Anomaly detection system and anomaly detection method

Abstract

It accurately detects anomalies in well drilling. [Solution] The anomaly detection system 10 is a system for detecting anomalies in well drilling, and comprises: a reference value acquisition unit 11 that acquires reference values, which are values ​​of multiple types of data related to drilling in a reference drilling that is different from the drilling to be detected as an anomaly; a target value acquisition unit 12 that acquires target values, which are values ​​of multiple types of data related to drilling in the drilling to be detected as an anomaly; a parameter calculation unit 13 that defines the relationships between multiple types of data and includes parameters, calculates the distribution of parameters in the reference based on the reference values, and calculates the distribution of parameters in the detection target based on the target values; and an anomaly detection unit 14 that compares the distribution of parameters in the reference with the distribution of parameters in the detection target, and detects an anomaly in the drilling to be detected based on the comparison result.

Description

Technical Field

[0001] The present invention relates to an abnormality detection system and an abnormality detection method for detecting abnormalities in well drilling.

Background Art

[0002] In the development of oil or natural gas, well drilling is an operation with high risks. Downhole troubles occurring during drilling operations are regarded as serious problems. The main downhole trouble is sticking. Sticking is a phenomenon in which drill pipes or the like become immovable. There are various sticking mechanisms, including differential pressure sticking, formation instability, deposition of drilling cuttings, and poor wellbore trajectory. When sticking occurs, countermeasures such as jarring, which vibrates the drilling assembly up and down to apply a large impact, and injecting a sticking release fluid are required. In more serious cases, it leads to sidetrack drilling or abandonment of the well. In addition to the increased costs due to such measures, the period for dealing with sticking requires interruption of operations, which also significantly contributes to an increase in NPT (non-productive time). As a result, sticking causes significant economic losses and is estimated to result in losses exceeding $200 million annually.

[0003] Conventionally, methods for detecting abnormalities in well drilling, including sticking, have been proposed. For example, Patent Document 1 shows that an estimation model for estimating values related to drilling is generated, and an estimated value estimated from the generated estimation model is compared with a measured value to detect an abnormality.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

[0006] The present invention has been made in view of the above, and an object thereof is to provide an abnormality detection system and an abnormality detection method capable of accurately detecting abnormalities in the excavation of a mine shaft.

Means for Solving the Problem

[0007] In order to achieve the above object, an abnormality detection system according to the present invention is an abnormality detection system for detecting an abnormality in the excavation of a mine shaft, and includes a reference value acquisition means for acquiring reference values which are values of a plurality of different types of data related to excavation in a reference excavation which is different from the excavation which is the detection target of the abnormality, a target value acquisition means for acquiring target values which are values of a plurality of different types of data related to excavation in the excavation which is the detection target of the abnormality, a parameter calculation means for calculating a distribution of the parameter in the reference based on the reference value acquired by the reference value acquisition means and calculating a distribution of the parameter in the detection target based on the target value acquired by the target value acquisition means for a preset model that defines the relationship of a plurality of types of data and includes a parameter, and an abnormality detection means for comparing the distribution of the parameter in the reference calculated by the parameter calculation means with the distribution of the parameter in the detection target and detecting an abnormality in the excavation of the detection target based on the comparison result.

[0008] In the abnormality detection system according to the present invention, the distribution of the parameter in the reference is compared with the distribution of the parameter in the detection target, and an abnormality is detected based on the comparison result and the reference. The distribution of the parameter appropriately reflects the state of the excavation. Therefore, according to the abnormality detection system according to the present invention, it is possible to accurately detect an abnormality in the excavation of a mine shaft.

[0009] The parameter calculation means may calculate the parameter distribution for both the standard and the target of detection using Bayesian estimation. With this configuration, the parameter distribution can be calculated appropriately and reliably, and anomalies in well drilling can be detected appropriately, reliably, and with high accuracy.

[0010] The distribution of parameters in the criteria and the distribution of those parameters in the detection target may both relate to a hierarchical Bayesian model in which at least one of the multiple types of data is a function of one of several types of data. With this configuration, anomalies in well drilling can be detected with even greater accuracy depending on one of the multiple types of data.

[0011] At least one of the distributions of the parameters in the standard and the distributions of the parameters in the detection target may be a function of drilling depth or drilling time, which are included in multiple types of data. With this configuration, anomalies in well drilling can be detected with even greater accuracy depending on the drilling depth or drilling time.

[0012] Multiple types of data may be data that can be measured outside the well. With this configuration, data values ​​can be easily acquired, and anomalies can be detected easily and accurately.

[0013] The data of multiple types may include data relating to the ground load of the drill bit used for drilling and data relating to the rotational torque of the drill pipe used for drilling, and the model may be a linear regression model. With this configuration, anomalies can be detected reliably and accurately.

[0014] The anomaly detection means may also calculate the distribution of estimated values ​​of a preset type of data from the target values ​​acquired by the target value acquisition means, based on the distribution of parameters and models in the standard calculated by the parameter calculation means, compare the calculated distribution with the target values ​​of the data of that type, and detect anomalies in the drilling target based on the comparison result. With this configuration, anomalies in well drilling can be detected with even greater accuracy according to the distribution of estimated values ​​of the data of the preset type.

[0015] The model may include an error term, and the anomaly detection means may calculate the distribution of the error term from the target values ​​acquired by the target value acquisition means, based on the distribution of parameters in the reference and the model calculated by the parameter calculation means, and then detect anomalies in the drilling target based on the calculated distribution. With this configuration, anomalies in well drilling can be detected with even greater accuracy depending on the distribution of the error term.

[0016] Incidentally, in addition to being described as an invention of an anomaly detection system as described above, the present invention can also be described as an invention of an anomaly detection method as follows. These are substantially the same invention, differing only in category, and produce similar functions and effects.

[0017] In other words, the anomaly detection method according to the present invention is an anomaly detection method which is an operation method of an anomaly detection system for detecting anomalies in well drilling, and includes: a reference value acquisition step of acquiring a reference value which is a value of multiple different types of data related to drilling in a reference drilling that is different from the drilling to be detected for an anomaly; a target value acquisition step of acquiring a target value which is a value of multiple different types of data related to drilling in the drilling to be detected for an anomaly; a parameter calculation step of calculating the distribution of the parameter in the reference based on the reference value acquired in the reference value acquisition step and the distribution of the parameter in the target for detection based on the target value acquired in the target value acquisition step for a preset model that defines the relationship between multiple types of data and includes parameters; and an anomaly detection step of comparing the distribution of the parameter in the reference calculated in the parameter calculation step with the distribution of the parameter in the target for detection and detecting an anomaly in the drilling to be detected based on the comparison result. [Effects of the Invention]

[0018] According to the present invention, abnormalities in well drilling can be detected with high accuracy. [Brief explanation of the drawing]

[0019] [Figure 1] This figure shows the functional configuration of an anomaly detection system according to an embodiment of the present invention. [Figure 2] This graph shows an example of the distribution of parameters calculated in an anomaly detection system. [Figure 3] This graph shows another example of the distribution of parameters calculated in an anomaly detection system. [Figure 4] This flowchart shows an anomaly detection method, which is a process performed in an anomaly detection system according to an embodiment of the present invention. [Figure 5] This graph shows the anomaly scores for each drilling depth for each pattern in the case study. [Figure 6]This graph shows the anomaly scores for each drilling depth for each pattern in the case study. [Figure 7] This graph shows the relationship between the number of correct and false positives for each pattern in the case study. [Modes for carrying out the invention]

[0020] Embodiments of the anomaly detection system and anomaly detection method according to the present invention will be described in detail below with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant explanations are omitted.

[0021] Figure 1 shows the functional configuration of the anomaly detection system 10 according to this embodiment. The anomaly detection system 10 is a system (device) that detects anomalies in well drilling. Drilling is performed, for example, on the seabed or on land. That is, drilling is, for example, offshore drilling or onshore drilling. The drilling itself can be the same as in the conventional method. For example, a drill bit connected to a drill pipe rotates due to the force transmitted from the drill pipe to drill. In addition, drilling fluid is injected into the well being drilled by a pump to remove drilling debris from the well. The drilling fluid is, for example, mud or seawater. Drilling may be performed in the vertical direction or in other directions (for example, horizontal direction). Note that drilling does not necessarily have to be performed as described above and may be performed by any method.

