Intelligent system integrating structure perception fuzzy division and modular neural regression modeling

Abstract

The invention provides an intelligent system fusing structure perception fuzzy division and modular neural regression modeling, belongs to the field of learning model construction, and constructs an intelligent learning system which is high in precision, explainable and efficient in calculation by fusing structure perception fuzzy division and hierarchical neural regression modeling. Wherein the structure sensing mechanism dynamically adapts to data local density distribution, and the naturalness and boundary adaptability of rule division are remarkably improved; the sparse superficial neural sub-network is used as a consequent model, and the interpretability is maintained through parameter constraint while the nonlinear mapping capability is reserved; the unified fusion framework supports heterogeneous model extension and dynamic weight optimization, so that the system still keeps low reasoning overhead in a complex scene. Finally, collaborative optimization of interpretability, precision and calculation efficiency is realized, and a deployable solution is provided for industrial control, real-time decision and other high-dimensional nonlinear tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of learning model construction, in particular to an intelligent system fusing structure-aware fuzzy partitioning and modular neural regression modeling. BACKGROUND

[0002] Fuzzy systems have been widely used in industrial control, predictive analysis and decision support for a long time due to their natural advantages in handling uncertainty and knowledge reasoning. Traditional fuzzy modeling methods (such as Takagi-Sugeno model or ANFIS framework) usually rely on fixed spatial partitioning algorithms to generate fuzzy rules, and the partitioning process does not consider the structural distribution characteristics of the samples, resulting in rigid rule boundaries. At the same time, the consequent of such models often uses linear expressions, which is difficult to fit complex nonlinear patterns, and there is a significant contradiction between interpretability, accuracy and computational efficiency.

[0003] In recent years, researchers have tried to introduce neural networks as the consequent of fuzzy rules to enhance the expression ability, such as embedding random feedforward networks or multilayer perceptrons into fuzzy systems. However, the existing fusion schemes still have three major defects: first, the fuzzy rule partitioning does not combine the local density characteristics of the data, resulting in rule redundancy or fuzzy boundaries; second, the complexity of the neural network consequent is uncontrollable, leading to a sharp increase in reasoning overhead and degradation of interpretability; third, the model structure and training strategy lack a unified framework, making it difficult to support flexible expansion of heterogeneous modules. Especially in high-dimensional input scenarios, existing methods are difficult to be practical due to the exponential growth of computational complexity, for example, patent CN118246483A proposes a granularity criterion to optimize sampling, but does not solve the core problems of rule adaptive generation and sparse modeling.

[0004] Therefore, it is urgent to develop a fuzzy-neural fusion framework that has structural adaptability, computational efficiency and strong interpretability, to balance the contradiction between complex nonlinear relationship modeling and system deployability. SUMMARY

[0005] The purpose of the present application is to provide an intelligent system fusing structure-aware fuzzy partitioning and modular neural regression modeling to solve the problems existing in the prior art.

[0006] To achieve the above purpose, the present application provides the following solutions: The present application provides an intelligent system fusing structure-aware fuzzy partitioning and modular neural regression modeling, comprising: A hybrid fuzzy partitioning module is used to generate an initial fuzzy rule framework through equidistant grid partitioning, and to refine and correct the rules, and a Gaussian membership function is used to construct a fuzzy excitation matrix. A hierarchical fuzzy neural network modeling module is used to correspond each fuzzy rule to a local shallow neural subnetwork as a consequent model, which contains randomly initialized hidden layer weights, a ReLU activation function and a linear output layer, and the network output weights are obtained through the least square closed-form solution of membership weighting; A fuzzy output fusion module is used to generate a system output through weighted fusion based on the excitation strength of each subnetwork.

[0007] Preferably, the refining and correcting of the rules comprises: A structure density guided center adjustment strategy is introduced to refine and correct the initial rule center according to the ordering structure or distribution characteristics of the samples; Or, a boundary adjustment based on kernel density estimation is adopted.

[0008] Preferably, the consequent model can be replaced by a long short-term memory network model, a convolutional neural network model or an attention mechanism model.

[0009] Preferably, the weighted fusion can be replaced by an attention mechanism, a maximum excitation strategy or a fuzzy integral method to replace dynamic weight updating.

[0010] Preferably, the equal-interval grid division can be replaced by GMM clustering, spectral clustering or information entropy driven division.

[0011] Preferably, the consequent model can be expanded into an SVR, a random forest or a multi-layer perceptron heterogeneous model, and is deployed in parallel or cascade through a unified fusion module.

[0012] Preferably, the fuzzy output fusion module is also used for a dynamic weight updating mechanism based on posterior probability or an attention mechanism to enhance decision consistency.

