Two-view collaborative tsk fuzzy classification method based on residual dynamic guidance, computer device and storage medium

Abstract

The application discloses a two-viewpoint collaborative TSK fuzzy classification method based on residual dynamic guidance, which comprises the following steps: obtaining two characteristic viewpoints for the same class of samples; constructing a two-viewpoint deep stack TSK fuzzy classification model; introducing a cross-viewpoint staggered semantic consistency constraint mechanism to train two sub-models in the TSK fuzzy classification model; calculating the output results of each current layer of the two viewpoints to construct a classification residual signal; activating the distribution stability by using a fuzzy information entropy evaluation rule, and constructing a dynamic residual weight by combining the classification residual; constructing a cross-viewpoint projection operator to map the weighted residual guidance information of one viewpoint to the original characteristic space of the other viewpoint, and completing the dynamic mutual guidance and feature updating of the two viewpoints layer by layer. The application retains the inherent explainability of the zero-order TSK model, is lightweight in structure design, and effectively enhances the cross-viewpoint collaboration ability and generalization performance.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and fuzzy inference, and in particular to a two-view collaborative TSK fuzzy classification method based on residual dynamic guidance, a computer device, and a storage medium. Background Technology

[0002] The TSK (Takagi-Sugeno-Kang) fuzzy classifier is a classic model based on fuzzy reasoning. Due to its powerful classification capabilities and high interpretability, it is widely used in fields such as medical diagnosis and image recognition. This classifier divides the input space into fuzzy subsets by constructing an interpretable fuzzy rule base and membership functions, and combines consequent parameters to achieve the decision output. The model retains the simulation characteristics of fuzzy logic for human cognition while possessing high interpretability.

[0003] In recent years, to address the challenges of high-dimensional and complex data, researchers have integrated deep learning with TSK models, proposing deep TSK fuzzy classifiers. These models achieve multi-level abstraction and iterative optimization of features through a stacked generalization architecture. Each layer of the TSK classifier generates fuzzy rules based on the higher-order features of the previous layer, forming a process of "fuzzification-rule reasoning-feature reconstruction." This structure not only inherits the interpretability of traditional TSK models but also significantly improves feature representation capabilities through deep stacking.

[0004] Most existing deep TSK models are designed for single-view data and struggle to effectively utilize complementary and synergistic information between multimodal or multi-view data. For example, in lung cancer diagnosis, although multiple data formats such as CT images and pathological slides can comprehensively describe the characteristics of the same patient's condition, single-view models may fail to capture the global characteristics of the data due to incomplete information, thus limiting their potential for application in real-world scenarios.

[0005] With advancements in data acquisition technology, the same research object can be represented using data from two or more perspectives. For example, in image analysis and processing tasks, pixel intensity features and edge features can be used to simultaneously represent the same image. However, existing methods still face certain challenges when dealing with two-view problems. Methods represented by deep neural networks, while capable of completing complex feature interactions through end-to-end learning, generally suffer from high training costs and high inference latency. Furthermore, their decision-making process exhibits typical "black box" characteristics, resulting in poor interpretability and difficulty in providing human-understandable decision-making basis, which is particularly critical in high-risk fields such as medical diagnosis.

[0006] While interpretable models, exemplified by the TSK fuzzy system, have explored multi-view learning—such as decoupling shared and private features, regularized incomplete two-view clustering, multi-view regularized subspace learning, and multi-view joint nonnegative matrix factorization—they generally suffer from three shortcomings in constructing two-view collaborative mechanisms: 1. Lack of lightweight layer-by-layer cross-view transfer mechanism: Existing multi-view TSK models rarely design layer-by-layer cross-view interaction logic under a stack structure, and cannot rely on multi-layer iterative training of the model to achieve continuous alternating correction and feature optimization of dual views.

[0007] 2. Static fusion strategies are difficult to dynamically capture the potential for collaborative optimization of classification boundaries: Most methods rely on fixed-weight fusion or rule-based aggregation, lacking the ability to dynamically adjust the information interaction between perspectives based on the classification difficulty at the sample level, and cannot accurately focus on the classification boundary regions that are shared across perspectives and have not been fully learned.

