Real-time drilling tool jamming early warning method based on multi-modal data fusion and graph neural network

Abstract

The invention provides a real-time drilling tool jamming early warning method based on multi-modal data fusion and a graph neural network. The method is suitable for drilling tool jamming monitoring in drilling operation. According to the method, multi-modal data such as vibration, pressure and temperature are collected in real time through multiple sensors, and key features are extracted and fused into unified data representation after denoising, missing value filling and standardization processing. And then constructing a graph neural network model, capturing the space-time correlation between the sensors by using a graph convolutional network, and predicting the drilling tool jamming risk. After real-time data access, corresponding early warning is triggered based on the risk probability of drilling tool jamming, and an early warning threshold value is dynamically adjusted to adapt to different geological conditions. According to the invention, the accuracy and the real-time performance of early warning of jamming of the drilling tool are improved, and the adaptability and the robustness of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration and development, and in particular to a real-time pipe-stuck warning method based on multimodal data fusion and graph neural network (GNN), which is suitable for monitoring, warning and risk management of pipe-stuck phenomena during drilling. Background Art

[0002] In oil and gas drilling operations, stuck drill is a common and complex downhole problem. Drill stuck refers to the phenomenon that the drill bit or drill tool is blocked in the well and cannot rotate or drill smoothly. It will cause delays in the drilling process, increase operating costs, and even cause equipment damage, posing safety hazards. Drill stuck usually occurs in complex formation conditions, such as hard formations, high-pressure formations, and high-temperature environments underground. With the gradual deepening of oil and gas resource exploitation, especially in deep and ultra-deep well drilling operations, the risk and difficulty of drill stuck are gradually increasing, so real-time warning of drill stuck becomes particularly important.

[0003] Patent CN114169656A discloses a drilling stuck pipe risk warning method and system based on historical data of adjacent wells, but it relies on static historical data and lacks the ability to integrate and dynamically adjust multimodal data. Therefore, a real-time stuck pipe warning method based on multimodal data fusion and graph neural network is needed to more accurately and flexibly deal with the risk of stuck pipe in complex formations.

[0004] The current methods for early warning of stuck pipe mainly include the following categories:

[0005] 1. Single parameter monitoring method: Most stuck pipe warning methods rely on monitoring a single drilling parameter, such as drilling pressure, torque, drill bit speed, etc. This method usually determines the risk of stuck pipe based on a set fixed threshold, is difficult to adapt to complex geological conditions, has a weak response to nonlinear factors, and has limited warning effect.

[0006] 2. Sensor monitoring method: Vibration sensors, acoustic sensors and other equipment are also used in drilling operations to detect drill bit vibration, well wall displacement, etc. This method requires high reliability of sensors. However, sensors are easily damaged in extreme underground environments, data is unstable and the cost is high, which affects the accuracy of early warning.

[0007] 3. Models based on traditional machine learning: Using historical data, combined with decision trees, support vector machines and other machine learning methods to build a stuck pipe prediction model. However, these models usually only use a small number of parameters, ignore the complex relationship between multi-dimensional data, and are difficult to effectively adapt to the variability of the underground environment.

[0008] These methods are difficult to cope with complex drilling environments, and their lack of understanding of the nonlinear and multidimensional relationships among multiple factors limits the accuracy and applicability of early warning.

[0009] To overcome the above problems, the present invention proposes an innovative real-time stuck pipe warning method based on multimodal data fusion and graph neural network (GNN). Through real-time acquisition, fusion processing and graph neural network modeling of multiple sensor data, potential information of stuck pipe risk can be extracted from complex spatiotemporal associations, accurate prediction and real-time warning of stuck pipe phenomenon can be achieved, and the efficiency and safety of drilling operations can be improved. Summary of the invention

[0010] The purpose of the present invention is to provide a real-time pipe-stuck warning method based on multimodal data fusion and graph neural network, which can realize high-precision warning of pipe-stuck phenomenon by real-time acquisition of downhole data by multiple sensors, multi-level feature fusion, and construction of graph neural network model to capture the complex spatiotemporal correlation of data. This method has good adaptability and can meet actual needs by dynamically adjusting the warning threshold under different geological conditions.

[0011] The purpose of the present invention is achieved through the following technical solutions:

[0012] Step 1: Multimodal data collection and preprocessing

[0013] Multiple sensors arranged on the drilling equipment collect multimodal data downhole, including vibration, pressure, temperature, drilling speed, etc., as well as downhole environmental data such as drilling fluid flow, density, well wall status (such as displacement and lithology changes, etc.).

[0014] Due to the complexity of the underground environment, the collected raw data is often affected by noise, missing and other problems, so it is necessary to clean and preprocess the data. The specific steps are as follows:

[0015] (1) Apply denoising algorithms such as wavelet transform and Fourier transform to remove random noise and ensure the stability of the data.

