Multi-source information decision fusion method and system of multi-granularity multi-view rough set

Abstract

The invention relates to the technical field of multi-source information processing, in particular to a multi-source information decision fusion method and system for a multi-granularity multi-view rough set, and the method comprises the following steps: obtaining multi-source heterogeneous data of an object to be decided, constructing a multi-source decision information table through cleaning and standardization processing, and dividing the information table into a plurality of view subsets; and constructing a multi-granularity hierarchical structure based on the view subset, calculating rough set approximation precision and attribute dependency under different granularity hierarchies, and generating a local decision rule set. And performing weighted aggregation on the local decision rule set by using a multi-source information fusion algorithm, and constructing decision evaluation parameters based on consistency measurement. According to the method, the problems of one-sided information under a single view angle and detail loss under a single granularity are effectively solved through the calculation of the multi-view and multi-granularity collaborative mechanism reverse driving decision cost, and the accuracy and robustness of decision analysis under a complex and uncertain environment are remarkably improved based on the feedback adjustment of the dynamic weight; and deep mining and efficient fusion of multi-source information are realized.

Description

Technical Field

[0001] This invention relates to the field of multi-source information processing technology, and in particular to a multi-source information decision fusion method and system for multi-granularity, multi-perspective rough sets. Background Technology

[0002] Multi-source information processing refers to the technical field of collecting, processing, and comprehensively analyzing heterogeneous data from different sensors, databases, or networks using computer technology. Its aim is to eliminate redundancy and contradictions among data and extract high-value decision-making information. With the development of the Internet of Things (IoT) and big data technologies, this field has been widely applied in scenarios such as medical diagnosis, financial risk control, and industrial fault detection, becoming a core means of solving complex system decision-making problems.

[0003] Traditional rough set decision-making methods refer to using a single granularity partitioning standard or a single data observation perspective to approximate the classification and rule extraction of incomplete or uncertain information, mainly relying on static attribute reduction algorithms for data dimensionality reduction.

[0004] However, existing traditional rough set methods have significant limitations when dealing with highly complex, multi-source, heterogeneous data. Because they only examine data from a single perspective or resolution, they often miss key local features and fail to effectively capture the deep structural relationships within the data. In situations with high data noise or non-linear attribute relationships, this one-dimensional approach is prone to generating biased decision rules, resulting in low classification accuracy and insufficient robustness, making it difficult to meet the practical needs of high-precision decision-making scenarios. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a multi-source information decision fusion method and system with multi-granularity and multi-perspective rough sets.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-source information decision fusion method using multi-granularity, multi-perspective rough sets, comprising the following steps: S1: Obtain the original multi-source data set of the target object, perform data cleaning and normalization processing on the original multi-source data set, construct a multi-source information decision table containing conditional attributes and decision attributes, and divide the multi-source information decision table into multiple independent view attribute subsets according to the data source characteristics; S2: For each subset of view attributes, a multi-level, multi-granularity spatial structure is constructed by setting different granularity criteria, and a binary relation matrix between objects is calculated at each granularity level to generate a multi-view, multi-granularity knowledge base; S3: Based on the multi-view, multi-granularity knowledge base, use rough set theory to calculate the lower approximation set and the upper approximation set at each granularity level, calculate the attribute importance index based on the lower approximation set and the upper approximation set, and then extract the local decision feature vector of each view at different granularities. S4: Construct a weighted fusion model based on attribute importance, input the local decision feature vectors of all views and granularities into the weighted fusion model for consistent aggregation calculation, generate a global decision score, and determine the final classification result of the target object based on the global decision score.

[0007] As a further aspect of the present invention, the process of constructing the multi-source information decision table specifically includes: detecting missing values ​​in the original multi-source data set and filling the missing values ​​with the mean of similar samples; The maximum-minimum normalization method is used to map numerical data to a standard interval. The conditional attributes and decision attributes are aligned column-wise to form a two-dimensional data matrix that serves as the multi-source information decision table.

[0008] As a further aspect of the present invention, the process of constructing a multi-level, multi-granularity spatial structure specifically includes: setting a set of incremental neighborhood radius threshold sequences; For each subset of the view attributes, the neighborhood similarity between samples is determined by each threshold in the neighborhood radius threshold sequence. Samples that meet the similarity conditions are grouped into the same information granularity, thereby forming multiple granularity levels from fine to coarse.

[0009] As a further aspect of the present invention, the process of calculating the attribute importance index specifically includes: calculating the dependency value of the decision attribute relative to the current view attribute subset at a specific granularity level; Calculate the change in dependency after removing a certain attribute from the current subset of view attributes; The change in dependency is used as the attribute importance value at the current granularity level.