[0008] 3. Complex coupling structure leads to a contradiction between model complexity and interpretability: Complex coupling mechanisms introduced to deeply explore perspective relationships (such as joint matrix factorization and latent variable inference) increase model complexity, obscure the fuzzy rule reasoning process, and undermine the interpretability advantage of the TSK model.

[0009] In summary, existing technologies struggle to achieve low-noise, lightweight, and interpretable cross-view iterative collaborative learning while maintaining the interpretability of the TSK fuzzy model. Therefore, a novel two-view fuzzy classification scheme is urgently needed to achieve dynamic alternating correction and adaptive noise reduction through mutual learning of dual-view parameters, balancing model accuracy, complexity, training stability, and interpretability. Summary of the Invention

[0010] The first objective of this invention is to provide a two-view collaborative TSK fuzzy classification method based on residual dynamic guidance to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a two-view collaborative TSK fuzzy classification method based on residual dynamic guidance, comprising the following steps: Step S1: Obtain two feature perspectives for the same type of samples, and use them as a two-perspective training dataset. Step S2: Construct a two-view depth stack TSK fuzzy classification model, which includes two parallel sub-models corresponding to two feature views respectively; use the least squares method to train the first layer of the two sub-models respectively, and fit their respective initial consequent parameters. Step S3: Perform inter-layer training on the two sub-models. During the training process, solve for the consequent parameters of each layer except the first layer. In the solution process, introduce cross-viewpoint inter-layer semantic consistency constraints to gradually enhance the semantic consistency between the two perspectives. Specifically, use the consequent parameters of the previous layer in one perspective sub-model that have converged as fixed constants to participate in the solution of the consequent parameters of the current layer in the other perspective sub-model. Decouple the bidirectional relationship between the two perspective parameters, realize mutual learning between the two perspective parameters and output closed-form solutions, and obtain the consequent parameters of the current layer in each of the two perspective sub-models respectively. Step S4: Each perspective sub-model calculates the output of the current layer based on its own consequent parameters, and then constructs a classification residual signal based on the difference between the output and the true label; constructs a sample-level fuzzy information entropy to quantify the activation stability of each sample's fuzzy rules, and constructs a dynamic residual weight by combining the classification residual signal, differentially weighting the residual information of each sample, adjusting the contribution intensity, realizing mutual guidance learning between the two perspectives layer by layer, and suppressing cross-perspective information transmission noise at the same time; Step S5: Construct an integrated cross-view projection operator that integrates dual-view semantic alignment, model parameter knowledge transfer, and cross-dimensional feature mapping functions to adapt to the differences in heterogeneous feature spaces of dual views; map the weighted residual information of one view to the original feature space of another view through the projection operator to complete the iterative update of dual-view features and realize the layer-by-layer cross-view feature collaborative optimization under the stack structure. Step S6: Repeat steps S3 to S5 until the preset stopping condition is met to obtain the output results of each perspective, and fuse the output results to obtain the final classification result.

[0012] As a specific implementation method, step S1 obtains two feature perspectives using any of the following methods: 1) Two characteristic viewpoint data were acquired using different sensors or measurement methods; 2) Two subsets obtained by semantically partitioning the single-view dataset based on domain knowledge; 3) Use feature clustering or feature selection algorithms to automatically divide the single-view dataset into two-view datasets.

[0013] As a specific implementation method, the formula for solving the consequent parameter in step S3 is as follows: (1) in, and Representing the lth layer v respectively i Viewpoint and layer l-1 v jThe consequent parameters of the perspective; λ1 represents the parameter regularization constraint coefficient, used to constrain the magnitude of the consequent parameters and suppress model overfitting; λ2 represents the inter-layer semantic constraint coefficient, used to adjust the consistency strength of inter-layer parameters between two perspectives, and the cross-perspective inter-layer semantic consistency constraint term. By constraining the parameter differences between adjacent layers, the semantic alignment effect between the two perspectives is strengthened layer by layer, suppressing the accumulation of perspective bias caused by independent training, and providing a low-noise parameter space for subsequent residual cross-perspective guidance. The closed-form solution corresponding to equation (1) is: (2) In equations (1) and (2) Let be the regular output matrix of the zero-order TSK, where each... For the nth sample in v i The hidden layer output from the perspective, i = 1 or 2, is used to reflect the sample. At the current layer, for the activation level of rule k, based on the output matrix of the aforementioned rule layer, construct the decision function corresponding to each perspective. And complete the sample error quantification calculation, as follows: (3).

