Drill bit wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM

Abstract

The invention relates to a drill bit wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM, and aims to realize real-time prediction of drill bit wear through an intelligent means. According to the method, drilling parameters collected on site are utilized, a physical model is fitted through a BP neural network, and a physical constraint module of a WGAN-GP generator is constructed to generate high-quality drilling data conforming to actual working conditions. And after verification of the generated data, combining the generated data with real data to form a complete data set. And then, a hybrid model combining a convolutional neural network (CNN) and a bidirectional long-short-term memory network (BiLSTM) with a multi-head attention mechanism is designed, spatial-temporal characteristics of drilling parameters are fully extracted, and classified prediction of the wear degree of the drill bit is realized. According to the method, the problem of insufficient field data is solved, the prediction accuracy is remarkably improved, the wear degree of the drill bit is predicted in advance, unnecessary premature replacement is avoided, the comprehensive cost of drilling operation is remarkably reduced, and the method has wide industrial application value.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas drilling engineering, and in particular to a drill bit wear prediction method based on physical constraints WGAN-GP and CNN-BiLSTM. Background Art

[0002] In drilling operations, the drill bit is a key component that directly contacts the formation, and its wear state directly affects drilling efficiency, equipment cost and operational safety. However, due to the complexity and variability of the drilling process, parameters such as formation hardness, bit pressure (WOB), rotation speed (RPM), torque (TOB), gamma ray (Gamma) and drilling speed (ROP) fluctuate greatly under different working conditions. Traditional drill bit wear prediction methods are difficult to accurately describe the nonlinear relationship between these parameters and wear state. This uncertainty often leads to premature replacement of the drill bit or accidents caused by excessive wear, thereby increasing operating costs and safety risks. Current drill bit wear detection methods mainly rely on on-site manual monitoring or simple models based on historical data. These methods have the following problems: data scarcity and noise problems, limited model expression capabilities, and lack of real-time early warning mechanisms.

[0003] With the rapid development of artificial intelligence technology, generative adversarial networks (GAN) and deep learning models have shown significant advantages in processing complex high-dimensional data. In particular, the generative adversarial network (WGAN-GP) can make up for the scarcity of field data by generating data, and the combination of convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM) can extract the spatiotemporal characteristics of drilling parameters, providing an efficient solution for drill bit wear prediction. However, existing research has not fully combined generative adversarial networks and deep learning technology to build an intelligent prediction system for drill bit wear.

[0004] In this context, the present invention proposes a drill wear prediction system based on physical constraint WGAN-GP and CNN-BiLSTM network, and its flow chart is as follows: Figure 1 As shown in the figure, it can not only generate high-quality drilling data to make up for the lack of field data, but also classify and predict the degree of drill bit wear in real time, providing an innovative solution for the development of intelligent drilling technology. Summary of the invention

[0005] The present invention mainly overcomes the shortcomings of the prior art. The present invention aims to solve the problems of data scarcity and noise, insufficient model expression ability, and lack of real-time warning mechanism in the existing drill bit wear detection methods, and provides a drill bit wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM network. By combining generative adversarial networks and deep learning technology, the present invention realizes high-quality data generation of drilling parameters, accurate classification prediction of drill bit wear degree, and real-time warning, which provides effective guarantee for the safety and economic benefits of drilling operations. In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] The implementation of a drill wear prediction system based on physical constraints WGAN-GP and CNN-BiLSTM includes the following steps:

[0007] S1: Collect key parameter data related to drill bit wear from the drilling site, including weight on bit (WOB), rotation speed (RPM), torque (TOB), gamma ray (Gamma) and rate of penetration (ROP).

[0008] S2: Normalize, denoise and perform feature engineering on the collected data.

[0009] S3: Use the BP neural network optimized by genetic algorithm (GA-BP) to fit the physical relationship of the parameters in S2 and build a physical model.

[0010] S4: Embed the physical model in S3 into the generator of WGAN-GP to ensure the physical consistency of the generated data.

[0011] S5: Compare and analyze the data generated by the WGAN-GP model in S4 with the field data, perform the Kolmogorov-Smirnov (KS) test, and calculate the Jensen-Shannon (JS) divergence. If the index requirements are met, the data enters S6, otherwise it returns to S3.

[0012] S6: The data in S5 is mixed with the real data in a certain proportion and then input into the trained CNN-BiLSTM neural network for wear prediction.

[0013] S7: The CNN-BiLSTM model outputs the classification results of the wear status and the model performance indicators: F1 score, recall rate, and precision rate, which are used to guide the decision-making of operators.

[0014] Furthermore, the normalization process in step S2 uses a min-max normalization method to map features of different dimensions to a unified scale for model training. The denoising process uses a wavelet transform combined with a threshold method.