[0010] As a further aspect of the present invention, the process of extracting local decision feature vectors specifically includes: selecting core attributes whose attribute importance values ​​are higher than a preset importance threshold; Based on the core attributes, a reduced local decision rule is constructed; Substitute the object to be decided into the local decision rule, output the local predicted classification label of the object in the current view and at the current granularity, and use the local predicted classification label as an element of the local decision feature vector.

[0011] As a further aspect of the present invention, the process of constructing a weighted fusion model based on attribute importance specifically includes: using the information entropy algorithm to calculate the entropy weight values ​​at different granularity levels under each view; A fusion weight coefficient is assigned to each local decision feature vector based on the entropy weight value; A weighted voting matrix is ​​constructed as the core of the weighted fusion model.

[0012] As a further aspect of the present invention, the process of generating a global decision score specifically includes: converting the local predicted classification labels of each view at each granularity into corresponding one-hot encoded vectors; The one-hot encoded vector is weighted and summed with the corresponding fusion weight coefficients to obtain the fusion probability vector. The maximum value in the fusion probability vector is extracted as the global decision score.

[0013] As a further aspect of the present invention, the method further includes a multi-granularity adaptive adjustment step: when the global decision score is lower than a preset confidence threshold, the value of the neighborhood radius threshold sequence is automatically reduced, and steps S2 to S4 are re-executed until the global decision score meets the requirements or reaches the upper limit of the number of iterations.

[0014] As a further aspect of the present invention, the method further includes a conflict detection step: calculating the conflict rate between the local predicted classification labels under different views; When the conflict rate exceeds the safety threshold, the subset of view attributes with the highest conflict rate is removed, and the aggregation calculation in S4 is performed only based on the remaining subset of view attributes.

[0015] To achieve the above objectives, the present invention adopts the following technical solution: a multi-source information decision fusion system for multi-granularity, multi-perspective rough sets, wherein the system is used to implement the above-mentioned multi-source information decision fusion method for multi-granularity, multi-perspective rough sets, and the system includes: The data preprocessing and view partitioning module is configured to acquire the original multi-source datasets, merge them, perform cleaning and normalization processing, construct a multi-source information decision table, and partition it into view attribute subsets. The multi-granularity space construction module is configured to construct a multi-level multi-granularity space structure based on the neighborhood radius threshold sequence and generate a binary relation matrix. The rough set approximation calculation and feature extraction module is configured to calculate the approximate set and attribute importance index, and extract local decision feature vectors; The multi-source weighted decision fusion module is configured to aggregate the local decision feature vectors using a weighted fusion model and output the final classification result.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention achieves comprehensive and in-depth analysis of multi-source information by constructing a multi-granularity, multi-perspective rough set model. First, it utilizes a multi-perspective mechanism to mine complementary information between different data sources, effectively compensating for the blind spots of a single perspective. Secondly, by employing a multi-granularity hierarchical structure, features are extracted at different levels of granularity, preserving both macroscopic structure and microscopic details, significantly improving the completeness of feature extraction. Finally, a weighted fusion strategy based on attribute importance automatically reduces the interference from low-quality views, greatly improving the accuracy and stability of decision-making results in uncertain environments. Attached Figure Description

[0017] Figure 1 This is the main flowchart of the multi-source information decision fusion method of the present invention; Figure 2 This is a flowchart illustrating the construction process of the multi-source information decision table in this invention. Figure 3 This is a flowchart illustrating the multi-perspective, multi-granularity knowledge base generation process of the present invention. Figure 4 This is a flowchart of the local decision feature vector extraction process of the present invention; Figure 5 This is a flowchart of the global decision score generation and optimization process of this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0019] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0020] Please see Figure 1 and Figure 2 This invention provides a technical solution: a multi-source information decision fusion method using multi-granularity, multi-perspective rough sets, comprising the following steps: S1: Obtain the original multi-source data set of the target object, perform data cleaning and normalization on the original multi-source data set, construct a multi-source information decision table containing conditional attributes and decision attributes, and divide the multi-source information decision table into multiple independent view attribute subsets according to the data source characteristics; The process of constructing a multi-source information decision table specifically includes: detecting missing values ​​in the original multi-source dataset and filling in the missing values ​​using the mean of similar samples; The maximum-minimum normalization method is used to map numerical data to a standard interval. The conditional attributes and decision attributes are aligned in columns to form a two-dimensional data matrix that serves as a multi-source information decision table.

