Intelligent operation decision analysis method and system based on cross-domain data fusion

Abstract

The invention relates to the technical field of data analysis, in particular to an operation decision intelligent analysis method and system based on cross-domain data fusion. The method comprises the following steps: firstly, based on an enterprise multi-domain ontology knowledge base, performing entity identification and relation mapping on heterogeneous data from different business systems through a semantic mapping-based multi-source heterogeneous data dynamic fusion algorithm, and establishing a unified data model; then, a causal reasoning and deep learning fused hybrid intelligent decision engine is adopted to analyze and process the model; then, a multi-level causal relationship network among business variables is constructed through a causal relationship discovery algorithm by utilizing an analysis result of the hybrid intelligent decision engine, and an adaptive business scene analysis model based on reinforcement learning is used to dynamically adjust an analysis strategy according to business environment changes; generating a Pareto optimal decision scheme set through a multi-objective optimization algorithm, and outputting operation decision suggestions; according to the invention, the comprehensiveness and accuracy of intelligent analysis of enterprise operation decisions are improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to an intelligent analysis method and system for operational decision-making based on cross-domain data fusion. Background Technology

[0002] The following technical problems are commonly found in current enterprise information systems: The company's various business systems operate independently, and there is a lack of effective data exchange mechanisms between systems such as OA, finance, human resources, and project management, resulting in data fragmentation and an inability to form a unified data view.

[0003] Existing systems mostly use static reports and single-dimensional analysis, lacking intelligent analysis capabilities based on multi-source data fusion, and thus cannot provide management with real-time and accurate decision support.

[0004] Traditional data analysis methods are limited to a single business area and cannot achieve cross-departmental, cross-business, and cross-time dimension correlation analysis, which restricts the overall optimization decision-making of enterprises.

[0005] Existing analytics systems are mostly based on offline data processing, which cannot meet the needs of enterprises to respond quickly to market changes.

[0006] To address this, a method and system for intelligent analysis of operational decisions based on cross-domain data fusion is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent analysis method and system for operational decision-making based on cross-domain data fusion.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An intelligent analysis method for operational decision-making based on cross-domain data fusion includes: Based on the enterprise multi-domain ontology knowledge base, a unified data model is established by performing entity recognition and relation mapping on heterogeneous data from different business systems through a dynamic fusion algorithm based on semantic mapping of multi-source heterogeneous data. Based on the data model, a hybrid intelligent decision engine that integrates causal reasoning and deep learning is used for analysis and processing; using the analysis results of the hybrid intelligent decision engine, a multi-level causal relationship network among business variables is constructed through a causal relationship discovery algorithm. Based on the aforementioned causal relationship network, an adaptive business scenario analysis model based on reinforcement learning is used to dynamically adjust the analysis strategy according to changes in the business environment. By combining a multi-level causal relationship network and an adaptive business scenario analysis model, a Pareto optimal decision set is generated through a multi-objective optimization algorithm, and operational decision recommendations are output.

[0009] Furthermore, the process of performing entity recognition and relation mapping on heterogeneous data from different business systems through a multi-source heterogeneous data dynamic fusion algorithm based on semantic mapping includes: Based on entity definitions in an enterprise multi-domain ontology knowledge base, literal similarity scores are calculated using edit distance and longest common subsequence algorithms to establish an initial candidate set for entity mapping. Utilizing cross-domain relationship mappings in the enterprise multi-domain ontology knowledge base, a pre-trained word vector model is used to calculate the semantic embedding distance of entity concepts, and semantic similarity scores are applied to the initial candidate set. According to semantic constraint rules in the enterprise multi-domain ontology knowledge base, the semantic matching results are verified and constraint consistency scores are generated through data type compatibility checks, numerical range constraint verification, and business rule consistency checks. Based on the weighted fusion results of the three-layer scores, high-confidence entity mapping relationships are confirmed and data transformation rules are established. The verified unified data model is then output to the hybrid decision engine for subsequent analysis and processing.

[0010] Furthermore, the analysis and processing using a hybrid intelligent decision engine that integrates causal reasoning and deep learning includes: The symbolic reasoning module receives entity mapping relationships and data transformation rules, and constructs a logical reasoning rule base in conjunction with enterprise business rules. It verifies data consistency and generates business-level logical analysis conclusions through a forward reasoning mechanism. The statistical reasoning module constructs a probability dependency graph between variables based on the logical analysis conclusions and historical data, and uses Bayesian reasoning to calculate the probability distribution and confidence interval of each business indicator. The neural reasoning module receives the probabilistic results from the statistical reasoning module and identifies nonlinear relationship patterns in the data through a deep neural network.

