CN115680645A - Rock mass characteristic real-time prediction method and system based on multi-source information fusion while drilling - Google Patents

Abstract

The invention belongs to the technical field of drilling, and in particular relates to a method and system for real-time prediction of rock mass characteristics based on multi-source while-drilling information fusion. The method includes the following steps: step 1, collecting four kinds of drilling information of drilling speed, torque, vibration Y and vibration Z in the drilling process in real time; step 2, inputting the drilling speed, torque, vibration Y and vibration Z into machine learning respectively model to obtain the classification prediction results of four single signal sources; step 3, based on the fusion theory of conditional probability and minimum risk decision-making level, fuse the classification prediction results of the four single signal sources obtained in step 2 to obtain the final rock mass characteristics forecast result. The present invention also provides a system for realizing the above method. The invention realizes more accurate real-time prediction of rock mass characteristics by selecting four suitable signal sources and multi-source data fusion based on decision-making level, and has good application prospects in the field of drilling.

Description

Method and system for real-time prediction of rock mass characteristics based on multi-source while-drilling information fusion

technical field

The invention belongs to the technical field of drilling, and in particular relates to a method and system for real-time prediction of rock mass characteristics based on multi-source while-drilling information fusion.

Background technique

With the demand for the development of underground space, many new engineering and technical problems have emerged. Among them, the real-time prediction of rock mass mechanical characteristics based on the characteristic signals of the drilling process is an urgent and extremely challenging problem. Real-time acquisition of the rock mass characteristics of the drilled strata based on drilling methods can effectively ensure drilling efficiency and coring quality, and can reduce the number of coring, which is of great significance to the development of the entire geological survey technology. Some scholars have carried out a series of studies on this part and its related contents.

Some scholars have studied the relationship between drilling parameters and rock mass properties from the perspective of bit cutting mechanism. Based on the balance theory of axial force and torque work and the principle of rock cutting energy conservation during drilling, Karasawa et al. derived the formulas of rock drillability strength, drilling rig torque and axial force, and described the uniaxial compressive strength and drilling Relationships between hole parameters. Song et al. introduced the rock cutting mechanism to analyze the stress on the cutting edge of the gusset drill bit, and established a mathematical model for the axial load, torque and rock and soil mechanical parameters of the drill bit. Wang Qi et al. established an indoor mechanical test model to establish a mechanical model related to the drilling parameters and rock cohesion and internal friction angle, and verified it through indoor test data.

With the development of artificial intelligence technology, some scholars have achieved a series of research results in the use of machine learning methods to carry out research on the related characteristics of rocks and rock masses. For example, inversion of rock mechanics parameters based on drilling rate, rotational speed and torque. Yue Zhongqi et al. developed a drilling process monitor (DPM) and deployed it in weathered volcanic rocks in Hong Kong, China. The results showed that the drilling speed of percussion rotary cutting drilling rigs for the same uniform continuous rock mass is fixed. Tan Zhuoying and others found through field drilling tests that monitoring parameters such as effective axial pressure, drilling tool speed, and penetration rate respond well to changes in rock strength at the boundary. AhmedGowida et al. collected rate of penetration (ROP), mud pumping rate (GPM), standpipe pressure (SPP), revolutions per minute (RPM), torque (T) and weight on bit (WOB), based on artificial neural network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS) and Support Vector Machine (SVM) established the relationship between drilling parameters and rock uniaxial compressive strength. In addition to predicting the uniaxial compressive strength and lithology of complete rocks based on drilling parameters, some scholars have also begun to pay attention to the prediction of rock mass quality using drilling parameters, mainly using real-time signals during the working process of TBM roadheaders based on machine learning. Method for inversion of rock mass quality. However, limited by the miniaturization and light weight of geological drilling equipment, there is almost no rock mass inversion research directly based on drilling parameters in the field of geological exploration.