[0015] Furthermore, the GA-BP network structure in step S3 is as follows: Figure 2 , the GA-BP network fits the physical relationship between the field data, such as the following formula (taking ROP as an example):

[0016]

[0017] Further, as a specific implementation of the present invention, the specific implementation of embedding the physical model into WGAN-GP in step S4 is: adding a penalty term in WGAN-GP, and the penalty term formula is as follows (taking ROP as an example):

[0018] L phys =λ×|ROP * -f GA-BP (WOB * , RPM * ,TOB * , Γ * )|

[0019] Among them, λ is a weight hyperparameter, and its size needs to be balanced during the training process to avoid excessive constraints that cause insufficient diversity of generated data. The penalty term formula should be added to the generator loss or discriminator loss. For example: when updating the generator gradient, the generator loss of ordinary WGAN-GP is modified to the following formula (taking ROP as an example):

[0020]

[0021] Further, as a specific implementation of the present invention, the implementation steps of the data comparison analysis in step S5 are as follows:

[0022] S1: Calculate multi-dimensional JS using the kernel density estimation algorithm. The formula is as follows:

[0023]

[0024] Where p k is the discrete distribution of real data, which can be obtained by multi-dimensional density function, q k The discrete distribution of the data generated by WGAN-GP can be obtained by the multidimensional density function, Js d is the multidimensional JS divergence.

[0025] S2: Calculate the multidimensional KS, perform a KS test on each dimension and then find the norm, perform the Frobenius norm difference between the generated data and the real data, normalize the norm difference, and then find the weighted sum of the two items to get the multidimensional KS divergence. The formula is as follows:

[0026]

[0027] ΔΣ =||Σ X -∑ Y || F

[0028]

[0029] KS d =α·D rms +(1-α)·Δ ∑,norm

[0030] Where D i KS test for each dimension, D rms is the norm, X is the real data set, Y is the generated data set, the dimensions of both data sets are d, α is the weight coefficient, its range is [0,1], KS d is the desired multidimensional KS divergence

[0031] S3: Normalize both the multidimensional KS test and the multidimensional JS divergence to the same dimension [0,1], the formula is as follows:

[0032]

[0033] KS d =KS d,norm

[0034] S3: Set weight θ, (the sum is 1), construct a comprehensive score M, the formula is as follows:

[0035]

[0036] Further, as a specific implementation of the present invention, the structure of the CNN-BiLSTM network in step S6 is as follows: the CNN convolution part is composed of 3 convolution layers, a Flatten smoothing layer, and a Dropout layer. The convolution layer includes a one-dimensional convolution layer, a maximum pooling layer (MaxPooling), and a BatchNormalization. The convolution kernel size is 3, the convolution step size is 1, and the number of convolution filters is 128, 256, and 512, respectively. The parameter in the maximum pooling layer is set to 2. The number of hidden units in the 2-layer BiLSTM is 256 and 128, respectively, and a multi-head attention mechanism layer (multiHeadAttentionLayer) is added after the BiLSTM layer to improve the expressive power of the model: the number is set to 4, and a Dropout layer is added after the multi-head attention mechanism to prevent overfitting. The optimizer selects adam, and the output layer uses a softmax function for classification prediction.

[0037] Beneficial effects:

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] WGAN-GP is combined with GA-BP to generate high-quality data that conforms to physical laws. GA-BP optimizes model parameters to ensure that the generated data meets actual needs in terms of physical consistency and diversity, enhancing the effectiveness and representativeness of training data. The indicator test further verifies the statistical consistency between the generated data and the real data, ensuring the reliability of data quality and model training.

[0040] In terms of feature learning and time series modeling, the CNN-BiLSTM network can accurately extract the local features and time series features of drilling parameters and accurately capture the key changes in the drill bit wear process. The multi-head attention mechanism further enhances the model's ability to focus on key moments and features, especially in time series data with long-term dependencies and complex changes, enhancing the model's attention ability and sensitivity to key time steps.

[0041] The method of the present invention can effectively learn the variation law of drill bit wear in multidimensional feature space, which not only improves the prediction accuracy, but also enhances the adaptability and generalization ability of the model to complex drilling conditions. Overall, the present invention provides an efficient and reliable drill bit wear prediction tool with wide application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the overall flow chart of the wear prediction system

[0043] Figure 2 This is the structure diagram of the GA-BP network

[0044] Figure 3 It is the performance training graph of GA-BP network

[0045] Figure 4 is the fitting error graph of the GA-BP network

[0046] Figure 5 This is the CNN-BiLSTM network structure diagram

[0047] Figure 6 This is the CNN-BiLSTM network performance graph DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and corresponding figures. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] Example:

[0050] A drill wear prediction system based on physical constraint WGAN-GP and CNN-BiLSTM network specifically includes the following steps:

[0051] S1: Export field data from the drilling sensors, integrate and name the data. Specifically, the data needs to be time synchronized, and for sensor data of different frequencies, the data needs to be aligned through interpolation and resampling, and the data of each sensor needs to be converted into a unified format;

[0052] S2: Data preprocessing, including denoising, filling missing values, feature engineering, standardization and normalization. Specifically, denoising uses wavelet transform and threshold method, which can be implemented by PyWavelets library in Python, filling missing values ​​can be implemented by fillna() method of pandas library in Python, and normalization is implemented by min-max method.