[0011] Furthermore, the process of constructing a multi-level causal relationship network among business variables using causal relationship discovery algorithms includes: By using the business logic conclusions of the symbolic reasoning module as prior constraints for causal relationships, a causal network topology that conforms to business logic is established. Based on the probability distribution results of the statistical reasoning module, statistical causal relationships between variables are identified on the basis of the topology through conditional independence testing. Combining the nonlinear relationship patterns identified by the neural reasoning module, Granger causality tests are performed using the temporal characteristics of time series data to determine the directionality of causal relationships. Through causal relationship strength assessment and network stability verification, a complete multi-level causal relationship network is formed.

[0012] Furthermore, the process of dynamically adjusting analysis strategies based on changes in the business environment using an adaptive business scenario analysis model includes: By using key causal nodes in the causal relationship network as feature dimensions of the environmental state, a multi-dimensional state space is established to represent the dynamic changes of the business environment. A state transition probability model is designed using relation weights in a causal network, and a sliding window and anomaly detection algorithm are used to monitor environmental change patterns and abnormal events in the state space. The environmental state is used as the observation input of the reinforcement learning agent, different analysis strategies are modeled as action choices, and the Q-learning algorithm is used to learn the state-action value function to optimize the strategy selection. The reward signal is calculated based on the business feedback after the strategy is executed. The Q network parameters are updated through the experience replay mechanism. At the same time, the learned optimal strategy and environment prediction results are output for decision optimization.

[0013] Furthermore, the process of generating a Pareto-optimal decision set through a multi-objective optimization algorithm and outputting operational decision recommendations includes: Using causal relationship networks as the decision constraint framework, and combining environmental state prediction and optimal strategy suggestions, a mathematical model for an optimization problem involving multiple conflicting objectives is constructed. The NSGA-II multi-objective genetic algorithm is used to search the decision space under constraints to generate a Pareto optimal solution set covering different trade-off schemes; and a hybrid intelligent decision engine and causal relationship network are invoked to conduct a comprehensive impact assessment of each Pareto solution. The optimal solution is selected from the solution set and output based on the current business priorities and the comprehensive impact assessment results.

[0014] An intelligent operational decision-making analysis system based on cross-domain data fusion includes: a data fusion unit, used to perform entity recognition and relationship mapping on heterogeneous data from different business systems based on an enterprise multi-domain ontology knowledge base and a dynamic multi-source heterogeneous data fusion algorithm based on semantic mapping, to establish a unified data model; an intelligent decision engine unit, used to perform data analysis based on the data model output by the data fusion unit, employing a hybrid intelligent analysis processing method that integrates causal reasoning and deep learning; a causal relationship unit, used to construct a multi-level causal relationship network between business variables using the analysis results of the intelligent decision engine unit and a causal relationship discovery algorithm; a business scenario analysis unit, used to dynamically adjust the analysis strategy according to changes in the business environment using an adaptive business scenario analysis model based on reinforcement learning, based on the causal relationship network output by the causal relationship unit; and a decision optimization output unit, used to combine the multi-level causal relationship network and the adaptive business scenario analysis model, and generate a Pareto optimal decision solution set through a multi-objective optimization algorithm to output operational decision recommendations.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using a dynamic fusion algorithm for multi-source heterogeneous data based on semantic mapping, and employing a three-layer scoring fusion mechanism of literal similarity, semantic similarity, and constraint consistency, it can accurately identify and map heterogeneous data entity relationships from different business systems such as OA, finance, human resources, and project management. This breaks through the limitations of traditional ETL methods in semantic understanding, realizes deep fusion and unified modeling of cross-domain data in enterprises, and thus improves the comprehensiveness and accuracy of intelligent analysis for enterprise operational decisions.

[0016] 2. By employing a hybrid intelligent decision engine that integrates symbolic reasoning, statistical reasoning, and neural reasoning modules, it can simultaneously handle deterministic business rules, probabilistic dependencies, and complex nonlinear patterns. This overcomes the limitations of single reasoning methods and constructs a multi-level causal network through causal relationship discovery algorithms. This provides interpretable causal evidence and a scientific analytical foundation for enterprise decision-making, thereby improving the comprehensiveness and accuracy of intelligent analysis for enterprise operational decisions.