[0013] The application also provides an intelligent modeling method of fusing structure perception fuzzy division and shallow neural regressor, comprising the following steps: S1. An initial fuzzy rule framework is generated through equal-interval grid division, and the rules are refined and corrected, and a fuzzy excitation matrix is constructed using a Gaussian membership function; S2. Each fuzzy rule corresponds to a local shallow neural subnetwork as a consequent model, which contains randomly initialized hidden layer weights, a ReLU activation function and a linear output layer, and the network output weights are obtained through the least square closed-form solution of membership weighting; S3. A system output is generated through weighted fusion based on the excitation strength of each subnetwork.

[0014] The application has the following beneficial technical effects relative to the prior art: The application provides an intelligent system of fusion structure perception fuzzy division and modular neural regression modeling, which constructs a high-precision, interpretable and computationally efficient intelligent learning system by fusing structure perception fuzzy division and hierarchical neural regression modeling. The structure perception mechanism dynamically adapts to local density distribution of data, significantly improving the naturalness and boundary adaptability of rule division; the sparse shallow neural subnetwork as the consequent model maintains interpretability while retaining nonlinear mapping capability through parameter constraint; the unified fusion framework supports heterogeneous model expansion and dynamic weight optimization, so that the system still maintains low reasoning overhead in complex scenarios. Finally, the interpretability, precision and computational efficiency are synergistically optimized, providing a deployable solution for high-dimensional nonlinear tasks such as industrial control and real-time decision. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 The principle diagram of the intelligent system of fusion structure perception fuzzy division and modular neural regression modeling provided by the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] The purpose of the present application is to provide an intelligent system of fusion structure perception fuzzy division and modular neural regression modeling to solve the problems in the prior art.

[0019] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0020] Embodiment 1 The present embodiment provides an intelligent system of fusion structure perception fuzzy division and modular neural regression modeling, as shown in Figure 1 The intelligent system of fusion structure perception fuzzy division and modular neural regression modeling comprises: The mixed fuzzy partition module is used for rough partitioning the input space by equal-interval grid partitioning to form an initial fuzzy rule framework, introducing a structure density guided center adjustment strategy, refining and correcting the initial rule center according to the ordering structure or distribution characteristics of the samples, to enhance the perceptivity and adaptability of the region partitioning, and using a Gaussian membership function to construct a fuzzy excitation matrix, to ensure the flexibility of the rule boundary and the locality of the expression. Specifically, first, the input data is normalized; For each dimension of the input feature , the minimum-maximum standardization is used: ; Secondly, each dimension is divided into equal-width fuzzy regions, and the initial center is: ; Further, a structure density guided mechanism is introduced, and the density representative point is obtained based on the sample ordering: ; wherein is an adjustment factor, which can be determined adaptively according to the data distribution; Finally, a Gaussian fuzzy membership function is constructed, and the membership degree of each sample to the th rule is calculated: ;

[0021] And the excitation matrix is obtained by normalization:

[0022] ;

[0023] This module realizes the adaptive response capability of the fuzzy region to the input space structure, while retaining the explainability of the fuzzy rule boundary; The hierarchical fuzzy neural network modeling module is used for corresponding each fuzzy rule to a local shallow neural subnetwork as the consequent model, which includes randomly initialized hidden layer weights, ReLU activation function and linear output layer, and the network output weight is obtained by the least square closed-form solution of the membership degree weighting, supporting different rules using different size or structure of the subnetwork to realize structure adaptation; Specifically, first, the local subnetwork structure is defined; Each fuzzy rule corresponds to a shallow neural network submodel: The input layer dimension is ; The number of hidden layer nodes is , which is determined by experience or a validation set. The activation function adopts ; The output layer is a linear unit; The hidden layer is represented as: ; Secondly, the enhanced feature matrix is constructed, and the formula is: ; Finally, the least square learning is carried out based on the fuzzy excitation weighted, and the formula is: ; wherein, is the sample weight matrix, is the regularization coefficient (to avoid ill-conditioned problems); The output function form is: ; This module uses the distributed guidance weighting mechanism to make each sub-network focus on modeling the high response area, improving the regional relevance of nonlinear modeling; The fuzzy output fusion module is used to generate the system output by weighted fusion based on the excitation strength of each sub-network. Firstly, local prediction generation is carried out, and all samples and rules : ; Secondly, the fuzzy weighted fusion is used to generate the system output, and the formula is: ; Further, the posterior probability is used to replace ; the attention weight learning mechanism is introduced to improve the fusion flexibility; the fuzzy integral is used to fuse the responses of different sub-models to enhance the decision stability.

[0024] As an embodiment, the rules can also be refined and corrected by using the boundary adjustment based on kernel density estimation.

[0025] As an embodiment, the consequent model can be replaced by a long short-term memory network model, a convolutional neural network model or an attention mechanism model.