[0022] The anomaly detected by the anomaly detection system 10 is a jamming (stack). The anomaly detected by the anomaly detection system 10 may also be a precursor to jamming. Furthermore, the anomaly detection system 10 may detect anomalies in well drilling other than jamming and precursors to jamming. The anomaly detection system 10 may also detect the risk of anomalies. As described later, the anomaly detection system 10 acquires values ​​of multiple different types of data related to drilling and detects anomalies based on the acquired values.

[0023] The anomaly detection system 10 is specifically composed of a computer including hardware such as a CPU (Central Processing Unit) and memory. The various functions of the anomaly detection system 10, as described later, are performed by these components operating through programs or the like. The anomaly detection system 10 may be implemented in a single computer, or it may be implemented in a computer system composed of multiple computers connected to each other by a network.

[0024] Next, the functions of the anomaly detection system 10 according to this embodiment will be described. As shown in Figure 1, the anomaly detection system 10 is configured to include a reference value acquisition unit 11, a target value acquisition unit 12, a parameter calculation unit 13, and an anomaly detection unit 14.

[0025] The reference value acquisition unit 11 is a reference value acquisition means that acquires reference values, which are values ​​of multiple different types of data related to drilling in a reference drilling that is different from the drilling that is the target of abnormality detection. The multiple types of data may be data that can be measured outside the well.

[0026] The anomaly detection system 10 uses the values ​​of multiple different types of data that can be measured during drilling to detect anomalies. The data values ​​are, for example, measured values ​​(actual values) (drilling data). These multiple types of data are related to each other. Specifically, the relationships between these multiple types of data can be defined by a model described later. For example, the model is an estimation model in which the values ​​of one type of data are explanatory variables and the values ​​of another type of data are dependent variables. In this case, the values ​​of the other type of data can be inferred from the values ​​of the first type of data.

[0027] Multiple types of data are pre-configured. Furthermore, there may be three or more types of data. For example, in the estimation model described above, the explanatory and dependent variables may each be multiple types of data. Multiple types of data may also be measurable outside the well being drilled, such as on the surface or on a ship (specifically, on a rig on the surface or on a ship). The data in the example shown below is measurable outside the well. Note that some or all of the multiple types of data may only be measurable inside the well (for example, by measuring information inside the well using sensors).

[0028] Multiple types of data may be measured in time series. For example, they may be measured at a predetermined sampling period (e.g., 0.25 Hz).

[0029] The values ​​of multiple types of data used in the anomaly detection system 10 (reference values ​​acquired by the reference value acquisition unit 11, and target values ​​acquired by the target value acquisition unit 12, which will be described later) are associated with each other. The association of the values ​​of multiple types of data is the same as the relationship between multiple types of data in the model described later. For example, the values ​​of multiple types of data are combinations of values ​​related to the same timing. The above association is performed in advance. In the anomaly detection system 10, the values ​​of multiple types of data are used in processing in units of the associated combinations.

[0030] Specifically, the multiple types of data include data relating to the ground load of the drill bit used for drilling, and data relating to the rotational torque of the drill pipe used for drilling. The data relating to the ground load of the drill bit is, for example, data that shows the installation load of the drill bit itself. Alternatively, the data relating to the ground load of the drill bit may not be the installation load of the drill bit itself, but data corresponding to the installation load of the drill bit. Alternatively, the data relating to the ground load of the drill bit may be data that shows the installation load of the drill bit itself, calculated from data corresponding to the installation load of the drill bit. The ground load of the drill bit is, for example, the ground load of the drill bit (surface-WOB) calculated from the suspension load. The ground load of the drill bit may also be the ground load of the drill bit (downhole-WOB) measured inside the well.

[0031] The data relating to the rotational torque of the drill pipe used for drilling may, for example, be data that represents the rotational torque of the drill pipe itself. Alternatively, the data relating to the rotational torque of the drill pipe may not be the rotational torque of the drill pipe itself, but rather data that corresponds to the rotational torque of the drill pipe. For example, the data relating to the rotational torque of the drill pipe may be the drilling torque or the rotational torque of the drill bit. Or, the data relating to the rotational torque of the drill pipe may be other data used to calculate the drilling torque (for example, the slurry injection pressure). Furthermore, the data relating to the rotational torque of the drill pipe may be data that represents the rotational torque of the drill pipe itself, calculated from data that corresponds to the rotational torque of the drill pipe.

[0032] In this case, the data may include multiple types, such as drilling depth or drilling time. Drilling depth is, for example, data relating to the length of the drill pipe in the ground. Data relating to the length of the drill pipe is, for example, the total length of the drill pipe or something equivalent (for example, the depth of the drill bit). Since the length of the drill pipe is often not included in the items measured by the drilling control device, the depth of the drill bit may be used as the data. Strictly speaking, the length of the drill pipe = the depth of the drill bit (= the length from the drill floor (the location outside the well where the drilling equipment is installed) to the bottom end of the drill pipe) + hook height (HookHeight) (the length from the drill floor to the top end of the drill pipe). However, since the hook height is not a large value, it may be ignored (= considered as 0) in the calculation. Drilling time is, for example, the elapsed time of drilling with the start of drilling set to 0.

[0033] Furthermore, other types of data may also be used for detecting anomalies in drilling, as long as their relationships can be defined by a model that includes parameters.

[0034] The values ​​of multiple types of data are measured, for example, by a measuring device installed in advance on the drilling equipment. The reference value acquisition unit 11 receives and acquires, for example, reference values ​​from the measuring device, which are measured values ​​of multiple different types of data related to drilling in a reference drilling operation. Alternatively, the reference value acquisition unit 11 may accept a reference value input operation from the user to the anomaly detection system 10 and acquire the reference value. Furthermore, the reference value acquisition unit 11 may acquire the reference value by any other method. The reference value acquisition unit 11 acquires a sufficient number of reference values ​​for processing by the parameter calculation unit 13, which will be described later.

[0035] The reference value acquired by the reference value acquisition unit 11 is the value obtained when it is assumed that no abnormality has occurred during well drilling. Furthermore, the reference value may be the value obtained at a time prior to the timing of detection of the abnormality in the well that is subject to abnormality detection. For example, a value that was not detected as abnormal by the abnormality detection system 10 (specifically, the abnormality detection unit 14) may be used as the reference value for subsequent detections.

[0036] For example, the reference value is a measurement taken during drilling in a pre-set well and time period where it is assumed that no abnormalities have occurred. For example, when using the above as multiple types of data, the reference value is the measurement taken during a certain period (e.g., 1 to 24 hours) in the well targeted for abnormality detection, several tens of minutes to several hours before the timing of abnormality detection (e.g., 2 hours before). Furthermore, the time period for measuring the reference value does not have to be a fixed period; it may be any time period during which measurement is possible for the reference value. Also, the time period related to the reference value is not necessarily limited to the above time period. For example, the time period may be several minutes to several days before the timing of abnormality detection. In addition, the reference value may be a measurement taken in a well other than the well targeted for abnormality detection (e.g., an offset well).

[0037] Furthermore, the values ​​related to the data used in the anomaly detection system 10 may be values ​​that have been corrected using conventional methods after being actually measured. Alternatively, the values ​​related to the data may be converted from measured values ​​of another type of data. In these cases, correction or conversion may be performed in the functional unit of the anomaly detection system 10. Also, the values ​​related to the data may not be measured values, as long as they relate to excavation and can be used in subsequent processing by the anomaly detection system 10. The reference value acquisition unit 11 outputs the acquired reference values ​​to the parameter calculation unit 13.

[0038] The target value acquisition unit 12 is a target value acquisition means that acquires target values, which are the values ​​of multiple different types of data related to excavation in the excavation that is the target of abnormality detection.

[0039] The target value acquisition unit 12 acquires data values ​​of the same type as the reference values ​​acquired by the reference value acquisition unit 11 as target values. The target value acquisition unit 12 acquires target values ​​in the same manner as the reference value acquisition unit 11 acquires reference values. The target value acquisition unit 12 acquires a sufficient number of target values ​​for processing by the parameter calculation unit 13, which will be described later.