[0014] As a specific implementation method, in step S4, starting from the second level iteration in any viewpoint sub-model, the following calculations are performed layer by layer: Step 41) Calculate the output residual of the current layer's view. Characterizing the sample classification fitting error: (4) Where Y represents the sample label vector; Step 42) The rule activation matrix is ​​row-normalized to unify the probability distribution scale of rule activation for each sample. : (5); Step 43) Solve the sample-level fuzzy information entropy based on equation (5). : (6) Step 44) Integrate the fuzzy information entropy of all samples to construct the information entropy vector of the current viewpoint layer. : (7); Step 45) Globally normalize the fuzzy information entropy and residuals, and construct dynamic residual weights using a negative exponential function. At the same time, it suppresses noise transmission: (8), Where μ and ν are configurable hyperparameters, used to adjust the influence of rule determinism and sample fitting accuracy on the guiding weights, respectively; all sample weights are integrated to construct a hierarchical dynamic residual weight vector. : (9).

[0015] As one specific implementation method, step S5 includes the following steps: 51) Constructing an integrated cross-view projection operator : (10) Where Z is a random projection matrix that follows a uniform distribution U(0,1) or a standard normal distribution, used to adapt to the dual-view differential feature dimensions and achieve cross-view feature space alignment. 52) Based on the constructed integrated projection operator and dynamic residual weight vector, cross-view residual guidance and feature iterative update are completed: (11) in, This represents the input of layer l+1; represents the initial sample; α is the scaling factor used to uniformly control the intensity of the guidance information injection; ⊙ represents the Hadamard operation of multiplying each sample row by row.

[0016] As a specific implementation, the elements of the random projection matrix follow a uniform distribution or a standard normal distribution to adapt to the differentiated feature dimensions of the dual-viewpoint system and complete the cross-viewpoint feature space alignment.

[0017] As a specific implementation method, in step S6, the final classification decision result is obtained by weighted fusion of the consequent parameter outputs from the two perspectives, specifically by generating the final output through the following global decision function F(X): (12) Where CS is the preset stacking iteration round of the model, and w∈(0,1) is the weighting parameter used to balance the contribution ratio of the dual-view classification output.

[0018] A second objective of this invention is to provide a computer device comprising a processor and a memory, wherein the memory is used to store instructions; and the processor is used to execute the instructions stored in the memory to implement the aforementioned two-view collaborative TSK fuzzy classification method.

[0019] A third objective of this invention is to provide a storage medium storing a computer program that, when executed by a processor, implements the aforementioned two-view collaborative TSK fuzzy classification method.

