JP7489043B2 - Detention prediction method and system - Google Patents

Description

The present invention relates to onshore and offshore drilling technology, and more specifically to anomaly detection during drilling.

While there is a shift towards renewable energy sources in energy supply due to considerations of the global environment, the proportion of dependency on fossil fuels remains high. Drilling technology for resource development must overcome significant challenges and be economically rational. Similarly, improvements in drilling technology are required in scientific surveys and research.

During drilling, underground troubles are a major issue. A typical underground trouble is known as detention. Detention occurs when the drill string is pressed down inside the well, making it impossible to rotate and/or pull out. Detention is a major risk as it can lead to the abandonment of the well, but it is believed that effective countermeasures can be implemented if the signs of detention can be detected.

Conventional anomaly detection techniques for drilling involve detecting outliers in measurement data and deviations from prior simulations, but the accuracy of these techniques is limited because drilling behavior varies greatly depending on the geological layer and geological information is often unknown. In fact, there are many false alarms, and workers' intuition often takes precedence.

Effective prediction methods have been proposed that utilize data obtained in real time during drilling. Non-Patent Document 1 discloses that in early detection of kicks, false alarms are reduced by dynamically setting alarm thresholds for the amount of mud water and flow rate during drilling. Non-Patent Document 2 discloses detection of detention states by learning from actual drilling data. Based on the results of these previous studies, it is expected that the accuracy of anomaly detection using data-driven models will improve.

There are also patent documents related to detention detection. Patent Document 1 discloses an ensemble prediction of detention using three or more machine learning programs. The machine learning model adopted in Non-Patent Document 2 is similar to the ensemble prediction described in Patent Document 1. Patent Document 2 also discloses a method of predicting detention using a linear regression model of hook load against bit depth.

In the data-driven approach, it is important to extract features such as the relationship and pattern of parameters that the measurement data shows when it is abnormal and normal. The inventors have found that general methods based on the time domain or frequency domain do not necessarily capture the features of abnormal conditions well. In addition, in order to learn efficiently, the length of the available measurement data is limited, and it is necessary to efficiently extract features from a limited amount of data.
WO2013/066746 WO2016/034945 Unrau, S., Torrione, P., Hibbard, M., Smith, R., Olesen, L., Watson, J., et al. (2017). Machine learning algorithms applied to detection of well control events. In SPE Kingdom of Saudi Arabia Annual Technical Symposium and Exhibition. Society of Petroleum Engineers. Alshaikh, A., Magana-Mora, A., Gharbi, SA, Al-Yami, A., et al. (2019). Machine learning for detecting stuck pipe incidents: Data analytics and models evaluation. In International Petroleum Technology Conference. International Petroleum Technology Conference.

The present invention aims to predict detention using features that are useful for predicting detention and input data of a limited size.

A first aspect of the present invention is a method for predicting detention, comprising: generating input data based on a plurality of excavation parameters obtained during excavation; and providing the input data to a detention prediction model to provide a detention risk, the method comprising:
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
Generate a 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on the current bit depth value as input data;
The retention risk is provided according to the input of the 2D histogram generated based on the current bit depth.

A second aspect of the present invention is a detention prediction system comprising a processor and a memory,
The memory includes:
an input data generation program that generates input data based on a plurality of drilling parameters obtained during drilling;
A detention prediction program that provides a detention risk in response to input of the input data;
Equipped with
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
The input data generation program is executed by the processor to generate, as input data, a 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on a current bit depth value;
The retention prediction program is executed by the processor to provide a retention risk as a function of input of the 2D histogram generated based on a current bit depth.

A third aspect of the present invention is a detention prediction program that generates input data based on a plurality of excavation parameters obtained during excavation, and provides a detention risk based on the input data,
The program, when executed by a processor, causes the processor to:
When acquiring the plurality of drilling parameters, the plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
generating, as input data, a 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on the current bit depth value;
The retention risk is provided in response to the input of the 2D histogram generated based on the current bit depth.
In one aspect, the detention prediction program is provided in a form stored on a computer-readable medium.

In one embodiment, the bit depth value and the one drilling parameter value are acquired as time series data. The time series data is sampled at a predetermined interval, and the bit depth value and the one drilling parameter value correspond to each other by time information.
In one aspect, the detention risk is provided as a detention prediction probability.
In one aspect, the detention risk is provided as an alarm generated based on a detention prediction probability.
In one aspect, the value of the one drilling parameter corresponding to the relative depth set with respect to the current bit depth value (which is data acquired before the current time) is:
The data is divided into first distribution data obtained within a predetermined period immediately preceding the current time point and second distribution data obtained prior to the first distribution data,
a first 2D histogram based on the first scatter data and a second 2D histogram based on the second scatter data are generated, the first 2D histogram and the second 2D histogram being distinguishable by color;
The input data is a superimposed 2D histogram in which the first 2D histogram and the second 2D histogram are superimposed.

In one embodiment, the one drilling parameter other than bit depth is selected from bit speed, drill pipe rotation torque, hook load, and mud injection pressure.
In one embodiment, the input data includes a 2D histogram of bit rotation speed versus relative depth, a 2D histogram of drill pipe rotation torque versus relative depth, a 2D histogram of hook load versus relative depth, and a 2D histogram of mud injection pressure versus relative depth.

In one aspect, the retention prediction model is machine-learned to provide retention risk in response to input of a 2D histogram generated based on training drilling data.
In one aspect, the retention prediction model is a supervised learning model.
In one embodiment, the detention prediction model uses a convolutional neural network (CNN) to provide detention risk.
In one aspect, the retention prediction model is an unsupervised learning model.

In one aspect, the detention risk is provided in response to an input of a plurality of 2D histograms obtained based on a plurality of drilling parameters;
The contribution of each drilling parameter to the detention risk is visualized and displayed on the display.
In one aspect, the area of the input data that contributed to the generation of the alarm is visualized using a gradient-based class activation map (Grad-CAM) to show the reason for the alarm generation.