[0040] For example, the target value is a measurement taken during drilling in a pre-defined well and time period that is subject to anomaly detection. When using the above as multiple types of data, the target value is the measurement taken in the well subject to anomaly detection over a certain period (e.g., 1 hour) that corresponds to the timing of anomaly detection.

[0041] To utilize the detection of anomalies by the anomaly detection system 10 for avoiding dangers during actual excavation, the acquisition of target values ​​by the target value acquisition unit 12 may be performed in real time or near real time. The target value acquisition unit 12 outputs the acquired target values ​​to the parameter calculation unit 13.

[0042] The parameter calculation unit 13 is a parameter calculation means that defines the relationships between multiple types of data and calculates the distribution of the parameter in the reference based on the reference value acquired by the reference value acquisition unit 11, and calculates the distribution of the parameter in the detection target based on the target value acquired by the target value acquisition unit 12, for a pre-set model that includes parameters. The parameter calculation unit 13 may calculate the distribution of the parameters in the reference and the detection target, respectively, by Bayesian estimation. The model may be a linear regression model. The model may include an error term.

[0043] As described above, the model defines the relationship between multiple types of data relating to the reference value and the target value. The model includes parameters whose distribution is calculated by the parameter calculation unit 13. The distribution of the calculated parameters is used for detecting anomalies. The model may be stored in the parameter calculation unit 13 in advance. The model may also reflect prior knowledge (for example, physical knowledge (equations of motion, etc.), experimental knowledge (experimental measurements, etc.), and operational knowledge (other knowledge obtained from actual operation and experience)).

[0044] For example, if the data types used for the reference value and target value are the ground load of the drill bit and the rotational torque of the drill pipe, a linear regression model can be used. The model is, for example, equation (1) below. Torque = w0 + w1 × WOB + ε (1) In the above equation, Torque represents the rotational torque value of the drill pipe (e.g., a value in [kN / m]). WOB represents the ground load value of the drill bit (e.g., a value in [kN]). w0 and w1 are parameters. The above model (Torque model) is an estimation model that estimates (predicts) the rotational torque value of the drill pipe (Torque) using the ground load value of the drill bit (WOB) and the parameters w0 and w1.

[0045] The parameters w0, w1, and ε in the above model represent the intercept, slope, and error term, respectively. Physically, parameter w0 is the sum of borehole wall friction, viscous friction with the drilling fluid, mechanical resistance, and friction losses, and is a parameter that depends on the characteristics of the geological formation, drilling mud, and drilling equipment. Parameter w1 is the sensitivity of the ground load of the drill bit to the rotational torque of the drill pipe, and is a parameter that varies depending on the coefficient of friction at the bottom of the borehole and the characteristics of the drill bit. ε is an error term that represents errors that cannot be expressed in the model, and includes the displacement of the rotational torque of the drill pipe due to dynamic effects such as data noise and vibration of the drill pipe.

[0046] The parameter calculation unit 13 receives reference values ​​from the reference value acquisition unit 11. Based on the input reference values, the parameter calculation unit 13 calculates the distribution of model parameters in the reference (excavation, which is the reference related to the reference values). The parameter calculation unit 13 receives target values ​​from the target value acquisition unit 12. Based on the input target values, the parameter calculation unit 13 calculates the distribution of model parameters in the detection target (excavation, which is the detection target related to the target values). The parameter calculation unit 13 has a method for calculating the parameter distribution stored in advance, and calculates the distribution of the two parameters described above using that calculation method (the same calculation method).

[0047] The parameter calculation unit 13 calculates a probability distribution as the distribution of the parameters. For example, the distribution of the parameters in the above model can be considered a normal distribution. In this case, the parameter calculation unit 13 calculates the mean and variance of the distribution as the distribution of the parameters. Furthermore, the parameter distribution calculated by the parameter calculation unit 13 may be other than those described above, as long as it can be used by the anomaly detection unit 14 described later.

[0048] For example, the parameter calculation unit 13 probabilistically calculates the parameter distribution using the Bayesian linear regression shown below. In estimating the rotational torque value (Torque) of the drill pipe using the linear regression model of equation (1) above, there are uncertainties due to the simplification and assumptions of the model, as well as uncertainties in the data used for estimation itself. It is necessary to appropriately handle and quantify these uncertainties. Therefore, we consider using the framework of Bayesian estimation (Bayesian inference) to assume a probability distribution for the parameters and reflect the uncertainties. Specifically, a prior distribution is given to the parameters, and by updating that distribution based on the data (e.g., measured values), a posterior distribution is derived, and the parameter distribution is calculated (estimated).

[0049] The linear regression model is given by equation (2) below. t n =w T x n +ε (2) t in the above equation nis the target variable (objective variable) (Torque). x n is the vector of explanatory variables (constant term and WOB). w T is the weight vector (vector of parameters w0 and w1). ε is the error term. Assuming that the error term ε follows a normal distribution with mean 0 and variance σ 2 the target variable t n follows the normal distribution of the following equation (3). t n ~N(w T x n ,σ 2 ) (3)

[0050] Here, a prior distribution is introduced for the parameter w. Assuming that w follows a multivariate normal distribution, it is expressed as in the following equation (4). w~N(μ0,Σ0) (4) In the above equation, μ0 is a vector of the dimension of the number of elements of w, with the mean value of each element (parameter) of w as the element value. Σ0 is the variance-covariance matrix of each element (parameter) of w.

[0051] According to Bayes' theorem, the posterior distribution of the parameter can be expressed in a form proportional to the product of the likelihood distribution and the prior distribution, and the posterior distribution p(w|X,t) can be derived from the following equation (5). [[ID=三十一]] p(w|X,t)∝p(t|w,X)p(w) (5) In the above equation, X and t are the training set consisting of the observed data (e.g., measurement values) x n ,t n (n = 1,..., N). N is the number of observed data. X is the observed data matrix (matrix obtained by arranging the observed data x n ). t is the vector of target variables (vector obtained by arranging the observed data t n ). Furthermore, by using the derived posterior distribution, when a new data point x´ is given, the range within which the target variable t´ can take values can be probabilistically obtained as the predictive distribution p(t´|x´,X,t), which can be expressed as in the following equation (6). p(t´|x´,X,t)=∫p(t´|x´,w)p(w|X,t)dw (6)

[0052] In this embodiment, in order to enable sequential estimation, a normal distribution is assumed as the conjugate prior distribution of the parameter w, as shown in equation (4), and the posterior distribution of the parameter w is analytically determined. This posterior distribution is the parameter distribution calculated by the parameter calculation unit 13. The accuracy of the noise (inverse variance) and the accuracy of the parameter prior distribution are given by the hyperparameters α and β shown in equations (7) and (8) below. t n ~N(w T x n ,β -1 ) (7) w~N(0,α -1 I) (8) In the above equation, 0 is the zero vector with dimension equal to the number of elements in w. I is the identity matrix with size equal to the number of elements in w × the number of elements in w.

[0053] Regarding the prior distribution, if there is generally no strong prior information, a normal distribution with a mean of 0 and a predetermined value (e.g., a large value) as the variance can be used as an uninformative prior distribution. The parameters w0 and w1 of the Torque model are also expected to fluctuate depending on the drilling conditions, and there is no domain knowledge that they take specific values. Therefore, it is considered appropriate to update the parameters w0 and w1 based on observed data without bias. If the prior distribution and likelihood function are normal distributions, the posterior distribution will also be a normal distribution. Therefore, the shape of the distribution can be analytically determined using equation (5). Specifically, it will be a normal distribution with the covariance matrix S in equation (9) and the mean m in equation (10). S=(αI+βX T X) -1 (9) m=βSX T t (10)

[0054] The predictive distribution of t', including the hyperparameters, can be expressed as shown in equation (11) below. In this case, the hyperparameters α and β are fixed during calculation. p(t´|t)=∫p(t´|x´,w,β)p(w|X,t,α,β)dw (11)

[0055] The hyperparameters α and β should be pre-set based on prior knowledge of the data or domain knowledge. If there is no prior knowledge or domain knowledge, α should be 10. -4 It should be pre-set to a small value like this. This setting provides a data-driven, flexible prior distribution.

