Method and system for detecting well wall in real time, electronic equipment and storage medium

Abstract

The invention discloses a method and system for detecting a well wall in real time, electronic equipment and a storage medium. The method and system are used for judging well drilling abnormity by analyzing the shape of rock debris. The method for detecting the well wall in real time comprises the steps that multi-mode underground real-time data are obtained; inputting the multi-modal underground real-time data into a pre-established anomaly recognition model, performing streaming reasoning, and outputting a recognition result; performing time sequence integration and trend analysis according to the identification result to obtain an analysis result; generating real-time alarm information according to a preset decision threshold rule and the analysis result; the anomaly recognition model is obtained through the following training method: acquiring multi-modal underground historical data; performing time synchronization processing and space positioning processing on the multi-mode underground historical data to obtain processing data; performing exception annotation on the processed data according to expert knowledge, and constructing an annotated data set; and training the multi-modal fusion anomaly recognition model according to the labeled data set to obtain an anomaly recognition model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drilling operation detection, and in particular to a method and system for real-time detection of well wall, an electronic device and a storage medium. BACKGROUND

[0002] In the field of downhole operation detection of oil and gas drilling engineering, mainly around a single data dimension or simple rules are developed to achieve preliminary identification of abnormalities such as well wall spalling, sticking precursor, fracture development, etc. Specifically, there are single-dimensional threshold alarms based on drilling operation parameters (such as drilling pressure, rotation speed, torque, standpipe pressure, etc.), which trigger alarms by setting a fixed safety range for a specific parameter; mud logging and debris inspection relying on human experience, or using simple image features for auxiliary judgment; using single logging curves such as acoustic wave, resistivity, gamma or caliper for independent anomaly identification; and constructing fixed judgment rules based on engineering experience, or using single-modal machine learning anomaly identification models (such as only based on acoustic data or only based on image data) for isolated anomaly judgment.

[0003] However, the visual data and acoustic data (such as drill string vibration, acoustic wave attenuation, etc.), electromagnetic resistivity data, downhole temperature and pressure data, and drilling operation parameters collected in the above prior art lack effective correlation mechanisms, cannot capture the implicit correlation between cross-modal (such as the corresponding relationship between the texture anomaly of the well wall fracture in the image and the enhanced acoustic attenuation at the same depth), and the anomaly features of a single data dimension are easily disturbed (such as camera mud blockage leading to visual failure, mud noise masking acoustic anomalies, etc.), which further leads to high false negative rate and high false positive rate, reducing the reliability and efficiency of downhole anomaly detection. SUMMARY

[0004] To solve the above technical problems, the present application provides a method and system for real-time detection of well wall, an electronic device and a storage medium for analyzing debris shape to judge drilling anomalies.

[0005] The technical solutions provided in the present application are described as follows: The first aspect of the present application provides a method for real-time detection of well wall, comprising: obtaining multi-modal downhole real-time data, the multi-modal downhole real-time data comprising at least two of current real-time collected debris and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data; inputting the multi-modal downhole real-time data into a pre-established anomaly identification model, performing stream inference and outputting an identification result; performing time series integration and trend analysis according to the identification result to obtain an analysis result; generating real-time alarm information according to a pre-set decision threshold rule and the analysis result; The anomaly identification model is obtained by the following training method: Obtain multi-modal downhole historical data, the multi-modal downhole historical data including at least two of the following: historical collected debris and borehole imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data; Perform time synchronization processing and spatial positioning processing on the multi-modal downhole historical data to obtain processed data; Annotate the processed data according to expert knowledge to construct a labeled data set; Train a multi-modal fusion anomaly identification model according to the labeled data set to obtain an anomaly identification model.

[0006] Optionally, the time synchronization processing and spatial positioning processing on the multi-modal downhole historical data comprises: According to a time synchronization protocol, write a respective timestamp to the multi-modal downhole historical data, the timestamp being a high-precision timestamp of milliseconds; According to a nearest neighbor interpolation or dynamic time warping algorithm, perform time alignment on the multi-modal downhole historical data based on the timestamp; Obtain a mapping relationship between the multi-modal downhole historical data and the vertical depth data of the target drilling; According to the mapping relationship, normalize the multi-modal downhole historical data to a preset borehole coordinate system.

[0007] Optionally, the annotating the processed data according to expert knowledge to construct a labeled data set comprises: Annotate the processed data twice independently according to expert knowledge to obtain two independent annotation results; When the two independent annotation results are different, merge and adjudicate the two independent annotation results to obtain an adjudication result; Integrate the two independent annotation results or the adjudication result, and construct a labeled data set.

[0008] Optionally, the inputting the multi-modal downhole real-time data into a pre-established anomaly identification model, performing stream inference and outputting an identification result comprises: According to a preset sliding window, segment the multi-modal downhole real-time data to obtain a plurality of segment results; Input the plurality of segment results into a pre-established anomaly identification model and perform forward inference to obtain an identification result.

[0009] Optionally, the time series integration and trend analysis according to the identification result to obtain an analysis result comprises: Smooth the identification result according to an exponential moving average method or a median filtering algorithm to obtain a smoothed result; Identify a turning point in the smoothed result according to a cumulative sum control chart method or a Bayesian change point detection algorithm, the turning point representing a change in an abnormal state; Calculate a trend indicator of the smoothed result according to the turning point, the trend indicator representing a trend of dynamic changes in an abnormal event of the smoothed result; Perform trend analysis according to the trend indicator to obtain an analysis result.

