Well drilling overflow early recognition method based on self-supervised learning

Abstract

The invention is suitable for the technical field of petroleum drilling engineering, and provides a drilling overflow early recognition method based on self-supervised learning, which comprises the following steps: constructing drilling multidimensional parameters into a time sequence sample by using a sliding time window, and inputting the time sequence sample into a multi-branch consistency constraint twinning network (MCSiam); the MCSiam firstly performs data enhancement on a sample, then extracts a local spatial-temporal feature and a long-term dependency relationship through an encoder composed of CNN-LSTM, and performs multi-branch comparative learning, consistency constraint optimization and predictor prediction and model optimization, and the finally trained model is used for early recognition of drilling overflow. The MCSiam is superior to a traditional method and other deep learning models in performance indexes such as accuracy, the calculation complexity is low, the applicability is wide, overflow early recognition is achieved, efficient and reliable technical support is provided for petroleum drilling safety and well control safety, and the method has good practicability and innovativeness in the field of drilling overflow early recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil drilling engineering, and particularly relates to a method for early identification of drilling fluid overflow based on self-supervised learning. Background Art

[0002] In oil drilling engineering, the early identification of drilling fluid overflow is crucial for reducing well control risks. Overflow refers to the situation where the density of the drilling fluid cannot effectively resist the pressure of formation fluids, causing oil, gas, and water in the formation to penetrate the wellbore wall, forcing the drilling fluid to overflow externally. Overflow not only seriously threatens the safety of operating personnel but also poses challenges to environmental and resource sustainability. The severity of overflow is affected by many factors such as the porosity and permeability of the formation and the magnitude of the negative pressure difference. High porosity, high permeability, and a large negative pressure difference will exacerbate the overflow phenomenon. If the invasion of oil, gas, and water cannot be detected in time and the density of the drilling fluid is not adjusted to restore balance, the pressure in the wellbore will further decrease, and the invasion speed of oil, gas, and water will accelerate, ultimately possibly leading to a surface blowout.

[0003] Traditional drilling fluid overflow detection methods establish physical models based on drilling data and issue early warnings by manually analyzing the impact of changes in physical parameters on risks. For example, Chen Ping et al. detected overflow in deepwater drilling by measuring acoustic waves, ultrasonic waves, and pressure, and could detect overflow within 3 - 8 minutes. Mei Dacheng et al. emphasized the importance of parameters such as drilling pressure, hook load, rotary table speed, and return flow rate for blowout prediction and developed a mathematical model to predict blowout events. Yue Weijie et al. analyzed the causes and manifestations of overflow and proposed an overflow monitoring scheme that comprehensively utilizes PWD (pressure while drilling), micro-flow monitoring, and comprehensive logging parameters. Zhang Xingquan et al. developed a fine drilling fluid overflow detection method based on the analysis of the changes in inlet and outlet micro-flows and constructed a flow calculation model during wellbore gas invasion based on the theory of gas-liquid two-phase flow. Liu Shujie et al. proposed to make judgments based on direct and indirect overflow precursor signals. Zhu Huangan et al. proposed to detect overflow and loss by means of drilling fluid pit level monitoring method and wellbore inlet and outlet flow monitoring method. Xia Qiang proposed to combine the flow difference between inlet and outlet, free gas content in the annulus, and logging, MWD (measurement while drilling), and LWD (logging while drilling) parameters for early detection of small wellbore overflow. In addition, with the development of machine learning technology, more and more scholars have used machine learning algorithms such as Bayesian models, support vector machines, and artificial neural networks to issue overflow risk warnings using logging parameters.

[0004] However, existing methods have obvious limitations: in terms of feature extraction, they overly rely on manual experience, and the manually selected drilling parameters cannot comprehensively and accurately present the full picture of overflow, which is extremely prone to misjudgment or missed judgment; the generalization ability is severely insufficient. Traditional physical model-based detection methods and common machine learning methods can often only adapt to specific well conditions, with poor generality and difficult promotion, and a large amount of expert resources are required for calibration; the data quality problem is prominent and the annotation cost is high. Drilling data not only has high dimensions, large noise, and contains invalid data, which urgently needs to be finely cleaned and optimized, but also the overflow samples are scarce, making the data annotation work difficult and the cost soar, which in turn leads to an unbalanced distribution of training samples and greatly weakens the robustness of the model; in addition, when dealing with drilling parameters that change dynamically over time, traditional methods are difficult to effectively capture the temporal dependence relationships therein and cannot accurately model temporal features and dynamic changes, which greatly restricts the early warning ability. To solve the above problems, the present invention proposes a method for early identification of drilling overflow based on self-supervised learning. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for early identification of drilling overflow based on self-supervised learning, aiming to solve the problems raised in the above background technology.