A fourth aspect of the present invention is a method for generating input data for a detention prediction model, comprising:
acquiring drilling data including a plurality of drilling parameters;
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
The relative depth set based on the bit depth value at a certain time t is set as a first axis, and the value of the one excavation parameter corresponding to the relative depth is set as a second axis, and the scatter data is acquired using the excavation data before a certain time t;
The scatter data is converted into a 2D histogram to generate input data.
The input data generated based on the drilling data is used as input data for both the training model and the trained model.
In one embodiment, the drilling data (the bit depth value and the one drilling parameter value) is time-series data. The time-series data is sampled at a predetermined interval, and the bit depth value and the one drilling parameter value correspond to each other by time information.
In one embodiment, the one drilling parameter other than bit depth is selected from bit speed, drill pipe rotation torque, hook load, and mud injection pressure.

In one aspect, the scatter data comprises:
The data is divided into first scatter data obtained within a predetermined period immediately before a certain time point t and second scatter data obtained before the first scatter data,
a first 2D histogram based on the first scatter data and a second 2D histogram based on the second scatter data are generated, the first 2D histogram and the second 2D histogram being distinguishable by color;
The input data is a superimposed 2D histogram in which the first 2D histogram and the second 2D histogram are superimposed.

A fifth aspect of the present invention is a method for generating a trained detention prediction model,
Acquire learning drilling data including a plurality of drilling parameters and detention information;
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
A 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on the bit depth value at a certain time point t is generated and used as input data;
Assuming that warning signs appear within a certain period of time before detention, the learning model is trained to provide detention risk according to input data generated based on the training drilling data.
In one aspect, the learning drilling data (the bit depth value and the one drilling parameter value) is time-series data. The time-series data is sampled at a predetermined interval, and the bit depth value and the one drilling parameter value correspond to each other by time information.
In one aspect, the learning drilling data includes time series data on bit rotation speed, drill pipe rotation torque, bit depth, hook load, and mud injection pressure, and detention information (detention start time and end time) is recorded, and one drilling parameter other than the bit depth is selected from the bit rotation speed, drill pipe rotation torque, hook load, and mud injection pressure.
In one embodiment, the input data includes a 2D histogram of bit rotation speed versus relative depth, a 2D histogram of drill pipe rotation torque versus relative depth, a 2D histogram of hook load versus relative depth, and a 2D histogram of mud injection pressure versus relative depth.

In one aspect, the value of the one drilling parameter corresponding to the relative depth set based on the bit depth value at a certain time t (data acquired before a certain time t) is
The data is divided into first scatter data acquired within a predetermined period immediately before a certain time point t and second scatter data acquired before the first scatter data,
a first 2D histogram based on the first scatter data and a second 2D histogram based on the second scatter data are generated, the first 2D histogram and the second 2D histogram being distinguishable by color;
The input data is a superimposed 2D histogram in which the first 2D histogram and the second 2D histogram are superimposed.

In one aspect, the learning model is trained by assuming that warning signs appear within a certain period of time before detention, and by assigning a correct answer label of 1 to input data within the certain period of time before detention, and a correct answer label of 0 to input data excluding the certain period of time before detention and the detention state.
In the experimental example described below, the certain period before detention is 6 minutes, but 6 minutes is merely an example, and the length of the "certain period" can be appropriately set by those skilled in the art.
In one embodiment, the training model is trained using a convolutional neural network (CNN).

In one aspect, the learning model learns input data generated based on learning excavation data, excluding input data for a certain period before detention and the detention state, as normal data so that the trained model can identify abnormal data and output a detention risk.
In one aspect, an autoencoder is used to generate a learning model that minimizes the error between the input data and the output data.

A fifth aspect of the present invention is a method for compressing drilling data, comprising the steps of:
acquiring drilling data including a plurality of drilling parameters;
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
Using the excavation data, a relative depth set based on the current bit depth value is set as a first axis, and the value of the one excavation parameter corresponding to the relative depth is set as a second axis, and scattering data is acquired using excavation data before the current time;
The scatter data is divided into first scatter data acquired during a predetermined period immediately preceding the current time point and second scatter data acquired prior to the first scatter data,
The first scatter data and the second scatter data are displayed on a display as one scatter diagram in a distinguishable manner.
In one embodiment, the first scatter data and the second scatter data are color-coded and displayed on a display as one scatter plot.
In one aspect, the drilling data is time series data.

In the present invention, by converting the drilling parameters into the depth domain, feature extraction is performed on the drilling parameters, and input data with such features is used to predict retention.

1 is an overall schematic diagram of a detention prediction system according to an embodiment of the present invention. A diagram showing the learning phase and detention prediction phase of the detention prediction system of this embodiment. 10A and 10B are diagrams illustrating supervised learning of a detention sign according to the present embodiment. FIG. 1 is a diagram showing an outline of comparative experiments 1 and 2. The left figure is a scatter plot showing hook load values for relative depths set based on bit depths at time t. In reality, the left figure is a color image, with the most recent data at time t displayed in red and historical data prior to the most recent data displayed in green. The right figure shows a superimposed 2D histogram in which a first 2D histogram generated based on the most recent data at time t is superimposed on a second 2D histogram generated based on historical data. In reality, the right figure is a color image, and the RGB value sets of the first 2D histogram and the RGB value sets of the second 2D histogram are distinguishable. FIG. 6 is a diagram showing the history data in the left diagram of FIG. 5 . FIG. 6 is a diagram showing the most recent data in the left diagram of FIG. 5 . The left image shows a first 2D histogram generated based on the most recent data, the center image shows a second 2D histogram generated based on historical data, and the right image shows an overlaid 2D histogram in which the first 2D histogram is overlaid on the second 2D histogram. 13 is a flow chart illustrating the generation of scatter plots by converting drilling parameters to the depth domain and the generation of overlaid 2D histograms based on the scatter data. 1 is a flowchart showing a method for generating input data (superimposed 2D histogram) to CNN according to the present embodiment. FIG. 13 is a diagram showing generation of input data (superimposed 2D histogram) to the CNN according to the present embodiment. FIG. 13 is a diagram illustrating input data (superimposed 2D histogram). FIG. 1 is a diagram illustrating the process of generating input data for a 3D-CNN model. The degrees of freedom/transposition of the 5D input tensors to the 3D-CNN model are shown, with the top showing (B, P, T, V, H) and the bottom showing (B, T, P, V, H). FIG. 13 shows real-time detention risk prediction using validation data (Case 3). FIG. 13 illustrates real-time detention risk prediction using learning data. A comparison of input data and the Grad-CAM heatmap is shown in Fig. 1. In practice, the heatmap is displayed as a color image.