[0021] In practical applications such as landslide geological hazard monitoring or fault diagnosis of complex industrial systems, the process first initiates a data acquisition command, acquiring a raw, multi-source dataset of the target object through a multi-type sensor network deployed in the monitoring area. The data sources are clearly divided into two main categories: space-based remote sensing monitoring and ground-based sensor monitoring. Space-based data is acquired through high-resolution optical satellites and synthetic aperture radar to obtain surface deformation rate and vegetation cover data, while ground-based data is acquired through deep inclinometers, piezoelectric rain gauges, and soil moisture sensors to collect deep displacement, rainfall, and seepage pressure data. The acquired raw data is transmitted to the central processing unit in timestamped JSON or CSV format. The data types cover continuous floating-point values, such as displacement and rainfall, as well as discrete status codes, such as equipment operating status bits. Before the data enters subsequent processing, a rigorous data cleaning operation must be performed. The system traverses the entire dataset to identify missing data items marked as NaN or Null. For detected missing values, the process executes imputation logic based on the mean of similar samples: First, it identifies the category label of the missing data, such as stable state or creeping state. Then, it retrieves the values ​​of all non-missing samples in that category for the corresponding attribute, sums these values, divides them by the total number of samples to obtain the mean, and finally writes this mean to the missing location to ensure data integrity. After handling missing values, given the significant differences in the units of data collected by different sensors (e.g., displacement is measured in millimeters while rainfall is measured in millimeters per hour), the process must perform data normalization. Here, a maximum-minimum normalization logic is used. By traversing the entire data column, the maximum and minimum values ​​for that attribute are identified, the difference between the current monitored value and the minimum value is calculated, and the range between the maximum and minimum values ​​is also calculated. The former is divided by the latter, thus linearly mapping all values ​​to a standard closed interval between 0 and 1. Subsequently, based on the source characteristics of the data, such as satellite views, ground sensor views, or physical attributes, such as displacement views and hydrological views, the process vertically divides the constructed two-dimensional data matrix, which contains conditional attributes (monitoring indicators) and decision attributes (landslide risk levels), into multiple independent view attribute subsets. Each subset contains only feature columns from a specific perspective, but shares the same set of decision attribute columns, laying the foundation for subsequent multi-view collaborative calculations.

[0022] In the aforementioned data preprocessing stage, the normalization process employs a range-based proportional mapping logic. Specifically, the system first reads the statistical extreme values ​​of a certain attribute column within the current time window, i.e., the maximum and minimum values ​​of the attribute. Next, for each original monitoring value in that column, the system performs a subtraction operation, calculating the difference between the original monitoring value and the attribute minimum value to obtain the offset. Simultaneously, the system calculates the difference between the attribute maximum and minimum values ​​to obtain the global range. Finally, the system performs a division operation, dividing the offset by the global range to obtain the normalized value corresponding to the original monitoring value. For example, when the maximum value of the deep displacement attribute at a certain monitoring point is 100 mm and the minimum value is 0 mm, and the currently collected original monitoring value is 45 mm, directly substituting these values ​​into the above logic, the offset is 45 mm, and the global range is 100 mm. Dividing these two values ​​yields a normalized value of 0.45. This normalized value of 0.45 is directly used as the basic input for subsequent granularity calculations, eliminating the influence of physical dimensions. The advantage of this operational logic is that it unifies physical quantities with different dimensions into the same dimensionless interval through range mapping, avoiding the masking of small numerical attributes by large numerical attributes in subsequent distance calculations, and significantly improving the balance of multi-source fusion. Table 1 lists the comparison of some monitoring data before and after cleaning and normalization.

[0023] Table 1 Comparison of Preprocessing Data from Multiple Sources at Monitoring Points As shown in Table 1, the original data of deformation rate for sample 003 was missing. The mean of similar "stable" samples was calculated and assumed to be 5 mm per day. The normalized result was then filled with 0.50.

[0024] Please see Figure 1 and Figure 3 S2: For each subset of view attributes, a multi-level, multi-granularity spatial structure is constructed by setting different granularity criteria, and a binary relation matrix between objects is calculated at each granularity level to generate a multi-view, multi-granularity knowledge base. The process of constructing a multi-level, multi-granularity spatial structure specifically includes: setting a set of incremental neighborhood radius threshold sequences; For each subset of view attributes, the neighborhood similarity between samples is determined by each threshold in the neighborhood radius threshold sequence. Samples that meet the similarity conditions are grouped into the same information granularity, thus forming multiple granularity levels from fine to coarse.