To sum up, research from the perspective of drill bit cutting rock mass can explain the relevant mechanism and laws of drilling parameters, but it is difficult to completely and accurately express the boundary conditions of complex geological objects and the real geological background. There are big differences. The inversion of machine learning methods based on measured data often ignores the deep mining of multi-type sensor signal sources due to the limitations of a single signal source or the limited fusion and mining of multi-source signals, resulting in large gaps in prediction results. contingencies and limitations. Therefore, there is an urgent need in this field for a machine learning method that integrates multiple signal sources, so as to achieve more accurate prediction of rock mass characteristics.

Contents of the invention

The invention belongs to the technical field of drilling, and specifically relates to a method and system for real-time prediction of rock mass characteristics based on multi-source while-drilling information fusion, with the purpose of more accurately predicting rock mass characteristics based on multi-source while-drilling information.

A method for real-time prediction of rock mass characteristics based on multi-source while-drilling information fusion, comprising the following steps:

Step 1, real-time collection of drilling speed, torque, vibration Y and vibration Z four kinds of information while drilling during the drilling process;

Step 2, input the drilling speed, torque, vibration Y and vibration Z into the machine learning model respectively, and obtain the classification and prediction results of four single signal sources;

Step 3. Based on the fusion theory of conditional probability and minimum risk decision-making level, the classification prediction results of the four single signal sources obtained in step 2 are fused to obtain the final prediction result of rock mass characteristics.

Preferably, in step 2, the machine learning model is a model based on the KNN algorithm.

Preferably, the sample size of the training data used to train the machine learning model in step 2 is 250.

Preferably, the training data used for training the machine learning model in step 2 is expanded in the following manner: set a subcycle at intervals of 0.06m from the starting point.

Preferably, in the training data used to train the machine learning model in step 2, a smote interpolation algorithm is used to process unbalanced samples.

Preferably, the classification prediction result or the rock mass feature prediction result is to classify rock masses of types II, III, IV, and V in the basic quality classification of rock masses.

Preferably, in step 3, the fusion theory based on conditional probability and minimum risk decision-making level adopts one of the following three types of risk matrices:

or,

II III IV Ⅴ II 0 0.7 1.2 1.2 III 0.9 0 0.7 0.5 IV 1 0.9 0 0.4 Ⅴ 1.5 1.5 1.2 0

or,

II III IV Ⅴ II 0 1 1.2 1.2 III 0.9 0 0.7 0.5 IV 1 0.8 0 0.4 Ⅴ 1.5 1.5 1.2 0

The present invention also provides a system for realizing the above-mentioned real-time prediction method of rock mass characteristics based on multi-source information fusion while drilling, including:

The input module is used to input four kinds of drilling information of drilling speed, torque, vibration Y and vibration Z;

A calculation module, configured to calculate and obtain a final rock mass feature prediction result according to the MWD information;

The output module is used to output the final prediction result of rock mass characteristics.

The present invention also provides a computer-readable storage medium, on which a computer program is stored, and the computer program is used to implement the above-mentioned method for real-time prediction of rock mass characteristics based on multi-source information fusion while drilling.

In the present invention, the "penetrating speed" reflects the degree of difficulty of drilling. For the same set of drilling equipment, in the actual construction process, when drilling into different formations, there will be different drilling speeds. The drilling speed can be It reflects the change of stratum lithology, and can comprehensively reflect the quality characteristics of the rock mass in the underground space. The "torque" is the rotational force provided by the drilling rig for the rotary drilling of the drilling equipment. The rotational torque enables the equipment to effectively cut into the rock mass, and the torque can better reflect the softness, hardness and fragmentation degree of the rock mass in the underground space. "Three-axis vibration acceleration" compared with the above direct monitoring sensors, the amount of information reflected by the vibration sensor is more abundant, but the monitoring of information is not direct enough, the vibration signal is the vibration of the drilling equipment and the rock mass excitation process and the additional vibration of the carrying machine Formed by superposition, due to the high frequency of the vibration sensor, the extremely high sampling frequency not only ensures rich information monitoring, but also brings great difficulties to the algorithm of inversion while drilling. Among them, "vibration Y" is the vibration acceleration in the drilling direction, and "vibration X" and "vibration Z" determine the direction according to the direction of "vibration Y" according to the right-handed coordinate system.