[0010] Optionally, after the training of the multi-modal fusion abnormality identification model according to the labeled data set to obtain the abnormality identification model, the method further includes: Obtain feedback data of the user on the real-time alarm information, the feedback data including confirmation of the alarm, rejection of the alarm, and / or correction of an abnormal category in the real-time alarm information; Incrementally fine-tune and update the abnormality identification model according to the feedback data to obtain an updated abnormality identification model.

[0011] Optionally, the decision threshold rule includes a hysteresis loop rule formed by a trigger threshold strategy, a duration threshold strategy, and a contact threshold strategy, and a multi-modal consistency voting rule.

[0012] The second aspect of the application provides a system for real-time detection of a well wall, including: A real-time data unit configured to obtain multi-modal downhole real-time data, the multi-modal downhole real-time data including at least two of currently collected debris and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological property data; An identification unit configured to input the multi-modal downhole real-time data into a pre-established abnormality identification model, perform stream-based reasoning, and output an identification result; An analysis unit configured to perform time series integration and trend analysis according to the identification result to obtain an analysis result; A generation unit configured to generate real-time alarm information according to a preset decision threshold rule and the analysis result; The abnormality identification model is obtained by the following training method: obtaining multi-modal downhole historical data, the multi-modal downhole historical data including at least two of historically collected debris and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological property data; performing time synchronization processing and spatial positioning processing on the multi-modal downhole historical data to obtain processed data; performing abnormal labeling on the processed data according to expert knowledge to construct a labeled data set; training a multi-modal fusion abnormality identification model according to the labeled data set to obtain an abnormality identification model.

[0013] The third aspect of the present application provides an electronic device, the electronic device comprising: a processor, a memory, an input / output unit and a bus; The processor is connected with the memory, the input / output unit and the bus; The memory stores a program, and the processor invokes the program to execute the method for detecting the well wall in real time in the first aspect and any optional one of the first aspect.

[0014] The fourth aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores a program, and the program executes the method for detecting the well wall in real time in the first aspect and any optional one of the first aspect when executed on a computer.

[0015] From the above technical solutions, the present application has the following beneficial effects: The application provides a method for real-time detection of well walls, which comprises the following steps: first, acquiring multi-modal downhole real-time data, wherein the multi-modal downhole real-time data comprises at least two of the following: real-time collected debris and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data; inputting the multi-modal downhole real-time data into a pre-established abnormality recognition model, performing stream inference and outputting a recognition result; performing time series integration and trend analysis according to the recognition result to obtain an analysis result; generating real-time alarm information according to a preset decision threshold rule and the analysis result; the abnormality recognition model is obtained through the following training method: acquiring multi-modal downhole historical data, wherein the multi-modal downhole historical data comprises at least two of the following: historically collected debris and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data; performing time synchronization processing and spatial positioning processing on the multi-modal downhole historical data to obtain processed data; performing abnormality labeling on the processed data according to expert knowledge to construct a labeled data set; training a multi-modal fusion abnormality recognition model according to the labeled data set to obtain the abnormality recognition model. The method for real-time detection of well walls provided by the application uses expert knowledge to construct a high-quality labeled data set in the training stage, and trains an abnormality recognition model that can autonomously learn the internal correlation among multi-modal data such as vision, acoustics and electromagnetism through mechanisms such as shared attention, effectively solving the problems of single data source being easily disturbed and characteristics being one-sided, and significantly improving the accuracy and robustness of recognition under complex working conditions. Based on the trained abnormality recognition model, at least two kinds of multi-modal real-time data are obtained from the following: debris and well wall imaging, acoustics, electromagnetism, temperature and pressure, drilling parameters and geological attributes, and stream inference is performed relying on the abnormality recognition model, then time series integration and trend analysis are combined, the development dynamics of abnormalities can be captured, and structured alarms can be generated through intelligent decision thresholds, which can provide more timely and accurate decision-making basis for drilling operations, reduce high false negative rate and high false positive rate, and significantly improve the reliability and efficiency of downhole abnormality detection. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the application, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0017] Figure 1 An embodiment flowchart of the method for real-time detection of well walls provided by the application; Figure 2 Another embodiment flowchart of the method for real-time detection of well walls provided by the application; Figure 3 Another embodiment flowchart of the method for real-time detection of well walls provided by the application; Figure 4 Another embodiment flow chart of the method for real-time detection of well wall provided by the present application is shown in FIG. 7; Figure 5 Another embodiment flow chart of the method for real-time detection of well wall provided by the present application is shown in FIG. 7; Figure 6 Another embodiment flow chart of the method for real-time detection of well wall provided by the present application is shown in FIG. 7; Figure 7 An embodiment structure diagram of the system for real-time detection of well wall provided by the present application is shown in FIG. 8; Figure 8 An embodiment structure diagram of the electronic device provided by the present application is shown in FIG. 9. DETAILED DESCRIPTION

[0018] The present application provides a method, system, electronic device and storage medium for real-time detection of well wall, which is used for judging drilling anomalies by analyzing the shape of cuttings.

[0019] It should be noted that the method for real-time detection of well wall provided by the present application can be applied to a terminal, a system, and a server. For example, the terminal can be a smart phone or a computer, a tablet computer, a portable computer terminal, or a desktop computer.

[0020] It should be noted that the terms "first", "second", etc. in the specification and drawings are used to distinguish similar objects, and do not necessarily indicate a specific order, sequence or importance.