[0017] 3. The adaptive business scenario analysis model based on reinforcement learning can take key nodes in the causal network as environmental state features, use the Q-learning algorithm to dynamically learn the optimal analysis strategy, adjust the analysis method in real time according to changes in the business environment, and generate a Pareto optimal decision solution set by combining multi-objective optimization algorithm. This significantly improves the agility and accuracy of enterprises in responding to market changes and business adjustments, thereby improving the comprehensiveness and accuracy of intelligent analysis of enterprise operation decisions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an intelligent analysis method for operational decision-making based on cross-domain data fusion, according to the present invention. Figure 2 This is a schematic diagram of the hybrid intelligent decision engine of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent analysis system for operational decision-making based on cross-domain data fusion according to the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1 to 3 This invention provides an intelligent analysis method and system for operational decision-making based on cross-domain data fusion, the technical solution of which is as follows: Example 1:

[0021] To improve the comprehensiveness and accuracy of operational decisions across multiple business sectors and departments, a company adopted an intelligent operational decision-making analysis method based on cross-domain data fusion, as proposed in this invention. The flowchart of this method is shown below. Figure 1 As shown, it specifically includes: Based on the enterprise multi-domain ontology knowledge base, a unified data model is established by performing entity recognition and relation mapping on heterogeneous data from different business systems through a dynamic fusion algorithm based on semantic mapping of multi-source heterogeneous data. Furthermore, the construction process of the enterprise multi-domain ontology knowledge base includes: analyzing the data structure and business processes of each business domain of the enterprise, identifying core business domains including human resources, financial management, project management, and office collaboration; defining a conceptual entity hierarchy for each core business domain, including entity categories, entity attributes, and inheritance relationships between entities; establishing cross-domain relationship mapping, defining the association, dependency, and constraint relationships between entities in different business domains; and formulating semantic constraint rules, including data type constraints, value range constraints, and business logic constraints, to form a complete enterprise multi-domain ontology knowledge base. Furthermore, the establishment of cross-domain relationship mapping is based on business logic analysis and expert knowledge. By analyzing business processes and data flows, the conceptual correspondence between different business domains is identified. Furthermore, relationships include the correspondence between "employees" in the human resources domain and "project members" in the project management domain; dependencies, such as the "budget" in the finance domain depending on the "project plan" in the project management domain; and constraints, such as the total employee salaries not exceeding the department's budget limit. Furthermore, the process of performing entity recognition and relation mapping on heterogeneous data from different business systems through a multi-source heterogeneous data dynamic fusion algorithm based on semantic mapping includes: Based on entity definitions in the enterprise multi-domain ontology knowledge base, literal similarity scores are calculated using edit distance and longest common subsequence algorithms to establish an initial entity mapping candidate set; By utilizing cross-domain relationship mapping in the enterprise's multi-domain ontology knowledge base, a pre-trained word vector model is used to calculate the semantic embedding distance of entity concepts, and semantic similarity scoring is performed on the initial candidate set; Based on the semantic constraint rules in the enterprise's multi-domain ontology knowledge base, the semantic matching results are verified and a constraint consistency score is generated through data type compatibility checks, numerical range constraint verification, and business rule consistency checks. Based on the weighted fusion results of the three-level scoring, the mapping relationship of high-confidence entities is confirmed and data transformation rules are established. The validated unified data model is then output to the hybrid decision engine for subsequent analysis and processing. Furthermore, a pre-trained word vector model is used to map entity concepts to a high-dimensional semantic space, and the Euclidean distance between concept vectors is calculated as a semantic difference measure. At the same time, the established cross-domain relationship mapping is used as prior knowledge, and the semantic calculation results are weighted and converted into similarity scores through distance normalization. The pre-trained word vector model can be Word2Vec or BERT. Furthermore, data type compatibility checks ensure that the data types of the matched fields can be converted to each other; numerical range constraint verification checks whether the data values ​​conform to business rules; and business rule consistency checks ensure that the matching results do not violate the enterprise's business logic. Furthermore, the literal similarity score, semantic similarity score, and constraint consistency score are weighted and fused, with the weights dynamically adjusted based on historical matching accuracy. The three-layer scores in the weighted fusion are normalized results, with values ​​ranging from [0, 1]. The weights are composed of historical matching accuracy, adjustment factors, and initial weights. For example, the semantic similarity weight is obtained by summing the weighted result of the adjustment factor and historical semantic matching accuracy with the complement of the adjustment factor and the initial weight. The adjustment factors for all three layers of scores are 0.7, and the initial weights are 0.3, 0.4, and 0.3, respectively.

[0022] By employing a dynamic fusion algorithm for multi-source heterogeneous data based on semantic mapping and a three-layer scoring fusion mechanism of literal similarity, semantic similarity, and constraint consistency, the algorithm can accurately identify and map heterogeneous data entity relationships from different business systems such as OA, finance, human resources, and project management. This breaks through the limitations of traditional ETL methods in semantic understanding, achieves deep fusion and unified modeling of cross-domain data for enterprises, and improves the comprehensiveness and accuracy of intelligent analysis for enterprise operational decisions.