The present invention uses the machine learning model to obtain the classification and prediction results of a single signal source based on the four types of drilling information of drilling speed, torque, vibration Y and vibration Z, and the prediction result matrix formed based on the four types of signal sources is based on the minimum risk theory of conditional probability By setting different risk levels, the classification results of heterogeneous signals can be fused from the decision-making level fusion, so as to realize the effective prediction of the quality level of several types of sample rock mass, and effectively improve the final overall classification accuracy and the classification accuracy of a small number of sample classes. Through experiments, it is found that after decision-level fusion, the identification accuracy of several types of rock masses can reach up to 97%, which is significantly improved compared with the inversion results based on a single signal source.

Apparently, according to the above content of the present invention, according to common technical knowledge and conventional means in this field, without departing from the above basic technical idea of the present invention, other various forms of modification, replacement or change can also be made.

The above-mentioned content of the present invention will be further described in detail below through specific implementation in the form of examples. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following examples. All technologies realized based on the above contents of the present invention belong to the scope of the present invention.

Description of drawings

Fig. 1 is the rock mass quality classification method;

Fig. 2 is a schematic diagram of the overlapping sampling method of drilling times;

Figure 3 is the data source correlation coefficient;

Figure 4 shows the ratio of the number of samples before and after smote interpolation;

Figure 5 shows the influence of k value on model prediction accuracy Pre;

Fig. 6 is the optimal prediction sample result based on the characteristics of each data source;

Figure 7 shows the increase in the number of samples to improve the prediction accuracy (taking pre as an example);

Fig. 8 is the conditional probability of various rock masses under the same prediction combination;

Fig. 9 is the prediction result based on conditional probability;

Figure 10 is a risk coefficient matrix;

Figure 11 is the prediction result of the sample after decision-level fusion;

Figure 12 is the prediction result based on the principle of minimum risk;

Figure 13 shows the key parameters of the prediction model for several types of rock masses under different fusion conditions.

Detailed ways

It should be noted that the algorithms for the steps of data collection, transmission, storage and processing not specifically described in the embodiments, as well as the hardware structures and circuit connections not specifically described can be realized by the disclosed content of the prior art.

Embodiment 1 Real-time Prediction Method and System of Rock Mass Characteristics Based on Multi-source Information Fusion While Drilling

The system of this embodiment includes:

The input module is used to input four kinds of drilling information of drilling speed, torque, vibration Y and vibration Z;

A calculation module, configured to calculate and obtain a final rock mass feature prediction result according to the MWD information;

The output module is used to output the final prediction result of rock mass characteristics.

Using the above system, the method for real-time prediction of rock mass characteristics includes the following steps:

Step 1, real-time collection of drilling speed, torque, vibration Y and vibration Z four kinds of information while drilling during the drilling process;

Step 2, input the drilling speed, torque, vibration Y and vibration Z into the KNN model respectively, and obtain the classification and prediction results of four single signal sources;

Step 3. Based on the fusion theory of conditional probability and minimum risk decision-making level, the classification prediction results of the four single signal sources obtained in step 2 are fused to obtain the final prediction result of rock mass characteristics.

Wherein, the classification prediction result or the rock mass characteristic prediction result is to classify rock masses of types II, III, IV, and V in the basic quality classification of rock masses.

Among them, the fusion theory of conditional probability and minimum risk decision-making level belongs to the prior art, and its specific principles are as follows:

Conditional probability refers to the probability of event ω occurring under the condition that another event X has occurred. Expressed as: P(ω|X). For rock mass quality classification prediction, P(ω|X) refers to the probability that the final decision is a certain type of rock mass under the combination of rock mass quality grades predicted by several types of sensors.

