Stratum multi-source information fusion interpretation model construction method based on extended belief rule base
Through the multi-source information fusion interpretation model based on the extended confidence rule base, the problem of inaccurate multi-source information fusion in deep stratigraphic exploration is solved, efficient and accurate acquisition of stratigraphic feature information is achieved, and exploration efficiency and accuracy are improved.
Patent Information
- Application Number
- CN202510373970.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-27
- Publication Date
- 2025-07-18
AI Technical Summary
The existing technology lacks methods to efficiently integrate multi-source information and accurately interpret stratigraphic feature information in deep strata exploration, resulting in limited accuracy and efficiency of exploration work.
The stratigraphic multi-source information fusion interpretation model based on the extended confidence rule base is adopted, and efficient fusion and accurate interpretation of multi-source information is achieved through multi-source information preprocessing, feature extraction, feature layer matching fusion, dynamic weight allocation and extended confidence rule base optimization.
It improves the fusion efficiency and accuracy of multi-source information, provides more comprehensive and accurate stratigraphic feature information for stratigraphic exploration, solves the differences, redundancy and uncertainty problems between multi-source information, and expands the application scope of the confidence rule base.
Smart Images
Figure CN120337125A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of deep formation exploration and information processing, and relates to a method for constructing a multi-source information fusion and interpretation model of strata. Background Art
[0002] In the field of formation exploration, comprehensively and accurately grasping formation characteristics is crucial for the successful development of exploration work. Although multi-source information such as oil pressure, drilling speed, torque, and drilling displacement collected by different sensors contains rich formation information, there are differences, redundancies, and uncertainties. Traditional belief rule bases (BRBs) have certain advantages in dealing with uncertain and fuzzy information, but when fusing multi-source formation information, they face problems such as difficult modeling due to the overly large scale of the rule base and insufficient accuracy of interpretation caused by inadequate parameter and structure optimization. Currently, there is a lack of a method that can efficiently fuse multi-source information of deep formations and accurately interpret it, which restricts the accuracy and efficiency of formation exploration work. Summary of the Invention
[0003] In order to overcome the deficiencies of the existing technology, the present invention provides a method for constructing a multi-source information fusion and interpretation model of strata based on an extended belief rule base, which improves the fusion efficiency and accuracy of multi-source information and provides more comprehensive and accurate formation characteristic information for formation exploration.
[0004] The technical solution adopted by the present invention to solve its technical problems is as follows:
[0005] A method for constructing a multi-source information fusion and interpretation model of strata based on an extended belief rule base, comprising the following steps:
[0006] Step 1, preprocessing and feature extraction of multi-source information: Preprocess and extract features from multi-source information such as oil pressure, drilling speed, torque, and drilling displacement obtained in deep formation exploration. The preprocessing includes removing obvious errors or invalid data using statistical analysis methods and removing noise using median filtering methods, and the feature extraction uses principal component analysis methods;
[0007] Step 2, matching and fusion at the feature level: Design a feature-level matching algorithm using the cosine similarity calculation method to achieve effective matching and fusion of multi-source information, and analyze redundancy by calculating the Pearson correlation coefficient and handle the uncertainty of multi-source information using a probability distribution method;
[0008] Step 3, dynamic weight allocation: Analyze the mapping relationship between drilling energy and rock mass parameters, collect data through drilling experiments in multiple different formation areas and use the least squares method for fitting to establish a mathematical model, and design a dynamic weight allocation algorithm based on this model to calculate and update the weights of each source of information during the fusion process;
[0009] Step 4: Construction and optimization of the extended belief rule base: Construct an extended belief rule base, and adopt a belief rule base optimization method based on parallel multiple populations and redundant gene strategies for parameter learning and structure learning to optimize its modeling accuracy and complexity. Among them, the expectation-maximization algorithm is used for online parameter learning in parameter learning, and the convergence condition is set as the change in the likelihood function value is less than a preset value. In structure learning, the contribution rate of premise attributes is calculated based on principal component analysis, and the attributes with a contribution rate greater than a preset percentage are selected as the dimension reduction technology for key premise attributes;
[0010] Step 5: Interpret the formation characteristics: Use the optimized extended belief rule base to interpret the fused information to obtain the deep formation characteristic information.