[0020] Compared with the prior art, the beneficial effects of this invention are as follows: 1) This invention innovatively uses the classification fitting residual as a cross-view guidance signal for two perspectives under a stacked structure. Existing multi-view TSK models can perform static post-processing information fusion after model training, but fail to build a perspective interaction correction path during model hierarchical iteration. The fitting residual can directly characterize the fitting defects and learning shortcomings of the current classification boundary of the model, accurately reflect the feature offset error of hard-to-classify samples, and has convenient data acquisition, simple calculation logic, and strong lightweight characteristics. Based on a multi-layer stacked training architecture, this invention relies on the residual signal to realize bidirectional alternating transmission of information and feature correction between the two perspectives during each layer of model training. In the continuous hierarchical iteration, it continuously corrects the classification boundary deviation of each perspective, iteratively optimizes feature expression, improves the fitting effect of common hard-to-classify samples and samples with overlapping features across perspectives, and effectively improves the model's adaptability to complex scenarios, classification accuracy and generalization performance. 2) This invention innovatively adopts a staggered cross-viewpoint semantic consistency constraint, which is different from the traditional same-layer bidirectional coupling constraint scheme. This invention uses the fixed viewpoint parameters of the previous layer to participate in the parameter solution of this layer, removes the mutual dependence of the two viewpoint parameters, and can directly solve the closed-form analytical solution of the consequent parameters without iterative approximation calculation. This reduces the model training complexity from a mathematical perspective and effectively improves the parameter solution efficiency and overall training stability. 3) This invention introduces a fuzzy information entropy constraint structure to address the problems of rule perturbation and layer-by-layer accumulation of fitting errors during stack-based iterative training. It uses information entropy to quantify the fuzzy rule activation state of each sample and performs differentiated weighted regulation on the sample residual information to suppress error propagation and accumulation during hierarchical iteration, thus ensuring the stability and effectiveness of the multi-layer cross-perspective collaborative learning process. 4) This invention constructs a lightweight collaborative architecture that combines hierarchical iteration with dynamic guidance. It integrates a dual optimization structure of staggered parameter constraints and entropy residual dynamic guidance, making it fully compatible with the basic inference system of zero-order TSK fuzzy models. It retains the model's native interpretability, ensuring that classification decisions are traceable and verifiable. This architecture has low computational cost and low deployment cost, overcoming the shortcomings of traditional multi-view TSK models in terms of training stability, cross-view collaborative effects, and lightweight deployment. It is suitable for classification scenarios with high reliability and high accuracy requirements, such as medical diagnosis, demonstrating good practicality and versatility. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the execution of the two-view collaborative TSK fuzzy classification method based on residual dynamic guidance described in this invention. Figure 2 This is an architecture diagram of the two-view collaborative TSK fuzzy classification method based on residual dynamic guidance described in this invention; Figure 3 This is a flowchart of the workflow in the embodiment. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention provides a two-view collaborative TSK fuzzy classification method based on residual dynamic guidance, comprising the following steps: Step S1: Obtain two feature perspectives for the same type of samples, as the two-perspective training dataset, assuming... The training dataset contains two perspectives, where , v1 and v2 represent two perspectives. , Let represent the feature vectors of the m-th sample in the first and second perspectives, respectively. For category labels, y m The true label corresponding to the sample, and Used for multi-class classification problems Used for binary classification problems.

[0024] Step S2: Using the dataset obtained in Step S1, construct and initialize a two-view depth-stacked TSK fuzzy classification model. This model contains two parallel zero-order TSK fuzzy sub-classifiers corresponding to the first and second views, respectively, serving as the basic classification unit for the overall classification method. Only the first layer of each view sub-model is trained, and the initial consequent parameters of the first layer of each view are obtained by fitting using the least squares method, thus completing the initialization of the model's first-layer parameters.

