Well cementation quality prediction method and device, computer equipment and storage medium

Abstract

The embodiment of the invention discloses a well cementation quality prediction method and device, computer equipment and a storage medium, and relates to the technical field of petroleum drilling engineering and the field of artificial intelligence. Comprising the following steps: acquiring well cementation data influencing well cementation quality; the well cementation data are input into the data normalization network, well cementation features corresponding to the well cementation data are obtained, and the well cementation features comprise dense features and sparse features; inputting the dense features into a dense feature processing model to obtain dense depth features; inputting the sparse features into a sparse feature processing model to obtain sparse depth features; inputting the dense depth features and the sparse depth features into a multi-head attention mechanism network to obtain attention scores corresponding to the dense depth features and the sparse depth features; and inputting the attention score into the full connection layer to obtain a first well cementation quality prediction result. By adopting the method provided by the embodiment of the invention, the accuracy of the well cementation quality prediction result can be improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the fields of oil drilling engineering technology and artificial intelligence, and in particular to a cementing quality prediction method, device, computer equipment and storage medium. Background Art

[0002] Cementing technology was pioneered in 1903 in the Lambers Oil Field in California, USA, by mixing 50 bags of Portland cement and sending it into the well using a sand bucket to seal the water layer above the oil layer in the oil well. Then, after 28 days, the cement plug inside the wellbore was drilled out and the oil layer was drilled. Since 1919, after the establishment of the API Standardization Committee, cementing-specific equipment and processes have been continuously developed, including cement trucks, testing equipment and cement admixtures, to meet cementing needs under various conditions.

[0003] Cementing quality is closely related to the productivity and life of oil and gas wells. In related technologies, cementing quality can be studied in areas such as cementing and logging through cementing software, testing software, interpretation methods, tools and accessories.

[0004] However, since there are many features that affect cementing quality, a detailed analysis of the factors affecting cementing quality is required. Existing machine learning and deep learning algorithms cannot well explore the relationship between features, resulting in poor prediction of cementing quality. Summary of the invention

[0005] The embodiment of the present application provides a cementing quality prediction method, device, computer equipment and storage medium. The technical solution is as follows:

[0006] On the one hand, an embodiment of the present application provides a cementing quality prediction method, the method comprising:

[0007] Inputting the cementing data into a data normalization network to obtain cementing features corresponding to the cementing data, wherein the data normalization network is used to extract normalized features from the cementing data, and the cementing features include dense features and sparse features;

[0008] Inputting the dense feature in the cementing feature into a dense feature processing model to obtain a dense depth feature;

[0009] Inputting the sparse feature in the cementing feature into a sparse feature processing model to obtain a sparse depth feature;

[0010] Inputting the dense depth feature and the sparse depth feature into a multi-head attention mechanism network to obtain attention scores corresponding to the dense depth feature and the sparse depth feature;

[0011] The attention score is input into the fully connected layer to obtain the first cementing quality prediction result.

[0012] On the other hand, an embodiment of the present application provides a cementing quality prediction device, the device comprising:

[0013] An acquisition module, used for acquiring cementing data affecting cementing quality, wherein the cementing data includes at least one of cementing layer characteristics, wellbore quality characteristics and cementing design characteristics;

[0014] A feature extraction module, used for inputting the cementing data into a data normalization network to obtain cementing features corresponding to the cementing data, wherein the data normalization network is used for performing normalized feature extraction on the cementing data, and the cementing features include dense features and sparse features;

[0015] A dense feature processing module, used for inputting the dense feature in the cementing feature into a dense feature processing model to obtain a dense depth feature;

[0016] A sparse feature processing module, used for inputting the sparse features in the cementing features into a sparse feature processing model to obtain sparse depth features;

[0017] An attention module, used to input the dense depth feature and the sparse depth feature into a multi-head attention mechanism network to obtain attention scores corresponding to the dense depth feature and the sparse depth feature;

[0018] A prediction module is used to input the attention score into a fully connected layer to obtain a first cementing quality prediction result.

[0019] On the other hand, an embodiment of the present application provides a computer device, comprising: a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the cementing quality prediction method as described in the above aspects.

[0020] On the other hand, an embodiment of the present application provides a computer-readable storage medium, in which at least one instruction is stored, and the instruction is loaded and executed by a processor to implement the cementing quality prediction method as described in the above aspects.

[0021] On the other hand, an embodiment of the present application provides a computer program product, which includes computer instructions, which are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the cementing quality prediction method as described in the above aspects.

[0022] In the embodiments of the present application, in view of the high-dimensional and dense nature of the factors affecting cementing quality and the shortcomings of the TabNet model in processing high-dimensional dense data, the present application proposes an improved TF-TabNet model, that is, dense features are processed by a dense feature processing model (such as the Transformer model) and sparse features are processed by a sparse feature processing model (such as TabNet), so as to enhance the model's adaptability to different feature densities, thereby improving its fitting ability and improving the effect of cementing quality prediction; a multi-head attention mechanism is added before the final output layer to calculate the attention score of each data feature, and more computing resources and weights are allocated to important features, so as to improve the expression and generalization capabilities of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 is a flow chart of a cementing quality prediction method provided by an exemplary embodiment of the present application;

[0025] Figure 2 is a data table of cementing data and cementing quality provided by an exemplary embodiment of the present application;

[0026] Figure 3 is a schematic diagram of the structure of a TF-TabNet model provided by an exemplary embodiment of the present application;

[0027] Figure 4 is a data table of some target hyperparameters provided by an exemplary embodiment of the present application;

[0028] Figure 5 is a schematic diagram of integrated learning provided by an exemplary embodiment of the present application;

[0029] Figure 6 is a schematic diagram of integrated learning provided by another exemplary embodiment of the present application;

[0030] Figure 7 is an evaluation index table of a base model provided by an exemplary embodiment of the present application;

[0031] Figure 8 An exemplary embodiment of the present application provides an evaluation index table for using four base models individually and integrating the base models;

[0032] Fig. 9is an evaluation index table of ablation experiment results provided by an exemplary embodiment of the present application;

[0033] Fig.10 is a structural block diagram of a cementing quality prediction device provided by an exemplary embodiment of the present application;

[0034] Fig.11 It is a structural diagram of a computer device provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0036] Among the related technologies, the analysis and research on the factors affecting cementing quality at home and abroad include:

[0037] In 1940, PH Jones and D. Berdin proposed the cement slurry displacement theory, believing that drilling fluid and cement slurry are both Newtonian fluids, and explored the factors affecting the displacement efficiency, such as cement slurry density, viscosity, casing centering, etc.

