CN118690271A - An intelligent identification method for early warning signals of pipe sticking accidents under long-tail distribution - Google Patents

Abstract

The present invention discloses an intelligent identification method for early sign signals of drill stuck accidents under long-tail distribution, which belongs to the field of artificial intelligence. The method comprises the following steps: step S10, collecting logging time series data, performing data cleaning and labeling, and constructing a drill stuck accident data set; step S20, establishing a channel-time series hybrid drill stuck prediction network according to the drill stuck accident data set; step S30, training the network using the loss function, fitting the data, and obtaining a trained channel-time series hybrid drill stuck prediction network model; step S40, predicting drill stuck accidents using the channel-time series hybrid drill stuck prediction network model. The beneficial effect of this method is that it can automatically learn the potential patterns of logging time series, perceive the changes in the precursor parameters of drill stuck, and promptly give early warnings for drill stuck, thereby reducing the economic losses and safety hazards caused by drill stuck accidents.

Description

An intelligent identification method for early warning signals of pipe sticking accidents under long-tail distribution

Technical Field

The present invention relates to a method for predicting a drill stuck accident. In particular, it relates to an intelligent method for identifying early sign signals of a drill stuck accident under a long-tail distribution. Specifically, it refers to using the data processing method of the present invention to process logging data, using the corresponding deep learning algorithm model to learn the correlation between features and early sign signals of a drill stuck, and using the long-tail problem loss function to improve the recognition accuracy of minority class samples, thereby realizing the recognition of early sign signals of a drill stuck, thereby achieving the purpose of predicting a drill stuck accident, and belongs to the field of artificial intelligence.

Background Art

Drilling is a core part of the oil industry. A stuck drill string refers to a drilling accident in which the drill string loses its free movement during the process of drilling, that is, the drill string cannot move up and down or rotate, and encounters an obstruction in a certain section of the wellbore. Nearly one-third of drilling time loss is caused by stuck drill string. A stuck drill string accident not only leads to extended drilling time, loss of drilling tools and scrapping of wellbore, but also seriously threatens the life safety of operators. The causes of stuck drill string are complex, including rock formation characteristics, wellbore stability, etc. At the same time, there are complex correlations between these factors. The complexity and diversity of stuck drill string make it a very challenging problem. Therefore, conducting research on the identification and prediction of stuck drill string accidents is of great significance to guiding drilling operations. The judgment of stuck drill string at the engineering site is basically based on manual experience, and there is no complete stuck drill string prediction software at the drilling site.

Drilling operations are affected by complex geographical environmental factors, and the drilling process varies. In complex drilling situations, the types of drill bit stuck accidents are also different. Typical drill bit stuck accidents include differential pressure drill bit stuck, sand settling drill bit stuck, keyway drill bit stuck, etc. The most common type of drill bit stuck accident is differential pressure drill bit stuck. There are three basic conditions for differential pressure drill bit stuck: 1. The static column pressure of the drilling fluid is greater than the formation pressure. 2. There is a permeable formation and there is a thick mud cake on the well wall. 3. The drill has been stationary for a period of time. We believe that the close contact between the drill tool and the mud cake is a process, not a one-time occurrence. This process is also the lead time of differential pressure drill bit stuck. The difference between the hydrostatic pressure and the pore pressure gradually increases, the mud cake gradually becomes thicker, and the pressure in the mud cake gradually decreases, which finally leads to drill bit stuck. Other types of drill bit stuck accidents also have their signs before they occur. If these early signs of drill bit stuck accidents can be detected, timely response measures can be taken on site to reduce losses. However, drill stuck accidents are always low-probability events, and the early warning signal samples of drill stuck accidents are also small samples. The number of samples that can be learned is scarce, which will lead to poor performance of the model in recognition. The current method for small sample recognition is mainly resampling, which is an intuitive way to balance the difference in the number of samples between categories. This method has a certain improvement effect but has the following problems: 1. Oversampling and repeated sampling of the precursor signal class will cause the model to overfit; 2. Undersampling discards most of the normal working condition samples, which will cause information loss and low model accuracy. Therefore, it is necessary to improve the recognition accuracy of the precursor signal samples of drill stuck without destroying the category distribution, so as to facilitate timely handling of accidents on site and avoid drill stuck.