[A] Overview of Drilling Technology [A-1] Drilling System The drilling system includes a drill string formed by connecting multiple drill pipes in the longitudinal direction, a bit provided at the lower end of the drill string, and a top drive provided at the upper end of the drill string. Drilling is the process of drilling the strata by rotating the drill string and the bit with the top drive, and drilling is performed to the target depth while adding drill pipes. In addition, drilling fluid is circulated in the well to balance the well pressure and the strata pressure, thereby preventing accidents in the well. Each element constituting the drilling system will be described in more detail below.

A drill string is a drill pipe assembly consisting of multiple drill pipes connected together, and the length of the drill string can reach several thousand meters. The drill string is hollow, allowing drilling mud to be supplied through the inside of the drill string. A mechanism called a Bottom Hole Assembly (BHA) is located at the bottom end of the drill string. The BHA is composed of multiple drill collars and multiple stabilizers connected together. The configuration of the BHA can be changed to suit the intended drilling operation. A bit is fixed to the bottom end of the BHA, and the rotating bit is used to penetrate the rock layer.

The bit is rotated by a top drive and/or a mud motor. The top drive is a device located on land or offshore that rotates the drill string and bit. The mud motor is a device located inside the BHA that is powered by the fluid power of the injected mud, and the bit can also be rotated using the mud motor. The bit rotation speed generated by the mud motor depends on the flow rate of the mud. The top drive and mud motor may be operated separately or simultaneously to rotate the bit.

The hook is the element of the lifting system or hoist from which the top drive hangs. The hook is attached to the lower end of the traveling block, which is a movable pulley. The traveling block can be raised and lowered by the drawworks.

The standpipe is the element that connects the mud pump to the top drive. The mud pump delivers drilling mud to the well through the standpipe and the drill string.

The derrick is a tower erected above the drilling floor. The height of the derrick determines the length of the drill pipe unit that is inserted into the mine at one time. The height of the derrick is usually several tens of meters, and the length of the drill pipe unit is usually 30-40 meters. The drilling floor is the human workspace and is located below the derrick.

The drilling system is equipped with a mud circulation system. Well pressure control is essential for safe drilling, so drilling mud is injected into the well at high pressure. The drilling mud flows through the standpipe and the drill string, and returns to the surface through the annulus formed between the drill string and the well wall. The drilling mud is circulated for reuse after undergoing the required treatment.

For an overall view of the drilling system, for example, refer to Patent Document 1 or "Wang, N. et al. (2015). A multibody dynamics model of contact between the drillstring and the wellbore and the rock penetration process. Advances in Mechanical Engineer." Note that while the above description of the drilling system is about land drilling, the drilling data that is the subject of the present invention also includes offshore drilling data, and it will be understood by those skilled in the art that offshore drilling uses offshore drilling rigs and risers. Since drilling systems are well known to those skilled in the art, further detailed description will be omitted.

[A-2] Drilling work Drilling work consists of multiple actions specific to the task. Typical actions include drilling (the work of rotating the bit to dig) and hoisting (the work of pulling out the drill string and lowering). The following are examples of typical actions in drilling work.

(a) Drilling
This involves pumping mud and rotating the drill pipe to increase the bit depth.
(b)Sliding
This is the process of increasing the bit depth using an underground mud motor without rotating the drill pipe.
(c)Reaming
In order to stabilize the mine, mud is pumped in and the drill pipe is rotated while it is raised and lowered repeatedly.
(d)Tripping
In order to stabilize the inside of the well, the drill pipe is raised and lowered repeatedly without rotating. The bit is rotated by the mud motor inside the well by the pumped mud fluid.
(e) Survey
This is the process of measuring the direction and inclination of a wellbore using a Measuring While Drilling (MWD) tool.
(f) Circulation
This involves pumping mud fluid to transport measured survey data from underground to the surface.
(g) Connection/Disconnection
This involves connecting and disconnecting drill pipes.
(h) RIH (Run-in-hole)
This is the process of connecting the drill pipe and lowering the pipe to reach the drill string to a certain depth in the wellbore that has been achieved by drilling.
(i) POOH (Pull-out-of-hole)
The process of removing the drill pipe and raising it in order to withdraw the drill string from a certain depth inside the mine.

[A-3] Drilling data (drilling parameters)
The drilling system is equipped with a number of sensors and measuring instruments at predetermined locations, and measures parameters such as those shown in Table 1 during drilling operations and monitors them to understand and predict the state of the underground environment. The values of these parameters are acquired as time-series data and stored in a database. Note that the parameters shown in Table 1 are merely examples and do not limit the drilling parameters.

(a)TD_trq
This is a value measured at the top drive and indicates the rotational torque of the drill string or drill pipe. TD_trq is strongly affected by the underground conditions, so it is considered to be one of the important indicators of detention.
(b)TD_spd
This is the top drive speed, i.e. the rotational speed of the drill bit. However, TD_spd (the rotational speed of the drill bit) and the actual top drive rotational speed are not equal when a downhole mud motor is used. TD_spd is the sum of the actual top drive speed and the downhole mud motor speed.
(c)Bitdepth
Depth to bit refers to the length from the well floor to the bottom of the drill string, not necessarily the vertical distance in the well, since wells are not always vertical.
(d) Hook height
The hook height represents the vertical position of the hook. The hook can move from the drilling floor to the top of the derrick, for example, the lower limit position is 0 [m] and the upper limit position is 40 [m].
(e) Hookload
The hook load represents the mass acting on the hook and is considered one of the important indicators of containment since it is affected by downhole conditions such as, for example, friction between the drill string and the formation.
(f) WOB
WOB is an estimate called Surface-WOB, which is distinct from Downhole-WOB, which is a measurement of the weight on the bit measured by an underground sensor and is more reliable than Surface-WOB, but it is difficult to measure in real time due to the need to transmit signals from the well to the surface.
(g)MRetFlow
MRetFlow is the percentage of mud return. Depending on the field instrument, MRetFlow is measured at the mud level in the return line or from the valve opening of the Gumbo separator.
(h) SPP_pressA, SPP_pressB
Standpipe pressure is a value that indicates the mud injection pressure observed at the standpipe manifold. Usually, there are two measurements, A and B, for the same tag name because there are multiple flowlines for redundancy. Standpipe pressure is considered one of the important indicators of containment. Mud injection pressure indicates reaction to downhole conditions such as flowline blockage.
(i) MPP_SPM1, MPP_SPM2, MPP_SPM3
MPP_SPM is the frequency of mud pump strokes observed in each mud tank. Typically, there are two or three mud tanks for redundancy, so there are multiple variables with the same tag name.
(j)FlowIn
FlowIn is an estimate that is the sum of products of MPP_SPM and the corresponding tank volume.