[0038] RCHant and RJCrook conducted a laboratory study in 1979 to simulate the factors affecting the wellbore annulus during cement slurry cementing and found that the displacement efficiency of cement slurry at high return rate was higher than that at low return rate. The study also provided displacement efficiency data under different flow states and the corresponding relationship between cement slurry contact time.

[0039] In his paper, Zhang Xingguo conducted a comprehensive and systematic analysis of the factors affecting cementing quality, thereby revealing the interrelationships between these factors. He elaborated on the impact of individual factors on cementing quality, and also pointed out the interactions between these factors.

[0040] Since there are many characteristics of factors affecting cementing quality, a detailed analysis of the factors affecting cementing quality is needed. The existing machine learning and deep learning algorithms cannot better explore the relationship between the characteristics, and a more powerful model is needed to predict cementing quality.

[0041] This application proposes a cementing quality prediction method. Based on the model structure proposed in this application, sparse features and dense features that affect cementing quality can be extracted separately to better explore the relationship between features and improve the prediction quality of cementing quality.

[0042] See also Figure 1 , Figure 1 : is a flow chart of a cementing quality prediction method provided by an exemplary embodiment of the present application. The method comprises the following steps:

[0043] Step 101, obtaining cementing data that affects cementing quality, wherein the cementing data includes at least one of cementing layer characteristics, wellbore quality characteristics, and cementing design characteristics.

[0044] Cementing data is data related to predicting cementing quality.

[0045] In some embodiments, cementing data can be obtained by writing an interface to read relevant data of a certain oil field in a database.

[0046] Optionally, cementing layer characteristics include nine data including depth, vertical depth, thickness, porosity, lithology description, layer type, maximum pore pressure, maximum oil and gas upwelling velocity, and overflow rate.

[0047] Optionally, the wellbore quality characteristics include 16 data including wellbore diameter, drill pipe diameter, displacement, pump pressure, surface casing depth, technical casing depth, well inclination, drilling fluid system, drilling fluid density, drilling fluid plastic viscosity, drilling fluid shear force, drilling fluid content, full-angle change rate, drill bit diameter, number of stabilizers, and stabilizer size.

[0048] Optionally, cementing design features include 25 data including drilling fluid density after completion, oil and gas upward velocity, drilling fluid density before cementing, annulus return velocity (displacement), 100BC thickening time of lead slurry, 100BC thickening time of tail rotor, annulus equivalent density after weightlessness, stabilizer type, stabilizer interval, dynamic shear force of drilling fluid before cementing, flushing fluid density, flushing fluid replacement amount, spacer fluid density, spacer fluid replacement amount, lead slurry system, lead slurry density, initial viscosity of lead slurry, displacement medium, displacement displacement, tail rotor system, compressive strength of tail rotor, construction time, circulation temperature, waiting time, whether the tail rotor loses water, etc.

[0049] Those skilled in the art can understand that cementing data may also include other data, such as single well basic data, drilling tool inspection data, formation fracture pressure data, etc. Such data that affect cementing quality are within the protection scope of this application.

[0050] See also Figure 2 , Figure 2 It is a data table of cementing data and cementing quality provided by an exemplary embodiment of the present application.

[0051] like Figure 2 As shown in the figure, the present embodiment collected cementing data corresponding to 7539 different well numbers. For example, the depth corresponding to the first well number is 3687.1m, the porosity is 11.16%, the wellbore diameter is 1.42m, and the drilling fluid density is 0.66g / cm 3The lithology is described as fine sandstone, and the predicted cementing quality grade is "good"; the depth corresponding to the second well number is 3715.4m, the porosity is 11.43%, the wellbore diameter is 1.27m, and the drilling fluid density is 0.71g / cm 3 The lithology is described as mudstone, and the predicted cementing quality grade is "medium"; the depth corresponding to well number 3 is 3739.4m, the porosity is 11.22%, the wellbore diameter is 1.42m, and the drilling fluid density is 1.77g / cm 3 The lithology is described as fine sandstone, and the predicted cementing quality grade is "medium"; the depth corresponding to well number 4 is 3845.4m, the porosity is 17.3%, the wellbore diameter is 1.42m, and the drilling fluid density is 0.18g / cm 3 , the lithology is described as fine sandstone, and the predicted cementing quality grade is "good"; ...; the depth corresponding to well number 7539 is 3948.6m, the porosity is 9.3%, the wellbore diameter is 1.27m, and the drilling fluid density is 1.98g / cm 3 The lithology is described as fine sandstone, and the predicted cementing quality grade is "poor".

[0052] Step 102, inputting the cementing data into a data normalization network to obtain cementing features corresponding to the cementing data. The data normalization network is used to extract normalized features from the cementing data. The cementing features include dense features and sparse features.

[0053] The data normalization network is used to scale the features extracted from the cementing data, map them into a smaller specific interval, and convert them into dimensionless pure values, so that indicators of different units or magnitudes can be compared and weighted.

[0054] Exemplarily, a data normalization network is used to uniformly map cementing features to the interval [0, 1].