[0056] Regarding β, it may be pre-set based on the variance of the observed data. Furthermore, the fit of the data to the model may be evaluated by sensitivity analysis, and the hyperparameters may be manually adjusted. Alternatively, the parameter calculation unit 13 may estimate the parameters using evidence approximation. Evidence approximation is a method that automatically estimates the optimal hyperparameters for the data by maximizing the evidence (marginal likelihood) shown in equation (12) below. p(t|α,β)=∫p(t|w,β)p(w|α)dw (12)

[0057] The parameter calculation unit 13 uses the Bayesian linear regression theory described above to calculate (estimate) the distribution of the Torque model parameters w0 and w1 (the posterior distribution described above) at the reference point, using the reference value input from the reference value acquisition unit 11 as the observed data described above. The parameter calculation unit 13 uses the Bayesian linear regression theory described above to calculate (estimate) the distribution of the Torque model parameters w0 and w1 (the posterior distribution described above) at the detection target, using the target value input from the target value acquisition unit 12 as the observed data described above.

[0058] The distribution of parameters in the criteria calculated by the parameter calculation unit 13 and the distribution of those parameters in the detection target may relate to a hierarchical Bayesian model in which at least one of the multiple types of data is a function of one of the multiple types of data. The distribution of parameters may be a function of drilling depth or drilling time, which are included in the multiple types of data.

[0059] The parameters w0 and w1 of the Torque model can be assumed to be generated by a combination of a global trend across the entire well and a local trend over a short period. This assumption is based on the physical grounds that there is an average trend common to the entire well that is unaffected by variations in formation characteristics and drilling parameters, and local fluctuations that reflect relatively short-term variations in drilling conditions (e.g., changes in formation, equipment wear, and operational adjustments).

[0060] The parameter distribution calculated by the parameter calculation unit 13 is made to relate to a hierarchical Bayesian model, that is, hierarchical Bayesian modeling is performed, taking this assumption into consideration. Hierarchical Bayesian modeling is a method of Bayesian estimation that assumes a hierarchical structure in parameters and takes into account the dependencies between each level. Therefore, it is assumed that the parameters estimated from each observation data are constrained by higher-level parameters (e.g., mean and variance) estimated from the population data. For example, in the case of a two-level model, the observed data y ij The lower parameter θ j Therefore, θ j The distribution follows a distribution based on the higher-level parameter φ. In this way, the parameters of each hierarchical level are simultaneously estimated using Bayesian estimation. Finally, the posterior distributions of the parameters θ and φ are calculated using equation (13) below. p(θ,φ|y)∝p(y|θ,φ)p(θ|φ)p(φ) (13)

[0061] In calculating (estimating) the parameter distribution in this embodiment, for example, a model is created assuming that there is a global parameter representing the overall well trend at a higher level, which influences lower-level local parameters on a local time scale (local trend). Specifically, the well is divided into layers based on a certain drilling depth or drilling time, and although the parameters w0 and w1 differ in each layer, it is assumed that they are generated under the influence of the global trend of the entire well and the local trend of each layer.

[0062] The rotational torque value of the drill pipe, Torque, is assumed to follow a normal distribution with mean μ and variance σ, as shown in equation (14) below. The standard deviation of the error for each layer is σ. layer For this, the prior distribution is given as a semi-normal distribution, as shown in equation (15) below. Torque~N(μ, σ layer ) (14) σ layer ~HalfNormal(10) (15)

[0063] μ is calculated using the parameters for each level of w0 and w1, as shown in equation (16) below. μ=w 0,layer + w 1,layer ×WOB (16)

[0064] Parameter w for each layer 0,layer ,w 1,layer Assuming that it is generated from a normal distribution following μ0, σ0, μ1, σ1 for common parameters in the well, it can be expressed as follows: w 0,layer ~N(μ0,σ0) (17) w 1,layer ~N(μ1,σ1) (18)

[0065] However, w 0,layer Regarding this, there is a global trend in which the value increases with increasing depth. Factors contributing to this include increased geological pressure, increased slurry pressure, and the accumulation of friction in the drill pipe. In the hierarchical Bayesian model, since it includes the trend for the entire well, depth and w 0,layer The relationship is explicitly given by the following equation (19). μ0 = a + b × Bit depth (19) Bit depth is the drilling depth (for example, a value in [m]) included in the reference value and target value. Thus, parameter w 0,layer The distribution is a function of the drilling depth.

[0066] For parameters a, b, and μ1 representing the overall wellhead trend based on all data, and for σ0 and σ1 representing their variances, a normal distribution and a semi-normal distribution are defined as prior distributions, respectively. The Markov chain Monte Carlo (MCMC) method is used for Bayesian estimation of these parameters. Note that, from a computational cost perspective, instead of using all of the baseline or target values, randomly sampled data from each layer (for example, 20 data points per layer) may be used for the above estimation.

[0067] In the example above, the parameter distribution is a function of drilling depth, and in this case, the reference value and target value include the drilling depth. Alternatively, the parameter distribution may be a function of drilling time, either in addition to or instead of drilling depth. In this case, the reference value and target value include the drilling time.

[0068] Furthermore, the calculation of the parameter distribution using the hierarchical Bayesian model may be performed only for the parameter distribution of the model in the reference state, and not for the parameter distribution of the model in the detection target. In this case, the calculation of the parameter distribution of the model in the detection target can be performed by the Bayesian estimation described above, without using the hierarchical Bayesian model. Alternatively, the calculation of the parameter distribution using the hierarchical Bayesian model may be performed for both the parameter distribution of the model in the reference state and the parameter distribution of the model in the detection target. Alternatively, the calculation of the parameter distribution using the hierarchical Bayesian model may be performed only for the parameter distribution of the model in the detection target, and not for the parameter distribution of the model in the reference state. In this case, the calculation of the parameter distribution of the model in the reference state can be performed by the Bayesian estimation described above, without using the hierarchical Bayesian model.

[0069] The parameter calculation unit 13 does not necessarily have to calculate the parameter distribution in the manner described above; it is sufficient if it is based on the reference value and the target value. The parameter calculation unit 13 outputs information to the anomaly detection unit 14 showing the distribution of the parameter in the calculated reference and the distribution of the parameter in the detection target.

[0070] The anomaly detection unit 14 is an anomaly detection means that detects anomalies in the excavation target by comparing the distribution of parameters in the reference, calculated by the parameter calculation unit 13, with the distribution of those parameters in the target, and based on the comparison result. The anomaly detection unit 14 may also detect anomalies in the excavation target by calculating the distribution of estimated values ​​of a preset type of data from the target values ​​acquired by the target value acquisition unit 12, based on the distribution of parameters and model in the reference, calculated by the parameter calculation unit 13, and comparing the calculated distribution with the target values ​​of that type of data, based on the comparison result. The anomaly detection unit 14 may also detect anomalies in the excavation target by calculating the distribution of error terms from the target values ​​acquired by the target value acquisition unit, based on the distribution of parameters and model in the reference, calculated by the parameter calculation unit 13, and based on the calculated distribution.

[0071] The anomaly detection unit 14 uses the parameter distributions for the standard and the target of detection, calculated by the parameter calculation unit 13, to detect anomalies in the drilling of the target. The Torque parameters w0 and w1 described above are parameters that summarize the characteristics of the constituent environment, such as the dynamic friction coefficients of the geological formation and fluid, after removing load fluctuations and data noise. Therefore, by sequentially tracking these parameters, friction fluctuations associated with changes in the well environment during drilling can be captured in detail.

[0072] If the reference excavation and the target excavation are carried out under similar conditions, that is, if the target excavation is carried out under the same normal conditions as the reference excavation, then it is assumed that the distribution of parameters in the reference and the distribution of parameters in the target excavation will be similar. Based on this assumption, the anomaly detection unit 14 detects anomalies in the target excavation based on the difference between the distribution of parameters in the reference and the distribution of parameters in the target excavation. For example, if this difference is greater than a certain level, the anomaly detection unit 14 determines that there is an anomaly in the target excavation, and if this difference is less than a certain level, it determines that there is no anomaly in the target excavation.

