Alarm threshold determination method for drill jamming early warning based on deep learning

Abstract

The invention discloses a method for determining an alarm threshold value of drill jamming early warning based on deep learning, and belongs to the technical field of petroleum drilling safety monitoring. Aiming at the problems of missing report and false report of double heights caused by large regional difference, scarcity of positive samples and incapability of a fixed threshold value to adapt to nonlinear high-dimensional data in the prior art, the method is based on logging data of six work areas including Ning, Lu, Yang, Huang, August and Foot and a jamming early warning model; three types of fluctuation ratio formulas of relative change rate, movement standard deviation and movement variation coefficient are innovatively provided and are matched with three threshold determination strategies of a double-side quantile warning method, a double-state quantile threshold method and an extreme value percentile warning method one by one, six sets of regionalization templates are formed, and accurate warning of one work area and one table is achieved. On-site verification shows that the early warning accuracy of drill jamming in each work area is improved by 15-35%, the highest early warning accuracy is 85%, and the false alarm amount is reduced by more than 30%.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and intelligent early warning technology for oil drilling engineering, specifically a characteristic parameter screening method based on stuck pipe early warning, applicable to the analysis and identification of characteristic parameters of stuck pipe accidents in drilling operations of various oil and gas reservoirs such as shale gas and conventional oil and gas. Background Technology

[0002] Stuck drill pipe is widely recognized as one of the most destructive downhole complexes in drilling operations. Causes can include abrupt formation changes, mud drift, drill string parameter mismatch, or wellbore environmental deterioration. Once the drill string is "locked up," drilling is immediately halted, potentially forcing sidetracking and even the abandonment of the entire reservoir. Statistics from major oil and gas regions in China in 2023 show that stuck drill pipe accounts for over 30% of all downhole failures, with an average handling time exceeding 48 hours. Extreme cases can result in direct losses of tens of millions of yuan, making it a core bottleneck restricting both efficiency and safety. Existing early warning methods remain in a "two-stage" blind spot: experience-based methods rely on engineers visually observing parameter curves, and the error rate soars due to human fatigue during complex horizontal well sections or long night shifts; traditional automatic early warning methods rely on fixed thresholds or simple statistical rules, which are powerless against the nonlinear coupling of multi-source high-dimensional data (drill pressure, torque, hook load, inlet and outlet flow rates, etc.), and the proportion of stuck drill samples is less than 0.1%, which is extremely unbalanced, causing latent precursors to be submerged in normal fluctuations, resulting in high rates of both missed and false alarms; Machine learning is highly anticipated, but it still faces a double bottleneck in drilling scenarios: On the data side, positive samples are scarce and have large geographical migration, with the AUC dropping by 30 percentage points when changing blocks with the same model; On the model side, blindly pursuing accuracy makes it difficult for black-box outputs to align with on-site experience, preventing engineers from making decisions based on this; At the same time, it is not sensitive to early weak signal-to-noise ratio features, triggering alarms at the slightest fluctuation, but the on-site team quickly turns off the warnings. For stuck drilling warning, a key technology related to drilling operation safety and efficiency, the most pressing problem the industry needs to solve is how to accurately capture abnormal features in the early stages of stuck drilling under complex data environments and unbalanced sample conditions, and select the most effective feature parameters for stuck drilling warning. This is the core problem that this invention aims to solve. Summary of the Invention