[0026] As an embodiment, the weighted fusion can be replaced by an attention mechanism, a maximum excitation strategy or a fuzzy integral method to replace the dynamic weight update.

[0027] As an embodiment, the equidistant grid division can be replaced by GMM clustering, spectral clustering or information entropy driven division.

[0028] As an implementation, the consequent model can be extended to an SVR, random forest or multi-layer perceptron heterogeneous model, build a rule diversity enhancement mechanism, and deploy in parallel or cascade through a unified fusion module, allowing to impose a rule structure regularization term, path sparsity constraint on each structure.

[0029] As an implementation, the fuzzy output fusion module is also used for dynamic weight updating mechanism or attention mechanism based on posterior probability to enhance decision consistency.

[0030] The intelligent system provided by the embodiment has the following advantages: 1. The modeling structure is more modular and scalable, facilitating system deployment and upgrading; 2. The fuzzy partition has self-adaptive structure recognition capability, and the modeling boundary is more natural; 3. The consequent model achieves a good balance between nonlinear modeling capability and interpretability; 4. The fusion strategy is flexible, supporting prediction, reasoning and control tasks in different complexity scenarios.

[0031] Embodiment 2 The embodiment provides an intelligent modeling method based on the fusion of structure-aware fuzzy partition and shallow neural regressor of the above system, comprising the following steps: S1. Generate an initial fuzzy rule framework by equal-interval grid partition, and refine and correct the rules, and use a Gaussian membership function to build a fuzzy excitation matrix; S2. Each fuzzy rule corresponds to a local shallow neural subnetwork as a consequent model, which includes randomly initialized hidden layer weights, ReLU activation function and linear output layer, and the network output weight is obtained by least squares closed-form solution of membership weight; S3. Weighted fusion based on the excitation strength of each subnetwork to generate system output.

[0032] The principles and implementation modes of the present application are described by specific examples, and the above examples are only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. An intelligent system that fuses structure-aware fuzzy partitioning with modular neural regression modeling, characterized by: The method comprises the following steps: The mixed fuzzy division module is used for generating an initial fuzzy rule framework through equal-interval grid division, refining and correcting the rules, and constructing a fuzzy incentive matrix by using a Gaussian membership function; The hierarchical fuzzy neural network modeling module is used for corresponding each fuzzy rule to a local shallow neural subnetwork as a consequent model, which contains randomly initialized hidden layer weights, a ReLU activation function, and a linear output layer, and the network output weight is obtained through a least square closed-form solution weighted by the membership degree; The fuzzy output fusion module is used for generating a system output through weighted fusion based on the incentive intensity of each subnetwork.

2. The intelligent system of claim 1, wherein the system is characterized by: The refining and correcting of the rules comprises: A center adjustment strategy guided by structural density is introduced, and the initial rule center is refined and corrected according to the ordering structure or distribution characteristics of the samples; Or, a boundary adjustment based on kernel density estimation is adopted. 3.The intelligent system of fusing structure-aware fuzzy partitioning and modular neural regression modeling of claim 1, wherein: The consequent model can be replaced by a long short-term memory network model, a convolutional neural network model or an attention mechanism model.

4. The intelligent system of claim 1, wherein the system is configured to perform the steps of: receiving a plurality of data sets; and determining a plurality of feature sets from the plurality of data sets. The weighted fusion can be replaced by an attention mechanism, a maximum incentive strategy or a fuzzy integral method to replace dynamic weight updating.

5. The intelligent system of claim 1, wherein the system is characterized by: The equal-interval grid division can be replaced by GMM clustering, spectral clustering or information entropy driven division.

6. The intelligent system of claim 1, wherein the system is configured to perform the steps of: receiving a plurality of data sets; and determining a plurality of feature sets from the plurality of data sets. The consequent model can be expanded into an SVR, a random forest or a multi-layer perceptron heterogeneous model, and is deployed in parallel or cascade through a unified fusion module.

7. The intelligent system of claim 1, wherein the system is configured to perform the steps of: receiving a plurality of data sets; and determining a plurality of feature sets from the plurality of data sets. The fuzzy output fusion module is also used for a dynamic weight updating mechanism based on posterior probability or an attention mechanism to enhance decision consistency.

8. An intelligent modeling method that fuses structure-aware fuzzy partitioning with shallow neural regressors, characterized by: The method comprises the following steps: S1. An initial fuzzy rule framework is generated through equal-interval grid division, the rules are refined and corrected, and a fuzzy incentive matrix is constructed by using a Gaussian membership function; S2. Each fuzzy rule is corresponding to a local shallow neural subnetwork as a consequent model, which contains randomly initialized hidden layer weights, a ReLU activation function and a linear output layer, and the network output weight is obtained through a least square closed-form solution weighted by the membership degree; S3. A system output is generated through weighted fusion based on the incentive intensity of each subnetwork.

Images