[0011] Furthermore, the process of step 1 is as follows:
[0012] 1.1 Data acquisition: Based on the formation exploration drill rig, install a high-precision pressure sensor to measure the drilling speed, a torque sensor to measure the torque, and a displacement sensor to measure the drilling displacement. Each sensor is integrated with the drill rig through a standard interface;
[0013] 1.2 Pretreatment: First, set the value range of the collected data according to physical meaning and historical experience to remove outliers; then use the median filtering method to remove noise according to the window size;
[0014] 1.3 Feature extraction: Use the principal component analysis method. First, standardize the preprocessed data to make the mean 0 and the variance 1, calculate the covariance matrix and perform eigen-decomposition, select the first n eigenvectors with large eigenvalues to retain the main features of the original data.
[0015] Furthermore, the process of step 2 is as follows:
[0016] 2.1 Calculate the similarity of feature vectors: Use the cosine similarity formula to calculate the similarity of feature vectors of each information source, and the similarity threshold is determined through multiple experiments;
[0017] 2.2 Matching and fusion: The feature vectors with a similarity higher than the similarity threshold are fused by the weighted average method. For example, the feature vectors of oil pressure and drilling speed are fused, and the fused feature vector C = xA + yB, where x and y are weights, to obtain the fused feature vector set.
[0018] Even further, the process of step 3 is as follows:
[0019] 3.1 Real-time monitoring of drilling parameters: During drilling, the sensors collect real-time drilling energy, rock hardness and strength, as well as oil pressure, drilling speed, torque, and drilling displacement parameters; the drilling energy E is measured by a power sensor, and the rock hardness H and strength S are obtained through acoustic logging and rock mechanics tests;
[0020] 3.2 Calculate weights: Assume there are m sources of multi-source information, and the weight calculation formula for the i-th source information is α i Determined by expert scoring combined with the analytic hierarchy process, k1, k2, and k3 are obtained by fitting from previous experiments, and g(E, H, S) represents a function related to the drilling energy E and rock mass parameters (rock mass hardness H and strength S).
[0021] 3.3 Update weights: According to the update mechanism, assume that the weights are recalculated every set drilling depth at a set distance or every set time period to reflect the formation changes and the changes in the importance of information sources.
[0022] The process of step 4 is as follows:
[0023] 4.1 Construction: Take the fused feature vector set and the corresponding weights as inputs, and construct a rule base according to the definition of the extended belief rule base and the rule generation method; generate rules according to the value range and logical relationship of the fused feature vectors. Each rule contains premise attributes, weights, and confidence information.
[0024] 4.2 Optimization: According to the optimization method based on the parallel multi-population and redundant gene strategy, first initialize multiple populations, and introduce redundant gene positions into each population; each population independently performs genetic operations and exchanges information every set number of generations; parameter learning uses the expectation-maximization algorithm for online learning, and assume that the convergence condition is that the change in the likelihood function value is less than the preset value; structure learning uses principal component analysis to calculate the contribution rate of premise attributes, and selects the attributes with a contribution rate greater than the preset percentage to optimize the structure of the rule base.
[0025] The technical concept of the present invention is as follows: The cosine similarity calculation method is used for feature matching, which can more accurately identify the feature vectors that can be fused, and improves the matching accuracy compared with the traditional simple matching method. The processing methods for redundancy and uncertainty are innovative. Through scientific quantification and screening mechanisms, the quality of the fused information is effectively improved. The established mathematical model of the mapping relationship between drilling energy and rock mass parameters is closely combined with the actual deep formation exploration. The weight calculation formula and update mechanism can adjust the weights of each source information in real time and accurately according to the drilling situation, which is an important difference from the traditional static weight allocation method. The introduced feature layer matching and fusion mechanism and dynamic weight allocation algorithm inject new vitality into the extended belief rule base, making it more adaptable to the complex requirements of multi-source information fusion and interpretation in deep formations. The optimization method based on the parallel multi-population and redundant gene strategy adopted can optimize the structure and parameters of the rule base at the same time, improve the optimization efficiency and modeling accuracy, and overcome the limitations of traditional optimization methods.
[0026] The beneficial effects of the present invention are mainly manifested in: improving the fusion efficiency and accuracy of multi-source information, providing more comprehensive and accurate formation characteristic information for formation exploration; effectively solving the problems of difference, redundancy and uncertainty among multi-source information through the feature layer matching fusion mechanism and the dynamic weight allocation algorithm; expanding the application scope of the belief rule base, making it have a broader application prospect in the field of formation exploration. Brief Description of the Drawings
[0027] Figure 1 is a flow chart of a method for constructing a formation multi-source information fusion and interpretation model based on an extended belief rule base. Specific Embodiments
[0028] The present invention will be further described below with reference to the accompanying drawings.