Summary of the invention

In response to the above problems, the present invention integrates multi-source data of on-site monitoring and proposes a drill stuck prediction method based on drill stuck precursor signal detection under long-tail distribution. The method can effectively sense the changes of various parameters, automatically discover complex feature patterns, and realize the detection of drill stuck precursor signals with higher accuracy and generalization ability.

To achieve the above object, the technical solution adopted by the present invention is as follows:

S1. Obtain logging time series data from different work areas; normalize the logging data and fill in missing values; use expert knowledge to mark the stuck drill accident and obtain a number of marked logging time series data; divide the logging time series data into a number of samples, including early sign signal samples of stuck drill accident and normal samples;

S2. Construct a channel-time series hybrid pipe-stuck prediction network; use cross-validation to train and test the model. When testing the data of a specific oil well, the data of the oil well will be removed from the data set during the training phase. Obtain the probability that the logging time series data sample is a precursor signal of a pipe-stuck accident;

S3. Use the long-tail distribution loss function to further optimize the network and solve the problem of mining stuck pipe samples when the positive and negative samples of logging data are unbalanced. Introduce the modulation coefficient to make the model pay more attention to the rare positive samples of early signs of stuck pipe accidents;

Furthermore, the specific process of implementing S10 is to obtain the historical and real-time updated logging data of different work areas and store them as time series data. The data parameters include drill bit position, torque, vertical pressure, etc. Check the data continuity and use missing fill; mark the stuck drill label based on the drilling log.

S11, the missing filling is a K nearest neighbor filling method;

S12. Based on the drilling log, the logging parameter time series obtained 15 minutes before the occurrence of the pipe sticking accident is divided into a sample at intervals of three minutes as an early sign signal sample of the pipe sticking.

S13. Randomly extract the logging parameter time series under normal working conditions from the drilling database, and divide it into a sample at intervals of three minutes as the normal drilling signal sample.

S14. Optimize engineering parameters related to the occurrence of stuck drill pipe accidents based on expert knowledge. The extracted parameters are time, well depth, drill bit position, drilling time, hook height, drilling pressure, rotation speed, casing pressure, riser pressure, inlet flow rate, and outlet flow rate.

Preferably, the specific steps to implement S2 are:

S21. Determine a sampling interval parameter s. Evenly downsample the original logging sequence data and decompose it into s subsequences.

S22. Perform temporal feature learning on each subsequence, using a two-layer fully connected network with GELU as the activation function, to obtain temporal dimension feature representation.

S23. Reorganize and merge the s subsequences according to the original order to achieve the fusion of time features.

S24. Based on the simplified linear model, learn the correlation between parameters and obtain channel dimension features.

S25, residual connection layer, deeply fuses the temporal features in step S23 and the channel features in step S24.

S26, linear mapping layer, maps the fused features to categories (normal/precursor signals).

S27, Softmax layer converts the output into the probability of the drill stuck accident sign signal.

The specific steps to implement S3 are:

S31. The model predicts the probability of each sample being a stuck diamond class. The cross entropy between the actual label of each sample and the predicted result is calculated as the basic loss.

S32, introduce the modulation coefficient, construct the Focal Loss, and modulate the basic loss.

S33, Focal Loss reduces the loss contribution of the model to the easy samples with accurate classification by setting the modulation factor.

S34. Use Focal Loss as the final loss and implement gradient descent optimization update of model parameters.

S4. Input the actual logging time series data, extract the specified parameter column, and detect whether there is missing data; if there is missing data, fill the missing values and build a prediction data set input model, and output the probability of whether it is a precursor signal of stuck drill.

Compared with the prior art, the present invention has the following beneficial effects:

(1) The present invention uses the drill stuck precursor signal detection to predict the equivalent drill stuck accident, thus avoiding the uncertainty caused by parameter prediction;

(2) The present invention combines logging time series data with a deep learning model to accurately identify precursor signals of drill sticking and issue an alarm in a timely manner;

(3) The present invention can effectively improve the recognition accuracy of precursor signals of drill stuck accidents in small samples without causing overfitting or underfitting.

The present invention is a channel-time series hybrid pipe-stuck accident precursor signal detection method, which can effectively perceive the potential change rules and implicit correlations of various parameters in logging data and extract the characteristics of pipe-stuck accident data. In the case of unbalanced distribution of pipe-stuck sample data, a long-tail distribution loss function is used to solve the problem of unbalanced data sample categories.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present invention, but are not intended to limit the present invention.