[0006] The purpose of the present invention is achieved through the following technical solutions: A method for early identification of drilling overflow based on self-supervised learning includes the following steps: Step 1: Input data; Using the sliding time window technique, multi-dimensional parameters in the drilling process are constructed into time series samples as input data; Step 2: Process the input data using a multi-branch consistency-constrained siamese network, and the specific steps are as follows: Data augmentation: Augment the input data, and the generated data views are used to train the model; Feature extraction: The augmented data is fed into an encoder composed of a CNN-LSTM network to extract local spatio-temporal features and long-term dependence relationships; Multi-branch contrastive learning: Construct a multi-branch contrastive learning structure, where each branch in the structure is composed of an encoder, and each encoder represents a time scale; for the time scale corresponding to each branch, perform feature learning on the data after feature extraction to obtain the feature representation under the time scale; each branch, in a contrastive learning manner, based on the data features obtained by itself, enhances the model's ability to represent features of multi-dimensional time series data at different time scales; Consistency Constraint: Introduce a consistency constraint module to calculate the similarity between the outputs of different branches in the multi-branch contrastive learning structure and use it as part of the loss function. By imposing a consistency constraint between the encoders corresponding to different augmented views, the feature representations learned by the model are kept consistent across different views; Prediction: The feature representations extracted from each encoder are fed into the predictor, and the predictor blocks the direct backpropagation of gradients to the encoder through a gradient clipping operation; Calculate Loss: Calculate the loss function through contrastive learning. The loss function combines the similarity between different augmented views and the consistency constraint term to optimize the feature representation of the model; Step 3: Downstream Task Application; Use the trained model to perform early identification of drilling fluid overflow through the learned feature representations.

[0007] Furthermore, in the data augmentation step, data augmentation methods such as smoothing, random cropping, and random imputation are used to augment the input data. The data generated by each augmentation method is fed into its corresponding encoder as a new training sample.

[0008] Furthermore, in the feature extraction step, the augmented data first enters the convolutional layer of the CNN-LSTM network. The convolutional layer is composed of multiple one-dimensional convolutional, batch normalization, and one-dimensional max-pooling modules connected in sequence. Data features are extracted through convolutional operations, batch normalization stabilizes the training process, and max-pooling performs downsampling; the data processed by the convolutional layer enters the long short-term memory network layer, and the long short-term memory network is used to process sequence data and capture long-term dependencies; subsequently, the data enters the fully connected layer, and the features are integrated using the fully connected layer, with dropout operations interspersed to prevent overfitting; finally, the data enters the mapping layer, and the mapping layer is used to map the features from the original space to a low-dimensional feature space for contrastive learning. The mapping layer contains a fully connected layer, batch normalization, and a rectified linear unit, and the final result is output after processing.

[0009] Furthermore, in the calculate loss step, the calculation formula of the loss function is as follows: ; where is the loss function, which is used to measure the difference between the model prediction value , , and the target value , , ; represents the distance metric function, which is used to calculate the distance between two inputs; is a weight coefficient used to balance the contributions of different distance metrics in the loss function.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: Enhanced feature capture ability: As an extension of the traditional contrastive learning model Simple Siamese Network (SimSiam), MCSiam increases the number of contrastive branches, with each branch representing a time scale. This multi-branch structure enables it to simultaneously process features at multiple time scales, comprehensively capture the early signals of overflow events in short-term fluctuations and long-term trends, and effectively solves the limitations of traditional methods in feature extraction.

[0011] Enhanced model performance: By introducing consistency constraints, it ensures that the feature representations learned by the model are consistent among different augmented views, not only preventing overfitting but also further enhancing the generalization ability of the model, while improving the robustness and stability of the model to features at different time scales.

[0012] Reduced annotation dependence: Self-supervised training using unlabeled data significantly reduces the dependence on labeled data, solving the problem of scarce and costly overflow data annotation. This method can also effectively utilize unlabeled samples, and through data augmentation techniques suitable for time series data, it improves the model's learning ability for time series features, thereby solving problems such as unbalanced data samples.

[0013] Obvious performance advantages: Experimental verification was carried out on multiple public datasets (such as UCR, UEA) and actual drilling data (CNPC dataset). The results show that MCSiam outperforms traditional methods and other deep learning models in performance metrics such as accuracy, demonstrating low computational complexity and wide applicability, and can perform well on different datasets. It not only improves the accuracy and generalization ability of overflow identification but also realizes the early identification of overflow, providing efficient and reliable technical support for oil drilling safety and well control safety, and having good practicality and innovation in the field of early identification of drilling overflow. Description of the Drawings

[0014] Figure 1 is the flowchart of the method of the present invention.