[0029] Referring to Figure 1 , a method for constructing a formation multi-source information fusion and interpretation model based on an extended belief rule base includes the following steps:
[0030] Step 1, preprocessing and feature extraction of multi-source information, the process is as follows:
[0031] 1.1 Data acquisition: Based on a formation exploration drill rig, install a high-precision pressure sensor (model: PT1000, measurement accuracy ±0.1 MPa) to measure the oil pressure, a speed sensor (model: LM393, measurement accuracy ±0.01 m / s) to measure the drilling speed, a torque sensor (model: TQ-200, measurement accuracy ±1 N·m) to measure the torque, and a displacement sensor (model: DW-50, measurement accuracy ±0.001 m) to measure the drilling displacement. The above-mentioned sensors are integrated with the drill rig through standard interfaces to ensure stable and accurate data transmission;
[0032] 1.2 Preprocessing: The collected data is first set with a value range according to physical meaning and historical experience to remove outliers; then the median filtering method is used, and the window size is set to 5 (which can be adjusted between 3 and 7 according to data fluctuations) to remove noise;
[0033] 1.3 Feature extraction: Using the principal component analysis method, first standardize the preprocessed data so that the mean is 0 and the variance is 1, calculate the covariance matrix and perform eigen-decomposition, select the first n eigenvectors with large eigenvalues, and retain the main features of the original data;
[0034] Step 2, feature layer matching and fusion, the process is as follows:
[0035] 2.1 Calculate the similarity of eigenvectors: Use the cosine similarity formula to calculate the similarity of eigenvectors of each information source. The similarity threshold is determined through multiple experiments. In this embodiment, it is set to 0.8 (which can be adjusted between 0.7 and 0.9 under different formation conditions);
[0036] 2.2 Matching and fusion: Feature vectors with a similarity higher than 0.8 are fused using the weighted average method. For example, the feature vectors of oil pressure and drilling speed are fused, and the fused feature vector C = xA + yB (the weights x and y can be adjusted according to the actual situation), obtaining the fused feature vector set;
[0037] Step 3: Dynamic weight assignment, the process is as follows:
[0038] 3.1 Real-time monitoring of drilling parameters: During drilling, the sensor collects in real time the drilling energy, rock hardness and strength, as well as parameters such as oil pressure, drilling speed, torque, and drilling displacement; the drilling energy is measured by a power sensor, and the rock hardness and strength are obtained through acoustic logging and rock mechanics tests;
[0039] 3.2 Calculate the weights: Calculate the weights of each information source according to the weight calculation formula Calculate the weight α i Determined by expert scoring combined with the analytic hierarchy process, k1, k2, and k3 are obtained by fitting in the previous experiment, and g(E, H, S) represents the function related to the drilling energy E and rock mass parameters (rock hardness H and strength S),
[0040] 3.3 Update the weights: According to the update mechanism, it is assumed that the weights are recalculated every 1 meter of drilling depth or every 10 minutes (which can be adjusted between 0.5 - 2 meters and 5 - 15 minutes according to the drilling situation) to reflect the formation changes and the changes in the importance of information sources;
[0041] Step 4: Construction and optimization of the extended belief rule base, the process is as follows:
[0042] 4.1 Construction: Take the fused feature vector set and the corresponding weights as inputs, and construct the rule base according to the definition and rule generation method of the extended belief rule base. For example, generate rules according to the value range and logical relationship of the fused feature vector. Each rule contains information such as premise attributes, weights, and confidence degrees.
[0043] 4.2 Optimization: According to the optimization method based on the parallel multi-population and redundant gene strategy, first initialize multiple populations (such as 3 populations, population 1 contains 20 rules, population 2 contains 30 rules, and population 3 contains 40 rules), and introduce 1 - 2 redundant gene positions into each population. Each population performs genetic operations independently and exchanges information every 5 generations (which can be adjusted between 3 - 10 generations according to the optimization effect). Parameter learning uses the expectation maximization algorithm for online learning, and the convergence condition is set as the change in the likelihood function value is less than 0.001; structure learning uses principal component analysis to calculate the contribution rate of the premise attributes, and selects the attributes with a contribution rate greater than 80% to optimize the structure of the rule base.