表２では、抑留のメカニズムを、パックオフ／ブリッジ、差圧抑留、幾何抑留（坑井の幾何学的要因による抑留）に分類しているが、抑留の原因には様々なものが考えられ、この分類は１つの態様を示すに過ぎないものである。 [A-3] DetentionDetention is an event in which the drill string is pressed down inside the well and cannot be rotated and/or pulled out. The detention mechanism and causes of detention are shown in Table 2.

In Table 2, the mechanisms of detention are classified into pack-off/bridging, differential pressure detention, and geometric detention (detention due to geometric factors of the wellbore), however, there are various possible causes of detention, and this classification represents only one aspect.

Pack-off/bridging occurs when a well wall collapses, blocking the annulus formed between the well string and the well wall. One early sign of this is an increase in mud injection pressure.

Differential pressure arrest can occur when the pressure between the wellbore and the permeability of the formation increases and balance is lost. If the drill string continues to be in contact with the mud wall, differential pressure arrest can occur. Compared to a sudden collapse of the wellbore, differential pressure arrest occurs gradually, so examples of signs that may appear include a gradual increase in the rotational torque and a gradual increase in the hook load (when lifting the pipe).

Retention due to the geometry of the wellbore can occur due to a change in the direction of the wellbore, etc. One typical cause is a key seat. In a section of the wellbore where there is a sudden bend (for example, a dogleg), the drill string scrapes the wellbore wall during drilling or lifting, and the cross-sectional shape of the wellbore resembles a keyhole, hence the name key seat. One example of a sign of geometric retention is a sudden increase in hook load during lifting.

Three indicators are known to predict impoundment: drag, torque, and mud pressure. Drag is the axial frictional force generated as the drill string moves up and down the wellbore and is a key indicator. Deteriorating wellbore conditions will increase drag and increase the risk of drill string impoundment. Drag can be observed by comparing the hook load values between run-in-hole (RIH) and pull-out-of-hole (POOH) operations. High drag indicates significant downhole contact and can be a key indicator of impoundment. Torque is the tangential frictional force generated as the drill string rotates. Excessive torque, like drag, indicates significant downhole contact. Mud pressure is the injection pressure of the drilling mud. Excessive mud pressure indicates mud line blockage, where pack-off/bridging may have occurred or is occurring.

Indications of detention are complex, and the way and timing of their appearance vary depending on the detention mechanism and excavation work. It is difficult to quantify detention indications, and it should be noted that the presence of indications is confirmed after detention has occurred. It should also be noted that there may be cases where indications are not observed (e.g. packing off due to sudden collapse of the mine wall).

[B] Detention prediction using actual excavation data (detection of detention signs)
[B-1] Overall Configuration The detention prediction system according to this embodiment is a computer system composed of one or more computers. As shown in FIG. 1, the computer includes an input for inputting drilling data from a drilling system, a memory for storing various data including drilling data and processing data, and a program for operating the computer, a processor for performing various calculations and control, and an output for providing various data including the calculation results. In the drilling system, each drilling parameter is measured by each sensor or measuring device and transmitted to the computer system by wire or wirelessly. The drilling data obtained from the drilling system is stored in the computer memory, and a predetermined calculation is performed in the processor according to a predetermined program, detention sign detection is performed, and a detention sign probability representative of the detention risk is provided. The drilling data is typically time series data of the values of the drilling parameters. In this specification, the time series data also includes secondary data including time information calculated based on the time series data (for example, depth-based data including time information). In one aspect, a program for generating an alarm based on the probability of a detention sign and a predetermined alarm generation condition (such as a threshold value or duration of the probability value) is stored in the memory of the computer, and a predetermined calculation is performed in the processor according to the alarm generation program, and an alarm is generated when the alarm generation condition is satisfied. In this case, the detention risk is provided in the form of an alarm. Various data including the measurement data and the calculation results may be displayed on a display as appropriate for monitoring. The location where the computer system is placed is not limited, and may be placed at the drilling site or in a management station remote from the drilling site. The computer system may be cloud-based.

As shown in FIG. 2, the detention prediction system according to this embodiment includes generating a trained model using excavation data, and outputting the detention risk by inputting the excavation data into the trained model. The detention prediction model of the detention prediction system is a data-driven prediction model, and measurement data is input into the model in real time during excavation to provide the detention risk in real time. As described below, in this embodiment, feature engineering is performed on the input data input into the data-driven prediction model to generate the trained model and output the detention risk.

Figure 3 shows a conceptual diagram of supervised learning according to this embodiment. The detention prediction system according to this embodiment is a detention sign detection system. Assuming that signs appear just before detention, the system learns the signal pattern just before detention. Data obtained just before detention (for example, the last 6 minutes) is classified as label 1 (positive), and data obtained before that is label 0 (negative), and the correct label is learned as a binary value. Data obtained after detention occurs is ignored. The detention sign detection model is trained with binary labeled data, and the trained model provides the detention sign probability, i.e., detention risk, for an unknown input.

掘削データの形式は19変数の時系列データであり（表１参照）、センサから直接得られたもの（一次データ）、およびそれらから計算された二次データの両方が含まれる。上記時系列データとともに、各抑留ケースの開始時刻・終了時刻の情報が与えられている。抑留時刻は、掘削日報のフォーマットに基づき15分精度で与えられた。 [B-2] Data Collection: Actual drilling data in a unified format provided by multiple companies will be used.

The drilling data is in the form of 19 variables of time series data (see Table 1), including both data obtained directly from the sensors (primary data) and secondary data calculated from them. Along with the time series data, the start and end times of each detention case are given. Detention times are given with 15-minute accuracy based on the format of the daily drilling report.

Data cleansing was performed as data preprocessing. The data sampling interval was standardized to 4 seconds. In addition, since it was thought that it would be impossible to learn the signs of detention that appear within a few minutes with time information with an accuracy of 15 minutes, in order to improve the accuracy of the labels, the time series of each detention case was analyzed and the detention time was identified with an accuracy of 4 seconds. After preprocessing, the excavation data was stored in the computer's memory.