[0021] Please refer to Figure 1 The present application first provides an embodiment of a method for real-time detection of well wall, which comprises: Step 101, acquiring multi-modal downhole real-time data, the multi-modal downhole real-time data comprising at least two of current real-time collected cuttings and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data; In the embodiment, the system synchronously collects various types of raw data by integrating various sensor arrays deployed in downhole, drill string and surface system. Specifically, the debris and borehole imaging data are captured in real time by downhole high-definition camera and borehole imaging instrument, both of which have the ability to resist mud pollution, and the borehole imaging instrument generates borehole images at a preset sampling frequency; the acoustic data are obtained by drill string vibration sensor and acoustic logging instrument, the drill string vibration sensor collects drill string acceleration, impact and lateral vibration signals in real time, and the acoustic logging instrument records full waveform acoustic data and parameters such as sound velocity and attenuation coefficient; the electromagnetic data are collected by inductive resistivity sensor to obtain inductive phase, amplitude and apparent resistivity data; the temperature and pressure data are detected by downhole pressure sensor and temperature sensor to obtain static pressure, dynamic pressure, temperature gradient and equivalent circulating density in real time; the drilling operation parameters are directly obtained by the surface control system, including drilling pressure, rotation speed, torque, mechanical drilling speed, pump speed / displacement, standpipe pressure, etc.; and the geological attribute data are recorded in synchronization with the real-time measurement results of natural gamma logging instrument and caliper to obtain natural gamma value, caliper data and interpreted horizon / lithology information. During the process of real-time data acquisition, various types of data are also uploaded to the ground processing unit in real time through the downhole data transmission link, ensuring that the data timeliness meets the real-time detection requirements.

[0022] Step 102, inputting the multi-modal downhole real-time data into the pre-established abnormality recognition model, performing stream inference and outputting the recognition result; In the embodiment, the multi-modal downhole real-time data can also be preprocessed before stream inference. Different preprocessing methods are used for different modal data. Specifically, the debris and borehole imaging data are subjected to distortion calibration, illumination and color normalization and denoising processing to avoid the influence of downhole light changes; the acoustic data are subjected to detrending and band-pass filtering; and the electromagnetic, temperature and pressure and drilling operation parameters are subjected to outlier detection, thereby ensuring the data quality of different modal data. Subsequently, the preprocessed multi-modal downhole real-time data are batch input into the pre-trained abnormality recognition model for one-time forward inference. Specifically, the stream inference is carried out using a sliding window processing mechanism, in which multiple sliding windows are set, and the length of each sliding window is 30-60 seconds and the step is 5-10 seconds, and then the preprocessed multi-modal data are synchronously input into the abnormality recognition model in each sliding window, the abnormality recognition model extracts features of each modal through the built-in multi-modal encoder, realizes feature correlation through the fusion layer, and finally outputs the recognition result containing abnormality category, abnormality probability and abnormality position.

[0023] It should be noted that the anomaly recognition model is obtained by the following training method: obtaining multi-modal downhole historical data, the multi-modal downhole historical data including at least two of historical collected debris and borehole imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data; performing time synchronization processing and spatial positioning processing on the multi-modal downhole historical data to obtain processed data; labeling the processed data according to expert knowledge to construct a labeled data set; training the multi-modal fusion anomaly recognition model according to the labeled data set to obtain the anomaly recognition model. The recognition ability of the anomaly recognition model comes from the deep mining of the multi-modal correlation law in its training stage. The training of the anomaly recognition model starts from obtaining multi-modal downhole historical data homologous to real-time data, and after strict time synchronization and spatial positioning processing, time-space aligned processed data are formed. The key step is that a joint team composed of geology and engineering experts will accurately label the anomalies in these processed data, including type, degree and spatial location, to construct a high-quality "labeled data set". Finally, a multi-modal fusion anomaly recognition model based on the Transformer architecture is supervised trained using the labeled data set. The anomaly recognition model learns the complex corresponding relationship between different modal data through its internal co-attention mechanism and other structures, thereby obtaining strong anomaly recognition ability.

[0024] Step 103, time series integration and trend analysis are performed according to the recognition result to obtain an analysis result; In this embodiment, first, the recognition result is cached according to the corresponding time sequence, then a lightweight time series network modeling can be used to capture the correlation of each data in the recognition result. Subsequently, on the basis of obtaining the correlation, trend analysis is performed, according to the correlation and the abnormal category, abnormal probability and abnormal position in the recognition result extracted in step 104, the situation that will occur at the place where the anomaly has occurred in the future preset time period and the possibility that the related anomaly will occur at the place where the anomaly has not occurred are inferred, and finally the above analysis and inference are integrated to form a comprehensive analysis result containing abnormal trend.

[0025] Step 104, generating real-time alarm information according to the preset decision threshold rule and the analysis result.

[0026] In this embodiment, it should be noted that the decision threshold rule includes a trigger threshold strategy, a continuous threshold strategy, a hysteresis loop rule formed by combining a contact threshold strategy, and a multi-modal consistency voting rule. Specifically, in step 103, the multi-modal downhole real-time data has been put into multiple sliding windows for stream inference, and the identification result includes an abnormal probability. Based on this, the trigger threshold T1, the release threshold T0, and the sliding window number threshold K of the abnormal probability are first set. If the abnormal probability in the continuous K sliding windows in the analysis result is greater than or equal to the trigger threshold T1, it is determined that the condition for triggering the alarm is reached, that is, it is indicated that the current multi-modal downhole real-time data can be used to determine that an abnormality has occurred or will occur in the well, and an alarm information is sent to the staff. If the abnormal probability is less than or equal to the release threshold T0, it is determined that the condition for releasing the alarm is reached, that is, it is indicated that the abnormality has been eliminated or the possible abnormality has been eliminated.

