noise reduction process of steps 2.1.1 to 2.1.3, reorganize them into the noise-reduced three-dimensional time series signals according to the original axial order. .

5. The drill string vibration identification and control method based on multimodal data fusion according to claim 4 is 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 transform respectively to obtain the one-dimensional vibration signal after noise reduction 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 processing flow is as follows: （4）； in, is the time-frequency domain matrix Middle Frame No. Complex spectrum of frequency points; is the serial number of the one-dimensional vibration signal after denoising; is the window length; represents the window function; is the step length; Step 2.2.2: Take the logarithm of the time-frequency domain matrix amplitude modulus as the image intensity to form a two-dimensional time-frequency image: （5）； in, is a two-dimensional time-frequency image; A small constant to prevent logarithmic divergence; Step 2.2.3: Process the three-axis vibration signals independently according to the process of steps 2.2.1-2.2.2 to obtain three time-frequency spectrum images. 、 and , corresponding to the two-dimensional time-frequency spectrum images in the X-axis, Y-axis and Z-axis directions respectively; Step 2.2.4: Use vertical splicing to combine the three time-frequency images in sequence to form a complete vibration time-frequency image. : （6）； in, represents 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: The specific process of step 3 is as follows: Step 3.1: Define the time step The corresponding multidimensional signal is ;in, 、 、 The three-axis vibration signals are respectively X-axis, Y-axis and Z-axis; the multidimensional signal corresponding to each time step is used as the input of the bidirectional gated recurrent unit network; The bidirectional gated recurrent unit network consists of two independent unidirectional gated recurrent units with no shared parameters; the forward gated recurrent unit arrive Sequential scanning, generating time steps The forward hidden state sequence ; Backward gated recurrent unit from arrive Scan in reverse order to generate time steps The backward hidden state sequence Each unidirectional gated recurrent unit controls the retention and forgetting of historical information through the update gate and reset gate. The unidirectional gated recurrent unit is The state update is calculated as follows: （7）； （8）； （9）； （10）； in, 、 and The time steps are Update door, reset door and hide door states; is the time step The hidden state of is the time step Candidate hidden states of ; For input To the update gate The weight matrix of For input To reset gate The weight matrix of For input To the candidate hidden state The weight matrix of Hidden To the update gate The weight matrix of Hidden To reset gate The weight matrix of Hidden go through Control to candidate hidden state The weight matrix of is the Sigmoid activation function; tanh is the hyperbolic tangent activation function; represents element-wise multiplication; Concatenate the forward hidden state and the backward hidden state of each time step to form a bidirectional hidden state vector, which is the time step Hidden door status , specifically: ,in, is the hidden layer dimension of the unidirectional gated recurrent unit; By applying a pooling operation to the bidirectional hidden state vector, the variable-length temporal feature sequence is compressed into a fixed-length feature vector to obtain the temporal modal feature ; Step 3.2: Use the ResNet18-SENet network to extract image modal features; the ResNet18-SENet network consists of two parts: the ResNet18 residual module and the SENet module.