[0015] Figure 2 is the framework diagram of the method of the present invention.

[0016] Figure 3 is the CNN-LSTM backbone network architecture in the present invention.

[0017] Figure 4 is the schematic diagram of the sliding window in Embodiment 1 of the present invention.

[0018] Figure 5This is the comparison between the actual value and the predicted value of the overflow in Embodiment 1 of the present invention. Detailed implementation manners

[0019] For a clearer understanding of the technical features, objectives, and beneficial effects of the present invention, the technical solution of the present invention will be described in detail below, but it should not be construed as a limitation on the scope of implementation of the present invention.

[0020] The present invention provides a method for early identification of drilling overflow based on self-supervised learning: a multi-branch consistency-constrained Siamese network (MCSiam). The method flow chart and the framework diagram are as Figure 1 and Figure 2 shown, and the specific steps are as follows: Step 1: Input data; Using the sliding time window technique, multi-dimensional parameters (such as weight on bit, flow rate, pressure, etc.) during the drilling process are constructed into time series samples as input data. These time series samples contain the observed values of multiple variables changing over time. The window size is selected according to the change frequency and characteristic scale of the drilling data, and the sliding step length is determined based on the computing resources and the fineness of capturing data changes.

[0021] Step 2: Process the input data using the multi-branch consistency-constrained Siamese network (MCSiam), including the following steps: Data augmentation: The input data is processed using three different data augmentation methods, namely smoothing (such as adding noise, etc.), random crop (such as time warping, etc.), and random interpolation (such as scaling, etc.). The data generated by each augmentation method (i.e., the original data after smoothing, the randomly cropped data, and the randomly interpolated data) are respectively sent into their corresponding encoders as new training samples. The data views generated by these augmentation methods are used to train the model, which increases data diversity while retaining the time series characteristics of the data, and improves the generalization ability and robustness of the model.

[0022] Feature extraction: The enhanced data is fed into an encoder composed of a CNN-LSTM network to extract the local spatio-temporal features and long-term dependencies of the enhanced data. In the present invention, CNN-LSTM is a deep learning model that combines a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is good at processing image data, and LSTM is good at processing time series data and can capture long-term dependencies. By inputting the features extracted by CNN into LSTM, data with spatio-temporal features can be effectively processed, such as video analysis, speech recognition, natural language processing, and drilling overflow prediction. In drilling overflow prediction, the CNN-LSTM network can fuse multi-dimensional time series data during the drilling process to predict the occurrence of overflow in advance and gain time for taking corresponding measures.

[0023] As Figure 3 shown, the enhanced data, as input data, first enters the convolutional layer (ConvLayer). The convolutional layer is composed of multiple one-dimensional convolutions (Conv1d), batch normalization (BatchNormalization), and one-dimensional max pooling (MaxPool1d) modules connected in sequence. Data features are extracted through convolutional operations, batch normalization stabilizes the training process, and max pooling performs downsampling. The data processed by the convolutional layer enters the long short-term memory network layer (LSTM Layer), and the long short-term memory network (LSTM) is used to process sequence data and capture long-term dependencies. Subsequently, the data enters the fully connected layer (Fully Connected Layer), and the features are integrated using fully connected (Fully Connected), with dropout operations interspersed to prevent overfitting. Finally, the data enters the projector layer (Projector). The projector layer is used to map the features from the original space to a low-dimensional feature space for contrastive learning. This layer includes fully connected (FullyConnected), batch normalization (BatchNormalization), and rectified linear unit (Relu), and the final result (Output) is output after processing.

[0024] In MCSiam, after the input data undergoes different data augmentation processes, features are extracted through the CNN-LSTM encoder, providing a basic feature representation for subsequent contrastive learning and consistency constraints.

[0025] Multi-branch contrastive learning: Construct a multi-branch contrastive learning structure. Each branch in the structure consists of an encoder, and each encoder represents a time scale. For the time scale corresponding to each branch, perform feature learning on the data after feature extraction in step 22 to obtain the feature representation at the time scale. Each branch, in a contrastive learning manner, based on the data features it obtains, enhances the model's ability to represent features of multi-dimensional time series data at different time scales, and thus comprehensively captures the key information of overflow events in short-term fluctuations and long-term trends.