[0027] When the alarm is triggered, a structured real-time alarm information is automatically generated, which covers the abnormal type, the occurrence depth and / or interval, the spatial positioning details, the severity, the abnormal development trend, the comprehensive confidence, and the recommended treatment scheme. The alarm information is displayed in real time through the board of the drilling operation control platform, and is pushed to the field operators in the form of sound and light alarm, short message notification, etc., to ensure that the abnormality is timely acquired and disposed. If the alarm release condition is met, the current alarm state is automatically cleared, and information such as alarm duration and disposal result is recorded.

[0028] The method for real-time detection of well wall in the application uses expert knowledge to construct a high-quality labeled data set in the training phase, and trains an abnormality identification model that can learn the internal correlation between multi-modal data such as vision, acoustics, and electromagnetism through a co-attention mechanism, effectively solving the problem of single data source being easily disturbed and feature being one-sided, and significantly improving the accuracy and robustness of identification under complex working conditions. Based on the trained abnormality identification model, at least two kinds of multi-modal real-time data are obtained from “cuttings and well wall imaging, acoustics, electromagnetism, temperature and pressure, drilling parameters, and geological properties”, and stream inference is performed relying on the abnormality identification model. Then, through time sequence integration and trend analysis, the development dynamics of the abnormality can be captured, and a structured alarm can be generated through intelligent decision threshold, which can provide more timely and accurate decision basis for drilling operations, reduce the high false alarm rate and high false alarm rate, and significantly improve the reliability and efficiency of downhole abnormality detection.

[0029] Reference Figure 2 According to some embodiments of the application, the time synchronization processing and spatial positioning processing of the multi-modal downhole historical data in the process of training the abnormality identification model can include, but is not limited to, the following: Step 201, according to the time synchronization protocol, the multi-modal downhole historical data is written into the corresponding time stamp, and the time stamp is a high-precision time stamp of millisecond level; In step 202, the multi-modal downhole historical data is time-aligned based on the timestamps according to the nearest neighbor interpolation or dynamic time warping algorithm. In step 203, a mapping relationship between the depth data corresponding to the multi-modal downhole historical data and the vertical depth data of the target drilling well is obtained. In step 204, the multi-modal downhole historical data is normalized into a preset borehole coordinate system according to the mapping relationship.

[0030] In this embodiment, at the source of data acquisition, a high-precision time server is deployed, and time synchronization protocols such as NTP or PTP are used to synchronize the clocks of data acquisition devices distributed in different physical locations such as downhole tools, ground sensors, and logging devices. When a data point is acquired, the corresponding acquisition device immediately marks it with a millisecond-level precision timestamp obtained from the time server. The time of the time server is usually selected as the high-precision clock of the ground drilling control system. For example, for each frame of image data acquired by the formation imaging data, a timestamp in the format of "2024-05-10 14:30:00.123" is synchronously written; for each 10 ms of acoustic data, a millisecond-level timestamp is also written, realizing the unified marking of the time dimension of various data.

[0031] Due to the different sampling frequencies of different sensors, the original data is a plurality of different time series, which should be aligned to the same time axis. The time series of the multi-modal downhole historical data with timestamps written above is combed, the sampling frequency difference of various data is counted, and the time interval characteristics of different modal data are determined. For data with small sampling frequency difference and no obvious time sequence drift, the nearest neighbor interpolation algorithm is used for time alignment: the timestamp sequence corresponding to the data with the highest sampling frequency is taken as the reference time axis, for other modal data, the nearest neighbor target data timestamp of each reference time point is found, the value of the target data is assigned to the corresponding reference time point, and the matching of different modal data at the same time node is realized. If the data has obvious time sequence drift, a dynamic time warping algorithm such as DTW algorithm is used, that is, by calculating the similarity between different modal data time series, the optimal time matching path is found, one of the time series is stretched or compressed, and the time nodes of the two sequences are accurately matched. For example, when the acoustic data and the electromagnetic data have a 50 ms time sequence deviation due to different transmission paths, the electromagnetic data time axis is adjusted by the DTW algorithm, so that they are completely aligned at millisecond-level time points such as "2024-05-10 14:30:00.123" and "2024-05-10 14:30:00.133", ensuring that the multi-modal data at the same time can be correlated and analyzed.

[0032] Then, from the MWD or wireline logging data, the depth of each multi-modal downhole history data corresponding to the length of the drill bit or instrument along the actual well trajectory is obtained. At the same time, through the well trajectory calculation anomaly recognition model, using the inclination and azimuth angle data, each depth point is calculated and mapped to the corresponding vertical depth, i.e. the depth of the point in the vertical direction relative to the wellhead. After obtaining the depth data and vertical depth data, a mapping relationship table or a fitting mapping function is constructed to ensure that any one depth data can quickly query or calculate the corresponding vertical depth data through the mapping relationship, and then a spatial index is constructed to "straighten" the curved well trajectory and project it to the vertical profile, so that all data can be effectively compared and analyzed in the vertical depth dimension.