[0025] In the multi-granularity spatial construction phase, the core of the process lies in simulating the mechanism by which humans perceive things at different levels of granularity by adjusting the observation scale, i.e., the neighborhood radius. First, the process defines a set of neighborhood radius thresholds that increase arithmetically, covering the entire range from fine differentiation to coarse clustering. For each subset of view attributes, such as the aforementioned ground-based sensor view, the process sequentially selects each neighborhood radius threshold in the sequence as the current judgment criterion. Under a specific threshold, the process iterates through each pair of sample objects in the sample set, performing a similarity measurement operation. This measurement is based on the verbal logic of Euclidean distance: for all conditional attribute values ​​of two sample objects under the current view, the difference between corresponding attribute values ​​is calculated one by one, and this difference is squared; then, the squared differences of all attributes are summed to obtain a sum of squares; finally, the square root of this sum of squares is performed to obtain the spatial distance between the two sample objects. After calculating the spatial distance, the process compares it numerically with the currently selected neighborhood radius threshold. If the calculated spatial distance is strictly less than or equal to the current neighborhood radius threshold, the two sample objects are determined to be indistinguishable or similar at the current granularity, meaning they belong to the same information particle or neighborhood class. Conversely, if the spatial distance is greater than the threshold, they are determined to be unrelated. Based on this determination, a binary relation matrix is ​​constructed, where rows and columns correspond to sample objects. If two samples satisfy the similarity condition, the corresponding element in the matrix is ​​marked as 1; otherwise, it is marked as 0. As the values ​​in the neighborhood radius threshold sequence increase sequentially, the similarity determination condition between samples gradually relaxes, the size of the formed information particles gradually increases, and the number gradually decreases, thus constructing a multi-level granularity spatial structure from fine-grained (small radius) to coarse-grained (large radius). The set of binary relation matrices for all views at all granularity levels together constitutes a multi-view, multi-granularity knowledge base, providing a structured foundation for subsequent rough set approximation calculations.

[0026] In this step, the neighborhood radius threshold sequence is generated using an arithmetic progression method. The system first sets a basic radius step size and a maximum radius limit, and then generates a series of thresholds by accumulating the step size. For example, if the basic radius step size is set to 0.05 and the maximum radius is 0.20, the generated neighborhood radius threshold sequence will be 0.05, 0.10, 0.15, and 0.20. In the distance calculation and similarity determination logic, the system first obtains the normalized attribute vectors of sample A and sample B in the current view and calculates the Euclidean distance between the two vectors. The specific calculation logic is as follows: the difference between the attribute value of sample A and the corresponding attribute value of sample B is taken, the squares are summed, and then the square root is taken. Subsequently, this distance is compared with the radius threshold at the current granularity. For example, assuming the normalized attribute vectors of sample A are 0.45 and 0.40, the normalized attribute vectors of sample B are 0.48 and 0.44, and the currently selected neighborhood radius threshold is 0.06. Substituting the values ​​into the above logical operation, the difference for attribute 1 is -0.03, which squares to 0.0009; the difference for attribute 2 is -0.04, which squares to 0.0016; summing the two yields 0.0025; taking the square root, the spatial distance between sample A and sample B is 0.05. Comparing this distance 0.05 with the threshold 0.06, since 0.05 is less than 0.06, sample A and sample B are determined to belong to the same neighborhood at the current granularity, and the corresponding position in the binary relation matrix is ​​recorded as 1. The advantage of this operational logic is that, through multi-level threshold determination, it can capture the structural features of data at different resolutions, avoiding the omission of various information caused by a single scale.

[0027] Please see Figure 1 and Figure 4 S3: Based on a multi-view, multi-granularity knowledge base, rough set theory is used to calculate the lower approximation set and the upper approximation set at each granularity level. The attribute importance index is calculated based on the lower approximation set and the upper approximation set, and then the local decision feature vector of each view at different granularities is extracted. The process of calculating attribute importance metrics specifically includes: calculating the dependency value of the decision attribute relative to the current view attribute subset at a specific granularity level; Calculate the change in dependency after removing a certain attribute from the current view's attribute subset; The change in dependency is used as the attribute importance value at the current granularity level; The process of extracting local decision feature vectors specifically includes: selecting core attributes whose attribute importance values ​​are higher than a preset importance threshold; Construct reduced local decision rules based on core attributes; Substitute the object to be decided into the local decision rule, output the local predicted classification label of the object in the current view and at the current granularity, and use the local predicted classification label as an element of the local decision feature vector.