[0053] S3: The processed data is fitted and trained using the GA-BP network to obtain a suitable physical model. The population size, crossover probability, mutation probability, and maximum number of iterations of the GA in the GA-BP network are 30, 0.7, 0.04, and 200, respectively. The number of neurons in the two hidden layers is 13 and 8, respectively. Figure 3 The performance training diagram of the GA-BP network is shown. In order to test the performance of the GA-BP fitting curve, a set of data is randomly selected to fit the curve. The fitting error is as follows Figure 4 ;

[0054] S4: The physical model fitted by GA-BP is added to the generator and discriminator loss functions of WGAN-GP, and the new WGAN-GP is used to generate data and perform indicator detection. Specifically, the activation function in WGAN-GP uses LeakyReLU with a slope of 0.2. Insert layerNormalizationLayer to enhance network stability. The hidden layers of the generator and the discriminator are both 3 layers, with the number of neurons in the generator hidden layer being 128, 256, and 256 respectively, and the number of neurons in the discriminator hidden layer being 256, 256, and 128 respectively. Both the generator and the discriminator use the Dropout layer with a ratio of 0.25.

[0055] S5: The data that meets the index is input into the CNN-BiLSTM network for wear prediction. The structure of CNN-BiLSTM is as follows Figure 5Specifically, the convolution part of CNN consists of 3 convolution layers, a Flatten smoothing layer, and a Dropout layer. The convolution layer includes a one-dimensional convolution layer, a maximum pooling layer (MaxPooling), and a BatchNormalization layer. The convolution kernel size is 3, the convolution step size is 1, and the number of convolution filters is 128, 256, and 512 respectively. The parameter in the maximum pooling layer is set to 2. The number of hidden units in the 2-layer BiLSTM is 256 and 128 respectively, and a multi-head attention mechanism layer (multiHeadAttentionLayer) is added after the BiLSTM layer to improve the expression ability of the model: the number is set to 4, and a Dropout layer is added after the multi-head attention mechanism to prevent overfitting. The optimizer selects adam, and the output layer uses the softmax function for classification prediction. The wear level of the drill bit is roughly divided into 3 categories: level 0, level 1-2, and level 2-3. Figure 6 The following is a graph showing the performance of CNN-BiLSTM in predicting time series data. Table 1 shows the prediction results of CNN-BiLSTM.

[0056] Table 1 Prediction results of CNN-BiLSTM

[0057] Sort precision recall f1-score 0 0.93 0.95 0.94 1-2 0.90 0.92 0.91 2-3 0.92 0.90 0.91

[0058] 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

1. A drill wear prediction method based on physical constraints WGAN-GP and CNN-BiLSTM, characterized in that: The following steps are involved: S1: During the drilling process, key parameters of the drill bit are collected through the drilling sensors; S2: preprocessing the data collected by the drilling sensor; S3: input the preprocessed data into the BP neural network (GA-BP) optimized by genetic algorithm to fit the physical constraints of the parameters; S4: embed the physical constraints generated by the GA-BP network into the WGAN-BP neural network to optimize the data generated by WGAN-BP; S5: Set evaluation indicators to evaluate the data generated by the WGAN-BP neural network, and the parameters that meet the indicators can be output; S6: Import the parameters that meet the indicators into the trained CNN-BiLSTM neural network for wear prediction.

2. A drill wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM as claimed in claim 1, characterized in that: The data preprocessing module in step S2 includes denoising, normalization, and feature engineering.

3. A drill wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM as claimed in claim 1, characterized in that: The GA-BP neural network in step S3 uses a genetic algorithm to optimize the BP neural network. The population size, crossover probability, mutation probability, and maximum number of iterations of the GA are 30, 0.7, 0.04, and 200, respectively, and the number of neurons in the two hidden layers are 13 and 8, respectively.

4. A drill wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM as claimed in claim 1, characterized in that: The WGAN-GP in step S4 adopts the following improvements: a physical model constraint module fitted by a GA-BP neural network is embedded in the generator or discriminator to ensure that the generated data conforms to the actual physical laws of drilling, the activation function adopts LeakyReLU with a slope of 0.2, and layerNormalizationLayer is inserted to enhance network stability. The hidden layers of the generator and the discriminator are both 3 layers, and the number of neurons in the hidden layer of the generator is 128, 256, and 256 respectively, and the number of neurons in the hidden layer of the discriminator is 256, 256, and 128 respectively. The generator and the discriminator both use a Dropout layer with a ratio of 0.

25.

5. A drill wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM as claimed in claim 1, characterized in that: In step S5, KS test and JS divergence comprehensive evaluation index are adopted.

6. A drill wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM as claimed in claim 1, characterized in that: The trained CNN-BiLSTM network in step S6 is trained using a historical drilling dataset labeled with wear levels.

7. A drill wear prediction method based on physical constraint WGAN-GP and CNN-BiLSTM as claimed in claim 1, characterized in that: The CNN-BiLSTM network in step S6 includes: 3 convolutional layers for extracting spatial sequence features; 2 BiLSTM layers for extracting time series features; Dropout is used to prevent overfitting; a multi-head attention mechanism is added before the fully connected layer to improve the model's expression ability, and a Softmax layer is used to output the classification results of the wear status.

Images