[0023] Based on the data model, a hybrid intelligent decision engine that integrates causal reasoning and deep learning is used for analysis and processing; using the analysis results of the hybrid intelligent decision engine, a multi-level causal relationship network among business variables is constructed through a causal relationship discovery algorithm. Furthermore, the structure of the hybrid intelligent decision engine is as follows: Figure 2 As shown, it includes a symbolic reasoning module, a statistical reasoning module, and a neural reasoning module; Furthermore, the analysis and processing using a hybrid intelligent decision engine that integrates causal reasoning and deep learning includes: The symbolic reasoning module receives entity mapping relationships and data transformation rules, combines them with enterprise business rules to build a logical reasoning rule base, and verifies data consistency and generates business-level logical analysis conclusions through a forward reasoning mechanism. Based on logical analysis conclusions and historical data, the statistical inference module constructs a probability dependency graph between variables and uses Bayesian inference to calculate the probability distribution and confidence interval of each business indicator. The neural inference module receives the probabilistic results from the statistical inference module and identifies non-linear relationship patterns in the data through a deep neural network. Furthermore, the symbolic reasoning module adopts a production rule representation method to convert business rules into the "IF-THEN" format, such as "IF Project schedule lag > 15% AND Budget utilization > 80% THEN Project risk level = high"; Furthermore, the statistical reasoning module receives the set of logical conclusions as prior knowledge, uses Bayesian network modeling to represent business variables as random variable nodes, and represents the dependencies between variables as directed edges; it learns the conditional probability distribution through Bayesian estimation methods and calculates the probability distribution and confidence interval of each business indicator. Furthermore, the neural inference module receives the probability distribution results of the Bayesian network as the input feature vector and outputs a nonlinear relationship matrix. The module adopts a three-layer fully connected network architecture, with two hidden layers: the first layer has 256 neurons and the second layer has 128 neurons. Both layers use the ReLU activation function and a Dropout rate of 0.2. Then, the Sigmoid activation function is used to output a relationship strength matrix in the range [0,1].

[0024] By leveraging the collaborative work of the three modules—symbolic reasoning, statistical reasoning, and neural reasoning—in the hybrid intelligent decision engine, data can be analyzed from three dimensions: logic, probability, and nonlinear relationships. This overcomes the limitations of single-dimensional analysis and provides more comprehensive business insights. The hierarchical and modular design makes the analysis process clear in structure, computationally efficient, and suitable for processing multidimensional data, meeting the real-time decision-making needs of enterprises.

[0025] Furthermore, the process of constructing a multi-level causal relationship network among business variables using causal relationship discovery algorithms includes: By using the business logic conclusions of the symbolic reasoning module as prior constraints on causal relationships, a causal network topology structure that conforms to business logic is established. Based on the probability distribution results of the statistical reasoning module, statistical causal relationships between variables are identified on the basis of topological structure through conditional independence tests; By combining the nonlinear relationship patterns identified by the neural reasoning module, Granger causality tests are performed using the temporal characteristics of time series data to determine the directionality of causal relationships. A complete multi-level causal relationship network is formed through causal relationship strength assessment and network stability verification; Furthermore, the business logic conclusions of the symbolic reasoning module are used as prior constraints for causal relationships to exclude causal connections that do not conform to business logic; based on domain expert knowledge and business rules, an initial causal network topology is established to identify variable pairs that may have causal relationships. Furthermore, based on the probability distribution results of the statistical inference module, the PC algorithm is used to test conditional independence; the statistical independence between variables is determined by the chi-square test or G test, and the real statistical causal relationship is identified on the basis of the network topology, and an undirected graph is constructed to represent the relationship between variables. Furthermore, the PC algorithm uses the following parameter settings: significance level of 0.05, maximum condition set size of 3, and G² statistic for conditional independence testing; the algorithm first constructs a completely undirected graph, then gradually deletes edges through conditional independence testing, and finally determines the direction of the edges through orientation rules; Furthermore, the Granger causality test uses the F-statistic with a significance level of 0.01, and the lag order is determined by the AIC criterion. The temporal relationship between variables is determined by regression analysis of lagged terms, and the directionality of the causal relationship is determined by directional rules such as the V-structure test, thus converting the undirected graph into a directed acyclic graph. Furthermore, the influence of each causal edge is quantified by assessing the strength of causal relationships, and the causal strength is calculated using mutual information or conditional mutual information methods. Cross-validation is used to evaluate the reliability of the network structure, forming a complete multi-level causal relationship network. By combining symbolic reasoning, statistical reasoning, and neural reasoning, a multi-level causal relationship network that conforms to business realities can be constructed, significantly improving the accuracy of causal analysis. By utilizing the temporal characteristics of time series data through Granger causality tests, the directionality of causal relationships can be accurately determined, overcoming the shortcomings of traditional methods in cross-time dimension analysis. Through causal relationship strength assessment and network stability verification, the robustness and reliability of the causal network are ensured. Based on the multi-level causal relationship network, cross-departmental and cross-business domain variable causal relationships can be revealed, enhancing overall decision-making capabilities.