[0025] Step S3 involves iteratively training the model at each layer except the first layer, solving for the consequent parameters of the current layer of each viewpoint sub-model. Unlike traditional TSK classifiers that use a minimum learning machine to independently optimize consequent parameters, this invention introduces cross-viewpoint inter-layer semantic consistency constraints during layer-by-layer iterative training. This gradually enhances the semantic consistency between the two viewpoints, constraining the difference in consequent parameters between the current and previous layers, and forcing the model to retain shared semantic consistency between viewpoints during layer-by-layer training—that is, forcing the two viewpoint consequent parameters to... and The difference is minimized to implicitly capture consistency information between perspectives. The corresponding parameter solution is as follows: (1) in, and Representing the lth layer v respectively i Viewpoint and layer l-1 v j The consequent parameters of the perspective; λ1 represents the parameter regularization constraint coefficient, used to constrain the magnitude of the consequent parameters and suppress model overfitting; λ2 represents the inter-layer semantic constraint coefficient, used to adjust the consistency strength of inter-layer parameters between two perspectives, and the cross-perspective inter-layer semantic consistency constraint term. By constraining the parameter differences between adjacent layers, the semantic alignment effect of the two perspectives is strengthened layer by layer, suppressing the accumulation of perspective bias caused by independent training, and providing a low-noise parameter space for subsequent residual cross-perspective guidance. Unlike the traditional same-layer bidirectional coupling constraint method, this invention uses the parameters trained in the previous layer as fixed constants to participate in the solution, thereby removing the bidirectional coupling relationship of the two perspective parameters. This allows formula (1) to be directly derived into a closed analytical solution without iterative approximation calculation, effectively reducing model training overhead and improving parameter solution efficiency and training stability. The closed solution corresponding to formula (1) is as follows: (2) In equations (1) and (2) Let be the regular output matrix of the zero-order TSK, where each... For the nth sample in the l-th layer of the model, v i The hidden layer outputs from the perspectives of (i=1,2) are used to reflect the samples. The activation level of rule k in the current layer. Based on the above rule layer output matrix, decision functions corresponding to each perspective can be constructed. And complete the sample error quantification calculation, as follows: (3) Step S4: After solving for the consequent parameters of each perspective in the current layer and obtaining the model decision output, calculate the classification residual signal between the model output and the true label of the sample. To improve the reliability of cross-perspective guidance information, this invention combines the dual dimensions of rule activation determinism and sample fitting accuracy, introduces sample-level fuzzy information entropy to quantify the stability of the rule activation distribution of each perspective, and constructs a dynamic residual weight based on the joint drive of entropy and residuals to regulate the guidance weight of each sample residual in cross-perspective guidance. The first layer of the model only completes the basic decision output. Starting from the second layer iteration, the following calculations are performed layer by layer: First, calculate the output residual of the current layer's viewpoint to characterize the sample classification fitting error: (4) Row normalization is performed on the rule activation matrix to unify the probability distribution scale of rule activation for each sample. : (5) Solving the sample-level fuzzy information entropy based on normalized activation probability The smaller the entropy value, the more concentrated the activation distribution of the current perspective blur rule for that sample, and the higher the stability of the rule mapping. (6) Integrate all sample entropy values ​​to construct the information entropy vector of the current viewpoint layer. : (7) Global normalization is applied to entropy and residual values, and dynamic residual weights are constructed using a negative exponential function. This allows low-entropy, low-residual samples to receive higher cross-perspective guidance weights, while suppressing noise propagation from inferior samples. (8), Here, μ and ν are configurable hyperparameters used to adjust the influence of rule determinism and sample fitting accuracy on the guidance weights, respectively. All sample weights are integrated to construct a hierarchical dynamic residual weight vector. : (9) Step S5 involves semantically aligning and projectively fusing the original input features from any viewpoint with residual guidance information from another viewpoint to generate the input features for the next layer of the stacked deep structure. This invention integrates dual-view rule association, model prior knowledge, and cross-space mapping capabilities to construct an integrated cross-view projection operator. This operator uniformly completes dual-view semantic alignment, rule knowledge transfer, and cross-feature space mapping. Combined with guidance information selected by dynamic residual weights, it achieves accurate and stable cross-view feature iterative updates. The specific steps are as follows: 51) Constructing an integrated cross-view projection operator : (10) in, A dual-view semantic alignment matrix is ​​formed to establish the association between the two viewpoint fuzzy rule spaces and complete the rule semantic matching; Z is a random projection matrix that follows a uniform distribution U(0,1) or a standard normal distribution, which is used to adapt the dual-viewpoint differentiated feature dimensions and realize cross-viewpoint feature space alignment. 52) Based on the constructed integrated projection operator and dynamic residual weight vector, cross-view residual guidance and feature iterative update are completed: (11) in, This represents the input of layer l+1; represents the initial sample; α is the scaling factor used to uniformly control the intensity of the guidance information injection; ⊙ represents the Hadamard operation of multiplying each sample row by row.

[0026] This mechanism uses sample-level dynamic weights in formulas (8) and (9) to filter high-quality guidance information, allowing only stable and accurately fitted samples to transmit cross-perspective knowledge, suppressing noise interference caused by random initialization of fuzzy rules, and improving the stability and accuracy of dual-perspective collaborative learning.