[0016] (2) Use spline interpolation to fill in missing data to ensure data continuity and integrity.

[0017] (3) Use Z-score normalization or Min-Max normalization to convert data from different sensors into the same dimensional range and reduce the scale differences between different data sources.

[0018] (4) The formula for Z-score standardization is:

[0019]

[0020] Where X, μ, and σ represent the individual observation value, the mean of the overall data, and the standard deviation of the overall data, respectively.

[0021] Step 2: Multimodal data fusion

[0022] The pre-processed multimodal data are fused to fully reflect the downhole drilling status.

[0023] (1) Data alignment and feature extraction

[0024] The data of various sensors are time-aligned to ensure data consistency. Then, methods such as principal component analysis (PCA) or linear discriminant analysis (LDA) are used to reduce the data dimension and extract features that have a significant impact on the prediction of stuck pipe.

[0025] (2) Data Fusion

[0026] A weighted fusion method based on the attention mechanism is used to calculate the importance weights of data of different modalities, ensuring that the features that play a key role in the prediction of stuck drills have a higher weight in the fusion, thereby forming a comprehensive multi-dimensional feature representation.

[0027] (3) The formula of the weighted fusion method is:

[0028]

[0029] Where qi is the i-th query vector; k j is the jth input information; a ij is the attention weight between the i-th query vector and the j-th key; context i is the ith output vector.

[0030] Step 3: Build a Graph Neural Network Model (GNN)

[0031] In order to accurately extract the spatiotemporal correlation between sensor data, the present invention introduces graph neural network (GNN) for modeling. GNN constructs sensor nodes and their associations into a graph structure, so that the model can better identify complex spatiotemporal relationships.

[0032] (1) Graph structure construction

[0033] The various sensors are regarded as nodes in the graph, and the associations between nodes (such as the relationship between temperature and pressure) are regarded as edges in the graph. The feature vector of the node represents the temporal characteristics of the sensor, and the characteristics of the edge describe the strength of the association between nodes.

[0034] (2) GNN model training

[0035] The model is trained through a graph convolutional network (GCN) or a graph attention network (GAT). The GNN model realizes the interaction of node features through layer-by-layer convolution operations and node information propagation, and can effectively extract the potential features of stuck pipe risk.

[0036] (3) Adaptive Optimization

[0037] During the training process, the optimal number of network layers and neurons is determined through hyperparameter tuning and cross-validation to ensure the generalization performance of the model under different well conditions.

[0038] (4) Output stuck drill risk probability

[0039] The GNN model outputs the probability of stuck drill risk at each time point and compares it with the preset threshold. When the risk value exceeds the threshold, the corresponding warning is triggered.

[0040] (5) The formula for inter-layer propagation of the neural network in the GCN is:

[0041]

[0042] in,

[0043]

[0044] Where W (l) is the learnable weight parameter of the lth layer; σ is the nonlinear activation function; H (l) is the feature of the lth layer; when l = 0, it represents the input layer; A is the adjacency matrix, I N is the identity matrix, It represents the adjacency matrix of the undirected graph G with self-connection added. yes The degree matrix of .

[0045] Step 4: Real-time warning and dynamic threshold adjustment

[0046] The present invention also includes real-time data access and dynamic threshold adjustment to enhance the adaptability of the early warning system.

[0047] (1) Receive the latest downhole data in real time through a data stream processing system (such as Apache Kafka) to ensure that the early warning model can respond in real time.

[0048] (2) When the probability of drill stuck predicted by the GNN model exceeds a low-level threshold (e.g., 10%), a low-level warning is triggered to alert the operator; if the risk probability exceeds a high-level threshold (e.g., 20%), a high-level warning is triggered, an alarm signal is automatically issued, and drilling is stopped.

[0049] (3) The warning threshold is dynamically adjusted according to real-time changing factors such as formation pressure, well depth, and drilling speed, so that the model can adapt to different geological conditions, further reduce the false alarm rate, and improve the practicality and flexibility of the system.

[0050] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0051] (1) Through the fusion of multimodal data, the data from different sensors are fully utilized, the limitations brought by a single data source are avoided, and the accuracy of the drill stuck warning is comprehensively improved.

[0052] (2) The introduction of graph neural networks can effectively identify and learn the complex spatiotemporal correlations between different sensor data, thereby understanding the mechanism of drill sticking at a deeper level.

[0053] (3) The warning threshold is dynamically adjusted according to real-time data and environmental changes, so that the warning system can adapt to different underground conditions, reduce false alarms and improve the timeliness of warnings.

[0054] (4) The present invention can issue an early warning within a few seconds to a few minutes before a drill stuck, providing operators with sufficient reaction time to avoid serious equipment damage and production delays.