[0003] To address the current pain points in the industry regarding stuck drill warnings—namely, a lack of samples, explanations, and generalization—this invention proposes a "stuck drill warning feature parameter screening method." This method can automatically identify the most sensitive and physically significant trace features in the early stages of a stuck drill under conditions of extremely imbalanced data and complex, variable operating conditions. It eliminates redundancy and noise, compressing the original high-dimensional parameter space into a low-dimensional core feature subset that is "understandable to experts and easy for the model to learn." This helps achieve "early, accurate, and understandable" stuck drill warnings and decision support. The specific steps of the stuck drill warning feature parameter screening method proposed in this invention are as follows: Step S10: Based on the logging data of six work areas—Ning, Lu, Yang, Huang, Zi, and Zu—all located in the Sichuan Basin, currently the most active area for shale gas and tight gas exploration and development, the logging response characteristics, oil and gas show levels, parameter variation patterns, and subsequent fracturing test effects of each work area differ significantly due to differences in structural location, target strata, burial depth, and preservation conditions. Predicted alarm information is obtained through a stuck drill warning model. Then, three volatility formulas (relative change rate + moving standard deviation + moving coefficient of variation) and three alarm threshold determination methods (two-sided quantile warning method + two-state quantile threshold method + extreme value percentile warning method) are used to screen and statistically analyze the accuracy of the prediction results for the six work areas, ultimately resulting in a set of volatility and alarm threshold determination methods applicable to each work area. Step S20: By comparing the accuracy results of three volatility formulas and corresponding three alarm threshold determination methods, the "moving standard deviation" volatility formula is used to screen alarm information for the Ninggong area, and the alarm threshold is determined by the "two-sided quantile warning method". The accuracy of alarm information in the Ninggong area can reach 84.16%, which is 20% higher than the previous method. Step S30: By comparing the accuracy results of the three volatility formulas and the corresponding three alarm threshold determination methods, the volatility formula of "moving coefficient of variation" is adopted for Huanggong District, and the alarm threshold is determined by the "bi-state quantile threshold method". The accuracy of alarm information in Huanggong District can reach 85%, which is 35% higher than before when the above methods were not used. Step S40: By comparing the accuracy results of three volatility formulas and corresponding three alarm threshold determination methods, the volatility formula of "relative change rate" is adopted for the four work areas of Zi, Zu, Yang, and Lu. The alarm threshold is determined by the "extreme value percentile warning method". The accuracy of alarm information in the four work areas of Zi, Zu, Yang, and Lu can reach 55%, 80%, 71.7%, and 70%, respectively, which is 15%, 24%, 28%, and 20% higher than before when the above methods were not used. Step S50: Finally, a method for screening the volatility characteristics and alarm thresholds for stuck drill warning is formed by integrating three volatility formulas and three alarm threshold determination methods.

[0004] A further technical solution is that, in step S10, the multi-source drilling data collected from the six work areas includes three types of core data: the first type is drilling engineering parameters, specifically including hook load, torque, standpipe pressure, rotation speed, drilling pressure, hook height, well depth, drilling time, inlet flow rate, outlet flow rate, inlet density, and outlet density; the second type is geological parameters, specifically including stratigraphy, lithology, and formation pressure; the third type is historical tag data, specifically including historical stuck pipe event occurrence time, stuck pipe type, and a total of 132 engineering characteristic parameters. After PEN correlation analysis, hook load, torque, rotation speed, and hook height, a total of four characteristic parameters, are used as input parameters for the stuck pipe early warning model. The early warning model is a fusion model composed of "precursor signal model + anomaly detection model + expert experience model," which can effectively provide stuck pipe early warning. The formula for calculating the accuracy of the stuck pipe early warning is as follows: A: Accuracy rate; TP: Number of correct alarms (i.e., the number of alarms within the two hours prior to the actual stuck drill time); False alarms (i.e., the number of cases outside the six hours prior to the actual time the drill got stuck).

[0005] A further technical solution is that, in step S20, the volatility formula of "moving standard deviation" is used to calculate and visualize the dataset of the four characteristic parameters for the Ninggong area. The "moving standard deviation" sliding window consists of 10 consecutive data points, and the formula is as follows: Moving standard deviation; x ᵢ : The i-th observation; μ: The sample mean (μ = Σx ᵢ / n); n: the number of observations in the sample; Σ(x ᵢ −μ)²: the sum of squares of the differences between all observations and the mean; n − 1: degrees of freedom; sqrt(): take the square root and restore the variance to the "standard deviation" with the same dimensions as the original data; Then, the alarm threshold for the Ning work area is determined using the "two-sided quantile warning method." The specific method of the "two-sided quantile warning method" is as follows: 1. First, take all the data sets of "stuck drill torque, hook load, hook height, and rotation speed" calculated using the above-mentioned moving standard deviation calculation method as samples. 2. Then, divide the high-risk area: directly take the 25%-75% quantile of the sample. 3. Absolute risk area: lower limit to the 5% quantile; upper limit raised to the 95th quantile: when at least three characteristic parameters are simultaneously in the high-risk area, the system issues a warning; if even one characteristic parameter exceeds the absolute risk area, an alarm is immediately triggered.