[0055] In some embodiments, the data normalization network may include multiple types of data normalization networks, such as BN layer (Batch Normalization), LN layer (Layer Normalization) or IN layer (Instance Normalization), etc. Among them, the BN layer normalizes vertically so that each neuron in the same layer has its own mean and variance, and the LN layer normalizes horizontally so that each neuron in the same layer has the same mean and variance, while different input samples have different means and variances.

[0056] Regarding the method of distinguishing dense features from sparse features, in some embodiments, a feature whose element value is 0 and whose proportion is lower than the classification threshold after being processed by the data normalization network is a dense feature, and a feature whose element value is 0 and whose proportion is higher than the classification threshold is a dense feature.

[0057] The classification threshold is a threshold used to classify the feature type of cementing data. For example, the classification threshold is 0.5.

[0058] In other embodiments, it is also possible to determine whether the features corresponding to the cementing data are dense features or sparse features based on the type of cementing data. Dense features are features corresponding to numerical type data in cementing data, such as cementing data such as depth, thickness, and drilling fluid density. Sparse features are features corresponding to category type data in cementing data, such as centralizer type, slurry system, lithology description, and other features.

[0059] Since different models have different feature extraction capabilities for dense features or sparse features, in order to enhance the model's adaptability to different feature densities and improve the model's fitting ability, in some embodiments, feature extraction can be performed separately by a dense feature processing model for processing dense features and a sparse feature processing model for processing sparse features.

[0060] Step 103: input the dense features in the cementing features into a dense feature processing model to obtain dense depth features.

[0061] In some embodiments, the dense feature processing model is a transformer model. The transformer model is a model that uses the attention mechanism to improve the model training speed, which is proposed by Google's paper Attention is All You Need. In some embodiments, the transformer model consists of two parts: an encoder and a decoder.

[0062] In other embodiments, the dense feature processing model can also be other models suitable for processing dense features, such as an LSTM model (Long Short Term Memory) or an RNN model (Recurrent Neural Network), etc., which is not limited to the embodiments of the present application.

[0063] Step 104: input the sparse features in the cementing features into a sparse feature processing model to obtain sparse depth features.

[0064] In some embodiments, the sparse feature processing model includes a TabNet model. The TabNet model is a model proposed by Google Cloud AI in 2020. It is a network structure designed specifically for tabular data and simulates the decision manifold of the tree model through a neural network.

[0065] In other embodiments, the sparse feature processing model may also include a tree model (Tree-based models), such as a decision tree, an XGBoost model, a lightGBM, or a Random Forest model.

[0066] Since cementing data is essentially tabular data, the current neural network algorithm is not as effective on tabular data as on image and text data. Usually, the tree model or the TabNet model that simulates the tree model through a neural network is more suitable for processing tabular data.

[0067] The TabNet model simulates the properties of decision trees and is more suitable for processing tabular data than other deep learning models. Currently, the TabNet model has been applied to a wide range of fields and has achieved success. TabNet is a deep learning model based on attention mechanism and sparse feature interaction proposed by the Google Brain team in 2020. It is mainly used for classification and regression tasks of tabular data. Compared with traditional neural network models, TabNet can better capture the relationship and importance between features, thereby improving prediction accuracy and stability. Compared with traditional tabular data algorithms such as XGBoost, LightGBM and other models, the TabNet model inherits the advantages of interpretability and coefficient feature selection of tree models, and performs best in classification and regression tasks of large data sets.

[0068] Cementing data contains both sparse features and dense features, and the TabNet model is usually better at processing sparse features, while the processing effect on dense features is relatively weak. Therefore, in view of the high-dimensional and dense nature of the factors affecting cementing quality and the shortcomings of the TabNet model in processing high-dimensional dense data, an improved TF-TabNet model is proposed in the embodiment of the present application. That is, dense features are processed by a dense feature processing model (such as the Transformer model) and sparse features are processed by a sparse feature processing model (such as TabNet), thereby enhancing the model's adaptability to different feature densities, thereby improving its fitting ability and improving the effect of cementing quality prediction.

[0069] Step 105, input the dense depth features and the sparse depth features into the multi-head attention mechanism network to obtain the attention scores corresponding to the dense depth features and the sparse depth features.

[0070] In some embodiments, the multi-headed self-attention network includes but is not limited to model structures such as the BERT model and the Transformer model.

[0071] Exemplarily, the attention scores corresponding to different features can be represented as vectors, such as [0.1, 0.5, 0.3...], where each dimension of the vector represents the weight of the influence of the feature value on the final decision when a certain condition is met.

[0072] By adding a multi-head attention mechanism network before the final output layer, it can be used to calculate the attention score of each data feature and allocate more computing resources and weights to important features, thereby improving the expressiveness and generalization capabilities of the model.

[0073] Step 106: input the attention score into the fully connected layer to obtain the first cementing quality prediction result.

[0074] Optionally, the first cementing quality prediction result may be a cementing quality score, such as 0 to 100 points.

[0075] Optionally, the first cementing quality prediction result may be cementing quality grades, such as three grades: good, and poor.

[0076] In summary, in view of the high-dimensional and dense nature of the factors affecting cementing quality and the shortcomings of the TabNet model in processing high-dimensional dense data, this application proposes an improved TF-TabNet model, that is, dense features are processed by dense feature processing models (such as the Transformer model) and sparse features are processed by sparse feature processing models (such as TabNet), so as to enhance the model's adaptability to different feature densities, thereby improving its fitting ability and improving the effect of cementing quality prediction; a multi-head attention mechanism is added before the final output layer to calculate the attention score of each data feature, and more computing resources and weights are allocated to important features, so as to improve the model's expressiveness and generalization capabilities.

[0077] See also Figure 3 , Figure 3 It is a structural diagram of the TF-TabNet model provided by an exemplary embodiment of the present application.

[0078] like Figure 3 As shown, cementing data 301 is input into the data normalization network 310 to obtain sparse features and dense features, wherein the sparse features are input into the TabNet model 330 for processing to obtain sparse depth features; and the dense features are input into the Transformer model 320 for processing to obtain dense depth features.