[0073] The above anomaly detection is based on the assumption of weak stationarity for the stochastic process under normal conditions. This assumption is that, under normal conditions, the drilling environment does not change much in the immediate vicinity of time, and therefore the statistical properties of parameters w0 and w1 (e.g., mean, variance, and covariance) are relatively stable and constant in the past and present within the immediate vicinity of time. Expressed mathematically, equation (21) holds for equation (20) below. However, in this case, the immediate past state is assumed to be normal. p(w|t,X,α,β)=N(w|m,S) (20) m(t1)=m(t2) and S(t1)=S(t2) ∀t1,t2(21)

[0074] At the same time, it is assumed that before an anomaly (e.g., a blockage) occurs, the well environment changes, the assumption of weak stationarity breaks down, and this manifests as a deviation in the posterior distribution of parameters. This behavior is then recognized as an anomaly. This method can also be described as change point detection by comparing the trends of parameters over time.

[0075] For example, the anomaly detection unit 14 detects anomalies in excavation as follows. The anomaly detection unit 14 receives information from the parameter calculation unit 13 indicating the distribution of parameters for both the standard and the detection target. The anomaly detection unit 14 calculates an anomaly score used for detecting anomalies from these distributions. The anomaly detection unit 14 calculates a value that quantifies the degree of deviation of these distributions as the anomaly score. For example, the anomaly detection unit 14 uses equation (22) below to calculate the Mahalanobis distance D of the distributions (95% confidence interval) of two parameters having the shape of a two-dimensional normal distribution as the anomaly score. M Calculate (N(μ1,Σ1), N(μ2,Σ2)). This allows us to determine the distance considering the correlation (variance and covariance) of the parameters. D M (N(μ1,Σ1),N(μ2,Σ2))=√{(μ1-μ2) T ((Σ1-Σ2) / 2) -1 (μ1-μ2)} (22) In the above equation, N(μ1,Σ1) and N(μ2,Σ2) are the distributions of the two parameters. μ1 and μ2 are vectors of the number of parameters, with the mean values ​​of each parameter that make up the distribution of each parameter as elements. Σ1 and Σ2 are the variance-covariance matrices of each parameter that make up the distribution of each parameter.

[0076] Figures 2 and 3 show examples of parameter distributions (95% confidence intervals) in the space of parameters w0 and w1. Figure 2 shows the distribution of two parameters (Posterior1,2). Figure 3 shows the distribution of three parameters (Posterior1-3).

[0077] This quantification allows for the calculation of a highly accurate anomaly score. A higher anomaly score indicates a greater degree of anomaly. Calculating the anomaly score involves comparing the distribution of the parameter in the baseline with the distribution of that parameter in the detected object. In other words, the calculated anomaly score reflects the results of this comparison.

[0078] The anomaly detection unit 14 compares the calculated anomaly score with a pre-set and stored threshold. If the anomaly score is equal to or greater than the threshold, the anomaly detection unit 14 determines that there is an anomaly in the excavation being detected. If the anomaly score is less than the threshold, the anomaly detection unit 14 determines that there is no anomaly in the excavation being detected.

[0079] Furthermore, the anomaly detection unit 14 may use the calculated anomaly score itself as the anomaly detection result, indicating the degree of risk of an anomaly. In this case, the above-mentioned judgment by comparison with a threshold is not necessarily required. Also, the anomaly detection unit 14 may perform anomaly detection by methods other than those described above, as long as it is based on a comparison between the distribution of parameters in the standard and the distribution of parameters in the detection target.

[0080] The anomaly detection unit 14 may detect anomalies based on factors other than a comparison of the distributions of the two parameters. For example, anomaly detection may be performed considering at least one of the following two factors.

[0081] The first is the detection of abnormal values ​​(outliers) in the rotational torque of the drill pipe. Generally, abnormal values ​​in the rotational torque of the drill pipe tend to be observed before detention. Therefore, when the measured value deviates above the predicted distribution (95% confidence interval) of rotational torque obtained from the posterior distribution of parameters in the normal range, this difference is calculated and used as an anomaly score. This is a probabilistic torque outlier detection. Specifically, the anomaly detection unit 14 performs the following processing.

[0082] In this case, the target value acquisition unit 12 outputs the acquired target value to the anomaly detection unit 14. The anomaly detection unit 14 receives the target value from the target value acquisition unit 12. In this case, the anomaly detection unit 14 also stores the model (for example, equation (1) above) in advance. The anomaly detection unit 14 uses the value of the ground load of the drill bit, which is an explanatory variable of the model among the target values, the distribution of parameters in the standard calculated by the parameter calculation unit 13, and the model to calculate the distribution of the value of the rotational torque of the drill pipe, which is the target variable of the model. For example, the anomaly detection unit 14 calculates the range of the 95% confidence interval as the distribution of the value of the rotational torque of the drill pipe. The anomaly detection unit 14 compares the distribution of the value of the rotational torque of the drill pipe with the value of the rotational torque of the drill pipe among the target values ​​to calculate an anomaly index value. For example, if the value of the rotational torque of the drill pipe among the target values ​​is above the range of the 95% confidence interval, the anomaly detection unit 14 uses the difference (for example, the average or maximum value of all target values) as the anomaly index value. In this case, a higher anomaly index value indicates a greater degree of abnormality.

[0083] The second is the detection of bias in the rotational torque of the drill pipe. The ε in equation (1) above includes parts that cannot be represented by the model. In particular, torque oscillations and imbalances induced by stick-slip phenomena, etc., can lead to suppression. Such torque bias can be detected from the change in the torque distribution from a normal distribution to a bimodal distribution and a biased distribution with broad tails. To quantify this, the skewness, which is the third moment of the Bayesian residue distribution (distribution of the error term), is used as an anomaly index.

[0084] The Bayesian residual distribution is the probability distribution of the prediction error ε between the observed value y, expressed by equation (23) below, and the predicted value based on the parameter posterior distribution (maximum likelihood value w^). ε = y - Xw^ (23) Skewness is a statistical measure that indicates the asymmetry of a distribution, with mean μ and variance σ. 2 For a random variable X, it is defined by equation (24) below. Skewness = E((X-μ)) 3) / σ 3 (twenty four)

[0085] In this case, the target value acquisition unit 12 outputs the acquired target value to the anomaly detection unit 14. The anomaly detection unit 14 receives the target value from the target value acquisition unit 12. In this case, the anomaly detection unit 14 also stores a model (for example, equation (1) above) in advance. The anomaly detection unit 14 uses the target value, the parameter distribution in the criteria calculated by the parameter calculation unit 13, and the model to calculate the distribution (probability distribution) of the error term ε. From the calculated distribution of ε, the anomaly detection unit 14 calculates the skewness of ε as an anomaly index value. In this case, a larger anomaly index value indicates a greater degree of anomaly.

[0086] When the anomaly detection unit 14 uses an anomaly index value as described above, it may calculate an anomaly score from a value based on a comparison of the distributions of two parameters (for example, the Mahalanobis distance described above) and the anomaly index value. For example, the anomaly detection unit 14 may use the product of these values ​​as the anomaly score. Alternatively, the anomaly detection unit 14 may use the sum of these values ​​as the anomaly score. In this case, each value may be normalized using min-max scaling or the standard deviation of previously accumulated anomaly scores, and the weights of each value may be calculated uniformly. Anomaly detection using the calculation of the anomaly score can be performed in the same manner as described above.

[0087] The anomaly detection unit 14 outputs information indicating the detection result of an anomaly. For example, the anomaly detection unit 14 may output this information in a format that is recognizable to the user of the anomaly detection system 10 (e.g., display). The user can refer to the output from the anomaly detection system 10 and take action if an anomaly has been detected in well drilling. The anomaly detection unit 14 may also output this information in other formats. For example, the anomaly detection unit 14 may transmit this information to another device. The above describes the functions of the anomaly detection system 10.

[0088] Next, the abnormality detection method, which is the process (operation method performed by the abnormality detection system 10) executed by the abnormality detection system 10 according to this embodiment, will be explained using the flowchart in Figure 4. In this process, the reference value acquisition unit 11 acquires reference values, which are values ​​of multiple different types of data related to excavation in a reference excavation that is different from the excavation that is the target of abnormality detection (S01, reference value acquisition step). In addition, the target value acquisition unit 12 acquires target values, which are values ​​of multiple different types of data related to excavation in the excavation that is the target of abnormality detection (S02, target value acquisition step).