[B-3] Experiment 1 (Comparative experiment 1 using time domain data)
Detention prediction was performed based on the hypothesis that signs of detention are characteristically present in the time series immediately prior to detention. The preprocessed data was divided into 13-variable time series for each hour and saved as input data. In detention prediction, signs of detention are defined as abnormal. The input data obtained in the 6 minutes immediately prior to detention was considered abnormal and classified as label 1 (positive), and the input data obtained before that was considered normal and classified as label 0 (negative). The input data must be of a fixed size, and the raw signal is processed and divided into multiple input data by determining the size Tc of the input data. In the experiment, the size Tc of the input data was 1 hour, and the sampling interval was 4 seconds, so the input data consisted of 900 data. The obtained input data was classified as normal or abnormal.

To determine the range of this precursor, the time region of a certain length immediately before the time of detention is defined as the Tr region. In the experiment, Tr is set to 6 minutes. Input data before detention that does not overlap with the Tr region is classified as label 0 (negative), and input data before detention that overlaps with the Tr region is classified as label 1 (positive) (see the left diagram in Figure 4). Data after detention occurs is ignored.

Since this involves pattern recognition of multivariate one-dimensional data, a 1D-CNN model was used for learning and verification using time-domain input data (Figure 4, right). The 1D-CNN model using time-domain input data did not perform sufficiently well in detecting signs of detention. When the model identifies signs of detention, the data that can be referenced is limited in time when using a method that uses time-domain input data. If the input data is for one hour, information from more than one hour prior cannot be referenced. This is also a limitation of the 1D-CNN model that uses time-domain input data.

[B-4] Experiment 2 (Comparative experiment 2 using frequency domain data)
We hypothesized that signs of detention appear in the frequency domain and performed detention prediction. The time series data from comparative experiment 1 was converted to a power spectrum for learning. FFT was applied to the time series data obtained by subtracting the mean from the time series data and deleting the constants, and high frequency components above the Nyquist frequency (0.125 Hz) were cut off. The mean and standard deviation of the data set were calculated, and z-score standardization was performed to create the input data (see the right diagram in Figure 4). The results are shown in Table 6. Signs of detention were detected in only one of the six cases.

In data-driven methods, it is important to extract features such as parameter relationships and patterns that measurement data shows when it is normal and when it is abnormal. As shown in the results of comparative experiments 1 and 2, general anomaly detection methods based on the time domain or frequency domain are unable to effectively grasp abnormal features. Furthermore, for efficient learning, the length of measurement data that can be input is finite, and it is necessary to efficiently extract features from a limited amount of data. In other words, feature engineering is important in determining what features to incorporate from drilling data into machine learning.

[B-5] Concept of Depth Region Approach From the perspective of drilling engineers, signs of detention are confirmed by comparing each current measurement value with past values at the same depth obtained in the same well. The "one hour" used in the comparison experiment is considered to be insufficient as a reference time. In feature engineering based on the knowledge of drilling engineers, we focused on depth information. In drilling work, the same depth is passed many times due to bit replacement, etc. This accumulates the history of values measured at each depth, so anomalies can be seen by comparing the current value with the history at nearby depths in the same well. By introducing the concept of depth region, we considered a method to efficiently compress long-term information obtained in the same well.

Drilling engineers often analyze drilling parameter values by comparing them with recorded data obtained at similar depths. For example, in a RIH, if the hook load is smaller than previously recorded hook loads at the same depth, the drilling engineer may detect an increase in friction drag, which may indicate a worsening of downhole conditions. Drilling engineers may also look at the torque at the bit during drilling operations. Even if the same depth data records are not available, a sudden change in torque behavior with depth may indicate a change in the downhole environment.

[B-6] Scattering data of drilling parameters in the depth range (scattering diagram of relative depth and drilling parameters)
A method of expressing drilling data in the depth domain will be described with reference to Fig. 9. During drilling, drilling data including time series data of multiple drilling parameters is acquired in real time and stored in a computer memory. The multiple drilling parameters include bit depth and multiple parameters other than bit depth (e.g., rotation torque of the drill string, hook load, bit rotation speed, mud injection pressure, etc.) (see Table 1).

Continuing to refer to FIG. 9, the relative depth set for the current bit depth value is set as the first axis, and the value of one drilling parameter X corresponding to the relative depth is set as the second axis, and the scattered data of the relative depth and drilling parameter X is obtained using time series data before the current time. The range of the relative depth of the first axis is set based on the bit depth value D(t) at the current time (t). The value of the drilling parameter X (second axis) corresponding to each relative depth (first axis) value is obtained from the time series data. The scattered data is the data of the parameter X organized in the depth domain.

The scatter data is divided into the first scatter data of the first length closest to the current time, and the second scatter data of the second length prior to the first scatter data. The first length is, for example, 2 hours, and the first scatter data is the most recent data at the current time t. The second length is, for example, 7 days (excluding the first length), and is the historical data prior to the most recent data at the current time t. The first scatter data and the second scatter data have distinguishable RBG values and are displayed on one scatter plot.

When a scatter plot is drawn with "bit depth" on the vertical axis and "another drilling parameter" on the horizontal axis, the values of each drilling parameter are organized by depth, but time history information is lost. In this case, the depth is not an absolute value, but a relative depth to the current bit depth. For example, the range of bit depth (vertical axis) is defined as current bit depth - α [m] to current bit depth + β [m]. The value of "another drilling parameter" corresponding to the relative depth can be obtained from historical data stored in memory. In a scatter plot, time information is lost, but by displaying past data and most recent data in different colors, it is possible to compare current data with past data.

Scatter data showing hook loads in the depth region is illustrated in the left figure of Figure 5. The left figure of Figure 5 shows a comparison of hook load data (depth region) for the most recent day and the past week. The range of the vertical axis (bit depth) is 4000 [m] to 5000 [m], and the range of the horizontal axis (hook load) is 100 [ton] to 300 [ton]. Please note that the scatter plot shown in the left figure of Figure 5 is a color image (RBG image), and in reality, the hook load data for the most recent day before the occurrence of detention (red) and the hook load data for the past week (green) are displayed in a distinguishable manner by color (there are restrictions on the images that can be used as patent drawings, and color images cannot be used). For convenience of explanation, the hook load data for the past week in the left figure of Figure 5 and the hook load data for the most recent day before the occurrence of detention are shown separately in Figures 6 and 7, respectively. In the left figures of Figures 5, 6, and 7, the horizontal line near 4600 [m] indicates the detention occurrence depth. In this specification, the most recent data at the current time is called the "signal," and the historical data prior to the most recent data is called the "history." The signal represents the current data, and the history represents the past data.