[0079] The TabNet model 330 may include an Attentive Transformer model for feature selection, a Feature Transformer model for feature processing, a Split model for segmentation, a MASK model for masking, and an activation function ReLU.

[0080] The sparse deep features and the dense deep features are input into the multi-head attention mechanism network 340 after feature fusion to obtain the attention scores corresponding to different features.

[0081] Exemplarily, the attention scores corresponding to different features can be represented as vectors, such as [0.1, 0.5, 0.3...], where each dimension of the vector represents the weight of the influence of the feature value on the final decision when a certain condition is met.

[0082] The attention score is input into the fully connected layer 350 and passes through the Softmax classification function 360 to obtain the first cementing quality prediction result 370.

[0083] In order to select appropriate hyperparameters for the TF-TabNet model, in some embodiments, a particle swarm algorithm may be used to determine appropriate target hyperparameters through multiple rounds of iterations.

[0084] In some embodiments, a particle swarm algorithm may be used to tune at least one hyperparameter corresponding to the Transformer model through multiple rounds of iterations to obtain a target hyperparameter.

[0085] Optionally, the hyperparameters include at least one of embedding dimension, number of attention heads, number of encoder layers, number of decoder layers, and learning rate.

[0086] In some embodiments, dense features may be input into a Transformer model using target hyperparameters to obtain dense deep features.

[0087] In some embodiments, a particle swarm algorithm can be used to tune at least one hyperparameter corresponding to the TabNet model through multiple rounds of iterations to obtain a target hyperparameter.

[0088] Optionally, the hyperparameters include at least one of the number of decision layers, the number of attention layers, the number of TabNet steps, the learning rate, and the optimizer.

[0089] In some embodiments, the sparse features can be input into a TabNet model using target hyperparameters to obtain sparse deep features.

[0090] The process of tuning at least one hyperparameter corresponding to the Transformer model or the TabNet model through multiple rounds of iterations using a particle swarm algorithm to obtain a target hyperparameter may include the following steps.

[0091] Step S1, randomly initialize a particle swarm. The particles in the particle swarm represent a combination of hyperparameters, and the particles have a current speed and a current position.

[0092] According to the number of parameters of the TF-TabNet model to be optimized, a certain number of particles are randomly initialized, each particle represents a parameter combination. At the same time, the current speed and current position of each particle are set.

[0093] Step S2, according to the hyperparameter combination characterized by each particle in the particle swarm, calculate the loss value corresponding to the particle on the training sample, and determine the current fitness of the particle based on the loss value.

[0094] Among them, the lower the loss value, the faster the convergence speed and the higher the current fitness of the particle.

[0095] Step S3, update the individual optimal solution. For each particle, if the current fitness is higher than the historical fitness, update the current position to the individual optimal position.

[0096] If the fitness value of the current position is better than the historical best fitness, the historical best fitness is updated.

[0097] Step S4, updating the swarm optimal solution. For a particle swarm, when the current global optimal fitness is higher than the historical global optimal fitness, the global optimal position is updated.

[0098] Select the global optimal position and fitness value among all particles. If the fitness value of the current position is better than the global optimal fitness value, update the global optimal position and fitness value.

[0099] Step S5, updating particle speed and position: Based on the particle's current position, individual optimal position, and global optimal position, the particle's current speed and current position are updated.

[0100] Optionally, the update formula includes acceleration coefficient, learning factor, weight, etc.

[0101] Step S6, determining the stopping condition. When the iteration stopping condition is met, the target hyperparameter is determined based on the hyperparameter combination corresponding to the global optimal position.

[0102] Optionally, check whether the maximum number of iterations is reached or the hyperparameter combination corresponding to the local optimal position satisfies certain stopping conditions.

[0103] Step S7, outputting the result. Determine the optimized TabNet model or Transformer model based on the hyperparameter combination corresponding to the global optimal position, and perform prediction.

[0104] Optionally, in the particle swarm algorithm, you can set the number of particles n_particles = 1000 to 2000, the particle dimension dimension = 5 to 10, the inertia weight w = 0.7, 0.8 or 0.9, the individual acceleration coefficient c1 = 2.0, the global acceleration coefficient c2 = 2.0, and set the maximum number of iterations to 100 or more. The obtained partial target hyperparameters are as follows: Figure 4 shown.

[0105] See also Figure 4 , Figure 4 This is a data table of some target hyperparameters provided by an exemplary embodiment of the present application.

[0106] like Figure 4 As shown in the figure, after multiple rounds of iterations of the particle swarm algorithm, the target hyperparameters obtained include the number of decision layers (N_d) of 5-10, the number of attention layers (N_a) of 5-20, the number of TabNet steps (N_steps) of 3, 4 or 5, the learning rate (Lr) of 0.01, and the optimizer of Adam algorithm.

[0107] The goal of machine learning is to search for a learning model with strong generalization ability and high robustness in the hypothesis space. Ensemble learning is to combine multiple trained single models to obtain a higher performance combined model to improve the performance of a single model.

[0108] Ensemble learning can be divided into homogeneous ensemble and heterogeneous ensemble according to whether the types of base models are consistent. Homogeneous ensemble refers to integrating base models of the same type according to a certain combination strategy. For example, the algorithms of the Boosting and Bagging series belong to homogeneous ensemble. Unlike homogeneous ensemble, the types of each base model in heterogeneous ensemble are different. Heterogeneous ensemble superimposes each base model and combines them through combination strategies, such as the Stacking series of algorithms and hybrid expert models. Among them, the hybrid expert model was proposed by Jacobs et al. as early as 1991. Then in 1992, Wolpert proposed the Stacking ensemble algorithm, which first trains multiple base models with high generalization performance, and then integrates the results through the stacking method to predict the final result.