[0089] Next, the parameter calculation unit 13 defines the relationships between multiple types of data and calculates the distribution of the parameter in the reference based on the reference values ​​acquired by the reference value acquisition unit 11, and calculates the distribution of the parameter in the detection target based on the target values ​​acquired by the target value acquisition unit 12 (S03, parameter calculation step). Note that the calculation of the distribution of the parameter in the reference and the calculation of the distribution of the parameter in the detection target can be performed independently of each other. Also, the acquisition of reference values ​​(S01) and the acquisition of target values ​​(S02) do not necessarily have to be performed in the above order, and can be performed before the calculation of the parameter distribution (S03).

[0090] Next, the anomaly detection unit 14 compares the distribution of parameters in the reference area, calculated by the parameter calculation unit 13, with the distribution of parameters in the target area, and detects an anomaly in the excavation of the target area based on the comparison result (S04, anomaly detection step). Subsequently, the anomaly detection result is output from the anomaly detection unit 14 (S05). The above is the processing performed by the anomaly detection system 10 according to this embodiment.

[0091] Next, a case study of anomaly detection in the anomaly detection system 10 according to this embodiment will be described. The anomalies to be detected in this case study are detention and signs of detention. In this case study, anomaly detection was performed in multiple patterns. The length of the time period (normal interval) in which the reference value acquired by the reference value acquisition unit 11 is measured was set to three patterns: (1) 1 hour, (2) 24 hours, and (3) the entire past data (more than 24 hours). For example, if the normal interval is set to 1 hour, the reference value becomes local, so the sensitivity of anomaly detection is expected to be high. On the other hand, if the normal interval is set to the entire past data, the reference value becomes global, so the sensitivity of anomaly detection may be low. In addition, the length of the time period (judgment interval) in which the target value acquired by the target value acquisition unit 12 is measured was set to 1 hour. A 2-hour interval was set between the normal interval and the judgment interval. This prevents the anomaly from becoming ambiguous due to the inclusion of anomalies in the normal interval. Furthermore, the normal interval and the judgment interval were treated as a single time window, and abnormalities were detected by sliding this time window every 10 minutes.

[0092] Furthermore, the parameter calculation unit 13 calculates the parameters using Bayesian estimation and does not use the hierarchical Bayesian model described above. The anomaly detection unit 14 detects anomalies using only an anomaly score based on the Mahalanobis distance between the distributions of the two parameters and does not use anomaly index values ​​based on the additional elements described above.

[0093] The patterns corresponding to the lengths of the normal intervals (1) to (3) are called Short-term Trend (Model 1), Middle-term Trend (Model 2), and Long-term Trend (Model 3).

[0094] Furthermore, in the three patterns described above, the three patterns in which the anomaly detection unit 14 detects anomalies using anomaly index values ​​based on the two additional elements mentioned above are called Short-term Trend, outlier, skew (Model 4), Middle-term Trend, outlier, skew (Model 5), and Long-term Trend, outlier, skew (Model 6). In these three patterns, when calculating the anomaly score, the values ​​used in the calculation are normalized using min-max scaling, as described above.

[0095] Furthermore, the parameter calculation unit 13 performed anomaly detection on patterns using the hierarchical Bayesian model described above. In this pattern, the length of the normal interval was set to the entire historical data, and anomaly detection was performed on two patterns with different layer divisions. One pattern uses 1 hour of data per layer, and the other uses 24 hours of data per layer. Otherwise, it is the same as Short-term Trend (Model 1). These two patterns are called Long*Short-term Trend (Model 7) and Long*Middle-term Trend (Model 8).

[0096] The Long*Short-term Trend (Model 7) compares the distribution of parameters in the nearest 1 hour, considering the overall well trend, with the distribution of parameters in the decision interval. The Long*Middle-term Trend (Model 8) compares the distribution of parameters in the nearest 24 hours, considering the overall well trend, with the distribution of parameters in the decision interval.

[0097] If the anomaly detection system 10 detects an anomaly within 24 hours prior to the actual time of detention, it will be considered to have detected the anomaly correctly. This is based on the observation from past detention cases that signs of detention often appear within 24 hours, often in conjunction with a deterioration of the excavation situation. If the anomaly detection system 10 detects an anomaly at any other time, it will be considered a false positive. Furthermore, once an anomaly is detected, it will be assumed that the anomaly will persist for the next 6 hours, during which time no further detection will be performed.

[0098] Next, we will explain the quantitative comparison method for each of the above patterns. In the problem of detecting signs of detention, there is a large imbalance in classes, and detention cases are a minority class. Therefore, when applying a normal evaluation metric for discrimination performance (for example, AUC (Area Under the Curve)), models that identify everything as normal tend to be evaluated relatively highly. However, the importance of detecting signs of detention is high, and it is necessary to reliably detect signs while suppressing false positives. Therefore, we will use a method that simultaneously compares the number of correct detections (the number of times anomalies were correctly detected) and the number of false positives to compare the effectiveness of each pattern in actual operation. Specifically, we will compare the patterns by plotting the number of correct detections and false positives simultaneously while changing the threshold used for detecting anomalies, and visualizing the relationship between them. Patterns that increase the number of detections while suppressing the increase in the number of false positives by setting the threshold are relatively superior.

[0099] Furthermore, simultaneously with pattern selection, it is necessary to set thresholds that not only improve detection accuracy but also enhance actual operational efficiency. Therefore, the following equation (25), which takes into account the reduction effect of NPT, is used. T red =T TP N TP -T FP N FP (twenty five) In the above formula, T red This represents the actual reduction in NPT time (hours) due to the detection of anomalies by the anomaly detection system 10. TP This is an NPT that can be avoided by detecting an anomaly in the anomaly detection system 10. TP Let's assume it's 37.5 hours. TPThis is the number of positive detections. FP This is the extra NPT incurred due to false detection of anomalies by the anomaly detection system 10. FP Let's assume it's 0.5 hours. FP This represents the number of false positives.

[0100] Using this formula, T red The threshold at which this value is maximized is determined to be the optimal threshold. This is a method of setting thresholds based on the on-site need to avoid productivity reduction due to detention, and is considered to be a realistic indicator for the practical application of the anomaly detection system 10.

[0101] The case study uses drilling data from 6 wells, including 10 detention cases. Figures 5 and 6 show the anomaly scores (vertical axis of the graph) for each drilling depth (horizontal axis of the graph) calculated for each pattern. From top to bottom in Figure 6, the graphs for each pattern are shown: Short-term Trend (Model 1), Middle-term Trend (Model 2), Long-term Trend (Model 3), Short-term Trend, outlier, skew (Model 4), Middle-term Trend, outlier, skew (Model 5), Long-term Trend, outlier, skew (Model 6), Long*Short-term Trend (Model 7), and Long*Middle-term Trend (Model 8).

[0102] The hatched areas on the right side of the graphs in Figures 5 and 6 represent the 24 hours immediately preceding internment, a period during which signs of internment are presumed to occur. In many patterns, there is a tendency for abnormal scores to increase in the 24 hours immediately preceding internment, but this is even more pronounced when compared to the few days prior to internment.

[0103] However, when looking at the entire well (operation over several months), peaks in anomaly scores are observed even outside of the 24 hours immediately preceding detention. The dashed line shows the maximum anomaly score in the 24 hours immediately preceding detention. If this value is used as a threshold, points (circles on the graph) where anomaly scores exceed this value outside of the immediate 24 hours before detention are considered false positives. Compared with Models 1-3, Models 4-6 show anomaly score responses only immediately before detention in the days leading up to detention, but a tendency for false positives to increase throughout the entire operation can be observed. On the other hand, Models 7 and 8 show the highest anomaly score immediately before detention throughout the entire operation, achieving both clarification of anomaly warnings and reduction of false positives. As can be seen from the comparison between Model 1 and Model 7, and between Model 2 and Model 8, this is likely due to the ability to keep anomaly scores low during normal operation.

[0104] The parameter distributions shown in Figure 2 are examples of the 95% confidence intervals (Posterior1) for the parameter distributions estimated from the 1-hour data representing the normal interval (posterior distribution) and the 95% confidence intervals (Posterior2) for the parameter distributions estimated from the 1-hour data representing the decision interval, as compared in Model 1. The Mahalanobis distance between Posterior1 and Posterior2 is shown in the upper left of Figure 2.