[0026] Furthermore, the multi-level causal relationship network also includes spatiotemporal causal mining and dynamic update mechanisms, specifically: Based on the analysis results of Granger causality test, a spatiotemporal causality mining module is constructed. A spatiotemporal graph convolutional network is used to model the identified causal relationships at multiple time scales. Short-term, medium-term and long-term windows are set to analyze the stability and intensity changes of causal relationships at different time scales. By utilizing the results of spatiotemporal causal relationship mining, a causal relationship confidence quantification mechanism is established. The causal strength obtained from spatiotemporal analysis is used as prior information. The posterior probability distribution of each causal edge is updated using the Bayesian causal inference method. The confidence interval is estimated based on the causal strength at multiple time scales using the Bootstrap resampling technique. Dynamic confidence scores are assigned to each edge in the multi-level causal relationship network. Based on the confidence quantification results, a dynamic update mechanism for the causal network is implemented. When the confidence of a causal edge is below the threshold of 0.6 for three consecutive time windows, it is marked as a declining state. When the confidence continuously rises and exceeds 0.8, it is marked as a growing state. The topology of the multi-level causal relationship network is automatically adjusted according to the trend of confidence changes.

[0027] Spatiotemporal causal mining and dynamic update mechanisms can enhance the ability of multi-level causal relationship networks to capture time-delayed causal effects, thereby supporting long-term strategic decision-making. At the same time, the causal relationship confidence quantification mechanism provides a credibility assessment of causal relationships, enhancing the scientific nature of decision-making. The dynamic update mechanism also improves the accuracy of the network in identifying causal relationships.

[0028] Based on causal relationship networks, an adaptive business scenario analysis model based on reinforcement learning is used to dynamically adjust the analysis strategy according to changes in the business environment. Furthermore, the process of dynamically adjusting analysis strategies based on changes in the business environment using an adaptive business scenario analysis model includes: By using key causal nodes in the causal relationship network as feature dimensions of the environmental state, a multi-dimensional state space is established to represent the dynamic changes of the business environment. A state transition probability model is designed using relation weights in a causal network, and a sliding window and anomaly detection algorithm are used to monitor environmental change patterns and abnormal events in the state space. The environmental state is used as the observation input of the reinforcement learning agent, different analysis strategies are modeled as action choices, and the Q-learning algorithm is used to learn the state-action value function to optimize the strategy selection. The reward signal is calculated based on the business feedback after the strategy is executed. The Q network parameters are updated through the experience replay mechanism. At the same time, the learned optimal strategy and environment prediction results are output for decision optimization. Furthermore, based on node centrality metrics such as degree centrality and betweenness centrality, as well as business importance assessment, the key business metrics that have the greatest impact on business results are selected; a multi-dimensional state space is constructed, with each dimension corresponding to the current state value of a key business metric. Furthermore, a state transition probability model is designed using the relation weights in the causal network, and a Markov state transition matrix is ​​established; a sliding window technique is used to analyze the historical state sequence, with the window size set to 30 time points; statistical process control methods and the isolated forest algorithm are used to monitor environmental change patterns and abnormal events in the state space. Furthermore, the environmental state is used as the observation input of the reinforcement learning agent, and different analysis strategies are modeled as discrete action spaces, including conservative, aggressive, and balanced strategies. The Q-learning algorithm is used to learn the state-action value function, and a greedy strategy is used to achieve a balance between exploration and exploitation. The reward function is evaluated based on business performance indicators such as ROI and efficiency improvement rate. The Q-learning algorithm parameters are set as follows: the learning rate is 0.1, the discount factor is 0.9, and the exploration rate adopts a decay strategy, with an initial value of 0.9, which decays by 0.01 every 1000 steps until it reaches 0.1. Furthermore, based on the business feedback after policy execution, immediate and cumulative rewards are calculated. An experience replay mechanism is used to store state transition quadruples, including state, action, reward, and next state. Then, mini-batch gradient descent is used to update the Q-network parameters, while simultaneously outputting the learned optimal policy recommendation and environment state prediction results. The experience replay buffer size is set to 10000, and 32 experience samples are randomly sampled each time training. When the buffer is full, a FIFO strategy is used to replace the oldest experience. The Q-network parameters are updated every 10 steps. The adaptive business scenario analysis model based on reinforcement learning can use key nodes in the causal network as environmental state features, dynamically learn the optimal analysis strategy using the Q-learning algorithm, and learn the optimal strategy through the state-action value function. It can dynamically adjust the analysis strategy according to business feedback, and the output environmental prediction results and optimal strategy provide data support for subsequent decision optimization, enhance the enterprise's ability to predict the future business environment, and thus improve the comprehensiveness and accuracy of intelligent analysis of enterprise operation decisions.