[0109] In the related art, the integrated learning fusion model is a simple averaging method, which cannot predict accurate cementing quality information to a greater extent. Therefore, the embodiment of the present application proposes an improvement measure for the Stacking algorithm. For the same base learner, a reasonable weight is assigned by calculating the error between the model prediction result and the true value in different k-fold cross validations. Then, these weights are used to perform a weighted average on the different prediction results obtained by the same base learner on the test set to obtain the final prediction result, so as to improve the prediction quality of cementing quality.

[0110] In some embodiments, Figure 3 The model proposed in the present application is called the TF-TabNet model, that is, the data normalization network, the dense feature processing model, the sparse feature processing model, the multi-head attention mechanism network and the fully connected layer constitute the TF-TabNet model.

[0111] Regarding the method of ensemble learning of multiple base learners, in some embodiments, the TF-TabNet model, MLP model, Wide&Deep model and DeepGBM model can be used as base learners, and the XGBoost model can be used as a meta-learner for integration.

[0112] In some embodiments, cementing data can be input into the TF-TabNet model, the MLP model, the Wide&Deep model and the DeepGBM model on the base training set and the base test set, respectively, to obtain a first cementing quality prediction result, a second cementing quality prediction result, a third cementing quality prediction result and a fourth cementing quality prediction result.

[0113] Among them, the base training set is a data set used to train the base model in ensemble learning, and the base test set is a data set used to test the base model.

[0114] In some embodiments, a meta-training set and a meta-testing set can be generated based on the first cementing quality prediction result obtained by the TF-TabNet model on the base training set and the base test set, as well as the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result; an XGBoost model is trained on the meta-training set; and cementing data is input into the XGBoost model on the meta-testing set to obtain cementing quality prediction results.

[0115] Among them, the meta-training set is a data set used to train the meta-model in ensemble learning, and the meta-test set is a data set used to test the meta-model.

[0116] See also Figure 5 , Figure 5 It is a schematic diagram of integrated learning provided by an exemplary embodiment of the present application.

[0117] like Figure 5 As shown, in some embodiments, the data set includes a base training set and a base test set, wherein the base training set is used for training base models M1, M2, M3 and M4, and the base test set is used for testing the above four base models.

[0118] Optionally, the base models M1, M2, M3 and M4 may be a TF-TabNet model, an MLP model, a Wide&Deep model and a DeepGBM model, respectively.

[0119] In some embodiments, the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result are all obtained by k-fold cross validation.

[0120] In some embodiments, the base training set may be k-folded, that is, the base training set may be divided into k subsets, and one of the subsets is used as the base validation subset and the remaining k-1 subsets are used as the base training subsets in turn.

[0121] In some embodiments, the TF-TabNet model, MLP model, Wide & Deep model and DeepGBM model can be trained based on the base training subset, and the prediction accuracy corresponding to the TF-TabNet model, MLP model, Wide & Deep model and DeepGBM model can be calculated based on the base validation subset.

[0122] For example, the F1-score can be used as the prediction accuracy of the base model on the base validation subset.

[0123] In some embodiments, k weighted weights corresponding to the TF-TabNet model, the MLP model, the Wide & Deep model, and the DeepGBM model can be determined according to the prediction accuracies corresponding to the k base validation subsets rotated in sequence, wherein the prediction accuracy is determined based on the F1-Score.

[0124] For example, the larger the F1-Score of the TF-TabNet model corresponding to the base validation subset, the larger the weighted weight corresponding to the base validation subset.

[0125] Optionally, taking the TF-TabNet model as an example, all F1-Scores corresponding to k base validation subsets can be added together, and then the proportion of the F1-Score of the model under different k values ​​to the total can be calculated to determine the weighted weight.

[0126] Exemplarily, the weighted weight is denoted as Ak.

[0127] In some embodiments, the prediction results of the TF-TabNet model, the MLP model, the Wide & Deep model and the DeepGBM model on the base test set can be weighted and summed according to the weighted weights to obtain the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result corresponding to the base test set.

[0128] Exemplarily, the prediction result of the base model on the base test set is denoted as Ptest, the first cementing quality prediction result obtained by weighted summing the prediction results of the base model M1 on the base test set is denoted as PtestM1, ..., and the fourth cementing quality prediction result obtained by weighted summing the prediction results of the base model M4 on the base test set is denoted as PtestM4.

[0129] In some embodiments, the prediction results of the TF-TabNet model, the MLP model, the Wide & Deep model and the DeepGBM model on the base training subset can be vertically superimposed to obtain the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result corresponding to the base training set.

[0130] Exemplarily, the prediction results Ptrain1-k of the base model M1 on the base training subset are vertically superimposed to obtain Ptrain M1, ..., and the prediction results Ptrain1-k of the base model M4 on the base training subset are vertically superimposed to obtain Ptrain M4.

[0131] In some embodiments, a meta-test set can be generated based on the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result obtained in the base test set; a meta-training set can be generated based on the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result obtained in the base training set. A meta-model (such as an XGBoost model) is trained on the meta-training set; cementing data is input into the meta-model on the meta-test set to obtain cementing quality prediction results.

[0132] It should be noted that the embodiment of the present application is only described with the number of base models being 4. In fact, the number of base models can be any number, and the embodiment of the present application does not limit the specific type of the base model.

[0133] See also Figure 6 , Figure 6 It is a schematic diagram of integrated learning provided by an exemplary embodiment of the present application.

[0134] like Figure 6As shown in the figure, the four base models (TF-TabNet model, MLP model, Wide&Deep model and DeepGBM model) are trained and tested respectively using data sets to form data sets generated by the base models, and then the meta-model (XGBoost model) is trained based on the generated data sets to obtain the output results.

[0135] In this embodiment, reasonable weighted weights are assigned by calculating the error between the model prediction result and the true value in different k-fold cross validations, and then these weighted weights are used to perform weighted averaging on the different prediction results obtained by the same base learner on the test set to obtain the final prediction result, which can improve the prediction quality of cementing quality.