[0006] A further technical solution is that, in step S30, the volatility formula of the "moving coefficient of variation" is used to calculate and visualize the dataset of the four characteristic parameters for the Huanggong area. The sliding window of the "moving coefficient of variation" is 10 consecutive data points, and the formula is as follows: : Moving coefficient of variation; Moving standard deviation; The average value of 10 consecutive data points in the sliding window of this feature value for this work area; Then, the "two-state quantile threshold method" is used to determine the value. The specific method of the two-state quantile threshold method is as follows: Divide the range of the moving coefficient of variation of each parameter during all normal operations in the entire well section into 20 segments. Then, convert the number of points where the moving coefficient of variation falls in each segment into probability density and calculate the 90th quantile P1. (norm) Then, the range of the moving coefficients of variation of each parameter during all stuck periods in the entire well section is divided into 20 segments. The number of points where the moving coefficients of variation of each parameter during the stuck period fall within each segment is then converted into probability density and the 10th quantile P1 is calculated. (stick) P2 of the 90th percentile (stick) P1 (norm) P1 represents the upper limit during normal operation; (stick) P2 represents the lower limit that the diamond can reach; (stick) This represents the most extreme 10% value when the drill string is stuck, i.e., the absolute danger line. This means that a certain characteristic parameter reaches P2. (stick) The system will then issue an alarm if at least three parameters reach P1. (stick) With P2 (stick) The system will issue an alarm if this happens.

[0007] A further technical solution is that, in step S40, the volatility formula of "moving standard deviation" is used to calculate and visualize the dataset of the four characteristic parameters for the four work areas of Zi, Zu, Yang, and Lu. The formula for the "relative rate of change" is as follows: Valᵗ: The value at the current moment; Val t-1: The value at the previous moment; Δ: The absolute value of the relative rate of change, unitless and non-negative; then determined by the "extreme value percentile warning method", the specific method of which is as follows: using the historical maximum relative rate of change of each well as the benchmark, the real-time relative rate of change is converted into "percentage of the maximum value". When the percentage exceeds the threshold on the left, the system judges it as "high risk". If no less than three characteristic parameters exceed this value, the system will immediately alarm; once it exceeds the higher threshold on the right, it is upgraded to "absolute risk". If any parameter exceeds this threshold, the system will immediately alarm. Different parameters are set with different thresholds according to the process characteristics: torque and hook height are at 70% for high risk and 90% for absolute risk; rotational speed is at 80% for high risk and 95% for absolute risk; hook load is at 80% for high risk and 90% for absolute risk in the positive range.

[0008] A further technical solution is that step S50 establishes a set of early warning characteristic parameters and threshold selection methods for different work areas. This includes three volatility calculation methods ("relative rate of change + moving standard deviation + moving coefficient of variation") and three alarm threshold determination methods ("two-sided quantile warning method + two-state quantile threshold method + extreme value percentile warning method"), which correspond one-to-one with each other. This method can effectively reduce the number of false alarms and invalid alarms caused by the stuck drill warning model, and can effectively improve the accuracy of the stuck drill warning model results.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention creates six sets of regional templates (Ningxia / Huanghua / Zibo / Zujiang / Yangzhou / Luzhou) with one-to-one correspondence for the "volatility formula-threshold determination method". Each template specifies four core input parameters and a dedicated threshold algorithm, realizing "one table for one work area", which effectively solves the problem of large regional differences in false alarms and missed alarms caused by the past "one standard for the whole country". This invention has been verified on field data from six major work areas, and the accuracy of the stuck drill warning has been improved by an average of more than 20 percentage points. Specifically, the accuracy rate in the Huang work area has increased from 50% to 85%, and in the Ning work area from approximately 64% to 84%. At the same time, false alarms (alarms triggered 6 hours before the stuck drill) have been reduced by more than 30%, directly reducing the losses from ineffective drilling stoppages. This invention uses a nested design of three warning lines: "double-sided quantile - double-state quantile - extreme value percentile". It sets extreme outliers (95% to 100% percentile range) as a separate "red line zone". As long as one parameter is exceeded, an alarm will be triggered, ensuring that the most dangerous signs of stuck drill are "zero missed". This invention makes up for the deficiency of traditional single mean ± variance model in sensitivity to extreme values. The entire method of this invention fully reuses 132 conventional engineering-geological parameters collected by existing logging sensors, without the need for additional logging or downhole tools; it has a small computational load, can be embedded in an edge computing box to run in real time, and can be put into operation on site with only a software upgrade, saving approximately RMB 300,000 to 500,000 in hardware and construction costs per well. Attached Figure Description

[0010] Figure 1 This is an example of a visualization of the moving standard deviation volatility of Ning 209H47-12 in the Ning Industrial Zone of this invention; Figure 2 This is a visualization example of the moving coefficient of variation volatility for the Huang 202H8-8 work area of ​​this invention; Figure 3 A visualization example of relative rate of change volatility is used for the Yang 101H32-5 work area of ​​this invention; Figure 4 This is a flowchart of the present invention. Detailed Implementation