[0029] Furthermore, reinforcement learning agents also include multi-agent collaboration and experience-sharing enhancement mechanisms, specifically: A multi-agent collaborative architecture is constructed, deploying four specialized Q-learning agents: a financial agent, a human resources agent, a project management agent, and a marketing agent. Each agent maintains an independent structure and performs local Q-value learning based on different dimensional subsets of the environmental state. By leveraging the collaborative results of multi-agent experience sharing, a knowledge transfer enhancement mechanism is established. Each Q-learning agent shares the learned Q-value knowledge through similar state mapping technology. When an agent encounters a similar business scenario, it can refer to the Q-value estimates of other agents as an initial reference. The weight of the referenced experience is adjusted through importance sampling to accelerate the Q-value convergence process in the new scenario. Based on the knowledge transfer-based learning acceleration capability, an agent decision fusion mechanism is realized. Each Q-learning agent maintains an independent Q-value learning process. In the action selection stage, a weighted voting method is used to integrate the action suggestions of each agent. The weights are dynamically adjusted according to the convergence degree of each agent's Q-value and the historical decision accuracy.

[0030] By enhancing the multi-agent collaboration and experience sharing mechanisms, the learning efficiency of agents in complex business scenarios can be significantly improved while maintaining the stability of the Q-learning algorithm. Experience sharing improves the convergence speed of each agent. At the same time, the knowledge transfer mechanism reduces the exploration time in new scenarios. The decision fusion mechanism also improves the stability and accuracy of the overall strategy and reduces the impact of decision-making errors by a single agent.

[0031] By combining a multi-level causal relationship network and an adaptive business scenario analysis model, a Pareto optimal decision set is generated through a multi-objective optimization algorithm, and operational decision recommendations are output.

[0032] Furthermore, the process of generating a Pareto-optimal decision set through a multi-objective optimization algorithm and outputting operational decision recommendations includes: Using causal relationship networks as the decision constraint framework, and combining environmental state prediction and optimal strategy suggestions, a mathematical model for an optimization problem involving multiple conflicting objectives is constructed. The NSGA-II multi-objective genetic algorithm is used to search the decision space under constraints to generate a Pareto optimal solution set covering different trade-off schemes; and a hybrid intelligent decision engine and causal relationship network are invoked to conduct a comprehensive impact assessment of each Pareto solution. The optimal solution is selected from the solution set based on the current business priorities and the comprehensive impact assessment results. Furthermore, using a causal relationship network as a constraint framework, decision variables are defined as controllable business parameters, such as resource allocation and budget allocation; objective functions include cost minimization, efficiency maximization, and risk minimization. Furthermore, the NSGA-II multi-objective genetic algorithm is used to search the decision space under constraints, generate new solution individuals through genetic operations, maintain population diversity using non-dominated sorting and crowding distance calculation, and generate a Pareto optimal solution set covering different trade-off schemes. The parameters of the NSGA-II multi-objective genetic algorithm are set as follows: population size is 100, crossover probability is 0.9, mutation probability is 0.1, and maximum number of generations is 500; the fast non-dominated sorting algorithm is used, and the crowding distance is calculated using Euclidean distance. Furthermore, the symbolic reasoning module in the hybrid intelligent decision engine is used to verify the logical feasibility of the decision scheme; the statistical reasoning module is used to assess the probabilistic risks of implementing the scheme; the neural reasoning module is used to identify the potential impact patterns of the scheme; and the causal relationship network is combined to simulate the propagation effect of the decision on the causal chain. Furthermore, a multi-criteria decision-making evaluation system is established, which combines current business priorities, risk tolerance, and resource constraints to select the most suitable decision-making scheme from the Pareto solution set; and a detailed implementation plan is output, including resource allocation recommendations, timelines, risk control measures, and expected results.

[0033] By using a multi-objective optimization model and the NSGA-II algorithm, Pareto optimal solution sets can be generated to balance conflicting objectives such as cost, efficiency, and revenue, providing diverse decision-making options. A hybrid intelligent decision engine and causal network are used to conduct a comprehensive impact assessment of the Pareto solution sets, ensuring that the potential impact of each option is fully considered, thus improving the scientific nature and reliability of decision-making. The optimal solution is selected from the solution set based on the current business priority, which can flexibly adapt to the needs of different business scenarios and provide customized operational decision-making suggestions.