The minimum risk decision theory is a basic method in statistical pattern recognition. This method considers the possible losses caused by misclassification of different categories on the basis of conditional probability analysis of data. This decision theory will consider the probability of each feature combination relationship for each decision category, and does not need to consider the number of features or affiliation of the fusion information, only that each feature combination is independent of each other. The risk of rock mass misclassification for different projects is not exactly the same. For example, the risk of misclassifying rock mass from Class III to Class V is compared with the risk of misclassifying rock mass from Class V to Class III. The former may lead to later engineering costs The latter may directly lead to insufficient rigor in engineering design, resulting in lower engineering quality and serious engineering accidents. Therefore, several different classifications have different risk factors. According to the actual engineering survey and prediction, the most suitable expected prediction result can be selected by setting different risk factors, and then the rock mass quality classification based on the fusion of input while drilling signal sources can be realized. The decision-level fusion adopted in this embodiment is to consider the misclassification matrix and calculate the minimum risk on the basis of calculating the graded probability of each type of decision combination, so as to complete the final decision.

R(α _i |X)＝minR(α _i |x), then x∈w _k (2)

In the above formula, R(α _i |X) represents the conditional risk. In this embodiment, it is specifically expressed as the risk of determining a certain type of rock mass under a certain prediction combination, and P(w _j |x) represents the conditional probability. This implementation In the example, it is specifically expressed as the probability of determining a certain type of rock mass under a certain prediction combination. λ(α _i , w _j ) is the corresponding risk coefficient under each prediction.

In this embodiment, according to the requirements for the accuracy of the classification results of different types of rock masses, the fusion theory based on conditional probability and minimum risk decision-making level adopts one of the following three types of risk matrices:

II III IV Ⅴ II 0 0.6 0.8 1 III 0.6 0 0.6 0.8 IV 0.8 0.6 0 0.6 Ⅴ 1 0.8 0.6 0

or,

II III IV Ⅴ II 0 0.7 1.2 1.2 III 0.9 0 0.7 0.5 IV 1 0.9 0 0.4 Ⅴ 1.5 1.5 1.2 0

or,

II III IV Ⅴ II 0 1 1.2 1.2 III 0.9 0 0.7 0.5 IV 1 0.8 0 0.4 Ⅴ 1.5 1.5 1.2 0

The technical solution of the present invention will be further described below through specific experimental examples. In the following experimental examples, the method steps not specifically described can be set with reference to Example 1.

Experimental example 1 selection of input data

1. Database construction

The data used in this experiment comes from a certain project, and the stratum that is crossed is T3xj. The lithology of the stratum mainly includes siltstone, argillaceous siltstone, and thin coal seam, and the geological conditions are extremely complex. The main drilling equipment is ROCK-600 full-hydraulic portable drilling rig, which uses rope coring mode for core drilling. The effective data collection length is 80m, including 9.6m of overburden (not used as the main research data of this experimental example). Therefore, in this experimental example, a total of 70.4m of the signal characteristics of the whole process while drilling and the characteristic information of the rock mass of the whole section were collected.

In this experimental example, 4 sensors are arranged on the drilling rig, and a total of 6 types of signal sources are monitored, including WOB, torque, drilling speed, vibration-X, vibration-Y, and vibration-Z. Among them, the first three are obtained by recording and interpolating the information of the instrument panel of the drilling rig, and the vibration signal is directly measured by the triaxial vibration acceleration installed at the front end of the drilling rig.

WOB: The pressure at the position where the drill bit is in direct contact with the rock mass can directly reflect the force acting on the drill bit, is directly related to the strength of the rock mass, and is affected by the total feed pressure.