[0055] (5) Through adaptive algorithms and dynamic optimization, the system can continue to work effectively under a variety of complex geological conditions, ensuring the stability of the early warning effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the flowcharts shown in these drawings without paying any creative work.

[0057] Figure 1 The figure is an overall flow chart of the method of the present invention.

[0058] Figure 2 Schematic diagram of the neural network structure of the present invention. DETAILED DESCRIPTION

[0059] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Embodiment 1:

[0062] (1) Multiple sensors are placed on the drilling equipment to collect vibration, pressure, temperature, drilling fluid and other data in real time. After denoising, filling and standardization, the data forms a consistent input;

[0063] (2) Temporally align the data, use PCA to reduce the dimensionality, and use weighted fusion based on the attention mechanism to generate a comprehensive feature representation;

[0064] (3) Construct the sensor data into a graph structure, calculate the relationship between nodes through the graph convolutional network, and determine the optimal model structure in experiments;

[0065] (4) Real-time access to data streams, triggering advanced warnings when the risk probability exceeds 20%, and dynamically adjusting thresholds based on well conditions.

[0066] Embodiment 2:

[0067] During the drilling process of high temperature and high pressure formations, early warning is achieved through the following steps:

[0068] (1) Due to the large fluctuations in data under high temperature and high pressure environments, a multi-stage filter is used to remove high-frequency noise, and missing values ​​are filled by dynamic interpolation.

[0069] (2) After feature extraction, a weighted attention mechanism is applied to fuse the data to improve the model’s response to key features such as temperature and pressure.

[0070] (3) A graph attention network (GAT) is used to further enhance the correlation between sensor data. The importance of each sensor node in the network is automatically adjusted through the attention mechanism to capture the potential risk of drill stuck in a high-pressure and high-temperature environment.

[0071] (4) During the real-time prediction stage of the model, after the real-time data is connected, if the probability of stuck drill reaches or exceeds 20%, the system automatically triggers the advanced warning, issues an alarm signal and advises the operator to take immediate action. At the same time, the system dynamically adjusts the warning threshold according to the field feedback and changes in geological conditions to adapt to the changing drilling environment.

Claims

1. A real-time drill stuck warning method based on multimodal data fusion and graph neural network, characterized in that: The following steps are involved: Step 1, data collection and preprocessing: collect multimodal data in real time through various sensors in drilling operations, including vibration, pressure, temperature, drilling fluid density and flow rate, etc.; denoise, fill in missing values ​​and standardize the data; Step 2, multimodal data fusion: extract features and reduce the dimension of multimodal data, and use the attention mechanism to weightedly fuse the features of each modality to form a unified multi-dimensional data representation; Step 3: Build a graph neural network model: treat the multimodal data as nodes in the graph, define edges based on the correlation between sensors, build a graph structure, and use a graph convolutional network (GCN) or a graph attention network (GAT) to train the model to extract the spatiotemporal features of the multimodal data and predict the risk of stuck drill bits. Step 4: Real-time data access and early warning: Establish a data transmission interface, input real-time data into the trained GNN model, calculate the probability of drill stuck risk, and trigger corresponding early warnings based on the size of the drill stuck risk probability.

2. The real-time pipe stuck warning method according to claim 1, characterized in that: The data preprocessing in step 1 includes: (1) De-noising vibration data using wavelet transform; (2) Use interpolation to fill missing data; (3) The Z-score method was used to standardize the data.

3. The real-time pipe stuck warning method according to claim 1, characterized in that: The data fusion in step 2 adopts a weighted fusion method based on an attention mechanism to weight the features of different modalities to form fused data.

4. The real-time pipe stuck warning method according to claim 1, characterized in that: The graph neural network model in step 3 includes a graph convolutional network (GCN) or a graph attention network (GAT), which extracts potential stuck drill risk features from the relationships between sensor nodes through graph convolution operations.

5. The real-time pipe stuck warning method according to claim 1, characterized in that: The step 4 includes the step of dynamically adjusting the warning threshold value, and dynamically adjusting the warning threshold value of the probability of stuck pipe risk according to different geological conditions, drilling depth and other parameters.

6. The real-time pipe stuck warning method according to claim 1, characterized in that: In step 4, when the probability of the risk of the drill being stuck exceeds 10% of the set threshold, a low-level warning is triggered; when it exceeds 20%, a high-level warning is triggered, and an alarm signal is directly issued or drilling is automatically stopped.

7. The real-time pipe stuck warning method according to claim 1, characterized in that: The GNN model is continuously optimized and adjusted according to the stuck pipe events and characteristic data fed back during the actual drilling process to improve the accuracy and real-time performance of the early warning.

Images