By displaying the signal and history in different colors in the scatter plot, the values of drilling parameters in the past and present depth regions can be overlaid and compared, and features can be extracted based on the knowledge of the drilling engineer. In the left figure of Figure 5 (Figures 6 + 7), it can be observed that the signal appears at the left end of the history. The signal is scattered within a vertical rectangle in the left figure of Figure 5. This means that the current hook load is relatively small compared to the past recorded data at the same depth. In addition, the history spreads horizontally near the detention depth. According to the drilling daily report, the detention occurred near 4600 [m] during RIH operation. A relatively small hook load during RIH operation is considered to be a general pattern of detention precursor. This is because it is thought that friction between the drilling string and the wellbore increases when the underground conditions deteriorate. In addition, the horizontal spread of the scattered data near 4600 [m] indicates that the underground conditions had already deteriorated in the past few days. Looking at the actual depth data, it is shown that the depth data extracts meaningful features that indicate impoundment. By displaying a scatter plot of drilling parameters in such a depth range on a display in real time, it can be used to understand the underground conditions during drilling.

[B-7] 2D histogram of relative depth and drilling parameters As shown in the latter half of the flowchart in FIG. 9, a 2D histogram (so to speak, a depth region 2D histogram) can be created by setting the number of bins on the vertical and horizontal axes for the scatter diagram of bit depth and each drilling parameter. 2D histograms are shown in the right diagram of FIG. 5, FIG. 8, and FIG. 11 to FIG. 13. Please note that these 2D histograms are actually color images (RBG images). With reference to FIG. 10, the creation step of the depth region 2D histogram according to this embodiment will be described. In this embodiment, the 2D histogram of the depth region is created using the values of the bit rotation speed, the rotation torque of the drill pipe, the bit depth, the hook load, and the mud injection pressure, which are parameters of interest to drilling engineers. That is, a 2D histogram of the bit rotation speed in the depth region, a 2D histogram of the rotation torque in the depth region, a 2D histogram of the hook load in the depth region, and a 2D histogram of the mud injection pressure in the depth region are created.

Using the drilling data (most recent data) immediately prior to time (t), first scatter data is obtained with the first axis (vertical axis): relative depth (obtained from bit depth) for the depth (Dt) at a certain time (t), and the second axis (horizontal axis): one piece of data selected from bit rotation speed, rotation torque, hook load, and mud injection pressure. By setting the number of bins on the vertical and horizontal axes, the first scatter data is converted into a 2D histogram to create a first 2D histogram (having a first RBG value set).

The second scatter data is obtained using drilling data (historical data) from before the most recent data, with the first axis (vertical axis) being the relative depth (obtained from the bit depth) for the depth (Dt) at a certain point in time (t), and the second axis (horizontal axis) being one data selected from bit rotation speed, rotation torque, hook load, and mud injection pressure. By setting the number of bins on the vertical and horizontal axes, the second scatter data is converted into a 2D histogram to create a second 2D histogram (having a second RBG value set). The first RBG value set and the second RBG value set are distinguishable.

The first 2D histogram (with the first set of RBG values) and the second 2D histogram (with the second set of RBG values) are overlaid to create a superimposed 2D histogram (Fig. 8 right, Fig. 11, Fig. 13). The drilling parameter values organized in the depth domain are converted to a 2D histogram by setting the number of bins (V) on the vertical axis and the number of bins (H) on the horizontal axis. The depth domain data is converted to a 3D tensor (superimposed 2D histogram) by generating 2D histograms for both the signal and the history and overlaying them along the new axis (see Fig. 8). Low frequency bins are emphasized and high frequency bins are suppressed (suppressed) by applying log(1+x) to all elements of the tensor. The superimposed 2D histogram can be processed by CNN like image data and used as input data to the CNN model (see Figs. 11 to 13). Please note that in this specification, the input data to the CNN model may be referred to as a clip.

In this embodiment, data of any time is compressed into a 2D histogram of a certain size by feature engineering based on the depth domain approach. In the 2D histogram, the vertical axis (depth domain) is not an absolute value but a relative depth to the current bit depth (see FIG. 12). This 2D histogram can be used as input data organized in the depth domain. By converting the data distribution (scattered data) in the depth domain into a 2D histogram and generating a data clip as input, infinitely long time data can be converted into a finite size and used. The time domain information that is discarded at this time can be divided into past and present using the RBG values of the image. Thus, the advantage of the depth domain approach is that information over a long period of time is compressed into a 2D histogram of a certain size, and current information can be compared with past data of the same well without increasing the model parameters. The time domain information is roughly divided into "signals" that represent current data and "history" that represent past data, and color information is assigned to each to make them identifiable. By comparing the "signals" and "history" expressed in the image as differences in color, data features that indicate signs of detention can be highlighted.

The data preprocessing for the 3D-CNN model is summarized below. The scatter plot of the data for each drilling parameter in the depth region is converted into input data (clips) that can be processed by CNN. Specifically, a 2D histogram is created by setting the number of bins on the vertical and horizontal axes of the scatter plot, and this is used as the input data. Although the data for each drilling parameter in the depth region loses time history information, it is possible to compare the data in the time direction by dividing it into the most recent data (signal) and the previous data (history). Furthermore, by extracting only the depth near the current time, the resolution in the depth direction is maintained and a clip focusing on the relative depth to the current depth is created (see Figure 12). In this embodiment, a 3D clip in which each 2D histogram generated based on the data of the last two hours and the past seven days is superimposed is learned by 3D-CNN (see Figure 11). A 3D clip was created for each of the four drilling parameters: rotation speed, rotation torque, hook load, and mud injection pressure (see Figure 13).

本実施形態では、水平方向の解像度は垂直方向の解像度よりも高く設定してあり、水平方向の偏差のより正確な情報を取得するようになっている。垂直方向の範囲は、現在のビット深度に対して設定することで、着目している深さにフォーカスしている。表４に示す値は例示であって、これらの値は、掘削エンジニアによって適宜設定ないしチューニングされ得る。 The signal and history lengths, spatial range and resolution of the 2D histograms employed in the feature engineering of this embodiment are shown in Table 4.