[0136] See also Figure 7 , Figure 7 It is an evaluation index table of a base model provided by an exemplary embodiment of the present application.

[0137] This application divides the processed data set into a training set and a test set in a 4:1 ratio, and inputs them into the Transformer model, TF-TabNet model, TabNet model, LSTM model, and CNN model respectively, and sets the maximum number of iterations to 40 times.

[0138] The loss values ​​of the five models tend to be stable after 30 iterations. Figure 6 It can be seen that the training error of the TF-TabNet model is the lowest (its F1-score is the highest, which is 0.7872), and the training error of the TF-TabNet model is lower than that of the unmodified TabNet model (its F1-score is the second highest, which is 0.7763). The training errors of the TF-TabNet model and the TabNet model are both lower than those of the Transformer model, LSTM, and CNN models.

[0139] From the above table, we can see that the precision, recall, and F1-Score of TF-TabNet are improved by 1.1%, 0.9%, and 1.1% respectively compared with the Tabnet model, and by 3.3%, 5.1%, and 4.2% compared with the Transformer model. TF-TabNet introduced the Transformer structure to process dense features, which improved the results. At the same time, compared with the LSTM and CNN models, the precision, recall, and F1-Score of the TF-TabNet, TabNet, and Transformer models have been greatly improved, which shows that the model with the attention mechanism can better focus on the features related to the current task, thereby improving the model performance.

[0140] See also Figure 8 , Figure 8An exemplary embodiment of the present application provides an evaluation index table for using four base models individually and integrating the base models.

[0141] like Figure 8 As shown in the figure, the precision, recall and F1-Score of the unimproved Stacking algorithm (whose F1-score is the second highest, 0.8190) and the improved Stacking algorithm (whose F1-score is the highest, 0.8305) are higher than those of the single model, which shows the rationality of the learner selected in this application. Among them, the fusion model after the improved Stacking algorithm is better than the fusion model without the improved Stacking algorithm, which shows that the fusion model after the improved Stacking algorithm in this application further improves the effect of cementing quality prediction.

[0142] In order to further prove the rationality of the selection of base learners and meta-learners, this application designed an ablation experiment, using one model from TF-TabNet, Wide&Deep, MLP, DeepGBM, and XGBoost as a meta-learner, and the other four models as base learners. The results of the fusion model are shown in the figure below.

[0143] See also Fig. 9 , Fig. 9 This is an evaluation index table of ablation experiment results provided by an exemplary embodiment of the present application.

[0144] pass Fig. 9 It can be clearly seen that the fusion model with XGBoost as the meta-learner and the other four models as the base learners has the highest precision, recall, and F1-Score. The experiment further verifies the rationality of the selection of base learners and meta-learners in this application. At the same time, it can also be found that the improved Stacking algorithm has good generalization ability.

[0145] See also Fig.10 , Fig.10 This is a structural block diagram of a cementing quality prediction device provided by an exemplary embodiment of the present application. The device includes:

[0146] An acquisition module 1001 is used to acquire cementing data that affects cementing quality, wherein the cementing data includes at least one of cementing layer characteristics, wellbore quality characteristics, and cementing design characteristics;

[0147] A feature extraction module 1002 is used to input the cementing data into a data normalization network to obtain cementing features corresponding to the cementing data. The data normalization network is used to extract normalized features from the cementing data. The cementing features include dense features and sparse features.

[0148] A dense feature processing module 1003 is used to input the dense feature in the cementing feature into a dense feature processing model to obtain a dense depth feature;

[0149] A sparse feature processing module 1004 is used to input the sparse features in the cementing features into a sparse feature processing model to obtain sparse depth features;

[0150] An attention module 1005 is used to input the dense depth feature and the sparse depth feature into a multi-head attention mechanism network to obtain attention scores corresponding to the dense depth feature and the sparse depth feature;

[0151] The prediction module 1006 is used to input the attention score into the fully connected layer to obtain a first cementing quality prediction result.

[0152] Optionally, the dense feature processing model is a Transformer model, and the device further includes a hyperparameter determination module, which is used to:

[0153] Using a particle swarm algorithm, tuning at least one hyperparameter corresponding to the Transformer model after multiple rounds of iterations to obtain a target hyperparameter, wherein the hyperparameter includes at least one of an embedding dimension, a number of attention heads, a number of encoder layers, a number of decoder layers, and a learning rate;

[0154] Optionally, the dense feature processing module 1003 is used to:

[0155] The dense features are input into the Transformer model using the target hyperparameters to obtain the dense deep features.

[0156] Optionally, the sparse feature processing model is a TabNet model, and the hyperparameter determination module is used to:

[0157] Through a particle swarm algorithm, at least one hyperparameter corresponding to the TabNet model is tuned through multiple rounds of iterations to obtain a target hyperparameter, wherein the hyperparameter includes at least one of the number of decision layers, the number of attention layers, the number of TabNet steps, the learning rate, and the optimizer;

[0158] Optionally, the sparse feature processing module 1004 is used to:

[0159] The sparse features are input into the TabNet model using the target hyperparameters to obtain the sparse deep features.

[0160] Optional, hyperparameter determination module, used to:

[0161] Randomly initialize a particle swarm, where particles in the particle swarm represent a hyperparameter combination, and the particles have a current speed and a current position;

[0162] According to the hyperparameter combination represented by each particle in the particle swarm, a loss value corresponding to the particle is calculated on a training sample, and a current fitness of the particle is determined based on the loss value;

[0163] For each particle, when the current fitness is higher than the historical fitness, updating the current position to the individual best position;

[0164] For the particle swarm, when the current global optimal fitness is higher than the historical global optimal fitness, updating the global optimal position;

[0165] Based on the current position of the particle, the individual optimal position, and the global optimal position, updating the current speed and the current position of the particle;

[0166] When the iteration stopping condition is met, the target hyperparameter is determined based on the hyperparameter combination corresponding to the global optimal position.