[0028] Based on the multi-level granular structure generated by S2, this process utilizes rough set theory to mine the causal logic between data. For each granular level of each view, the process first identifies decision attributes, such as landslide risk level, and the resulting decision classification, that is, dividing the samples into several decision classes, such as high-risk and low-risk classes, according to different decision values. Subsequently, using the neighborhood information particles at the current granularity, the lower approximation set and upper approximation set of the decision class are calculated. The logic for constructing the lower approximation set is to find the union of neighborhood information particles that are completely contained in a certain decision class, representing the set of samples that can be deterministically classified by the attributes of the current view; the upper approximation set is the union of information particles that have a non-empty intersection with the decision class, representing the set of samples that may belong to that decision class. On this basis, the process calculates the dependency value of the decision attribute relative to the condition attribute set. This value is obtained by counting the number of samples contained in the lower approximation set and dividing it by the total number of samples, i.e., the cardinality of the universe of discourse, reflecting the explanatory power of the current view on the decision results at the current granularity. To quantify the contribution of each specific attribute, the process performs an attribute removal test: one attribute is temporarily removed from the current view, and the dependency value of the remaining attribute set is recalculated. The change in dependency value is obtained by subtracting the dependency value after removing the attribute from the complete dependency value; this change is defined as the attribute importance value. If removing an attribute causes a significant drop in dependency, it indicates that the attribute is crucial. After obtaining the importance of all attributes, the process sets an importance threshold, such as 0.02, and selects all attributes with importance values ​​greater than this threshold as core attributes, removing redundant attributes. Based on these core attributes and their corresponding attribute value ranges and decision labels, the process constructs a local decision rule set in the form of "IF-THEN". Finally, the target object data to be decided is substituted into these rules for matching, outputting the local predicted classification label of the object in the current view and at the current granularity, such as 1 representing high risk and 0 representing low risk. The labels output by all views at all granularities are arranged sequentially, forming the local decision feature vector of the object.

[0029] In this step, the calculation logic for attribute importance is based on dependency difference. The system first obtains the dependency value for the entire attribute set, which is equal to the number of samples in the approximate set divided by the total number of samples. Then, the system obtains the dependency value for the reduced set after removing the target attribute. Finally, a subtraction operation is performed to obtain the attribute importance. For example, assuming a total of 100 samples, at a granularity level with a radius of 0.05, a complete view including attributes displacement, rainfall, and osmotic pressure can deterministically classify 80 samples, so the total dependency value is 80 divided by 100, which equals 0.80. If, after removing the rainfall attribute, the remaining attribute can only deterministically classify 70 samples, then the dependency value after removal is 70 divided by 100, which equals 0.70. Substituting 0.80 and 0.70 into the subtraction operation, the attribute importance value for rainfall at the current granularity is obtained as 0.10. The value of 0.10 was then compared with a preset importance threshold of 0.05. Since 0.10 is greater than 0.05, rainfall was determined to be a core attribute and retained. The advantage of this calculation logic is that by quantifying the loss of classification ability caused by attribute removal, it accurately identifies features that play a key role in decision-making, achieving dimensionality reduction and optimization of the feature space. Table 2 shows the importance calculation results of different attributes at a specific granularity.

[0030] Table 2. Results of Attribute Importance Calculation and Screening As shown in Table 2, the importance of ambient temperature is only 0.01, which is below the threshold, so it is not considered when constructing local decision rules.

[0031] Please see Figure 1 and Figure 5 S4: Construct a weighted fusion model based on attribute importance, input all views and local decision feature vectors at all granularities into the weighted fusion model for consistent aggregation calculation, generate a global decision score, and determine the final classification result of the target object based on the global decision score; The method also includes a conflict detection step: calculating the conflict rate between locally predicted classification labels under different views; When the conflict rate exceeds the safety threshold, the subset of view attributes with the highest conflict rate is removed, and the aggregation calculation in S4 is performed only based on the remaining subset of view attributes. The process of constructing a weighted fusion model based on attribute importance specifically includes: using the information entropy algorithm to calculate the entropy weight values ​​at different granularity levels under each view; A fusion weight coefficient is assigned to each local decision feature vector based on the entropy weight value; A weighted voting matrix is ​​constructed as the core of the weighted fusion model. The process of generating a global decision score specifically includes: converting the local predicted classification labels of each view and each granularity into corresponding one-hot encoded vectors; The one-hot encoded vector is weighted and summed with the corresponding fusion weight coefficients to obtain the fusion probability vector. Extract the maximum value from the fusion probability vector as the global decision score; The method also includes a multi-granularity adaptive adjustment step: when the global decision score is lower than the preset confidence threshold, the value of the neighborhood radius threshold sequence is automatically reduced, and steps S2 to S4 are re-executed until the global decision score meets the requirements or the upper limit of the number of iterations is reached.