[0027] In this scheme, the residual dynamic guidance mechanism corresponding to formula (11) and the cross-perspective inter-layer semantic consistency constraint corresponding to formula (1) form a dual complementary optimization mechanism: the inter-layer parameter constraint implicitly unifies the dual-perspective semantic distribution in the parameter space, and the residual dynamic guidance mechanism accurately transmits high-quality classification boundary information in the feature space. The two mechanisms work together without introducing redundant operations such as complex matrix factorization and latent variable iteration. The overall algorithm structure is simplified, effectively controlling the computational load of the model, reducing the algorithm complexity, and balancing the model's discriminative performance and lightweight capability.

[0028] Step S6: Repeat the iterative training from steps S3 to S5 until the preset stopping condition is met, and output the final classification decision result. The final classification decision result is obtained by weighted fusion of the consequent parameter outputs from the two perspectives, specifically through the following global decision function. Generate the final output: (12) Where CS is the preset stacking iteration number of the model, and w∈(0,1) are weighting parameters used to balance the contribution ratio of the dual-view classification output. A complete residual-guided stacking depth model can be achieved through... Figure 2 The model architecture diagram shown and Figure 1 The flowchart shown is easy to understand. Example

[0029] This embodiment uses the Robotnavigation dataset as an example to demonstrate the practical application of the proposed residual dynamic-guided two-view collaborative TSK fuzzy classification method in robot environmental perception and decision-making. This verifies the feasibility and effectiveness of the technical solution of this invention. The dataset was collected by 24 distance sensors mounted on the SCITOS G5 robot platform and can simulate complex outdoor and indoor obstacle navigation scenarios. It is used to identify the distribution of surrounding obstacles and generate corresponding robot motion commands.

[0030] First, the 24-dimensional sensor input features are divided into two perspectives, v1 and v2. Perspective v1 includes front and left-side sensor data, while perspective v2 includes right and rear-side sensor data, representing environmental information from different directions of the robot, thus achieving dual-perspective feature decomposition of the same source data.

[0031] During the model initialization phase, a zero-order TSK fuzzy classifier is constructed for each viewpoint. The number of cluster centers, c, is set to 5, and the corresponding cluster center value is... Ten fuzzy rules are generated based on Gaussian membership functions to achieve fuzzy partitioning of the input feature space. The activation level of each rule is then determined. Calculated using the Gaussian membership function, for example, the rule output for viewpoint v1 in the first layer is: ; Perspective The rule output is: ; This was followed by a multi-round iterative training process. In the first round of training, the least squares method was used to fit the consequent parameters from the two perspectives. and Obtain the initial decision output and .

[0032] Starting from the second round of iterative training, the cross-view inter-layer semantic consistency constraint of this invention is introduced to optimize the consequent parameters of each layer. Taking the parameter update of the l-th view vi as an example, the parameter solution is completed through the aforementioned formula (1). This constraint method can effectively reduce the difference in dual-view consequent parameters between adjacent layers, unify the dual-view semantic distribution layer by layer, and avoid the problem of accumulated view semantic bias caused by independent training.

[0033] Furthermore, after each layer of training is completed, the decision residual for the current viewpoint is calculated using formula (4). Taking the first round as an example, the residual vector for viewpoint v1 is: , The residual vector of viewpoint v2 is: , Based on formulas (5) and (6), the information entropy vectors of the samples corresponding to the two sets of viewpoints are calculated as follows: , . The entropy and residual are scaled and normalized, and then substituted into formula (8) to calculate the guidance weights for each sample, thus obtaining the guidance weight vectors for the two-view hierarchical samples. , as follows: , , Following formula (12), the dual-view feature semantic alignment and information fusion are completed, adaptively achieving high-quality residual information transfer across views. Finally, after a preset number of rounds of CS iteration training, the model outputs the final global decision result: , This output vector corresponds to one of the robot's four operations [Move-Forward, Slight-Right-Turn, Sharp-Right-Turn, Slight-Left-Turn]. The maximum decision output value of 0.4311 corresponds to position 2 of the vector, which is the action command that the robot should execute: "Slight Right Turn," that is, to turn slightly to the right to avoid the obstacle in front and continue moving forward.