[0044] Step 5: Interpret the formation characteristics, the process is as follows:
[0045] Interpret and fuse the information using the optimized extended belief rule base, and judge according to the rules and confidence levels in the rule base to obtain formation feature information such as formation structure and rock type; if the premise attributes of a certain rule match the fused information and the confidence level is greater than 0.8, determine the formation rock type according to the conclusion of this rule.
[0046] The feature layer matching and fusion mechanism of this embodiment is as follows:
[0047] For multi-source information such as oil pressure, drilling speed, torque, and drilling displacement in formation exploration, preprocessing is first performed. Using a statistical analysis-based method, set a reasonable threshold range for the data to remove obvious incorrect data outside the range; adopt a median filtering algorithm to remove noise interference and ensure data quality. Then, use the principal component analysis (PCA) method for feature extraction. Taking the oil pressure, drilling speed, torque, and drilling displacement data to form a multi-dimensional data set, PCA projects it into a low-dimensional space through the eigen-decomposition of the data covariance matrix to obtain eigenvectors that are uncorrelated with each other for each source of information, and these eigenvectors retain the main information of the original data.
[0048] In terms of the feature matching algorithm, the cosine similarity calculation method is adopted. The reason for choosing this method is that it pays more attention to the consistency of vector directions and can better reflect the similarity degree of the eigenvectors of multi-source formation information in the feature space, and is not affected by the vector length. For two eigenvectors A = (a1, a2,..., a n ) and B = (b1, b2,..., b n ), their similarity Set a reasonable similarity threshold (such as 0.8), and fuse the eigenvectors with similarity higher than the threshold. For example, if the similarity between the eigenvector extracted from the oil pressure information and the eigenvector of the drilling speed information is higher than the threshold, fuse them by the weighted average method to achieve effective matching of multi-source information and improve the accuracy and reliability of the fused information.
[0049] Methods for dealing with the redundancy and uncertainty of multi-source information: For the identification of redundant information, calculate the Pearson correlation coefficient between eigenvectors. If the correlation coefficient is greater than 0.9 (which can be adjusted between 0.8 - 0.95 according to the characteristics of actual exploration data), it is considered that redundancy exists. During fusion, compare the variances of the redundant eigenvectors and retain the eigenvector with the largest variance because it contains more abundant information, thereby eliminating redundancy. For uncertainty quantification, based on historical data and expert experience, set a probability distribution range for the value of each eigenvector. For example, according to past exploration experience in a certain type of formation and combined with expert judgment, determine the probability of a certain drilling displacement eigenvector taking values in a specific interval, and comprehensively process this uncertainty during fusion.
[0050] The dynamic weight allocation algorithm for coupling the drilling energy - rock mass parameter mapping relationship in this embodiment is as follows:
[0051] Weight calculation formula: Deeply analyze the mapping relationship between the drilling energy and rock mass parameters (such as hardness, strength, etc.) during the drilling process. By conducting drilling experiments in multiple different formation areas, collect data samples of drilling energy E, rock mass hardness H, strength S, as well as oil pressure, drilling speed, torque, drilling displacement, etc. Use the least squares method to fit and establish a mathematical model E = f(H, S). Through fitting the experimental data, we get E = k1H + k2S + k3 (k1, k2, k3 are fitting coefficients).
[0052] Based on this mapping relationship, design a dynamic weight allocation algorithm. Suppose there are m sources of multi-source information, and the weight w of the i-th source information (such as information sources like oil pressure, drilling speed, etc.) i The calculation formula is where α i is a coefficient related to the characteristics of the i-th source information itself. Its value range is determined by expert scoring combined with the analytic hierarchy process (AHP) according to factors such as the importance and reliability of the information source, and is generally between 0.1 - 0.9. g(E, H, S) is a function related to the drilling energy and rock mass parameters During the drilling process, as the drilling energy E and rock mass parameters H, S change in real time, and information such as oil pressure, drilling speed, torque, and drilling displacement changes, recalculate the weights every 1 meter of drilling depth or every 10 minutes (which can be adjusted between 0.5 - 2 meters and 5 - 15 minutes according to the actual drilling situation and the formation change speed) to ensure that the weights accurately reflect the importance of each source information under different conditions.