[0032] After obtaining the local predicted classification labels from each view and granularity level, the process enters the multi-source decision fusion stage. First, to assess the reliability of each information source, an information entropy algorithm is introduced to calculate weights. For the prediction results of all samples at a certain granularity for a particular view, the information entropy of its prediction distribution is calculated. The more chaotic the prediction distribution, i.e., the larger the entropy value, the lower the certainty of the information source, and the smaller the weight should be assigned; conversely, a larger weight should be assigned. The specific weights are calculated by normalizing the complementary values ​​of the entropy weights. Before formal fusion, the process performs a conflict detection step: the proportion of inconsistencies in the predicted labels given by different views for the same object is statistically analyzed and defined as the conflict rate. For example, if two out of three views predict safety and one predicts danger, a conflict exists. If the calculated conflict rate exceeds a preset safety threshold, such as 0.6, the process will identify and remove the subset of view attributes that differs most from the other majority of views; this is usually a data source with an extremely low signal-to-noise ratio, and only the remaining views will be used in subsequent calculations. Subsequently, a weighted fusion model is constructed, converting the local predicted classification labels of each retained view and at each granularity into one-hot encoded vectors. For example, if there are 3 categories, label 2 is converted into a vector form of 0, 1, 0. Next, these one-hot encoded vectors are weighted and summed with their corresponding fusion weight coefficients. That is, each component of each vector is multiplied by its corresponding weight, and the values ​​at the corresponding component positions of all vectors are summed to generate a fusion probability vector. Each value in this vector represents the overall confidence that the target object belongs to the corresponding category. The process extracts the maximum value in this vector as the global decision score, and the category index corresponding to this maximum value is determined as the final classification result of the target object. If the global decision score is lower than a preset confidence threshold, such as 0.7, it indicates that the current granularity setting is too coarse or too fine, resulting in insufficient consensus. The system triggers an adaptive adjustment mechanism: automatically, proportionally (e.g., by 10%), the neighborhood radius threshold sequence value in step S2 is reduced, and the fusion steps S2, S3, and S4 are rolled back and re-executed until the score meets the requirements or reaches the preset iteration limit, such as 5 times, thereby ensuring the high robustness of the decision result.

[0033] In this step, the calculation of the global decision score involves a weighted summation of one-hot encodings. The system first obtains the local predicted labels for each component and converts them into standard unit vectors. Next, it obtains the fusion weight coefficients for each component, calculated using the aforementioned entropy weighting method, with the sum of all coefficients being 1. Then, it performs vector weighting: multiplying each unit vector by its corresponding weight coefficient to obtain a weighted vector; then, it sums all weighted vectors bit-by-bit to obtain the fusion probability vector. Finally, it searches for the maximum value in the fusion probability vector. For example, suppose the system has two decision components: component A has a weight of 0.6 and a predicted label of category 1; component B has a weight of 0.4 and a predicted label of category 2. The system has three categories: category 0, category 1, and category 2. The one-hot vector for component A is 0,1,0, and the one-hot vector for component B is 0,0,1. Substituting these into the weighting logic: The weighted vector A = 0 multiplied by 0.6, 1 multiplied by 0.6, 0 multiplied by 0.6 = 0, 0.6, 0; The weighted vector B = 0 multiplied by 0.4, 0 multiplied by 0.4, 1 multiplied by 0.4 = 0, 0, 0.4; The fusion probability vector = 0 + 0, 0.6 + 0, 0 + 0.4 = 0, 0.6, 0.4; The maximum value of this vector, 0.6, is extracted and becomes the global decision score. Since 0.6 is less than the preset confidence threshold of 0.7, the system determines that the current decision lacks credibility, triggering adaptive adjustment logic to reduce the neighborhood radius from 0.05 to 0.045 (a 10% decrease) and restart the calculation. The advantage of this logic is that it preserves the confidence information of each view through soft fusion of probability vectors rather than simple hard voting, and solves the problem of decision uncertainty at a single granularity through a closed-loop feedback mechanism. Table 3 shows specific numerical examples of the fusion calculation.

[0034] Table 3. Example of Multi-Source Decision Fusion Calculation Results As shown in Table 3, the final fusion vector is 0, 0.6, 0.4. Based on this, the system outputs the final classification result as category 1. However, because the score of 0.6 does not meet the standard, a recalculation is required.