[0167] Optionally, the data normalization network, the dense feature processing model, the sparse feature processing model, the multi-head attention mechanism network and the fully connected layer constitute a TF-TabNet model, and the device further includes an integrated learning module for:

[0168] On a base training set and a base test set, the cementing data are respectively input into an MLP model, a Wide&Deep model and a DeepGBM model to obtain a second cementing quality prediction result, a third cementing quality prediction result and a fourth cementing quality prediction result; the base training set is a data set used to train a base model in ensemble learning, and the base test set is a data set used to test the base model;

[0169] Based on the first cementing quality prediction result obtained by the TF-TabNet model on the base training set and the base test set, as well as the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result, a meta-training set and a meta-testing set are generated; the meta-training set is a data set for training a meta-model in ensemble learning, and the meta-testing set is a data set for testing the meta-model;

[0170] Training an XGBoost model on the meta-training set;

[0171] The cementing data is input into the XGBoost model on the meta-test set to obtain cementing quality prediction results.

[0172] Optionally, the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result are all obtained by k-fold cross validation, and the integrated learning module is used to:

[0173] Divide the base training set into k subsets, and take turns to use one of the subsets as the base validation subset and the remaining k-1 subsets as the base training subsets;

[0174] The MLP model, the Wide & Deep model and the DeepGBM model are trained based on the base training subset, and the prediction accuracy corresponding to the MLP model, the Wide & Deep model and the DeepGBM model is calculated based on the base validation subset;

[0175] According to the prediction accuracies corresponding to the k base validation subsets rotated in sequence, k weighted weights corresponding to the MLP model, the Wide & Deep model, and the DeepGBM model are determined, wherein the prediction accuracy is determined based on F1-Score;

[0176] Performing weighted summation on the prediction results of the MLP model, the Wide & Deep model, and the DeepGBM model on the base test set according to the weighted weights to obtain the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result, and the fourth cementing quality prediction result corresponding to the base test set;

[0177] The prediction results of the MLP model, the Wide & Deep model and the DeepGBM model on the base training subset are vertically superimposed to obtain the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result corresponding to the base training set.

[0178] Optional, integrated learning modules for:

[0179] generating the meta-test set based on the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result obtained on the base test set;

[0180] The meta-training set is generated based on the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result obtained on the base training set.

[0181] Optionally, the dense feature is a feature whose element value is 0 and the proportion is lower than the classification threshold, and the sparse feature is a feature whose element value is 0 and the proportion is higher than the classification threshold, or,

[0182] The dense features are features corresponding to numerical type data in the cementing data, and the sparse features are features corresponding to category type data in the cementing data.

[0183] See also Fig.11 , Fig.11 It is a structural diagram of a computer device provided by an exemplary embodiment of the present application.

[0184] The computer device can execute the cementing quality prediction method of the above embodiment. The computer device can also include one or more of the following components: a processor 1110 and a memory 1120 .

[0185] Optionally, the processor 1110 uses various interfaces and lines to connect various parts of the entire electronic device, and executes various functions of the electronic device and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 1120, and calling data stored in the memory 1120. Optionally, the processor 1110 can be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 1110 can integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), and a baseband chip. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing the content that needs to be displayed on the touch display; the NPU is used to implement artificial intelligence (AI) functions; and the baseband chip is used to process wireless communications. It is understandable that the above-mentioned baseband chip may not be integrated into the processor 1110, but may be implemented by a separate chip.

[0186] The memory 1120 may include a random access memory (RAM) or a read-only memory (ROM). Optionally, the memory 1120 includes a non-transitory computer-readable storage medium. The memory 1120 may be used to store instructions, programs, codes, code sets, or instruction sets. The memory 1120 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the following various method embodiments, etc.; the data storage area may store data created according to the use of the electronic device (such as audio data, a phone book), etc.

[0187] In addition, those skilled in the art will appreciate that the structure of the electronic device shown in the above figures does not constitute a limitation on the electronic device, and the electronic device may include more or fewer components than shown in the figures, or a combination of certain components, or a different arrangement of components.

[0188] The embodiment of the present application also provides a computer-readable storage medium, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement the method described in the above embodiment. Optionally, the computer-readable storage medium may include: ROM, RAM, solid state drive (SSD, Solid State Drives) or optical disk, etc. Among them, RAM may include resistance random access memory (ReRAM, Resistance Random Access Memory) and dynamic random access memory (DRAM, Dynamic Random Access Memory).

[0189] The embodiment of the present application also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in various optional implementations of the above aspects.

[0190] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A cementing quality prediction method, characterized in that: The method comprises: Acquiring cementing data that affects cementing quality, wherein the cementing data includes at least one of cementing layer characteristics, wellbore quality characteristics, and cementing design characteristics; Inputting the cementing data into a data normalization network to obtain cementing features corresponding to the cementing data, wherein the data normalization network is used to extract normalized features from the cementing data, and the cementing features include dense features and sparse features; Inputting the dense feature in the cementing feature into a dense feature processing model to obtain a dense depth feature; Inputting the sparse feature in the cementing feature into a sparse feature processing model to obtain a sparse depth feature; Inputting the dense depth feature and the sparse depth feature into a multi-head attention mechanism network to obtain attention scores corresponding to the dense depth feature and the sparse depth feature; The attention score is input into the fully connected layer to obtain the first cementing quality prediction result.

2. The method according to claim 1, characterized in that: The dense feature processing model is a Transformer model. Before inputting the cementing data into the data normalization network, the method further includes: Using a particle swarm algorithm, tuning at least one hyperparameter corresponding to the Transformer model after multiple rounds of iterations to obtain a target hyperparameter, wherein the hyperparameter includes at least one of an embedding dimension, a number of attention heads, a number of encoder layers, a number of decoder layers, and a learning rate; The step of inputting the dense feature in the cementing feature into a dense feature processing model to obtain a dense depth feature includes: The dense features are input into the Transformer model using the target hyperparameters to obtain the dense deep features.