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments, so that the objectives, technical solutions and advantages of the present invention will be clearer. Specifically, it includes the following steps: In step S10, the multi-source drilling data collected from the six work areas includes three core data categories: the first category is drilling engineering parameters, specifically including hook load, torque, standpipe pressure, rotation speed, drilling pressure, hook height, well depth, drilling time, inlet flow rate, outlet flow rate, inlet density, and outlet density; the second category is geological parameters, specifically including formation location, lithology, and formation pressure; the third category is historical tag data, specifically including historical stuck pipe event occurrence time, stuck pipe type, and a total of 132 engineering characteristic parameters. After PEN correlation analysis, hook load, torque, rotation speed, and hook height, a total of four characteristic parameters, are used as input parameters for the stuck pipe early warning model. The early warning model is a fusion model composed of a "precursor signal model + anomaly detection model + expert experience model," which can provide effective stuck pipe early warning. The formula for calculating the accuracy of the stuck pipe early warning is as follows: A: Accuracy rate; TP: Number of correct alarms (i.e., the number of alarms within the two hours prior to the actual stuck drill time); False alarms (i.e., the number of cases outside the six hours prior to the actual time the drill got stuck); Table 1. Description of Fields in Multi-Source Drilling Data Collect multi-source drilling data (as shown in Table 1).

[0012] Step S20: For the Ninggong area, the volatility formula of "moving standard deviation" is used to calculate and visualize the dataset of the four characteristic parameters. The "moving standard deviation" sliding window consists of 10 consecutive data points. The formula is as follows: Moving standard deviation; x ᵢ : The i-th observation; μ: The sample mean (μ = Σx ᵢ / n); n: the number of observations in the sample; Σ(x ᵢ −μ)²: the sum of squares of the differences between all observations and the mean; n − 1: degrees of freedom; sqrt(): take the square root and restore the variance to the "standard deviation" with the same dimensions as the original data; Then, the alarm threshold for the Ning work area is determined using the "two-sided quantile warning method." The specific method of the "two-sided quantile warning method" is as follows: 1. First, take all the data sets of "stuck drill torque, hook load, hook height, and rotation speed" calculated using the above-mentioned moving standard deviation calculation method as samples. 2. Then, divide the high-risk area: directly take the 25%-75% quantile of the sample. 3. Absolute risk area: lower limit to the 5% quantile; upper limit raised to the 95th quantile: when at least three characteristic parameters are simultaneously in the high-risk area, the system issues a warning; if even one characteristic parameter exceeds the absolute risk area, an alarm is immediately triggered.

[0013] Step S30: For the Huanggong area, the volatility formula of the "moving coefficient of variation" is used to calculate and visualize the dataset of the four characteristic parameters. The sliding window of the "moving coefficient of variation" is 10 consecutive data points. The formula is as follows: : Moving coefficient of variation; Moving standard deviation; The average value of 10 consecutive data points in the sliding window of this feature value for this work area; Then, the "two-state quantile threshold method" is used to determine the value. The specific method of the two-state quantile threshold method is as follows: Divide the range of the moving coefficient of variation of each parameter during all normal operations in the entire well section into 20 segments. Then, convert the number of points where the moving coefficient of variation falls in each segment into probability density and calculate the 90th quantile P1. (norm) Then, the range of the moving coefficients of variation of each parameter during all stuck periods in the entire well section is divided into 20 segments. The number of points where the moving coefficients of variation of each parameter during the stuck period fall within each segment is then converted into probability density and the 10th quantile P1 is calculated. (stick) P2 of the 90th percentile (stick) P1 (norm) P1 represents the upper limit during normal operation;(stick) P2 represents the lower limit that the diamond can reach; (stick) This represents the most extreme 10% value when the drill string is stuck, i.e., the absolute danger line. This means that a certain characteristic parameter reaches P2. (stick) The system will then issue an alarm if at least three parameters reach P1. (stick) With P2 (stick) The system will issue an alarm if this happens.