Drilling speed: reflects the difficulty of drilling. For the same set of drilling equipment, in the actual construction process, when drilling into different formations, there will be different drilling speeds. The drilling speed can reflect the change of formation lithology, At the same time, it can comprehensively reflect the quality characteristics of the rock mass in the underground space.

Torque: The rotary force provided by the drilling rig for the rotary drilling of the drilling equipment. The rotary torque enables the equipment to effectively cut into the rock mass. The torque can better reflect the softness, hardness and crushing degree of the rock mass in the underground space.

Three-axis vibration acceleration: Compared with the above direct monitoring sensors, the vibration sensor reflects more abundant information, but the monitoring of information is not direct enough. The vibration signal is the vibration of the drilling equipment and the rock mass excitation process and the additional vibration of the carrying machine. As a result, due to the high frequency of the vibration sensor, the extremely high sampling frequency not only ensures rich information monitoring, but also brings great difficulties to the algorithm of inversion while drilling. Among them, "vibration Y" is the vibration acceleration in the drilling direction, and "vibration X" and "vibration Z" determine the direction according to the direction of "vibration Y" according to the right-handed coordinate system.

2. Construction of machine learning sample library

1. Construction of rock mass labels

For the collected 70.4m drilling process at the site, based on the RMR quality classification method, according to the in-situ test and the rock mass catalogue, the refined description of the data and the rock mass quality classification of the entire research section were realized. The groundwater is not exposed, so the influence of several factors such as uniaxial compressive strength, RQD, joint condition, and joint spacing are mainly considered. Figure 2 is a schematic diagram of the method of rock mass cataloging this time.

Every 1.5m from the starting point is used as a return period, and 47 return units can be extracted from a single return period. The total amount of data has been expanded by 26 times without considering the single cycle of overlap rate. This method fully and effectively mines the drilling information. The overlapping sampling method of drilling times is shown in Fig. 3.

2. Preprocessing of monitoring signals

振动信号采样频率最高，但过多的特征点又会给反演带来很大的困难，基于传感器首先获取的是振动信号的时域信息，横坐标为时间(s),纵坐标为振动加速度(g)；基于傅里叶变换将其生成为振动频域信息，横坐标为频率(HZ)，纵坐标为振幅强度。基于以上生成的时频域信息，按照每2s分别提取一次峰值频率与峰值强度，对回次内的峰值频率值与峰值强度分别计算它的最大值、最小值、平均值、方差，最终作为该回次内的振动信号特征。几类信号的提取特征如下表所示。Regarding the feature extraction of data sources, since there are great differences in the sampling frequency and signal characteristics of each type of signal, different features are extracted according to the differences in signal characteristics of different data sources. The size of the rotary cutting force reflected by the torque. In this experiment example, the value collected by the sensor includes the specific size of each rotary force in the entire cycle, so its numerical characteristics are directly extracted. For a single research cycle, select this cycle The inner maximum value, minimum value, average value, and variance reflect the characteristics of the torque in this round; the drilling speed record data is relatively small, and it is mainly calculated indirectly by recording the time within each fixed interval distance, so the drilling speed data To characterize the selection of the drilling speed v and the change of the speed for each round

The sampling frequency of the vibration signal is the highest, but too many feature points will bring great difficulties to the inversion. Based on the sensor, the first thing to obtain is the time domain information of the vibration signal. The abscissa is time (s), and the ordinate is the vibration acceleration (g); generate it as vibration frequency domain information based on Fourier transform, the abscissa is frequency (HZ), and the ordinate is amplitude intensity. Based on the above-generated time-frequency domain information, the peak frequency and peak intensity are extracted every 2s, and the maximum value, minimum value, average value, and variance of the peak frequency value and peak intensity within each cycle are calculated, and finally used as the The characteristics of the vibration signal within a round. The extracted features of several types of signals are shown in the table below.