[0105] The parameter distributions shown in Figure 3 are examples of the parameter distributions (posterior distributions) used in Model 8 for data in the same time period as the parameter distributions shown in Figure 2. The parameter distributions shown in Figure 3 are the global trend estimated from the entire well data (Posterior2), the parameter distribution estimated from the past 24-hour data while considering this (posterior distribution) (Posterior1), and the parameter distribution estimated from the data in the decision interval (posterior distribution) (Posterior3). All three parameter distributions (posterior distributions) in Figure 3 are 95% confidence intervals. The Mahalanobis distance between Posterior1 and Posterior3 is shown in the upper left of Figure 3.

[0106] Comparing these, Model 1 shows that the posterior distribution of Posterior1 has an elongated shape, and the variability of simultaneous increases or decreases in the same direction of parameters w0 and w1 is evaluated as small. Therefore, when w0 and w1 move in the same direction, the Mahalanobis distance is calculated to be large. As a result, the sensitivity of anomaly detection is high, but false positives are more likely to occur.

[0107] In contrast, in Model 8, Posterior1 has a relatively large distribution due to the influence of Posterior2, and the anomaly score (Mahalanobis distance) is also relatively low. This indicates that when considering the overall well trend including diverse well environments, the variability of parameters w0 and w1 is estimated to be large, and as a result, the Mahalanobis distance tends to be calculated to be small.

[0108] Figure 7 shows the relationship between the number of correct detections (horizontal axis of the graph) and the number of false detections (vertical axis of the graph) for each pattern when Models 1 to 8 are applied, using drilling data from 6 wells, including 10 detention cases. The graph in Figure 7 shows the results for multiple thresholds used to detect anomalies. Each point in the graph in Figure 7 corresponds to one threshold.

[0109] As shown in Figure 7, there is a trade-off relationship between the number of correct detections and the number of false detections. Adjusting the threshold to increase the number of correct detections tends to increase the number of false detections as well. The rate of increase in the number of false detections differs for each pattern. In particular, there is almost no difference in the relationship between the number of correct detections and false detections in Models 1 to 3. On the other hand, the number of false detections decreases slightly in Models 4 to 6 compared to Models 1 to 3. In Model 8, the increase in the number of false detections relative to the increase in the number of correct detections is significantly suppressed compared to the other patterns, demonstrating high accuracy.

[0110] Furthermore, using the threshold set based on the above NPT for Model 8, the highest effective reduction time of NPT was obtained at 237.5 hours with 9 positive detections.

[0111] The case study results shown in Figures 5 and 6 confirm that Models 1-3 exhibit relatively good anomaly detection immediately before detention. These patterns were compared by varying the time scale of the normal interval compared to the judgment interval, and all patterns captured signs of detention. This suggests that both long-term global trends across the entire well and short-term local trends in the recent past are important for estimating the normal range of parameters.

[0112] In particular, as seen in Figures 5-7, Model 8, which uses a hierarchical Bayesian model that considers the overall well trend, demonstrated high accuracy in predicting potential problems. This is likely because, by capturing the global trend for the entire well while estimating the local trend in the vicinity, the normal range of parameters could be accurately estimated. Furthermore, as can be seen from the comparison between Figures 2 and 3, the uncertainty of the normal range of parameters tends to be underestimated when only short-term data is used. This suggests that this could be improved by considering long-term trends. However, the higher accuracy of Model 8 compared to Model 3, which only considers long-term trends, highlights the importance of utilizing data with similar drilling conditions (near-time data) in the estimation process.

[0113] Next, considering the difference in predictive detection accuracy between Model 7 and Model 8, it is suggested that 24-hour data is more suitable than 1-hour data for estimating the normal range. This suggests that approximately 24 hours of data is necessary to estimate normal parameter fluctuations and their uncertainties due to variations in geological and drilling parameters.

[0114] Furthermore, Models 4-6, which utilize abnormal values ​​of drill pipe rotational torque and biases in the residual distribution of rotational torque, show better behavior in the days prior to detention compared to Models 1-3. This suggests the possibility of clarifying anomalies by using multiple anomaly indicators. However, false positives increase when considering the entire operation. This is because temporary increases in drill pipe rotational torque push up the anomaly score, but in many cases this does not lead to detention, resulting in false positives. To reduce false positives, it is considered necessary to continuously grasp the overall trend of the data rather than just focusing on abnormal torque values.

[0115] As described above, the advantages and disadvantages of combining multiple anomaly indicators versus using them complementaryly become clear when it comes to detecting signs of detention using multiple anomaly indicators. Specifically, when multiple anomaly indicators are combined, the accuracy of sign detection improves, but a challenge is that certain anomaly indicators may be suppressed by other indicators, potentially causing signs to be missed. On the other hand, when indicators are used complementaryly, the range of signs that can be detected broadens and the detection coverage improves, but this also leads to the problem of increased false positives.

[0116] In this respect, methods that capture parameter distributions based on the overall trend of the data, such as Models 1-3 and Models 7 and 8, are likely to reduce false positives, which are predictive detection methods based on a single anomaly indicator. In particular, approaches using hierarchical Bayesian models further improve the accuracy of predictive detection by considering both global and local trends.

[0117] Detention has a significant impact not only on NPT but also on operating costs. Therefore, when setting the threshold used for anomaly detection, it is advisable to consider not only NPT but also operating costs, as described above. In the case study above, a certain number of false positives were observed, but based on physical knowledge, it is assumed that these false positives occurred at times when parameters fluctuated significantly at operational and geological change points. Therefore, from an engineer's perspective, some false positives may be acceptable, and in that case, it is thought that the accuracy of predictive detection can be further improved, leading to more efficient operations.

[0118] In this embodiment, the parameter distribution in the reference is compared with the parameter distribution in the target of detection, and an anomaly is detected based on the comparison result and the reference. The parameter distribution appropriately reflects the drilling conditions. Therefore, according to this embodiment, anomalies in well drilling can be detected with high accuracy.

[0119] As in this embodiment, the parameter calculation unit 13 may calculate the parameter distributions for both the standard and the detection target using Bayesian estimation. With this configuration, the parameter distribution can be calculated appropriately and reliably, and anomalies in well drilling can be detected appropriately, reliably, and with high accuracy. However, the parameter calculation unit 13 may calculate the parameter distribution using a method other than Bayesian estimation.

[0120] As in this embodiment, at least one of the parameter distributions in the reference and the parameter distribution in the detection target may relate to a hierarchical Bayesian model in which the distribution is a function of one of several types of data. With this configuration, anomalies in well drilling can be detected with even greater accuracy depending on one of the several types of data.

[0121] Furthermore, in this case, the parameter distribution may be a function of drilling depth or drilling time, which are included in multiple types of data. With this configuration, anomalies in well drilling can be detected with even greater accuracy depending on the drilling depth or drilling time. However, the parameter distribution may be a function of data other than drilling depth or drilling time. Also, the parameter distribution does not necessarily have to relate to a hierarchical Bayesian model.

[0122] Furthermore, as in this embodiment, the multiple types of data may be data that can be measured outside the well. With this configuration, the data values ​​can be easily acquired, and anomalies can be detected easily and accurately. However, some or all of the multiple types of data do not have to be data that can be measured outside the well.

[0123] Furthermore, as in this embodiment, the multiple types of data may include data relating to the ground load of the drill bit used for drilling and data relating to the rotational torque of the drill pipe used for drilling, and the model may be a linear regression model. With this configuration, anomalies can be detected reliably and accurately.

[0124] However, the data and models of multiple types do not necessarily have to be as described above. For example, the data of multiple types may include the injection flow rate of the drilling fluid used in drilling, data related to the length of the drill pipe used in drilling, and the injection pressure (head pressure) of the drilling fluid used in drilling. Here, the data related to the length of the drill pipe is as described above. In this case, the model may be the following linear regression model. (Injection pressure of drilling fluid) = w 11 (Injection flow rate of drilling fluid) w12 + w 13 (Drill bit depth) (Injection flow rate of drilling fluid) w14 + w 15 +ε

[0125] Furthermore, as in this embodiment, the anomaly detection unit 14 may also calculate the distribution of estimated values ​​of a preset type of data from the target values ​​acquired by the target value acquisition unit 12, based on the distribution of parameters and models in the reference calculated by the parameter calculation unit 13, compare the calculated distribution with the target values ​​of the data of that type, and detect anomalies in the drilling target based on the comparison result. For example, as described above, an anomaly index value may be calculated based on the comparison result to detect anomalies. With this configuration, anomalies in well drilling can be detected with even greater accuracy according to the distribution of estimated values ​​of the data of the preset type. However, it is not always necessary to perform the above in detecting anomalies.