In this embodiment, the horizontal resolution is set higher than the vertical resolution to obtain more accurate information of the horizontal deviation. The vertical range is set relative to the current bit depth to focus on the depth of interest. The values shown in Table 4 are examples, and these values can be set or tuned appropriately by the drilling engineer.

In the experimental example, the signal length Ts is the last two hours, and the history Th is the last seven days, but these time lengths are examples. For example, Ts may be the last hour or the last day, or Th may be set longer. Th may be the entire past period (excluding abnormal times such as detention). Th does not necessarily have to be a continuous period, and may be a combination of multiple periods. The number of bins V on the vertical axis is 32, and the number of bins H on the horizontal axis is 128, but these numbers are examples. Either or both of V and H may be changed to change the resolution of the 2D histogram. The range of the vertical axis (relative depth) of the histogram is α = 100, β = 10 from the current bit depth -α [m] to the current bit depth +β [m], but the values of α and β can be set appropriately by those skilled in the art. In the range of the horizontal axis of the histogram, in the experimental example, the bit speed (TD_spd) is [0, 300] rpm, the rotational torque (TD_trq) is [0, 100] kNm, the hook load (Hookload) is [0, 500] ton, and the mud injection pressure (SPP_pressA) is [0, 50] MPa, but these values are also examples and can be set appropriately by those skilled in the art.

In this embodiment, four parameters, namely, bit rotation speed (TD_spd), rotation torque (TD_trq), hook load (Hookload), and mud injection pressure (SPP_pressA), are used to avoid overly complex models. These parameters are typically investigated by drilling engineers when drilling abnormalities are suspected. In this embodiment, model training is performed using all four of the above parameters, but model training may also be performed using any one, or a combination of two or three of the above parameters. Furthermore, parameters other than the above four parameters may be used in addition to one or more parameters selected from the above four parameters.

[B-8] 3D-CNN model The 3D-CNN model according to this embodiment is a data-driven detention prediction model that makes it possible to utilize historical information of the same well by performing pattern learning of depth-area data obtained by feature engineering based on the knowledge of drilling engineers. In the 3D-CNN approach, it is assumed that detention signs appear locally in the depth data, and the 3D-CNN model uses the localized features to perform detention sign detection where features appear in the depth area.

The process of generating input data for the model is shown in Figure 13. For each drilling parameter, the signal and history are converted into 2D histograms, respectively, and stacked along new dimension axes, so that the input data for the model becomes a 4D tensor of size (T, P, V, H), where T (time) = 2 (signal and history), P (drilling parameter) = 4, V (vertical axis resolution) = 32, and H (horizontal axis resolution) = 128. Multiple clips are further stacked along new axes to form mini-batches, so that the input data for the 3D-CNN becomes a 5D tensor of size (B, T, P, V, H), where B is the batch size (=16).

3D-CNN models can take 4D tensors (single channel) or 5D tensors (multi-channel) as input. Here, we adopt a multi-channel 3D-CNN and feed the 5D-clip to the model. 3D-CNN applies convolution to a tensor of shape (B, P, T, V, H) to extract features in three dimensions later. Since the input data is 5D, we need to select the three dimensions for which local feature extraction is required. If we ignore the batch dimension and assume that local feature extraction is required for V and H, but not for P and T, the 3D-CNN model can work with either (B, P, T, V, H) or (B, T, P, V, H).

In this embodiment, the Grad-CAM analysis is considered, and the shape of the tensor is transposed to (B, T, P, V, H) (see FIG. 14). Since the second dimension is compressed to a single feature in the first layer, when P (the drilling parameter dimension) is set to the second dimension, the contribution of each drilling parameter cannot be analyzed in the final convolutional layer. The 3D-CNN model uses a 3D kernel to learn the pattern of the 5D tensor in the (P, V, H) axis. The model architecture of the 3D-CNN model according to this embodiment is shown in Table 5. Since local feature extraction of the P axis is not required, the 3D kernel is effectively a 2D kernel. The purpose of using 3D-CNN is to enable multi-parameter Grad-CAM analysis rather than 3D pattern recognition.

[B-9] Experiment 3 (Experiment using depth region data)
A multivariate 3D clip (4D clip) was used as input, and detention prediction was performed using a 3D-CNN model (see Figure 13). The 4D clip was input to the 3D-CNN model after applying emphasis and normalization of low-frequency parts. 4D clips obtained within 6 minutes before detention were classified as label 1 (positive), and those obtained before that were classified as label 0 (negative), and learning and validation were performed. The model output is the probability (continuous value from 0 to 1) that the input clip is a detention predictor.

[B-10] Experimental Results Table 6 shows the success or failure of the detention mechanism and detention prediction detection for the six cases used in the experiment, compared between the 3D-CNN model (depth domain) and the 1D-CNN model (frequency domain). The detention prediction detection is within 6 minutes immediately before detention. Each model calculates the detention risk at 4-second intervals, that is, there is an opportunity to issue 90 alarms in the 6 minutes before detention. If at least one alarm is generated within 6 minutes before detention, it is determined that the prediction detection is successful. The 1D-CNN model (frequency domain) was able to detect the detention prediction in only one case out of six cases, whereas the 3D-CNN model (depth domain) was able to detect the detention prediction in three cases out of six cases. It is considered that the depth domain approach was effective for differential pressure detention, whose predictions appear gradually, because the characteristics are likely to appear in the depth data. Please note that the success rate of 3/6 does not allow us to evaluate the performance of this model. For example, it should be noted that some detentions may occur suddenly and that the warning signs may be difficult or impossible to detect.

Figure 15 shows real-time predictions of the 1D-CNN model and the 3D-CNN model for Case 3. In Case 3, the 1D-CNN model was unable to detect the signs of detention, whereas the 3D-CNN model was able to detect them. In the output of the 3D-CNN model, the high risk of detention continued from 30 minutes before the occurrence of detention (1). The 3D-CNN model captured the signs of detention at the time of lifting, and the lifting was still in progress when the detention occurred. It is believed that the signs could have been detected from the work several hours earlier.