Based on the correlation test results of the mean value of each element and the label, the correlation test results are shown in Figure 4. The correlation between the weight on bit and the vibration acceleration in the X direction and the specific scoring value of the final rock mass quality is relatively low. The correlation of tag values is relatively high, and the correlation between signal sources is low, so drilling speed, torque, vibration Y, and vibration Z are finally selected as input signal sources to participate in machine learning prediction.

3. How to deal with unbalanced samples

The types of rock masses processed by the above methods have high imbalance, so it is necessary to rationalize the interpolation of small sample data based on effective data interpolation methods. The idea of Smote algorithm (Synthetic Minority Oversampling Technique) is to synthesize New minority samples, for each minority class sample a, select b from its nearest neighbor (according to Euclidean distance), and then select a new point from a, b as a new minority class sample.

Based on the smote interpolation algorithm, the finally obtained samples have a good performance in balance. The specific sample ratio before and after interpolation is shown in Figure 5.

3. Prediction results of a single signal source

The original sample and the sample after data enhancement are randomly divided into training set and test set according to the ratio of 8:2. There are 1222 original samples in total, and 2580 samples obtained after interpolation according to the smote data. The original sample and the enhanced data 240 and 539 test set samples were obtained respectively. For various signal sources, the extracted features are trained and predicted based on the KNN algorithm, the minimum distance is measured based on the Euclidean distance, and the accuracy of k = 3, 4, 5, 6, and 7 is tested separately. According to the experimental results It shows that the choice of K value has different adaptability to the prediction results of different signal sources, and the samples after data enhancement have better performance in prediction accuracy. It can be seen from Figure 6 that the choice of k value has a significant impact on the data results of the same sample of the same type of signal. The prediction accuracy of Type II and Type V rock masses varies greatly with the change of k value, while the prediction accuracy of Type III and Type V rock masses is relatively stable. relatively small impact.

Figure 7 reflects the optimal prediction results based on each sensing signal, and intuitively shows whether the prediction of each verification sample is accurate or not under the optimal prediction results based on the KNN algorithm for various sensing signal sources. As shown in the table below:

Among all single signal sources, there are obvious complementarities and differences in the prediction accuracy of several types of signal sources for various rock masses. in terms of overall prediction accuracy. The prediction accuracy of the torque signal for the original data is 70.3%, and the prediction accuracy after sample expansion is 71.1%. The prediction accuracy of the ROP signal for the original data is 78.9%, and the prediction accuracy after sample expansion is 73.5%. The prediction accuracy of the vibration-Y signal for the original data is 78.1%, and the prediction accuracy after sample expansion is 78.4%. The prediction accuracy of the vibration-Z signal for the original data is 65.6%, and the prediction accuracy after sample expansion is 71.2%. Overall, the prediction accuracy of several types of signal sources differs little except vibration-Z signal. Specific to the specific classification of several types of rock masses, the inversion model based on vibration signals (Y, Z) has a high classification accuracy for Type II and Type V rock masses. Taking the vibration-Y signal as an example, After the expanded samples are classified by the KNN algorithm, the PRC, REC and F1 values of the predicted Class II rock mass are 0.85, 0.69 and 0.76 respectively, and the predicted results of the Class V rock mass are 0.89, 0.74 and 0.81 respectively. . Compared with the two types of drilling parameters, torque and drilling speed, the prediction accuracy has been significantly improved. Taking the inversion results of the torque signal as an example, the predicted results PRC, REC and F1 of class Ⅱ rock mass based on the torque signal of the expanded sample after KNN algorithm classification are 0.67, 0.56 and 0.61, and the prediction results of class Ⅴ rock mass PRC , REC and F1 values were 0.66, 0.7 and 0.68, respectively. However, from the evaluation results of Type III and Type IV rock masses, the inversion results based on the vibration signal source do not show better characteristics than the other two types of signal sources. The PRC, REC and F1 values of the predicted results are 0.75, 0.793 and 0.771, respectively, while the PRC, REC and F1 values of the type IV rock mass based on the vibration-Y signal source are 0.63, 0.75 and 0.69, respectively.