[0033] After obtaining the mapping relationship between the depth data and the vertical depth data, all modal data is uniformly converted into a preset three-dimensional wellbore coordinate system with the wellhead as the origin, which usually takes the vertical depth as the Z axis and the north and east directions as the X and Y axes. Finally, the positioning information of each multi-modal downhole history data is converted from the original local coordinate system dependent on specific instrument tools to this global and unified wellbore coordinate system, realizing the normalization of spatial dimension, ensuring the comparability and correlation of different modal data in the same spatial position, and laying a foundation for providing accurate abnormal position alarm information in the future.

[0034] Reference Figure 3 According to some embodiments of the present application, the process of training the anomaly recognition model to label the processed data according to expert knowledge can specifically include, but is not limited to, the following: Step 301, according to expert knowledge, the processed data is labeled twice independently, and two independent labeling results are obtained; Step 302, when there is a difference between the two independent labeling results, the two independent labeling results are combined and adjudicated to obtain an adjudication result; Step 303, integrate the two independent labeling results or the adjudication result, and construct a labeling data set.

[0035] In the embodiment, two identical processing data are respectively provided to two different annotation expert groups, and one independent annotation is respectively performed to obtain two independent annotation results. In the process of independent annotation, a unified annotation guide and example library should be used. The annotation guide clearly defines the abnormal class definition, abnormal degree division standard, and abnormal region marking specification. For example, "well wall peeling" needs to meet the conditions of an area of ≥5% of the falling area in the well wall imaging graph and a sudden change in acoustic data. The light abnormality is an abnormal influence range <10% well section / parameter fluctuation <15%, the moderate abnormality is 10%~30% / 15%~30%, and the serious abnormality is >30% / >30%. The image class data is marked by pixel-level mask, and the time series class data is marked by time window / depth window starting and ending point, etc. The example library contains typical data samples and standard annotation results of various abnormalities.

[0036] When the system detects that the annotations for the same data segment or region are inconsistent, a ruling process should be started, that is, the two independent annotation results are provided to a third-party ruling group that is different from the above two annotation expert groups. The third-party ruling group can make a comprehensive judgment combined with the correlation characteristics of multi-modal data. The final judgment made by the third-party ruling group is the ruling result, which is considered as the most authoritative annotation. Finally, if there is no difference between the two independent annotation results, the two independent annotation results are integrated into one data, thereby constructing an annotation data set. If there is a difference between the two independent annotation results, a ruling result is generated by ruling, and the ruling result is integrated to construct an annotation data set.

[0037] Referring to Figure 4 According to some embodiments of the present application, the inputting of the multi-modal downhole real-time data into the pre-established abnormality recognition model in step 102 for stream inference and outputting of the recognition result can specifically include, but is not limited to, the following: Step 401, segmenting the multi-modal downhole real-time data according to a pre-set sliding window to obtain a plurality of segmentation results; Step 402, inputting the plurality of segmentation results into the pre-established abnormality recognition model and performing forward inference to obtain a recognition result.

[0038] In this embodiment, a fixed time length sliding window is set in the memory, that is, a sliding sliding window, and it is continuously pushed forward with a shorter step, for example, the time length of each sliding window mentioned in step 102 is 30-60 seconds, and the step is 5-10 seconds, which can balance the data volume and timeliness, too short may lead to incomplete features, too long increases the delay of early warning, and ensures that there is data overlap between adjacent sliding windows to avoid missing abnormal features across sliding windows. Each time the sliding window slides, the system intercepts the latest multi-modal real-time data in the sliding window as an independent data segment. During segmentation, the system synchronously performs data integrity verification: if there is a missing in the current sliding window, the missing part is completed through the data compensation algorithm; if the missing data accounts for more than 20%, mark the segmentation result as “to be completed”, and trigger the data retransmission request, to ensure that each segmentation result finally output contains complete multi-modal data, and each type of data is time-aligned in the sliding window, and finally a plurality of segmentation results are obtained.

[0039] Then the plurality of segmentation results are sequentially input into the deployed abnormality recognition model, first the data of different modalities in the segmentation result is adapted and processed, for example, for the imaging data of cuttings, the image frames in the sliding window are normalized according to the preset size, and the preprocessing parameters are loaded for pixel value standardization, and converted into a tensor format that can be received by the visual encoder of the abnormality recognition model; for acoustic, electromagnetic and other time series data, the time series signals in the sliding window are truncated or zero-padded according to a fixed length, converted into a two-dimensional feature map through time-frequency transformation, and then input into the waveform encoder after standardization processing; for structured data such as drilling parameters, the parameter values in the sliding window are normalized according to the statistical quantity, and converted into a vector form. After the adaptation processing is completed, the forward inference is performed, that is, each modality encoder in the abnormality recognition model extracts features, for example, the visual encoder extracts texture features and shape features of the imaging data through convolution layers or Transformer blocks, the waveform encoder extracts frequency features and amplitude mutation features of the time series data through one-dimensional convolution and time series attention, and the structured encoder extracts trend features of the parameters through a multi-layer perceptron; then the correlation between the features of each modality is established through a cross-modality co-attention mechanism, for example, the crack features of the imaging data are associated with the attenuation features of the acoustic data, and a fusion feature vector is output through a gating weighting, for example, the weight of the visual feature is reduced when the camera is contaminated based on the modality reliability; then the fusion feature vector is input into the classifier and the regressor in the abnormality recognition model, and finally a comprehensive recognition result is output. The recognition result is a structured output, which usually includes the type probability of the anomaly, the predicted severity score, and the spatial positioning information of the anomaly in the well section.