3. The method according to claim 1, characterized in that The sparse feature processing model is a TabNet model. Before inputting the cementing data into the data normalization network, the method further includes: Through a particle swarm algorithm, at least one hyperparameter corresponding to the TabNet model is tuned through multiple rounds of iterations to obtain a target hyperparameter, wherein the hyperparameter includes at least one of the number of decision layers, the number of attention layers, the number of TabNet steps, the learning rate, and the optimizer; The step of inputting the sparse feature in the cementing feature into a sparse feature processing model to obtain a sparse depth feature includes: The sparse features are input into the TabNet model using the target hyperparameters to obtain the sparse deep features.

4. The method according to claim 2 or 3, characterized in that: The particle swarm algorithm is used to tune at least one hyperparameter corresponding to the Transformer model or the TabNet model through multiple rounds of iterations to obtain a target hyperparameter, including: Randomly initialize a particle swarm, where particles in the particle swarm represent a hyperparameter combination, and the particles have a current speed and a current position; According to the hyperparameter combination represented by each particle in the particle swarm, a loss value corresponding to the particle is calculated on a training sample, and a current fitness of the particle is determined based on the loss value; For each particle, when the current fitness is higher than the historical fitness, updating the current position to the individual best position; For the particle swarm, when the current global optimal fitness is higher than the historical global optimal fitness, updating the global optimal position; Based on the current position of the particle, the individual optimal position, and the global optimal position, updating the current speed and the current position of the particle; When the iteration stopping condition is met, the target hyperparameter is determined based on the hyperparameter combination corresponding to the global optimal position.

5. The method according to claim 1, characterized in that The data normalization network, the dense feature processing model, the sparse feature processing model, the multi-head attention mechanism network and the fully connected layer constitute a TF-TabNet model, and the method further includes: On a base training set and a base test set, the cementing data are respectively input into an MLP model, a Wide&Deep model and a DeepGBM model to obtain a second cementing quality prediction result, a third cementing quality prediction result and a fourth cementing quality prediction result; the base training set is a data set used to train a base model in ensemble learning, and the base test set is a data set used to test the base model; Based on the first cementing quality prediction result obtained by the TF-TabNet model on the base training set and the base test set, as well as the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result, a meta-training set and a meta-testing set are generated; the meta-training set is a data set for training a meta-model in ensemble learning, and the meta-testing set is a data set for testing the meta-model; Training an XGBoost model on the meta-training set; The cementing data is input into the XGBoost model on the meta-test set to obtain cementing quality prediction results.

6. The method according to claim 5, characterized in that The first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result are all obtained by k-fold cross validation. On the base training set and the base test set, the cementing data are respectively input into the MLP model, the Wide & Deep model and the DeepGBM model to obtain the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result, including: Divide the base training set into k subsets, and take turns taking one of the k subsets as a base validation subset and the remaining k-1 subsets as base training subsets; The MLP model, the Wide & Deep model and the DeepGBM model are trained based on the base training subset, and the prediction accuracy corresponding to the MLP model, the Wide & Deep model and the DeepGBM model is calculated based on the base validation subset; According to the prediction accuracies corresponding to the k base validation subsets rotated in sequence, k weighted weights corresponding to the MLP model, the Wide & Deep model, and the DeepGBM model are determined, wherein the prediction accuracy is determined based on F1-Score; Performing weighted summation on the prediction results of the MLP model, the Wide & Deep model, and the DeepGBM model on the base test set according to the weighted weights to obtain the second cementing quality prediction result, the third cementing quality prediction result, and the fourth cementing quality prediction result corresponding to the base test set; The prediction results of the MLP model, the Wide & Deep model and the DeepGBM model on the base training subset are vertically superimposed to obtain the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result corresponding to the base training set.

7. The method according to claim 6, characterized in that The generating of a meta-training set and a meta-testing set based on the first cementing quality prediction result obtained by the TF-TabNet model on the base training set and the base test set, and the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result, comprises: generating the meta-test set based on the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result obtained on the base test set; The meta-training set is generated based on the first cementing quality prediction result, the second cementing quality prediction result, the third cementing quality prediction result and the fourth cementing quality prediction result obtained on the base training set.

8. The method according to claim 1, characterized in that The dense feature is a feature whose element value is 0 and the proportion is lower than the classification threshold, and the sparse feature is a feature whose element value is 0 and the proportion is higher than the classification threshold, or, The dense features are features corresponding to numerical type data in the cementing data, and the sparse features are features corresponding to category type data in the cementing data.

9. A cementing quality prediction device, characterized in that: The device comprises: An acquisition module, used for acquiring cementing data affecting cementing quality, wherein the cementing data includes at least one of cementing layer characteristics, wellbore quality characteristics and cementing design characteristics; A feature extraction module, used for inputting the cementing data into a data normalization network to obtain cementing features corresponding to the cementing data, wherein the data normalization network is used for performing normalized feature extraction on the cementing data, and the cementing features include dense features and sparse features; A dense feature processing module, used for inputting the dense feature in the cementing feature into a dense feature processing model to obtain a dense depth feature; A sparse feature processing module, used for inputting the sparse features in the cementing features into a sparse feature processing model to obtain sparse depth features; An attention module, used to input the dense depth feature and the sparse depth feature into a multi-head attention mechanism network to obtain attention scores corresponding to the dense depth feature and the sparse depth feature; A prediction module is used to input the attention score into a fully connected layer to obtain a first cementing quality prediction result.

10. A computer device, characterized in that: The computer device comprises: a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the cementing quality prediction method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and the instruction is loaded and executed by the processor to implement the cementing quality prediction method according to any one of claims 1 to 8.

12. A computer program product, characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the cementing quality prediction method according to any one of claims 1 to 8.

Images