From the overall training results, increasing the number of samples can improve the accuracy of the original data to a certain extent, but it is not absolute. As shown in Figure 8, when the number of training samples is less than 250, increasing the number of training samples can significantly improve the prediction accuracy. When the number of training samples exceeds 500, the prediction accuracy of some categories decreases instead. For example, after nearly doubling the classification effect of type IV rock mass, the PRC for torque, drilling speed, and vibration-Y types of information sources has shown and for the original small sample categories such as Type II rock mass and Type V rock mass, the classification accuracy is still unsatisfactory. Although the new samples generated after smote interpolation have better Good classification prediction effect, but the prediction effect of type Ⅲ and type Ⅳ rock mass is negatively inhibited. The sensitivity of several types of sensing information to the data samples is also different. For example, the prediction accuracy of the torque signal will increase first and then decrease with the increase of the sample size, while the vibration-Y signal will be more sensitive to the samples involved in the experiment. The small amount basically still maintains the trend that the accuracy increases with the increase of the sample size. It can be seen that the prediction of a single signal source is difficult to take into account the correct classification of several types of rock mass at the same time, and the method of simply enhancing the sample data can improve the classification accuracy of certain types of small samples to a certain extent, but for its accuracy The improvement is very limited, and even reduces the classification accuracy of multiple samples in the original data.

Experimental example 2 decision-level fusion results

In this experimental example, the input data selected in Experimental Example 1 is used to study the results of decision-level fusion.

Based on the conditional probability, the maximum conditional probability under each combination is used as the final decision category. According to the classification results predicted by the KNN classifier, the final results obtained by viewing the four sensors constitute a prediction matrix A. Theoretically, the prediction matrix A has 4 ×4×4×4=256 different types, the final result shows that there are 127 forecast combinations in the forecast data in this study, and the corresponding conditional probability of each forecast combination is shown in Figure 9. The decision-level fusion results of several types of rock mass based on the maximum conditional probability are shown in Fig. 10.

The symmetric risk matrix and the asymmetric risk matrix are respectively set, and the cost of different types of rock mass misclassification is set through the risk coefficient matrix, so that the minimum risk under the judgment matrix is selected as the final decision under this condition. In this experimental example, the four signal sources of drilling speed, torque, vibration-Y, and vibration-Z are independently used to predict the rock mass quality classification by the KNN algorithm, and determine the risk coefficient under different decisions. The risk coefficient matrix selected in this experimental example includes Three types, where the matrix of type a is a symmetric risk matrix (λ _ij =λ _ji ), and the type b and type c are asymmetric risk matrices (λ _ij ≠λ _ji ). Figure 11 shows the risk coefficient matrix corresponding to the misclassification of Type II, III, IV, and V rock masses. The symmetric risk matrix is set based on the assumption that the cost of misclassification of every two types of rock masses is the same, while the asymmetric risk matrix takes into account the misclassification of rock masses with higher quality grades to rock masses with lower quality grades. The time cost is less than the misclassification from lower quality rock mass to higher quality rock mass. Combined with field work experience, three different types of risk matrices are set up. The prediction results of the three different risk matrices are shown in Figure 12.

From the final prediction results, compared with the fusion method directly determined by conditional probability, the fusion method based on the minimum risk has significantly improved the prediction accuracy of several types of rock masses. As shown in Figure 13, it intuitively shows the distribution of wrongly predicted samples under centralized fusion conditions. fusion1, fusion2, fusion3, and fusion4 represent fusion directly based on conditional probability, fusion of the first type of risk matrix, and fusion of the second type of risk matrix The fusion of the third type of risk matrix.