[0026] Consistency constraint: Introduce a consistency constraint module to calculate the similarity between the outputs of different branches in the multi-branch contrastive learning structure and use it as part of the loss function. By imposing consistency constraints between the encoders corresponding to different augmented views, ensure that the feature representations learned by the model are consistent between different views, improve the model's generalization ability, and prevent overfitting. The consistency constraint introduced by MCSiam effectively suppresses the situation where the model is too sensitive to specific data augmentations or noises during training through a regularization mechanism, and improves the stability of the model.

[0027] Prediction: The feature representations extracted from each encoder are fed into a predictor. The predictor prevents the gradient from directly backpropagating to the encoder through a Gradient Stop operation to avoid model degradation.

[0028] Calculate the loss: Calculate the loss function in a contrastive learning manner. The loss function combines the similarity between different augmented views and the consistency constraint term to optimize the feature representation of the model. The specific formula is as follows: ; where, is the loss function, used to measure the difference between the model prediction value , , and the target value , , ; represents a distance metric function, used to calculate the distance between two inputs; is a weight coefficient, used to balance the contributions of different distance metrics in the loss function. By combining the distances between different predicted values and the target values, as well as the distances between predicted values through weighted combination, calculate the final loss to guide the training and optimization of the model.

[0029] Step 3: Application of downstream tasks; The trained model can be used for downstream tasks such as time series classification and anomaly detection. In the drilling overflow prediction scenario, the model can predict the occurrence of overflow in advance, buy time for taking corresponding measures, and better complete various tasks through the learned feature representation.

[0030] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0031] Example 1: 11 overflow samples were collected from the drilling history data of a certain oil field of PetroChina, of which 9 were labeled and 2 were unlabeled. Each sample contained 56 features, and the sampling frequency was once per second. After sorting, the obtained sample statistical information is shown in Table 1.

[0032] Table 1 Sample statistics

[0033] use Figure 4 The sliding time window shown (window size is 500 time steps, sliding step size is 1) constructs time series samples, and normalizes the data to the range of [0,1] as the data preprocessing step of the experimental part.

[0034] During the experiment, the SimCLR method was used as a comparison. The SimCLR method first used three labeled overflow samples for initial contrastive learning training, generated positive and negative samples through data enhancement, calculated the loss value and back-propagated to update the model. Then, the trained model was used to extract features from the unlabeled data to generate pseudo labels, and then some negative samples were selected through hard negative sampling for semi-supervised contrastive learning, and the model and pseudo labels were iteratively updated.

[0035] The MCSiam method proposed in the present invention first performs multiple data enhancements (smoothing, random cropping, random interpolation) on the input data, then inputs the encoder and predictor for feature extraction and prediction, calculates the contrast loss and consistency constraints, and back-propagates to update the model. In the self-supervised training process, it is compared with multiple methods, and all files are discarded as training sets. At the end of each training cycle (epoch), 3 or 5 labeled files are used to train a K-nearest neighbor (KNN) model as a monitor to observe the effect of model training. A total of 50 epochs are trained, and the accuracy is used as the evaluation index. The experimental results are shown in Table 2. The data in brackets in Table 2 represent the gap with the baseline model, and the bold data represent the best results in each comparison.

[0036] Table 2 Self-supervised learning results

[0037] As can be seen from the self-supervised learning results shown in Table 2, the MCSiam method performs excellently on different datasets. On the CNPC dataset, when using 3 overflows as the training set, the accuracy of the MCSiam method reaches 73%; when using 5 overflows as the training set, the accuracy of the MCSiam method reaches 85.4%. These results prove the high efficiency, reliability of the present invention in the overflow identification task and its wide applicability on different datasets, providing strong technical support for safety management and risk prevention in oil drilling engineering.

[0038] In addition, this embodiment also uses the UCR and UEA datasets to verify the MCSiam method. The UCR dataset contains multiple univariate and multivariate time series datasets, and the UEA dataset contains multiple multivariate time series datasets. These datasets are widely used in time series classification and related research, covering problems in different fields. When dealing with these two datasets, the MCSiam method also performs various data augmentations on the input data, and then inputs them into the encoder and predictor for feature extraction and prediction, calculates the contrastive loss and the consistency constraint term, and updates the model through backpropagation. The results show (Table 2) that on the UCR dataset, the accuracy of the MCSiam method reaches 84%; on the UEA dataset, the accuracy reaches 71.2%. These results indicate that the present invention not only performs excellently on specific drilling datasets, but also has good generalization ability on a wider range of time series datasets, further verifying its effectiveness and reliability in different application scenarios.