[0053] The construction and optimization method of the extended belief rule base in this embodiment is as follows:
[0054] On the basis of the traditional belief rule base, introduce a feature layer matching and fusion mechanism and a dynamic weight allocation algorithm to construct an extended belief rule base. Take the multi-source information after feature layer matching and fusion, including the fused feature vectors such as oil pressure and drilling speed, and the corresponding dynamic weights as the input of the extended belief rule base, and determine the weights of each input information in the rule base.
[0055] Optimization is carried out using a belief rule base optimization method based on a parallel multi-population and redundant gene strategy. The parallel multi-population strategy means initializing multiple populations simultaneously, with different numbers and structures of rules in each population. For example, 3 populations are initialized, where population 1 has 20 rules, population 2 has 30 rules, and population 3 has 40 rules. Each population independently performs genetic operations (selection, crossover, mutation), exchanges information regularly, and shares excellent rules and parameters. The redundant gene strategy is to introduce redundant genes when encoding the rule base. For example, 1 - 2 redundant attribute bits are added to the encoding of the rule premise attributes. These do not participate in rule judgment initially but change during genetic operations, providing more possibilities for the evolution of the rule base and ensuring the smooth optimization operation of the extended belief rule base with different numbers of rules. By generating the optimal solutions of the extended belief rule bases with different numbers of rules, the Pareto front is obtained, and decision-makers can screen the optimal solutions on the Pareto front according to their own preferences and actual needs.
[0056] For the parameter learning of the extended belief rule base, the expectation-maximization (EM) algorithm is used for online parameter learning. The specific steps are as follows: First, initialize the rule base parameters (such as rule confidence, weight, etc.); according to the current input data, calculate the likelihood function value of the data under the current parameters; update the parameters through iterative calculation to increase the likelihood function value until the convergence condition is met (such as the change in the likelihood function value is less than 0.001), providing an online modeling method for formation exploration decisions with high timeliness requirements.
[0057] In terms of structure learning, key premise attributes and their reference values are identified and screened. Dimension reduction techniques based on principal component analysis and others are used. For example, the contribution rate of each premise attribute is calculated using PCA, and attributes with a contribution rate greater than 80% are selected as key premise attributes to optimize the structure of the extended belief rule base, reduce the scale, and improve the calculation efficiency.
[0058] Refer to Figure 1 For formation exploration, the drilling rig is equipped with a pressure sensor (measuring oil pressure), a speed sensor (measuring drilling speed), a torque sensor, and a displacement sensor to collect multi-source information such as oil pressure, drilling speed, torque, and drilling displacement. The collected data first enters the preprocessing module, where statistical analysis is used to remove outliers and median filtering is used to remove noise. The preprocessed data enters the feature extraction module, and feature vectors are extracted through principal component analysis. The feature vectors enter the feature-level matching and fusion module, where the cosine similarity formula is used to calculate the similarity, and matching and fusion are performed according to the threshold to obtain a set of fused feature vectors. The set of fused feature vectors and real-time drilling parameters (drilling energy, rock mass hardness and strength, etc.) enter the dynamic weight allocation module to calculate and update the weights. Finally, the set of fused feature vectors and weights are input into the extended belief rule base construction and optimization module. After construction, parameter learning, and structure learning optimization, the optimized rule base is used to interpret the formation feature information, and results such as formation structure and rock type are obtained.
[0059] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept and is only for illustrative purposes. The protection scope of the present invention should not be regarded as limited to the specific forms stated in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be conceived by those of ordinary skill in the art based on the inventive concept of the present invention.