Looking at Figure 16, which uses learning data, it can be seen that in this embodiment, even though only the six minutes immediately prior to detention were used to learn the signs of detention (no data prior to 23:54 was used), the risk of detention begins to rise from 23:00, one hour before the occurrence of detention. The model according to this embodiment is able to detect signs that gradually appear from six minutes prior to detention.

In a detention prediction system using a data-driven detention prediction model, an alarm is generated based on the detention prediction probability output from the model. In reality, since it is humans who make operational decisions in response to alarm generation, it is important to be able to trace the cause of alarm generation in order to make appropriate decisions. In this embodiment, the basis of the detention prediction probability on which the alarm was generated is traceable by using Grad-CAM (Gradient-based Class Activation Map). The area of the input data that contributed to the detention prediction probability on which the alarm was generated is visualized by Grad-CAM to show the basis of the alarm generation. Specifically, a feature map in which the input features are extracted by the convolution layer is weighted by the sensitivity to the detention prediction probability and drawn. Grad-CAM has a lower resolution than the input because it is after convolution. As shown in Figure 17, it reacts particularly strongly to the hook load (during the dismounting work), and it can be seen that a reasonable alarm has been generated.

[C] In the unsupervised learning model experiment, a supervised learning model was used to give the data set of 6 minutes just before the start of detention a label of 1 (positive). This 6-minute period during which it is assumed that detention premonitions occur is arbitrary. It is difficult to identify when the detention premonitions actually began, and it is not possible to give a label to all detention premonitions.

Using the input data (clip) according to this embodiment, it is also possible to perform unsupervised learning by focusing on the negative data set. That is, by combining the depth region data that captures the characteristics of the detention sign well with the label-independent unsupervised learning, only the normal pattern is learned by the unsupervised learning, and the sign is detected by deviation from the pattern. In one aspect, the normal data targeted by the unsupervised learning is data for a period excluding the "detention state" and the "period in which the detention sign is assumed to occur". The learning model is input data generated based on the learning excavation data so that the learned model can discriminate abnormal data and provide the detention risk, and learns the input data within a certain period before the detention and the detention state as normal data. In one aspect, a learning model that minimizes the error between the input data and the output data is generated by an autoencoder. The autoencoder is an example, and any method may be used to discriminate abnormal data by methods such as feature extraction, classification, and error discrimination by having the unsupervised model learn the characteristics of normal data.

Claims (12)

1. A method for predicting detention risk, comprising: generating input data based on a plurality of drilling parameters obtained during drilling; and providing the input data to a detention prediction model to provide a detention risk, the method comprising:
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
Generate a 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on the current bit depth value as input data;
providing a retention risk in response to the input of the 2D histogram generated based on a current bit depth;
Internment prediction methods. The value of the one drilling parameter corresponding to the relative depth set based on the current bit depth value is:
The data is divided into first distribution data obtained within a predetermined period immediately preceding the current time point and second distribution data obtained prior to the first distribution data,
a first 2D histogram based on the first scatter data and a second 2D histogram based on the second scatter data are generated, the first 2D histogram and the second 2D histogram being distinguishable by color;
The input data is a superimposed 2D histogram in which the first 2D histogram and the second 2D histogram are superimposed.
The method for predicting detention according to claim 1 . The one drilling parameter other than the bit depth is selected from a bit rotation speed, a rotation torque of a drill pipe, a hook load, and a mud injection pressure.
The method for predicting detention according to claim 1 or 2. The detention prediction model is machine-learned to provide detention risk in response to an input of a 2D histogram generated based on training drilling data.
The method for predicting detention according to any one of claims 1 to 3. A processor and a memory,
The memory includes:
an input data generation program that generates input data based on a plurality of drilling parameters obtained during drilling;
A detention prediction program that provides a detention risk in response to input of the input data;
Equipped with
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
The input data generation program is executed by the processor to generate, as input data, a 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on a current bit depth value;
The retention prediction program is executed by the processor to provide a retention risk in response to an input of the 2D histogram generated based on a current bit depth.
Detention prediction system. the detention prediction program provides a detention risk in response to an input of a plurality of 2D histograms obtained based on a plurality of drilling parameters;
The memory includes a program that, when executed by the processor, visualizes and displays the contribution of each drilling parameter to the detention risk.
The detention prediction system according to claim 5 . A detention prediction program that generates input data based on a plurality of excavation parameters obtained during excavation, and provides a detention risk based on the input data,
The program, when executed by a processor, causes the processor to:
When acquiring the plurality of drilling parameters, the plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
generating, as input data, a 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on the current bit depth value;
providing a retention risk in response to the input of the 2D histogram generated based on a current bit depth;
A program that enables a computer to perform detention predictions. acquiring drilling data including a plurality of drilling parameters;
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
The relative depth set based on the bit depth value at a certain time t is set as a first axis, and the value of the one excavation parameter corresponding to the relative depth is set as a second axis, and the scatter data is acquired using the excavation data before a certain time t;
converting the scatter data into a 2D histogram to generate input data;
A method for generating input data for the detention prediction model. Acquire learning drilling data including a plurality of drilling parameters and detention information;
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
A 2D histogram of values of the one drilling parameter corresponding to a relative depth set based on the bit depth value at a certain time point t is generated and used as input data;
Assuming that a warning sign appears within a certain period of time before detention, the learning model is trained to provide a detention risk according to input data generated based on the training drilling data;
How to generate a trained detention prediction model. The learning model learns input data generated based on the learning excavation data, excluding input data for a certain period before detention and input data for a detention state, as normal data so that the learned model can distinguish abnormal data and provide detention risk.
A method for generating a trained detention prediction model as described in claim 9 . acquiring drilling data including time series data of a plurality of drilling parameters;
The plurality of drilling parameters include a bit depth and at least one drilling parameter other than the bit depth, and the value of the bit depth and the value of the one drilling parameter correspond to each other by time information;
Using the excavation data, a relative depth set based on the current bit depth value is set as a first axis, and the value of the one excavation parameter corresponding to the relative depth is set as a second axis, and scattering data is acquired using excavation data before the current time;
The scatter data is divided into first scatter data acquired during a predetermined period immediately preceding the current time point and second scatter data acquired prior to the first scatter data,
The first scatter data and the second scatter data are displayed on a display as one scatter diagram in an identifiable manner.
A method for compressing drilling data that focuses on the depth domain and compresses information obtained from the same well over a long period of time . An excavation rig comprising the retention prediction system according to any one of claims 5 and 6 .

Images