[0034] Furthermore, the Pareto optimal solution set generation process also includes a preference learning mechanism, specifically: Based on the Pareto optimal solution set generated by the NSGA-II multi-objective genetic algorithm, a decision preference learning mechanism is constructed. The feature distribution of each solution in the solution set in dimensions such as cost, risk, and efficiency is extracted as the input features for preference learning. Candidate solutions with significant differences in feature distribution are shown to the decision-maker and compared in pairs. The preference prediction model based on support vector machine is trained through the comparison results to learn the decision-maker's trade-off preference pattern among different objectives. By utilizing the prediction results of the preference learning model, a preference-oriented solution set optimization mechanism is established. The Pareto optimal solution set is reordered and screened according to the learned preference weights. The weighted distance method is used to calculate the closeness of each scheme to the ideal preference point. The hybrid intelligent decision engine and causal relationship network are invoked to conduct a comprehensive impact assessment on the Pareto optimal solution set after preference screening, and to verify the logical feasibility, probabilistic risk and potential impact mode of the scheme.

[0035] By implementing a preference learning mechanism on the Pareto optimal solution set, the personalization of decision-making schemes can be significantly improved. Solution set selection based on preference learning can improve decision-making efficiency, thereby ensuring the practicality and acceptability of decision-making schemes.

[0036] This embodiment proposes an intelligent analysis method and system for operational decision-making based on cross-domain data fusion. First, based on an enterprise multi-domain ontology knowledge base, a dynamic fusion algorithm based on semantic mapping is used to perform entity recognition and relationship mapping on heterogeneous data from different business systems, establishing a unified data model. Then, a hybrid intelligent decision engine integrating causal reasoning and deep learning is used to analyze and process this model. Next, utilizing the analysis results of the hybrid intelligent decision engine, a multi-level causal relationship network among business variables is constructed through a causal relationship discovery algorithm. An adaptive business scenario analysis model based on reinforcement learning is used to dynamically adjust the analysis strategy according to changes in the business environment. A Pareto optimal decision set is generated through a multi-objective optimization algorithm, outputting operational decision recommendations. This invention improves the comprehensiveness and accuracy of intelligent analysis for enterprise operational decisions.

[0037] Example 2:

[0038] This embodiment takes a large retail enterprise as an example, applying the intelligent operational decision-making analysis system based on cross-domain data fusion proposed in this invention to intelligently analyze and optimize operational data across multiple business areas, thereby improving the efficiency and accuracy of inventory management, sales forecasting, and cost control; the system structure is as follows. Figure 3 As shown, the specific implementation process is as follows: The data fusion unit is used to perform entity recognition and relation mapping on heterogeneous data from different business systems based on the enterprise's multi-domain ontology knowledge base and a dynamic fusion algorithm based on semantic mapping to establish a unified data model. In this case study, the retail enterprise's business systems include a Work Management System (WMS), a Sales and POS (POS) system, an Enterprise Resource Planning (ERP) system, and a Customer Relationship Management (CRM) system. First, based on the enterprise's multi-domain ontology knowledge base, the data structure and business processes of each business system are analyzed to identify core business domains, including inventory management, sales management, financial management, and customer management. The entity hierarchy for each domain is then defined, such as "products" and "inventory quantity" in inventory management, "orders" and "sales revenue" in sales management, "costs" and "profits" in financial management, and "customers" and "purchase preferences" in customer management. Literal similarity scores are calculated using edit distance and longest common subsequence algorithms. For example, matching "product ID" in WMS with "product number" in POS establishes an initial candidate set of entity mappings. A pre-trained BERT model is used to calculate the semantic embedding distance of entity concepts, such as identifying the semantic association between "inventory quantity" and "available inventory," and semantic similarity scores are applied to the candidate set. The matching results are verified using semantic constraint rules from the ontology knowledge base, such as verifying whether "sales revenue" data conforms to the value range constraints of the financial system, and the business logic consistency between "orders" and "customers." Based on a weighted fusion of literal similarity, semantic similarity, and constraint consistency, high-confidence entity mapping relationships are confirmed, data transformation rules are generated, and a unified retail enterprise data model is finally output, containing entities such as products, orders, customers, and finances, and their relationships.

[0039] The intelligent decision engine unit is used to perform data analysis based on the data model output by the data fusion unit, employing a hybrid intelligent analysis and processing approach that integrates causal reasoning and deep learning. The symbolic reasoning module receives entity mapping relationships and data transformation rules, combines them with the business rules of retail enterprises, constructs a logical reasoning rule base, and verifies data consistency through forward reasoning; for example, it checks whether the sales revenue in the POS system is consistent with the revenue record in the ERP system, and generates logical analysis conclusions at the business level, such as "abnormal sales of a certain product".