Claims
1. A method for constructing a formation multi-source information fusion interpretation model based on an extended belief rule base, characterized in that, It includes the following steps: Step 1, Multi-source information preprocessing and feature extraction: Preprocess and extract features from multi-source information such as oil pressure, drilling speed, torque, and drilling displacement obtained in deep formation exploration. The preprocessing includes using statistical analysis methods to remove obvious errors or invalid data, and using median filtering to remove noise. The feature extraction uses the principal component analysis method; Step 2, Feature layer matching and fusion: Design a feature layer matching algorithm using the cosine similarity calculation method to achieve effective matching and fusion of multi-source information, and analyze redundancy by calculating the Pearson correlation coefficient and handle the uncertainty of multi-source information using a probability distribution method; Step 3, Dynamic weight allocation: Analyze the mapping relationship between drilling energy and rock mass parameters, collect data through drilling experiments in multiple different formation areas and use the least squares method for fitting to establish a mathematical model. Based on this model, design a dynamic weight allocation algorithm to calculate and update the weights of each source information during the fusion process; Step 4, Construction and optimization of the extended belief rule base: Construct an extended belief rule base, and use a belief rule base optimization method based on parallel multiple populations and redundant gene strategies for parameter learning and structure learning to optimize its modeling accuracy and complexity. Among them, parameter learning uses the expectation maximization algorithm for online parameter learning, and the convergence condition is set as the change in the likelihood function value is less than a preset value. Structure learning uses the principal component analysis to calculate the contribution rate of premise attributes, and filters out the attributes with a contribution rate greater than the preset percentage as the dimension reduction technology of key premise attributes; Step 5, Interpret formation features: Use the optimized extended belief rule base to interpret the fusion information to obtain deep formation feature information.
2. The method for constructing a formation multi-source information fusion interpretation model based on an extended belief rule base according to claim 1, characterized in that The process of Step 1 is as follows: 1.1 Data acquisition: Based on the formation exploration drill rig, install a high-precision pressure sensor to measure the drilling speed, a torque sensor to measure the torque, and a displacement sensor to measure the drilling displacement. Each sensor is integrated with the drill rig through a standard interface; 1.2 Preprocessing: First, set the value range of the collected data according to physical meaning and historical experience to remove outliers; then use the median filtering method to remove noise according to the window size; 1.3 Feature extraction: Use the principal component analysis method. First, standardize the preprocessed data so that the mean is 0 and the variance is 1, calculate the covariance matrix and perform eigen decomposition, select the first n eigenvectors with large eigenvalues to retain the main features of the original data.
3. The method for constructing a formation multi-source information fusion interpretation model based on an extended belief rule base according to claim 1 or 2, characterized in that, The process of Step 2 is as follows: 2.1 Calculate the similarity of feature vectors: Use the cosine similarity formula to calculate the similarity of feature vectors of each information source, and the similarity threshold is determined through multiple experiments; 2.2 Matching and fusion: Feature vectors with a similarity higher than the similarity threshold are fused using the weighted average method. For example, the feature vectors of oil pressure and drilling speed are fused, and the fused feature vector C = xA + yB, where x and y are weights, to obtain a set of fused feature vectors.
4. The method for constructing a formation multi-source information fusion interpretation model based on an extended belief rule base according to claim 1 or 2, characterized in that, The process of Step 3 is as follows: 3.1 Real-time monitoring of drilling parameters: During drilling, the sensors collect real-time drilling energy, rock hardness and strength, as well as oil pressure, drilling speed, torque, and drilling displacement parameters. The drilling energy E is measured by a power sensor, and the rock hardness H and strength S are obtained through acoustic logging and rock mechanics tests; 3.2 Calculate the weights: Assume that there are m sources of multi-source information, and the weight calculation formula for the i-th source information α i Determined by expert scoring combined with the analytic hierarchy process, k1, k2, and k3 are obtained by fitting with previous experiments, and g(E, H, S) represents a function related to the drilling energy E, rock mass hardness H, and strength S. 3.3 Update weights: According to the update mechanism, the weights are recalculated every set distance of drilling depth or every set period to reflect the formation changes and the changes in the importance of information sources.
5. The method for constructing a formation multi-source information fusion interpretation model based on an extended belief rule base according to claim 1 or 2, characterized in that, The process of step 4 is as follows: 4.1 Construction: Using the fused feature vector set and the corresponding weights as inputs, construct a rule base according to the definition of the extended belief rule base and the rule generation method; generate rules according to the value range and logical relationship of the fused feature vectors, and each rule contains premise attributes, weights, and confidence information; 4.2 Optimization: According to the optimization method based on parallel multiple populations and redundant gene strategies, first initialize multiple populations, and introduce redundant gene positions into each population; each population performs genetic operations independently and exchanges information every set number of generations; Parameter learning is performed online using the expectation-maximization algorithm, and the convergence condition is set that the change in the likelihood function value is less than the preset value; Structure learning calculates the contribution rate of premise attributes using principal component analysis, and filters out the attributes with a contribution rate greater than the preset percentage to optimize the structure of the rule base.