Compared with the prediction results of a single signal source, the multi-classification prediction results after decision-level fusion have a better improvement. Among them, the decision-level fusion directly based on conditional probability can reach 0.8, 0.88, 0.96, and 0.9 for the prediction accuracy of types II, III, IV, and V rock masses respectively; , Ⅳ, and Ⅴ rock mass prediction accuracy are 0.9, 0.85, 0.98, 0.84 respectively; considering the two different risk matrices of the cost of mutual misclassification, the prediction accuracy of Ⅱ, Ⅲ, Ⅳ, Ⅴ rock mass are respectively It is 0.8, 0.9, 0.93, 0.96 and 0.9, 0.81, 0.93, 0.97. It can be seen that different risk matrices set according to different experience judgments have a greater impact on the classification and prediction accuracy of several different rock masses. Compared with the fusion method model based solely on conditional probability, the fusion method considering the minimum risk has a certain improvement in various indicators, especially in small samples and important categories of samples. According to the needs of different projects, different risk judgment matrices are set so that the prediction results can adapt to different project needs while maintaining a certain accuracy.

It can be seen from the above-mentioned embodiments and experimental examples that the present invention realizes a more accurate signal by selecting four suitable signal sources of drilling speed, torque, and vibration acceleration (Y axis, Z axis), and through multi-source data fusion based on decision-making level. The real-time prediction of rock mass characteristics has a good application prospect in the field of drilling.

Claims (9)

1. A rock mass feature real-time prediction method based on multi-source MWD information fusion, is characterized in that, comprises the steps: Step 1, real-time collection of drilling speed, torque, vibration Y and vibration Z four kinds of information while drilling during the drilling process; Step 2, input the drilling speed, torque, vibration Y and vibration Z into the machine learning model respectively, and obtain the classification and prediction results of four single signal sources; Step 3. Based on the fusion theory of conditional probability and minimum risk decision-making level, the classification prediction results of the four single signal sources obtained in step 2 are fused to obtain the final prediction result of rock mass characteristics. 2. according to the rock mass feature real-time prediction method according to claim 1, it is characterized in that: in step 2, described machine learning model is the model based on KNN algorithm. 3. according to the rock mass characteristic real-time prediction method of claim 1, it is characterized in that: the sample size of the training data that is used to train the machine learning model in step 2 is 250. 4. according to the rock mass feature real-time prediction method according to claim 1, it is characterized in that: the training data that is used for the machine learning model in the training step 2 adopts the following mode to expand: from starting point every 0.06m is that interval is set back subcycle. 5. According to the rock mass feature real-time prediction method according to claim 1, it is characterized in that: in the training data for the machine learning model in the training step 2, the smote interpolation algorithm is used to process the unbalanced samples. 6. The method for real-time prediction of rock mass characteristics according to claim 1, characterized in that: the classification prediction result or the rock mass characteristic prediction result is the classification of the basic quality of rock mass in categories II, III, IV, and V rock mass classification. 7. According to the rock mass feature real-time prediction method according to claim 6, it is characterized in that: in step 3, the described fusion theory based on conditional probability and minimum risk decision-making level adopts one of the following three types of risk matrices: II III IV Ⅴ II 0 0.6 0.8 1 III 0.6 0 0.6 0.8 IV 0.8 0.6 0 0.6 Ⅴ 1 0.8 0.6 0 or,

or, II III IV Ⅴ II 0 1 1.2 1.2 III 0.9 0 0.7 0.5 IV 1 0.8 0 0.4 Ⅴ 1.5 1.5 1.2 0 . 8. A system for realizing the rock mass feature real-time prediction method based on multi-source MWD information fusion described in any one of claims 1-7, characterized in that, comprising: The input module is used to input four kinds of drilling information of drilling speed, torque, vibration Y and vibration Z; A calculation module, configured to calculate and obtain a final rock mass feature prediction result according to the MWD information; The output module is used to output the final prediction result of rock mass characteristics. 9. A computer-readable storage medium, characterized in that: a computer program is stored thereon, and the computer program is used to realize the rock mass feature based on multi-source information fusion while drilling according to any one of claims 1-7 real-time forecasting method.

Images