Fig. 1 is a schematic diagram of the process of the present invention;

FIG2 is a schematic diagram of a factorized time channel hybrid network of the present invention;

DETAILED DESCRIPTION

The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

The channel-sequence hybrid pipe sticking prediction method of the present invention comprises the following steps:

S1. Collect logging data from different work areas. Some parameter statistics are shown in Table 1.

Table 1 Statistics of some parameters

After inputting the data, check whether there are missing values in the data. The present invention adopts the "k nearest neighbor method" to fill the missing values. First, determine the number of nearest neighbors k used to fill the missing values. Then, calculate the distance between the sample where the missing value is located and other samples, using Euclidean distance or other distance measurement.

Among them, dxy is the Euclidean distance between the data with missing values and the complete data, x is the data with missing values, and y is the complete data. After calculating the Euclidean distance between the data with missing values and other complete data, select the k samples closest to the missing value sample. Use the average or weighted average of these k nearest neighbors to fill the missing value. Finally, apply the filled value to the location where the missing value is located.

According to the on-site construction log, the data 15 minutes before the drill stuck accident occurred was marked as the precursor signal of the drill stuck, and the samples were divided into intervals of three minutes. Similarly, the normal data samples were extracted with a sample time span of three minutes.

When predicting stuck drill, the overall trend of each parameter is more important than the specific value, and the values of each parameter vary greatly. Therefore, in order to improve the convergence speed of the model, each parameter is normalized to a value of 0 to 1 to eliminate the influence of different parameter dimensions. Suppose the original value of a parameter is x, and the normalized value is x', the conversion formula is:

The method is described in detail below with reference to the accompanying drawings. FIG1 is a general framework diagram of a network.

S1. Construct a factorized time channel hybrid network to model the variation rules and parameter correlation of each logging parameter. Extract features from the time and space dimensions respectively, and obtain the mapping relationship between the features and the precursor signal of the stuck drill.

With reference to FIG2 , step S2 is described in detail below.

The structure of the factorized time-channel mixed prediction network is shown in Figure 2. It consists of two parts: temporal interaction and channel mixing.

First, a temporal interaction module is designed to fully capture the subtle changes in different parameters.

The original logging time series data The downsampling is divided into s interleaved subsequences, each subsequence learns the time dimension features separately through the linear layer, and finally the features of these subsequences are reassembled in the original order to avoid feature loss caused by decomposition. Suppose the input sequence is Then the interleaved sampling is implemented as follows:

1. Uniform downsampling: Downsample _Xh by interval s and divide it into s subsequences:

X _h,i =X _h [i-1::s,:]，1≤i≤s (2)

Here, i indexes the i-th subsequence, and [::s] means sampling every s steps. The value of s is usually selected based on experience.

Feature learning: Apply a temporal feature extractor to each subsequence X _h,i . This paper uses linear interaction to capture the temporal pattern of stuck drill data, and uses a multilayer perceptron with two hidden layers as a feature extraction module. The design of two hidden layers avoids the risk of overfitting. The high-level features extracted by the multilayer perceptron provide more discriminative input for subsequent time series prediction tasks, improving the prediction accuracy of the model. The GELU activation function is used to avoid the situation where the output of the activation unit is zero for a long time when the input is negative. Linear model learning parameters The time dimension feature representation _Xh,i ′ is obtained by learning:

in is the addition in the column direction.

3. Feature aggregation: As shown in formula (3), the features of s subsequences _Xh,i ′ are re-aggregated in the original order to obtain a new time series representation _Xh ^*

X _h ^* =[X _h,1 ′[0],X _h,2 ′[0],...,X _h,1 ′[1],X _h,2 ′[1],...] ( 4)

After the above processing, the original redundant sequence is downsampled into multiple subsequences, and features are learned separately on each subsequence. Finally, the features of each subsequence are reaggregated in the original order to reduce redundancy and encode time dimension information. By adjusting the downsampling step size s, the degree of downsampling can be controlled to balance the amount of information and redundancy.

Secondly, a channel mixing module is designed to capture the correlation between channels.

Furthermore, we hope to obtain the relationship between parameters such as torque and drilling speed. Therefore, a channel mixing capture module is designed.

The relationship between channels can be expressed as follows:

in

represents noise, Indicates the channel dependency (association) relationship after denoising. and represents the channel interaction after decomposition. Formula (5) is similar to the linear expression, so the channel dependency of time series data can be learned using a linear model as follows

in σ(·) is the activation function.

This can effectively extract the correlation between channels in a very simple way, effectively reducing the time complexity and space complexity of the model.