[0040] Referring to Figure 5 According to some embodiments of the present application, the time series integration and trend analysis according to the recognition result in step 103 can specifically include, but is not limited to, the following: Step 501, according to the exponential moving average method or median filtering algorithm, the identification result is smoothed to obtain a smoothed result; Step 502, according to the cumulative sum control chart method or Bayesian change point detection algorithm, the turning point in the smoothed result is identified, and the turning point represents the change of abnormal state; Step 503, according to the turning point, the trend index of the smoothed result is calculated, and the trend index represents the trend of the dynamic change of the abnormal event of the smoothed result; Step 504, according to the trend index, the trend analysis is carried out to obtain an analysis result.

[0041] In this embodiment, in order to suppress the instantaneous fluctuation of the identification result in a single sliding window, so as to retain the real abnormal change trend, the identification result should be smoothed. If the fluctuation is mainly caused by Gaussian noise, the exponential moving average method is used, which gives higher weight to recent data, and can better reflect the latest change trend while smoothing, for example, the formula of the exponential moving average method is: current smoothed result = smoothing coefficient x current identification result + (1-smoothing coefficient) x previous smoothed result, which can be: the "fracture development" abnormal probability of a certain sliding window is 0.7, the previous smoothed result is 0.6, and the smoothing coefficient is 0.2, then the current smoothed result is calculated as: 0.2x0.7+0.8x0.6=0.62; if there are obvious impulse outliers in the sequence, the median filtering algorithm is used, that is, the median of all values in the neighborhood of a certain data point is used to replace the value of the point, which can effectively filter out such singular values, for example, the "well wall collapse" abnormal probability of 5 sliding windows is 0.5, 0.8, 0.6, 0.7, 0.9, and the order is 0.5, 0.6, 0.7, 0.8, 0.9, and the median 0.7 is the smoothed result of the third sliding window.

[0042] After obtaining the smoothed result, whether an abnormality occurs is mainly judged according to the turning point in the smoothed result. The cumulative sum control chart method or the Bayesian change point detection algorithm can be used, wherein the cumulative sum control chart method calculates the sum of the small deviations between the sequence values and the target values, and when the cumulative sum exceeds the control limit, it is determined that a change point occurs, which is very sensitive to small and continuous drift. The Bayesian change point detection algorithm calculates the posterior probability of each point in the sequence becoming a change point from the probability point of view, and can give the uncertainty measure of the change point position in the form of probability, which is more suitable for processing complex change patterns.

[0043] After the turning point is identified, the latest data segment after the detected turning point is selected or compared with a previous stable period, and a series of quantifiable trend indicators are calculated. These indicators include but are not limited to: linear regression is performed on the abnormal probability sequence in this time period, and its slope represents the speed and direction of change. If the slope is greater than 0, it means that the abnormal probability increases over time, i.e. the trend is deteriorating. If the slope is less than 0, it means that the abnormal probability decreases over time, i.e. the trend is mitigating. If the slope is equal to 0, it means that the abnormal probability is stable; calculate the expansion rate of the abnormal area in space, for example, for image classification results, extract the abnormal area area or depth interval length of each sliding window in the trend segment, calculate the change between adjacent sliding windows, and then take the average to get the expansion rate. A positive rate indicates that the abnormal area is expanding, a negative rate indicates that the abnormal area is shrinking, and a rate close to 0 indicates that the area is stable; or analyze the coordination mode between different modal abnormal indicators, for example, calculate the correlation coefficient of different modal smoothing results in the trend segment. The correlation coefficient close to 1 indicates that the multi-modal anomaly is synchronized to enhance, i.e. the coordination is deteriorating, the correlation coefficient close to 0 indicates that the multi-modal anomaly is unrelated, and the correlation coefficient is negative. The trend of multi-modal anomaly is opposite.

[0044] Finally, compare the calculated trend indicators with the preset empirical threshold or rule to perform trend analysis. For example, if the abnormal probability slope is positive and exceeds the threshold, and the abnormal area expansion rate is accelerated, the analysis result is "trend deterioration"; if the slope is close to zero, it is determined to be "trend stable". The final analysis result is a structured judgment that combines the direction, speed and intensity of the trend, and clearly indicates the current development stage of the abnormal event.

[0045] Reference Figure 6 In practice, after obtaining the abnormality recognition model through the first training, in order to further improve the accuracy and reliability of the detection process, the abnormality recognition model should be optimized and updated, that is, after training the multi-modal fusion abnormality recognition model according to the labeled data set to obtain the abnormality recognition model, it can specifically include but is not limited to the following: Step 601, obtaining feedback data of the user on the real-time alarm information, the feedback data including confirming the alarm, rejecting the alarm, and / or correcting the abnormal category in the real-time alarm information; Step 602, incrementally fine-tuning and updating the abnormality recognition model according to the feedback data to obtain an updated abnormality recognition model.

[0046] In the embodiment, after generating the real-time early warning information, the early warning details are pushed to the user through the interactive interface of the drilling operation control platform, including the abnormal type, position, severity, trend analysis result and multi-modal data corroboration. The interface sets a special feedback entrance and provides three standardized feedback options: "confirm alarm" indicates that the real-time early warning information is correct, "dismiss alarm" indicates that the real-time early warning information is a false alarm, and "correct abnormal category" to correct the recognition result. After the user submits the feedback, the feedback content, feedback timestamp, feedback personnel identity information are automatically recorded, and the feedback data are associated with the corresponding early warning information, corresponding multi-modal downhole real-time data and abnormal recognition model version for storage.