[0035] A multi-source information decision fusion system based on multi-granularity and multi-perspective rough sets, wherein the system is used to execute the aforementioned multi-source information decision fusion method based on multi-granularity and multi-perspective rough sets, the system comprising: The data preprocessing and view partitioning module is configured to acquire the original multi-source datasets, merge them, perform cleaning and normalization processing, construct a multi-source information decision table, and partition it into view attribute subsets. The multi-granularity space construction module is configured to construct a multi-level multi-granularity space structure based on the neighborhood radius threshold sequence and generate a binary relation matrix. The rough set approximation calculation and feature extraction module is configured to calculate the approximate set and attribute importance index, and extract local decision feature vectors; The multi-source weighted decision fusion module is configured to aggregate local decision feature vectors using a weighted fusion model and output the final classification result.

[0036] This system, as a high-performance dedicated computing architecture, integrates various modules that operate collaboratively via a high-speed data bus. The data preprocessing and view partitioning module, serving as the system's input front-end, integrates multi-protocol interface units and is configured to receive raw multi-source data sets from different communication protocols, such as Modbus and TCP / IP, in parallel. Internally, this module runs a fixed cleaning algorithm, automatically performing mean-filling logic to repair data breakpoints, and using a hardware accelerator to perform max-min normalization operations, mapping heterogeneous data to a multi-source information decision table structure in a unified memory space. Subsequently, based on metadata tags, the memory blocks are logically partitioned into multiple view attribute subsets. Following this, the multi-granularity space construction module is configured to read the neighborhood radius threshold sequence from the configuration register, use vector processing units to calculate the Euclidean distance between large-scale sample matrices in parallel, and quickly generate the corresponding binary relation matrix based on comparator logic, constructing a multi-level granular space. The rough set approximation calculation and feature extraction module is the core computational unit of the system. It is configured to perform bitwise operations on the binary relation matrix based on set theory logic, such as AND, OR, and NOT, to quickly derive the lower and upper approximation sets. It then calculates attribute importance based on dependency differential logic, selects core features, and generates local decision feature vectors. Finally, the multi-source weighted decision fusion module is configured to aggregate all feature vectors, perform one-hot encoding fusion operations using a built-in weighted summation circuit, and includes a comparison and judgment circuit for conflict detection and adaptive threshold comparison. If the judgment result is unsatisfactory, this module sends an interrupt signal and new parameter configuration to the multi-granularity space construction module via a feedback line, triggering iterative operation of the system until the final high-confidence classification result is output.

[0037] In the operational logic of the system modules, data flow and parameter calculation are highly coupled. The rough set approximation calculation and feature extraction module executes the following logic when calculating dependency: The module first counts the number of rows in the binary relation matrix that satisfy the inclusion relationship, denoted as the number of positive domain samples; simultaneously, it reads the value from the total number of rows register, denoted as the total number of samples. The divider within the module performs the operation of dividing the number of positive domain samples by the total number of samples to obtain the dependency value. For example, when the number of positive domain samples is 80 and the total number of samples is 100, the divider outputs 0.80. In the multi-source weighted decision fusion module, the system executes adaptive threshold comparison logic: The module reads the content of the maximum value register in the fusion probability vector, such as 0.6, and compares it with the content of the confidence threshold register, such as 0.7, and inputs this comparison into the digital comparator. If the comparator outputs a "less than" signal (logic low), the parameter adjustment logic is triggered: the multiplier reads the current neighborhood radius register value (e.g., 0.05) and the attenuation coefficient register value (e.g., 0.9), performs a multiplication operation to obtain 0.045, writes this new value back to the neighborhood radius register, and simultaneously sets the "recalculate" flag. This not only achieves a closed-loop logic but also ensures the real-time performance and accuracy of the operation through hardware-level data flow control. The advantage of this logic is that it solidifies complex iterative optimization algorithms into signal interactions between modules, significantly reducing system latency.