[0034] like Figure 3 As shown, this embodiment fully demonstrates the entire process from sensor data input, perspective segmentation, fuzzy rule generation, residual calculation and dynamic injection, to final decision output. The entire process is based on explicit mathematical expressions and an interpretable fuzzy reasoning mechanism, achieving not only high classification performance but also maintaining good model transparency, making it suitable for intelligent robot control systems with high interpretability requirements.

[0035] To verify the effectiveness and superiority of this invention, detailed experimental comparisons were conducted on the Robotnavigation dataset. This dataset contains real-world environmental information collected by 24 distance sensors from the SCITOS G5 robot platform, covering a variety of complex navigation scenarios.

[0036] Five existing two-view algorithms were selected as benchmarks for comparison: IMC_GRMF, DAIMC, Tvd-TFC, IMV_TSK, and AMVMED. All algorithms were trained and tested in the same experimental environment to ensure a fair comparison.

[0037] The experimental results are shown in Table 1.

[0038] The results above demonstrate that the fuzzy classification method proposed in this invention performs excellently on both the training and test sets, especially achieving an accuracy of 82.73% ± 1.85% on the test set, significantly outperforming other comparative algorithms. This fully demonstrates the advantages of this invention in processing multi-source perceptual data and achieving interpretable decision-making.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A two-view collaborative TSK fuzzy classification method based on residual dynamic guidance, characterized in that, Includes the following steps: Step S1: Obtain two feature perspectives for the same type of samples, and use them as a two-perspective training dataset. Step S2: Construct a two-view depth stack TSK fuzzy classification model, which includes two parallel sub-models corresponding to two feature views respectively; use the least squares method to train the first layer of the two sub-models respectively, and fit their respective initial consequent parameters. Step S3: Perform inter-layer training on the two sub-models. During the training process, solve for the consequent parameters of each layer except the first layer. In the solution process, introduce cross-viewpoint inter-layer semantic consistency constraints to gradually enhance the semantic consistency between the two perspectives. Specifically, use the consequent parameters of the previous layer in one perspective sub-model that have converged as fixed constants to participate in the solution of the consequent parameters of the current layer in the other perspective sub-model. Decouple the bidirectional relationship between the two perspective parameters, realize mutual learning between the two perspective parameters and output closed-form solutions, and obtain the consequent parameters of the current layer in each of the two perspective sub-models respectively. Step S4: Each perspective sub-model calculates the output of the current layer based on its own consequent parameters, and then constructs a classification residual signal based on the difference between the output and the true label; constructs a sample-level fuzzy information entropy to quantify the activation stability of each sample's fuzzy rules, and constructs a dynamic residual weight by combining the classification residual signal, differentially weighting the residual information of each sample, adjusting the contribution intensity, realizing mutual guidance learning between the two perspectives layer by layer, and suppressing cross-perspective information transmission noise at the same time; Step S5: Construct an integrated cross-view projection operator that integrates dual-view semantic alignment, model parameter knowledge transfer, and cross-dimensional feature mapping functions to adapt to the differences in heterogeneous feature spaces of dual views; map the weighted residual information of one view to the original feature space of another view through the projection operator to complete the iterative update of dual-view features and realize the layer-by-layer cross-view feature collaborative optimization under the stack structure. Step S6: Repeat steps S3 to S5 until the preset stopping condition is met to obtain the output results of each perspective, and fuse the output results to obtain the final classification result.

2. The two-view collaborative TSK fuzzy classification method based on residual dynamic guidance according to claim 1, characterized in that, In step S1, two feature perspectives are obtained using any of the following methods: 1) Two characteristic viewpoint data were acquired using different sensors or measurement methods; 2) Two subsets obtained by semantically partitioning the single-view dataset based on domain knowledge; 3) Use feature clustering or feature selection algorithms to automatically divide the single-view dataset into two-view datasets.