The specific method is to transpose the original data, pass the linear projection layer, and the GELU activation function, and finally get the correlation between the parameters. The transposition makes the entire sequence of each parameter as a feature channel. The linear projection layer maps the features to a lower dimensional space in order to compress the features and reduce redundant components. The GELU activation function plus the linear projection layer realizes the dimensional transformation to ensure that the dimension of the final output is consistent with the input.

Use residual connections to further deeply fuse the time mixing features and channel mixing features. Use linear mapping to map the original input length to the target prediction length to avoid the error accumulation caused by iterative prediction. So far, the encoding representation of the parameter space has been obtained.

Finally, a linear layer mapping and a Softmax layer are used to obtain the probability of whether the drill is stuck, and the larger index of the two is output as the output, completing the detection of the drill stuck precursor signal sample.

S3. Considering the problem that the proportion of stuck drill event samples to all samples is extremely low, the Focal Loss function is used to significantly improve the network's ability to identify stuck drill samples. Finally, the prediction and early warning of stuck drill accidents are achieved.

The Focal Loss loss function is as follows:

FL(p _t )=α(1-p _t ) ^γ log(p _t ) (7)

Where _pt is the model's predicted probability for positive samples, and 1- _pt represents the probability of the model's prediction error. γ is an adjustment factor. When γ increases, the loss function can pay more attention to the difficult-to-classify stuck drill samples.

In order to optimize model performance and prevent overfitting, this paper adopts the early stopping strategy to train the model. Early stopping is an iterative termination method based on the performance of the validation set, which aims to improve the generalization ability of the model. At the end of each training cycle (epoch), the performance of the model on an independent validation dataset is evaluated. If the performance on the validation set (such as the loss function value or accuracy) does not improve significantly in several consecutive training cycles, the training process will be terminated early.

S4: Input the test data sample, check whether the data is missing, fill the data if missing, and then input the complete test data into the model, and output the identification result of the drill stuck precursor signal, which is the probability that the sample belongs to the drill stuck precursor signal.

The above description is not intended to impose any form of limitation on the present invention. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims (4)

1. A method for intelligently identifying early warning signs of a pipe-stuck accident under a long-tail distribution, the method being implemented by a computer and comprising: S1: Obtaining an initial sample, where the initial sample is logging data; The logging data is input into a channel-time series hybrid pipe-stuck precursor signal detection network model. The network model includes a factorized time channel hybrid prediction network. The network model outputs the probability that the sample is an early sign signal of a pipe-stuck accident. S2: Training channel-time series hybrid stuck drill precursor signal detection method. This method predicts stuck drill accidents by identifying early sign signals of stuck drill accidents. This method fully captures various nonlinear characteristics and their correlations of stuck drill accidents from the two dimensions of time and channel of logging data, and designs a corresponding loss function training model based on the scarcity of stuck drill accidents to realize the identification of early sign signals of stuck drill accidents, so as to achieve early warning prediction of stuck drill accidents. S3: Use the long-tail problem loss function to train the network, fit the data, and obtain a trained channel-time series hybrid stuck drill prediction network model. S4: Predicting pipe stuck accidents using a channel-time hybrid pipe stuck precursor signal detection network. 2. According to the method for intelligently identifying early sign signals of a stuck drill accident under a long-tail distribution in claim 1, it is characterized in that: the step S1 is to obtain time series data of logging in different work areas; normalize the logging data and fill in missing values; use expert knowledge to mark the stuck drill accident and obtain a number of marked logging time series data; divide the logging time series data into a number of samples, which are divided into samples of early sign signals of stuck drill accidents and normal samples. 3. The method for intelligently identifying early warning signals of a stuck pipe accident under a long-tail distribution according to claim 1, characterized in that the step S2 constructs a channel-time series hybrid stuck pipe prediction network; uses a cross-validation training test model, and when testing data of a specific oil well, the data of the oil well will be removed from the data set during the training phase. The probability that the logging time series data sample is a precursor signal of a stuck pipe accident is obtained. 4. The method for intelligently identifying early warning signals of a stuck pipe accident under a long-tail distribution according to claim 1 is characterized in that the step S3 further optimizes the network using the loss function of the long-tail distribution to solve the problem of mining stuck pipe samples under the imbalance of positive and negative samples in the logging data. The modulation coefficient is introduced to make the model pay more attention to the rare positive samples of early warning signals of a stuck pipe accident.