[0126] Furthermore, as in this embodiment, the model may include an error term, and the anomaly detection unit 14 may calculate the distribution of the error term from the target values ​​acquired by the target value acquisition means, based on the distribution of parameters in the reference and the model calculated by the parameter calculation unit 13, and detect anomalies in the drilling target based on the calculated distribution. For example, as described above, an anomaly index value may be calculated based on the distribution of the error term ε, and anomalies may be detected. With this configuration, anomalies in well drilling can be detected with even greater accuracy depending on the distribution of the error term. However, it is not always necessary to perform the above in detecting anomalies.

[0127] The anomaly detection system and anomaly detection method of this disclosure have the following configuration. [1] An anomaly detection system for detecting anomalies in well drilling, A means for acquiring reference values, which are values ​​of multiple different types of data related to drilling in a reference drilling that is different from the drilling that is the target of anomaly detection, A target value acquisition means for acquiring target values ​​which are the values ​​of multiple different types of data related to excavation in which anomalies are to be detected, A parameter calculation means that defines the relationships between the multiple types of data and includes a preset model that includes parameters, calculates the distribution of the parameter in the standard based on the standard value acquired by the standard value acquisition means, and calculates the distribution of the parameter in the detection target based on the target value acquired by the target value acquisition means, An anomaly detection means that compares the distribution of the parameter in the standard calculated by the parameter calculation means with the distribution of the parameter in the target to be detected, and detects an anomaly in the excavation of the target to be detected based on the comparison result, An anomaly detection system equipped with the following features. [2] The anomaly detection system according to [1], wherein the parameter calculation means calculates the distribution of the parameters in the criterion and the target to be detected by Bayesian estimation. [3] An anomaly detection system according to [1] or [2], relating to a hierarchical Bayesian model in which at least one of the distribution of the parameters in the criteria and the distribution of the parameters in the target to be detected is a function of one of the plurality of types of data. [4] The anomaly detection system according to [3], wherein at least one of the distribution of the parameter in the criteria and the distribution of the parameter in the target of detection is a function of the drilling depth or drilling time included in the plurality of types of data. [5] An anomaly detection system as described in any of [1] to [4], wherein the aforementioned multiple types of data are data that can be measured outside the well. [6] The aforementioned multiple types of data include data relating to the ground load of the drill bit used for drilling, and data relating to the rotational torque of the drill pipe used for drilling, The aforementioned model is an anomaly detection system described in any of the linear regression models [1] to [5]. [7] An anomaly detection system according to any one of [1] to [6], wherein the anomaly detection means calculates the distribution of estimated values ​​of a preset type of data from the target values ​​acquired by the target value acquisition means based on the distribution of the parameters in the standard and the model calculated by the parameter calculation means, compares the calculated distribution with the target values ​​of the data of that type, and detects an anomaly in the excavation to be detected based on the comparison result. [8] The above model includes an error term, An anomaly detection system according to any one of [1] to [7], wherein the anomaly detection means calculates the distribution of the error term from the target value acquired by the target value acquisition means based on the distribution of the parameters in the standard and the model calculated by the parameter calculation means, and detects anomalies in the excavation of the target to be detected based on the calculated distribution. [9] An anomaly detection method, which is a method for operating an anomaly detection system for detecting anomalies in well drilling, A reference value acquisition step involves acquiring reference values, which are values ​​of multiple different types of data related to drilling in a reference drilling that is different from the drilling that is the target of anomaly detection, and which are reference values ​​in a reference drilling. A target value acquisition step involves acquiring target values, which are data values ​​of multiple different types related to the excavation that are the target of anomaly detection in the excavation, and A parameter calculation step that defines the relationships between the multiple types of data and includes a pre-defined model including parameters, calculates the distribution of the parameter in the standard based on the standard value acquired in the standard value acquisition step, and calculates the distribution of the parameter in the detection target based on the target value acquired in the target value acquisition step, An anomaly detection step involves comparing the distribution of the parameter in the standard calculated in the parameter calculation step with the distribution of the parameter in the target to be detected, and detecting an anomaly in the excavation of the target to be detected based on the comparison result. An anomaly detection method including [Explanation of Symbols]

[0128] 10...Anomaly detection system, 11...Reference value acquisition unit, 12...Target value acquisition unit, 13...Parameter calculation unit, 14...Anomaly detection unit.

Claims

1. An anomaly detection system for detecting abnormalities in well drilling, A means for acquiring reference values, which are values ​​of multiple different types of data related to drilling in a reference drilling that is different from the drilling that is the target of anomaly detection, A target value acquisition means for acquiring target values ​​which are the values ​​of multiple different types of data related to excavation in which anomalies are to be detected, A parameter calculation means that defines the relationships between the multiple types of data and includes a preset model that includes parameters, calculates the distribution of the parameter in the standard based on the standard value acquired by the standard value acquisition means, and calculates the distribution of the parameter in the detection target based on the target value acquired by the target value acquisition means, An anomaly detection means that compares the distribution of the parameter in the standard calculated by the parameter calculation means with the distribution of the parameter in the target to be detected, and detects an anomaly in the excavation of the target to be detected based on the comparison result, An anomaly detection system equipped with the following features.

2. The anomaly detection system according to claim 1, wherein the parameter calculation means calculates the distribution of the parameters in the criterion and the target to be detected, respectively, by Bayesian estimation.

3. An anomaly detection system according to claim 1 or 2, wherein at least one of the distribution of the parameters in the criteria and the distribution of the parameters in the target to be detected is a function of one of the plurality of types of data, relating to a hierarchical Bayesian model.

4. The anomaly detection system according to claim 3, wherein at least one of the distribution of the parameters in the criteria and the distribution of the parameters in the target to be detected is a function of the drilling depth or drilling time included in the plurality of types of data.

5. The anomaly detection system according to claim 1 or 2, wherein the aforementioned multiple types of data are data that can be measured outside the well.

6. The aforementioned multiple types of data include data relating to the ground load of the drill bit used for drilling, and data relating to the rotational torque of the drill pipe used for drilling. The anomaly detection system according to claim 1 or 2, wherein the model is a linear regression model.

7. An anomaly detection system according to claim 1 or 2, wherein the anomaly detection means calculates the distribution of estimated values ​​of a preset type of data from the target values ​​acquired by the target value acquisition means based on the distribution of the parameters in the standard and the model calculated by the parameter calculation means, compares the calculated distribution with the target values ​​of the data of that type, and detects an anomaly in the excavation to be detected based on the comparison result.

8. The aforementioned model includes an error term, The anomaly detection system according to claim 1 or 2, wherein the anomaly detection means calculates the distribution of the error term from the target value acquired by the target value acquisition means based on the distribution of the parameters in the standard and the model calculated by the parameter calculation means, and detects anomalies in the excavation target to be detected based on the calculated distribution.

9. An anomaly detection method, which is a method for operating an anomaly detection system for detecting anomalies in well drilling, A reference value acquisition step involves acquiring reference values, which are values ​​of multiple different types of data related to drilling in a reference drilling that is different from the drilling that is the target of anomaly detection, and which are reference values ​​in a reference drilling. A target value acquisition step involves acquiring target values, which are data values ​​of multiple different types related to the excavation that are the target of anomaly detection in the excavation, and A parameter calculation step that defines the relationships between the multiple types of data and includes a pre-defined model including parameters, calculates the distribution of the parameter in the standard based on the standard value acquired in the standard value acquisition step, and calculates the distribution of the parameter in the detection target based on the target value acquired in the target value acquisition step, An anomaly detection step involves comparing the distribution of the parameter in the standard calculated in the parameter calculation step with the distribution of the parameter in the target to be detected, and detecting an anomaly in the excavation of the target to be detected based on the comparison result. An anomaly detection method including

Images