3. The two-view collaborative TSK fuzzy classification method based on residual dynamic guidance according to claim 1, characterized in that, The formula for solving the consequent parameter in step S3 is as follows: （1） in, and Representing the lth layer v respectively i Viewpoint and layer l-1 v j The consequent parameters of the perspective; λ1 represents the parameter regularization constraint coefficient, used to constrain the magnitude of the consequent parameters and suppress model overfitting; λ2 represents the inter-layer semantic constraint coefficient, used to adjust the consistency strength of inter-layer parameters between two perspectives, and the cross-perspective inter-layer semantic consistency constraint term. By constraining the parameter differences between adjacent layers, the semantic alignment effect between the two perspectives is strengthened layer by layer, suppressing the accumulation of perspective bias caused by independent training, and providing a low-noise parameter space for subsequent residual cross-perspective guidance. The closed-form solution corresponding to equation (1) is: （2） In equations (1) and (2) Let be the regular output matrix of the zero-order TSK, where each... For the nth sample in v i The hidden layer output from the perspective, i = 1 or 2, is used to reflect the sample. At the current layer, for the activation level of rule k, based on the output matrix of the aforementioned rule layer, construct the decision function corresponding to each perspective. And complete the sample error quantification calculation, as follows: (3)。 4. The two-view collaborative TSK fuzzy classification method based on residual dynamic guidance according to claim 3, characterized in that, In step S4, starting from the second level iteration in any viewpoint sub-model, the following calculations are performed layer by layer: Step 41) Calculate the output residual of the current layer's view. Characterizing the sample classification fitting error: (4) Where Y represents the sample label vector; Step 42) The rule activation matrix is ​​row-normalized to unify the probability distribution scale of rule activation for each sample. : （5）； Step 43) Solve the sample-level fuzzy information entropy based on equation (5). : （6） Step 44) Integrate the fuzzy information entropy of all samples to construct the information entropy vector of the current viewpoint layer. : （7）； Step 45) Globally normalize the fuzzy information entropy and residuals, and construct dynamic residual weights using a negative exponential function. At the same time, it suppresses noise transmission: （8）； Where μ and ν are configurable hyperparameters, used to adjust the influence of rule determinism and sample fitting accuracy on the guiding weights, respectively; all sample weights are integrated to construct a hierarchical dynamic residual weight vector. : （9）。 5. The two-view collaborative TSK fuzzy classification method based on residual dynamic guidance according to claim 4, characterized in that, Step S5 includes the following steps: 51) Constructing an integrated cross-view projection operator : （10） Where Z is a random projection matrix that follows a uniform distribution U(0,1) or a standard normal distribution, used to adapt to the dual-view differential feature dimensions and achieve cross-view feature space alignment. 52) Based on the constructed integrated projection operator and dynamic residual weight vector, cross-view residual guidance and feature iterative update are completed: （11） in, This represents the input of layer l+1; represents the initial sample; α is the scaling factor used to uniformly control the intensity of the guidance information injection; ⊙ represents the Hadamard operation of multiplying each sample row by row.

6. The two-view collaborative TSK fuzzy classification method based on residual dynamic guidance according to claim 5, characterized in that, The elements of the random projection matrix follow a uniform or standard normal distribution to adapt to the dual-view differentiated feature dimensions and complete cross-view feature space alignment.

7. The two-view collaborative TSK fuzzy classification method based on residual dynamic guidance according to claim 1, characterized in that, In step S6, the final classification decision result is obtained by weighted fusion of the consequent parameter outputs from the two perspectives, specifically by generating the final output through the following global decision function F(X): (12) Where CS is the preset stacking iteration round of the model, and w∈(0,1) is the weighting parameter used to balance the contribution ratio of the dual-view classification output.

8. A computer device, characterized in that, Including processor and memory, in, The memory is used to store instructions; The processor is used to execute the instructions stored in the memory to implement the two-view collaborative TSK fuzzy classification method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the two-view collaborative TSK fuzzy classification method as described in any one of claims 1 to 6.

Images