7. The drill string vibration identification and control method based on multimodal data fusion according to claim 6, characterized in that: The specific process of step 3.2 is as follows: Step 3.2.1, the vibration time-frequency image It is fed into the ResNet18-SENet network as input; first, a 7×7 convolutional layer is used to perform the first stage of convolution output, and the formula is: （11）； in, represents the convolution output of the first stage; The convolution output is Samples, Channels, positions The value of is the convolution kernel; is the position index in the convolution operation output matrix, is the vertical direction, is the horizontal direction; and are the height and width of the convolution kernel respectively; and is the index of the convolution kernel in the horizontal and vertical directions; For the The bias term of each channel; is the convolution operation; The convolution output of the first stage is batch normalized, and the calculation formula on each channel is: （12）； （13）； （14）； in, and Respectively The mean and variance of each channel; and All are Learnable parameters of channels; is a small constant that prevents division by zero; is the batch size; The convolution output of the first stage is normalized and linearly transformed. Use the ReLU activation function on the normalized output to enhance the nonlinear expression ability: （15）； in, The convolution output of the first stage is the result after ReLU activation; Then use maximum pooling to reduce the spatial size of the feature map: （16）； in, is the feature map after pooling; It is the maximum pooling operation in the first stage; is the pooling window; Step 3.2.2, the pooled feature map The network is fed into the ResNet18 residual module for deep feature extraction. The 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 first residual units connected in series. The second, third, and fourth residual blocks have the same structure, and all include a second residual unit and a first residual unit connected in series. The first residual unit and the second residual unit both use the standard unit structure, and perform 3×3 convolution, the first batch normalization, the ReLU activation function, the 3×3 convolution, and the second batch normalization in sequence to form the convolution map output: （17）； in, is the convolution map output; The input of the residual unit is the pooled feature map. , the input of 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 normalization operation; The skip connection of the first residual unit is the identity mapping: （18）； in, Mapping for skip connections; The second residual unit adds one more 1×1 convolution in the skip connection to match the number of channels or achieve downsampling: （19）； in, It is a 1×1 convolution; Each residual unit outputs for: （20）； The output of each residual unit is recalibrated through the SENet module for channel attention; the SENet module consists of two stages: compression and excitation; The compression stage performs global average pooling on each channel: （21）； in, To indicate the The global average pooling value of the channels, , is the number of channels in the convolutional network; is the global average pooling operation; is the output of the residual unit Feature map corresponding to each channel; Then the vector obtained in the compression phase is Input two fully connected layers for excitation, adopt a bottleneck structure, first reduce the dimension and then increase the dimension, and generate channel weights : （22）； in, and are the learnable weight matrices of the two fully connected layers; and is the bias vector of the two fully connected layers; Finally, the channel weighting operation is used to multiply the channel weights obtained in the excitation stage back to the original features channel by channel to achieve adaptive recalibration of the feature map: （23）； in, is the first Feature maps of channels; Weighting operation for channels; For the The channel weight of each channel; After four residual blocks, the spatial dimensions of the feature map are compressed into tensor values. ;in, 、 are all spatial dimensions, is the number of channels; Step 3.2.

3. Next, use the global average pooling operation to compress each channel: （24）； in, For the Channels at position The tensor value of ; For the The eigenvalues ​​after channel compression; The eigenvalues ​​after compression of all channels are spliced ​​together to obtain the global average pooling feature : （25）； Step 3.2.4: Project the fully connected layer to obtain the final image modality features : （26）； in, is the 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: The temporal modal features and image modality features Perform splicing operations to generate fusion features : （27）； Introduce the self-attention mechanism to calculate the attention weight: （28）； （29）； in, is the feature vector after attention mapping; is the ReLU activation function; 、 All are learnable parameter matrices; 、 is the learnable bias vector; Represents the attention weights of fusion features in different dimensions; The attention weight Multiply with the fusion feature point to complete the weighted fusion and obtain the final weighted fusion 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 fusion features are sent to the fully connected layer, and the Softmax activation function is used for nonlinear mapping to output the corresponding vibration state classification results: （31）； in, 、 All are hierarchical parameters; Indicates belonging The probability distribution of the vibration state class is calculated; the class corresponding to the highest probability is taken as the drill string vibration type under the current drilling state. The vibration types include general vibration, vortex 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 characteristic values ​​include acceleration mean, peak difference, and stick-slip ratio; Normal vibration is harmless and does not require any treatment. Just continue drilling and monitor the vibration status of the drill string. For vortex vibration, by analyzing the three-dimensional acceleration signal, the acceleration mean range and peak difference within a specific time window are extracted as the judgment basis; when the acceleration mean is between 1-2g and the peak difference is between 1-5g, it is judged as mild vortex; when the acceleration mean is between 3-5g and the peak difference is between 5-10g, it is judged as moderate vortex; and when the acceleration mean is greater than 5g and the peak difference is greater than 10g, it is judged as severe vortex. 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; between 0.5 and 1.0, it is considered medium-level stick-slip. 1.0 to 1.5 is considered advanced stick-slip; more than 1.5 is considered severe stick-slip; Based on the evaluation results, control recommendations are given: in case of moderate or severe vortex, it is recommended to reduce the rotation speed and increase the drilling pressure. When the severe vortex exceeds the limit, drilling should be stopped. In case of advanced or severe stick-slip, it is recommended to increase the rotation speed and reduce the drilling pressure. When the severe stick-slip exceeds the limit, drilling should be stopped.

Images