[0047] The automatically recorded content is used as a new training sample, and incremental fine-tuning and updating are regularly performed according to the new training sample, that is, the current deployed abnormal recognition model is used as the basis, the bottom feature extraction layer of the abnormal recognition model is frozen, so as to retain the learned general features, and only the top classification head and fusion layer parameters are fine-tuned; at the same time, a smaller learning rate is used to avoid damaging the existing capabilities of the abnormal recognition model; and the training target focuses on correcting the bias reflected by the feedback data. During the fine-tuning process, the performance change of the abnormal recognition model is observed in a timely manner to ensure that the corresponding error rate of the feedback data is reduced while the recognition accuracy of other abnormal categories is not reduced.

[0048] After fine-tuning, the indicators of the updated model after updating are compared with those of the model before updating. If the comprehensive performance of the updated model is improved and the error rate of the feedback data corresponding to the scene is reduced by ≥30%, the updated model is determined as the updated abnormal recognition model, and the update content is recorded through the model version management system. Then, the updated model is replaced with the original model, and the old model version is retained for backtracking analysis. By continuously receiving feedback and incremental fine-tuning, the model can continuously adapt to new drilling scenes and abnormal features, and gradually improve the recognition accuracy.

[0049] The above embodiment describes the method for detecting the well wall in real time provided in the application. The system for detecting the well wall in real time, the electronic device and the storage medium provided in the application are described as follows: Please refer to Figure 7 An embodiment of a system for detecting the well wall in real time is provided in the application, and the system / device comprises: The real-time data unit 701 is configured to acquire multi-modal downhole real-time data, and the multi-modal downhole real-time data comprises at least two of the currently collected rock debris and well wall imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters and geological attribute data. The identification unit 702 is configured to input the multi-modal downhole real-time data into a pre-established abnormality identification model, perform stream inference, and output an identification result; the abnormality identification model is obtained by the following training method: obtaining multi-modal downhole historical data, the multi-modal downhole historical data including at least two of historical collected debris and borehole imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological attribute data; performing time synchronization processing and spatial positioning processing on the multi-modal downhole historical data to obtain processed data; performing abnormal labeling on the processed data according to expert knowledge to construct a labeled data set; training the multi-modal fusion abnormality identification model according to the labeled data set to obtain the abnormality identification model; The analysis unit 703 is configured to perform time series integration and trend analysis according to the identification result to obtain an analysis result. The generation unit 704 is configured to generate real-time alarm information according to a preset decision threshold rule and the analysis result.

[0050] Optionally, the training method in the identification unit 702 is specifically configured to: According to a time synchronization protocol, the multi-modal downhole historical data is written with respective corresponding time stamps, and the time stamps are all high-precision time stamps at the millisecond level; According to a nearest neighbor interpolation or dynamic time warping algorithm, the multi-modal downhole historical data is time-aligned based on the time stamps; A mapping relationship between the multi-modal downhole historical data and the vertical depth data of the target well is obtained; According to the mapping relationship, the multi-modal downhole historical data is normalized into a preset borehole coordinate system.

[0051] Optionally, the training method in the identification unit 702 is specifically configured to: According to expert knowledge, the processed data is labeled twice independently to obtain two independent labeling results; When there is a difference between the two independent labeling results, the two independent labeling results are combined and adjudicated to obtain an adjudication result; The two independent labeling results or the adjudication result are integrated to construct a labeled data set.

[0052] Optionally, the identification unit 702 is specifically configured to: According to a preset sliding window, the multi-modal downhole real-time data is segmented to obtain a plurality of segmented results; The plurality of segmented results are input into a pre-established abnormality identification model, and forward inference is performed to obtain an identification result.

[0053] Optionally, the analysis unit 703 is specifically configured to: According to an exponential moving average method or a median filter algorithm, the identification result is smoothed to obtain a smoothed result. According to the cumulative sum control chart method or the Bayesian change point detection algorithm, a turning point in the smoothing result is identified, and the turning point represents a change in the abnormal state; According to the turning point, a trend index of the smoothing result is calculated, and the trend index represents a trend of dynamic changes of the abnormal event of the smoothing result; According to the trend index, trend analysis is performed to obtain an analysis result.

[0054] Optionally, the training method in the identification unit 702 is specifically used for: Obtaining feedback data of the user on the real-time alarm information, the feedback data including confirming the alarm, rejecting the alarm, and / or correcting the abnormal category in the real-time alarm information; According to the feedback data, the incremental fine-tuning and updating of the abnormal identification model are performed to obtain an updated abnormal identification model.

[0055] Optionally, the decision threshold rule includes a hysteresis loop rule formed by combining a trigger threshold strategy, a duration threshold strategy, and a contact threshold strategy, and a multi-modal consistency voting rule.

[0056] Please refer to Figure 8 The application also provides an electronic device, including: The processor 801, the memory 802, the input and output unit 803, and the bus 804; The processor 801 is connected with the memory 802, the input and output unit 803, and the bus 804; The memory 802 stores a program, and the processor 801 calls the program to perform the method for detecting the well wall in real time as shown in any one of the embodiments. Figures 1 to 6 The method for detecting the well wall in real time as shown in any one of the embodiments.

[0057] The application also relates to a computer readable storage medium, and the computer readable storage medium stores a program, and when the program runs on a computer, the computer performs the method for detecting the well wall in real time as shown in any one of the embodiments. Figures 1 to 6 The method for detecting the well wall in real time as shown in any one of the embodiments.