[0014] Step S40: For the four work areas of Zi, Zu, Yang, and Lu, the volatility formula of "moving standard deviation" is used to calculate and visualize the dataset of the four characteristic parameters. The formula for the "relative rate of change" is as follows: Valᵗ: The value at the current moment; Val t-1 : The value at the previous moment; Δ: The absolute value of the relative rate of change, unitless and non-negative; further determined by the "extreme value percentile warning method", the specific method of which is as follows: using the historical maximum relative rate of change of each well as the benchmark, the real-time relative rate of change is converted into a percentage of the maximum value. When the percentage exceeds the threshold on the left, the system judges it as "high risk". If no less than three characteristic parameters exceed this value, the system will immediately alarm; once it exceeds the higher threshold on the right, it is upgraded to "absolute risk". If any one of them exceeds this threshold, the system will immediately alarm. Different parameters are set with different thresholds according to the process characteristics: torque and hook height are considered high-risk thresholds when they exceed 70% of the maximum historical relative change rate of the well, and absolute risk thresholds when they exceed 90% of the maximum historical relative change rate of the well; rotational speed is considered high-risk threshold when it exceeds 80% of the maximum historical relative change rate of the well, and absolute risk threshold when it exceeds 95% of the maximum historical relative change rate of the well; hook load is considered high-risk threshold when it exceeds 80% of the maximum historical relative change rate of the well in the positive range, and absolute risk threshold when it exceeds 90% of the maximum historical relative change rate of the well.

[0015] Step S50 establishes a set of early warning characteristic parameters and threshold selection methods for different work areas. It includes three volatility calculation methods ("relative rate of change + moving standard deviation + moving coefficient of variation") and three alarm threshold determination methods ("two-sided quantile warning method + two-state quantile threshold method + extreme value percentile warning method"), which correspond one-to-one with each other. This method can effectively reduce the number of false alarms and invalid alarms caused by the stuck drill warning model, and can effectively improve the accuracy of the stuck drill warning model results.

[0016] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for determining an alarm threshold of a deep learning-based alert of a stuck pipe, characterized in that, The method comprises the following steps: Step S10, for the logging data of the six work areas of Ning, Lu, Yang, Huang, Zi and Zu, the predicted alarm information is obtained through the stuck drill warning model, and then the three volatility formulas of "relative change rate + moving standard deviation + moving coefficient of variation" and the three alarm threshold determination methods of "double-sided quantile warning method + double-state quantile threshold method + extreme percentile warning method" are used to screen and statistically analyze the prediction results of the six work areas, and finally a set of volatility and alarm threshold determination methods suitable for each work area is obtained; Step S20, by comparing the accuracy rate results of the three volatility formulas and the corresponding three alarm threshold determination methods, the "moving standard deviation" volatility formula is used to screen the alarm information for the Ning work area, and the alarm threshold is determined by the "double-sided quantile warning method"; Step S30, by comparing the accuracy rate results of the three volatility formulas and the corresponding three alarm threshold determination methods, the "moving coefficient of variation" volatility formula is used to screen the alarm information for the Huang work area, and the alarm threshold is determined by the "double-state quantile threshold method"; Step S40, by comparing the accuracy rate results of the three volatility formulas and the corresponding three alarm threshold determination methods, the "relative change rate" volatility formula is used to screen the alarm information for the Zi, Zu, Yang and Lu work areas, and the alarm threshold is determined by the "extreme percentile warning method"; Step S50, finally a stuck drill warning volatility feature and alarm threshold screening method is formed by the fusion of the three volatility formulas and the three alarm threshold determination methods.

2. The method of claim 1, wherein the method is based on deep learning. In the step S10, the collected multi-source drilling data of the six work areas includes three types of core data: the first type is drilling engineering parameters, specifically including hook load, torque, standpipe pressure, rotary speed, drilling pressure, hook height, well depth, drilling time, inlet flow, outlet flow, inlet density, and outlet density; the second type is geological parameters, specifically including horizon, lithology, and formation pressure; and the third type is historical label data, specifically including historical stuck drill event occurrence time, stuck drill type, etc., totaling 132 engineering characteristic parameters. After person correlation analysis, hook load, torque, rotary speed, and hook height, totaling 4 characteristic parameters, are used as input parameters of the stuck drill warning model. The warning model is a fusion model composed of "precursor signal model + anomaly check model + expert experience model", which can effectively warn about stuck drill. The stuck drill warning accuracy rate calculation formula is as follows: A: accuracy rate; TP: True Positives (i.e., the number of times the alert was correct within two hours of the actual time of the stuck pipe); : False Positives (i.e., the number of times the alert was incorrect outside of six hours of the actual time of the stuck pipe).