[0038] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A multi-source information decision fusion method using multi-granularity, multi-perspective rough sets, characterized in that, Includes the following steps: S1: Obtain the original multi-source data set of the target object, perform data cleaning and normalization processing on the original multi-source data set, construct a multi-source information decision table containing conditional attributes and decision attributes, and divide the multi-source information decision table into multiple independent view attribute subsets according to the data source characteristics; S2: For each subset of view attributes, a multi-level, multi-granularity spatial structure is constructed by setting different granularity criteria, and a binary relation matrix between objects is calculated at each granularity level to generate a multi-view, multi-granularity knowledge base; S3: Based on the multi-view, multi-granularity knowledge base, use rough set theory to calculate the lower approximation set and the upper approximation set at each granularity level, calculate the attribute importance index based on the lower approximation set and the upper approximation set, and then extract the local decision feature vector of each view at different granularities. S4: Construct a weighted fusion model based on attribute importance, input the local decision feature vectors of all views and granularities into the weighted fusion model for consistent aggregation calculation, generate a global decision score as the decision evaluation parameter of the target object, and map the decision evaluation parameter to the object's priority, selection probability or ranking weight, thereby determining the final classification result and decision response strategy of the target object.

2. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 1, characterized in that, The process of constructing the multi-source information decision table specifically includes: Detect missing values ​​in the original multi-source dataset and fill the missing values ​​using the mean of similar samples; The maximum-minimum normalization method is used to map numerical data to a standard interval. The conditional attributes and decision attributes are aligned column-wise to form a two-dimensional data matrix that serves as the multi-source information decision table.

3. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 1, characterized in that, The process of constructing a multi-level, multi-granular spatial structure specifically includes: Define a set of incremental neighborhood radius thresholds; For each subset of the view attributes, the neighborhood similarity between samples is determined by each threshold in the neighborhood radius threshold sequence. Samples that meet the similarity conditions are grouped into the same information granularity, thereby forming multiple granularity levels from fine to coarse.

4. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 3, characterized in that, The process of calculating the importance index of the attribute specifically includes: Calculate the numerical value of the dependency of the decision attribute relative to the current view attribute subset at a specific granularity level; Calculate the change in dependency after removing a certain attribute from the current subset of view attributes; The change in dependency is used as the attribute importance value at the current granularity level.

5. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 4, characterized in that, The process of extracting local decision feature vectors specifically includes: Filter out core attributes whose importance values ​​are higher than a preset importance threshold; Based on the core attributes, a reduced local decision rule is constructed; Substitute the object to be decided into the local decision rule, output the local predicted classification label of the object in the current view and at the current granularity, and use the local predicted classification label as an element of the local decision feature vector.

6. The multi-source information decision fusion method for multi-granularity, multi-view rough sets according to claim 5, characterized in that, The process of constructing a weighted fusion model based on attribute importance specifically includes: The entropy weight values ​​at different granularity levels under each view are calculated using the information entropy algorithm; A fusion weight coefficient is assigned to each local decision feature vector based on the entropy weight value; A weighted voting matrix is ​​constructed as the core of the weighted fusion model.

7. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 1, characterized in that, The process of generating a global decision score specifically includes: The local predicted classification labels of each view and each granularity are converted into corresponding one-hot encoded vectors; The one-hot encoded vector is weighted and summed with the corresponding fusion weight coefficients to obtain the fusion probability vector. The maximum value in the fusion probability vector is extracted as the global decision score.

8. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 3, characterized in that, The method further includes a multi-granularity adaptive adjustment step: The system monitors changes in external user demand parameters in real time. When a new demand instruction is received or when the global decision score is lower than the preset confidence threshold, the system dynamically resets the decision benchmark set and automatically adjusts the values ​​of the neighborhood radius threshold sequence. Steps S2 to S4 are then re-executed to achieve adaptive updates of the fusion weights and decision results as demand changes.

9. The multi-source information decision fusion method for multi-granularity, multi-perspective rough sets according to claim 8, characterized in that, The method also includes a collision detection step: Calculate the conflict rate between the locally predicted classification labels under different views; When the conflict rate exceeds the safety threshold, the subset of view attributes with the highest conflict rate is removed, and the aggregation calculation in S4 is performed only based on the remaining subset of view attributes.

10. A multi-source information decision fusion system with multi-granularity and multi-perspective rough sets, characterized in that, The system is used to implement the multi-source information decision fusion method of multi-granularity, multi-view rough set as described in any one of claims 1-9, and the system includes: The data preprocessing and view partitioning module is configured to acquire the original multi-source datasets, merge them, perform cleaning and normalization processing, construct a multi-source information decision table, and partition it into view attribute subsets. The multi-granularity space construction module is configured to construct a multi-level multi-granularity space structure based on the neighborhood radius threshold sequence and generate a binary relation matrix; The rough set approximation calculation and feature extraction module is configured to calculate the approximate set and attribute importance index, and extract local decision feature vectors; The multi-source weighted decision fusion module is configured to aggregate the local decision feature vectors using a weighted fusion model and output the final classification result.

Images