[0058] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the system, the device, and the unit described above can refer to the corresponding processes in the foregoing method embodiments, and will not be described here.

[0059] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the described device embodiments are merely schematic. The division of the units is merely a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0060] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0061] In addition, each functional unit in the embodiments of the present application can be integrated in a processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware, or in the form of software function units.

[0062] When the integrated unit is implemented in the form of software function units and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such an understanding, the technical solutions of the present application essentially, or the part that contributes to the prior art, or all or a part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), magnetic disk or optical disk, and various other media that can store program codes.

Claims

1. A method for real-time detection of wellbore walls, characterized in that, The method includes: Acquire multimodal downhole real-time data, which includes at least two of the following: currently acquired cuttings and wellbore imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological attribute data; The multimodal downhole real-time data is input into a pre-established anomaly identification model, stream inference is performed, and the identification results are output. Based on the identification results, time-series integration and trend analysis are performed to obtain the analysis results; Real-time alarm information is generated based on preset decision threshold rules and the analysis results; The anomaly detection model is obtained through the following training method: Acquire multimodal downhole historical data, which includes at least two of the following: historically collected cuttings and wellbore imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological property data; The multimodal downhole historical data is subjected to time synchronization processing and spatial positioning processing to obtain processed data; Anomaly labeling is performed on the processed data based on expert knowledge, and a labeled dataset is constructed. The multimodal fusion anomaly recognition model is trained based on the labeled dataset to obtain the anomaly recognition model.

2. The method according to claim 1, characterized in that, The time synchronization and spatial positioning processing of the multimodal downhole historical data includes: According to the time synchronization protocol, the multimodal downhole historical data are written with their respective timestamps, all of which are high-precision timestamps at the millisecond level. Based on the timestamp, the multimodal downhole historical data is time-aligned according to the nearest neighbor interpolation or dynamic time alignment algorithm. Obtain the mapping relationship between the depth measurement data corresponding to the multimodal downhole historical data and the vertical depth data of the target well; The multimodal downhole historical data is normalized to a preset wellbore coordinate system based on the mapping relationship.

3. The method according to claim 1, characterized in that, The step of annotating the processed data based on expert knowledge and constructing an annotated dataset includes: Based on expert knowledge, the processed data was independently labeled twice to obtain two independent labeling results; When there are differences between the two independent annotation results, the two independent annotation results are merged and adjudicated to obtain the adjudication result; The two independent annotation results or the adjudication results are integrated to construct an annotation dataset.

4. The method according to claim 1, characterized in that, The step of inputting the multimodal downhole real-time data into a pre-established anomaly identification model, performing streaming inference, and outputting the identification results includes: The multimodal downhole real-time data is segmented according to a preset sliding window to obtain multiple segmentation results; The multiple segmented results are input into a pre-established anomaly recognition model, and forward reasoning is performed to obtain the recognition result.

5. The method according to claim 1, characterized in that, The step of performing time-series integration and trend analysis based on the identification results to obtain the analysis results includes: The recognition results are smoothed using the exponential moving average method or the median filtering algorithm to obtain a smoothed result. Based on the cumulative sum control chart method or the Bayesian change point detection algorithm, the inflection points in the smoothing results are identified, and the inflection points represent changes in the abnormal state. A trend index for the smoothing result is calculated based on the inflection point, and the trend index characterizes the trend of the dynamic changes of abnormal events in the smoothing result. Trend analysis is performed based on the aforementioned trend indicators to obtain the analysis results.

6. The method according to any one of claims 1 to 5, characterized in that, After training the multimodal fusion anomaly detection model based on the labeled dataset to obtain the anomaly detection model, the method further includes: Obtain user feedback data on the real-time alarm information, including alarm confirmation, alarm rejection, and / or correction of the anomaly category in the real-time alarm information; The anomaly detection model is incrementally fine-tuned and updated based on the feedback data to obtain an updated anomaly detection model.

7. The method according to any one of claims 1 to 5, characterized in that, The decision threshold rules include a hysteresis loop rule formed by combining trigger threshold strategy, persistence threshold strategy, and contact threshold strategy, as well as a multimodal consensus voting rule.

8. A system for real-time detection of wellbore walls, characterized in that, The system includes: The real-time data unit is used to acquire multimodal downhole real-time data, which includes at least two of the following: currently acquired cuttings and wellbore imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological attribute data. The identification unit is used to input the multimodal downhole real-time data into a pre-established anomaly identification model, perform streaming inference, and output the identification results. The analysis unit is used to perform time-series integration and trend analysis based on the identification results to obtain analysis results; The generation unit is used to generate real-time alarm information based on preset decision threshold rules and the analysis results; The anomaly identification model is obtained through the following training method: acquiring multimodal downhole historical data, which includes at least two of the following: historically collected cuttings and wellbore imaging data, acoustic data, electromagnetic data, temperature and pressure data, drilling operation parameters, and geological attribute data; performing time synchronization and spatial positioning processing on the multimodal downhole historical data to obtain processed data; labeling the processed data for anomalies based on expert knowledge to construct a labeled dataset; and training the multimodal fusion anomaly identification model based on the labeled dataset to obtain the anomaly identification model.

9. An electronic device, characterized in that, The electronic device includes: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor calls to execute the method for real-time wellbore monitoring as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the method for real-time wellbore monitoring as described in any one of claims 1 to 7.

Images