3. The method of claim 1, wherein the method is based on deep learning. In the step S20, the "moving standard deviation" volatility formula is used to calculate and visualize the data set of the 4 characteristic parameters for the Ning work area. The "moving standard deviation" sliding value window is 10 consecutive data points, and the formula is as follows: : moving standard deviation; x ᵢ : i-th observation μ: sample mean (μ = Σx ᵢ / n); n: number of observations in the sample; Σ(x ᵢ − μ)²: sum of squared differences from the mean; n−1: degrees of freedom; sqrt(): square root to reduce the variance back to the same units as the original data, the "standard deviation"; Then the alarm threshold for the Ning work area is determined by the "double-sided quantile warning method". The specific method of the "double-sided quantile warning method" is as follows:

1. Take all the "stuck drill segment torque, hook load, hook height, rotary speed, and data set calculated by the above moving standard deviation calculation method" as samples. Then divide the high-risk area: directly take the 25% to 75% quantile of the samples.

3. Absolute risk zone: the lower limit is the 5th percentile, and the upper limit is raised to the 95th percentile. When there are no less than 3 characteristic parameters in the high risk zone, the system gives a warning. If there is only 1 characteristic parameter that breaks the absolute risk zone, the system immediately gives an alarm.

4. The method of claim 1, wherein the method is based on deep learning. In step S30, the data set of the 4 characteristic parameters is calculated and visualized by using the "moving coefficient of variation" volatility formula for the yellow work area. The "moving coefficient of variation" sliding value window is 10 consecutive data points, and the formula is as follows: : moving coefficient of variation; : moving standard deviation; : average value of 10 consecutive data points of the feature value sliding window of the work area Then, the "two-state quantile threshold method" is used to determine the value. The specific method of the two-state quantile threshold method is as follows: Divide the range of the moving coefficient of variation of each parameter during all normal operations in the entire well section into 20 segments. Then, convert the number of points where the moving coefficient of variation falls in each segment into probability density and calculate the 90th quantile P1. (norm) Then, the range of the moving coefficients of variation of each parameter during all stuck periods in the entire well section is divided into 20 segments. The number of points where the moving coefficients of variation of each parameter during the stuck period fall within each segment is then converted into probability density and the 10th quantile P1 is calculated. (stick) With 90th percentile P2 (stick) P1 (norm) P1 represents the upper limit during normal operation; (stick) P2 represents the lower limit that the diamond can reach; (stick) This represents the most extreme 10% value when the drill string is stuck, i.e., the absolute danger line. This means that a certain characteristic parameter reaches P2. (stick) The system will then issue an alarm if at least three parameters reach P1. (stick) With P2 (stick) The system will issue an alarm if this happens.

5. The method of claim 1, wherein the method is based on deep learning. In step S40, the data set of the 4 characteristic parameters is calculated and visualized by using the "moving standard deviation" volatility formula for the self, foot, Yang, and Lu work areas. The formula for the "relative change rate" method is as follows: Valt: the value of the previous moment; Val t-1 : the value of the previous moment; Δ: the absolute value of the relative change rate, unitless, non-negative; further determined by "extreme percentile warning method", the specific way of which is as follows: taking the maximum value of the historical relative change rate of each well as the benchmark, the real-time relative change rate is converted into "percentage of the maximum value". When the percentage exceeds the left threshold value, the system determines that it is "high risk", and if there are not less than three characteristic parameters breaking through the value, the system will immediately alarm; once breaking through the higher right threshold value, it is upgraded to "absolute risk", and if there is one item exceeding the threshold value, the system will immediately alarm. Different parameters have different threshold values according to the process characteristics: torque and hook height are high-risk threshold values when they exceed 70% of the maximum historical relative change rate of the well, and are absolute risk threshold values when they exceed 90% of the maximum historical relative change rate of the well; the rotating speed is a high-risk threshold value when it exceeds 80% of the maximum historical relative change rate of the well, and is an absolute risk threshold value when it exceeds 95% of the maximum historical relative change rate of the well; and the hook load is a high-risk threshold value when it exceeds 80% of the maximum historical relative change rate of the well in the positive range, and is an absolute risk threshold value when it exceeds 90% of the maximum historical relative change rate of the well.

6. The method of claim 1, wherein the method is based on deep learning. Step S50 is a set of characteristic parameter and threshold screening methods for different work areas. It includes three volatility calculation methods: "relative change rate + moving standard deviation + moving coefficient of variation", and three alarm threshold determination methods: "double-sided percentile warning method + double-state percentile threshold method + extreme percentile warning method". They are one-to-one corresponding. This method can effectively reduce the number of false alarms and invalid alarms caused by the pipe sticking warning model, and can effectively improve the accuracy of the results of the pipe sticking warning model.

Images