[0039] Figure 5 Shows the comparison between the actual value and the predicted value of the overflow. Figure 5The abscissa represents the index of time slices; the left vertical axis represents the predicted results and the true label values; the right vertical axis represents the values of the riser pressure and the outlet flow rate; the red line represents the riser pressure. From the trend of the curve, it is relatively stable, which means that the pressure change is small; the green dashed line represents the outlet flow rate, and the curve fluctuates, reflecting the change of the outlet flow rate; the orange line represents the predicted overflow events by the model. The time points with a value of 1 are the moments when the model predicts the occurrence of overflow; the blue dashed line represents the actual overflow event label, and the time points with a value of 1 also show the occurrence moments of the actual overflow events, which can be compared with the predicted results. It can be seen from the figure that the actual value starts to become 1 around the 2700th time slice, while the predicted value has started to become 1 around 2500 in advance. At the same time, it is also observed that the outlet flow rate starts to rise around the 2700th time slice, and it is speculated that an overflow may occur at this time, which means that the model recognized the impending overflow about 200 time slices before the start of the overflow. If an alarm is issued and corresponding preventive measures are taken at this time, the occurrence of the overflow can be effectively prevented. However, due to the possible false alarms and missed alarms in actual applications, when actually used, the model can be fine-tuned and appropriate thresholds can be set to reduce the occurrence of false alarms and missed alarms.

[0040] The above are only the preferred embodiments of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicability of the patent.

Claims

1. A method for early identification of drilling overflow based on self-supervised learning, characterized in that, It includes the following steps: Step 1: Input data; Using the sliding time window technique, multi-dimensional parameters in the drilling process are constructed into time series samples as input data; Step 2: Process the input data using a multi-branch consistency-constrained siamese network. The specific steps are as follows: Data augmentation: Augment the input data to generate data views for training the model; Feature extraction: The augmented data is fed into an encoder composed of a CNN-LSTM network to extract local spatio-temporal features and long-term dependencies; Multi-branch contrastive learning: Construct a multi-branch contrastive learning structure. Each branch in the structure consists of an encoder, and each encoder represents a time scale; For each time scale corresponding to a branch, perform feature learning on the data after feature extraction to obtain the feature representation at the time scale; Each branch, in a contrastive learning manner, based on the data features obtained by itself, enhances the model's feature representation ability for multi-dimensional time series data at different time scales; Consistency constraint: Introduce a consistency constraint module to calculate the similarity between the outputs of different branches in the multi-branch contrastive learning structure and use it as part of the loss function. By imposing consistency constraints between the encoders corresponding to different augmented views, the feature representations learned by the model are kept consistent between different views; Prediction: The feature representations extracted from each encoder are fed into a predictor. The predictor blocks the direct backpropagation of gradients to the encoder through a gradient cut-off operation; Calculate the loss: Calculate the loss function in a contrastive learning manner. The loss function combines the similarity between different augmented views and the consistency constraint term to optimize the feature representation of the model; Step 3: Apply to downstream tasks; Use the trained model to perform early identification of drilling fluid loss by means of the learned feature representation.

2. The early identification method for drilling overflow based on self-supervised learning according to claim 1, wherein In the data augmentation step, the input data is augmented using smoothing, random cropping, and random imputation data augmentation methods. The data generated by each augmentation method is fed into its corresponding encoder respectively as new training samples.

3. The early identification method of drilling overflow based on self-supervised learning according to claim 1, wherein In the feature extraction step, the augmented data first enters the convolutional layer of the CNN-LSTM network. The convolutional layer is composed of multiple one-dimensional convolutions, batch normalization, and one-dimensional max pooling modules connected in sequence. Data features are extracted through convolutional operations, batch normalization stabilizes the training process, and max pooling performs downsampling; The data processed by the convolutional layer enters the long short-term memory network layer. The long short-term memory network is used to process sequence data to capture long-term dependencies; Subsequently, the data enters the fully connected layer, and the features are integrated using the fully connected layer, with random inactivation operations interspersed to prevent overfitting; Finally, the data enters the mapping layer. The mapping layer is used to map the features from the original space to a low-dimensional feature space for contrastive learning. The mapping layer includes a fully connected layer, batch normalization, and a rectified linear unit, and the final result is output after processing.

4. The early identification method for drilling overflow based on self-supervised learning according to claim 1, wherein In the step of calculating the loss, the calculation formula of the loss function is as follows: ； Among them, is the loss function, which is used to measure the difference between the predicted value of the model , , and the target value , , ; represents the distance metric function, which is used to calculate the distance between two inputs; is a weight coefficient, which is used to balance the contributions of different distance metrics in the loss function.

Images