[0040] Based on logical analysis conclusions and sales and inventory data from the past 12 months, the statistical inference module constructs a probability dependency graph between variables and uses Bayesian inference to calculate the probability distribution of business indicators. For example, it calculates the probabilistic impact of "promotional activities" on "sales revenue" and outputs a probability distribution with a confidence interval of [0.75, 0.95].

[0041] The neural reasoning module receives statistical reasoning results and uses a three-layer fully connected neural network to identify nonlinear relationship patterns in the data; for example, it discovers a complex nonlinear relationship between "customer purchase preferences" and "promotional activities" and outputs a relationship strength matrix.

[0042] The causal relationship unit is used to construct a multi-level causal relationship network between business variables by utilizing the analysis results of the intelligent decision engine unit and through causal relationship discovery algorithms. Based on the business logic conclusions of the symbolic reasoning module, a topological foundation for the causal network is established. Utilizing the probability distribution results from the statistical reasoning module, a PC algorithm is used to test conditional independence and identify statistical causal relationships between variables, such as the association between "inventory level" and "replenishment frequency." Combining the nonlinear relationship patterns of the neural reasoning module, Granger causality tests are performed using the temporal characteristics of time series data to determine causal directionality, such as "promotional activities" preceding changes in "sales volume." The strength of causal relationships is evaluated using mutual information methods, and network stability is ensured through cross-validation, forming a multi-level causal network that reveals the causal chain of "promotional activities → sales volume → inventory level → replenishment frequency."

[0043] The business scenario analysis unit is used to dynamically adjust the analysis strategy according to changes in the business environment based on the causal relationship network output by the causal relationship unit and the adaptive business scenario analysis model based on reinforcement learning. By using key nodes in the causal relationship network (such as "sales volume", "inventory", "promotional activities" etc.) as feature dimensions of the environmental state, a multi-dimensional state space is constructed. A state transition probability model is designed using the relation weights in the causal network, and sliding window technology and isolated forest algorithm are used to monitor environmental changes, such as detecting abnormal demand fluctuations during peak sales seasons.

[0044] The environmental state is used as the observation input of the reinforcement learning agent, and the analysis strategy is modeled as action selection. The Q-learning algorithm is used to learn the state-action value function, and the reward function is calculated based on business performance indicators, such as a positive reward for a 10% increase in sales after a promotional activity. Based on the feedback after the strategy is executed, the Q network parameters are updated through an experience replay mechanism, and the optimal strategy and environmental prediction results are output.

[0045] The decision optimization output unit combines a multi-level causal relationship network and an adaptive business scenario analysis model to generate a Pareto optimal decision set through a multi-objective optimization algorithm, and outputs operational decision recommendations.

[0046] Using a causal network as a constraint framework, decision variables such as promotion budget and inventory allocation ratio are defined to construct a multi-objective optimization model. The objectives include minimizing costs, maximizing sales revenue, and maximizing inventory turnover. The NSGA-II multi-objective genetic algorithm is used to generate Pareto optimal solution sets through fast non-dominated sorting and crowding distance, which include different trade-off schemes, such as "high promotion budget + high inventory turnover" and "low cost + medium sales revenue".

[0047] The hybrid intelligent decision engine comprehensively evaluates each Pareto solution: the symbolic reasoning module verifies the logical feasibility of the solution, the statistical reasoning module assesses the probabilistic risk, and the neural reasoning module identifies potential impact patterns; it also combines causal relationship networks to simulate the propagation effect of decisions, such as how increasing the promotional budget affects sales and inventory levels.

[0048] Based on current business priorities, select the optimal solution from the Pareto solution set and output a detailed implementation plan, including promotional activity arrangements, inventory allocation suggestions, risk control measures, and expected results.

[0049] In this case, the system successfully integrated multi-source heterogeneous data from retail enterprises, constructed a unified data model, and revealed the causal relationship between promotional activities, sales, and inventory levels. Based on reinforcement learning, the adaptive analysis model dynamically adjusted promotional and inventory strategies to cope with demand fluctuations during peak sales seasons. The final Pareto optimal decision set provided enterprises with diversified operational suggestions. After implementation, sales increased by 10.5%, inventory turnover rate increased by 8%, and costs were controlled within budget, significantly improving the comprehensiveness and accuracy of operational decisions.

[0050] 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 method for intelligent analysis of operational decisions based on cross-domain data fusion, characterized in that, include: Based on the enterprise multi-domain ontology knowledge base, a unified data model is established by using a semantic mapping-based multi-source heterogeneous data dynamic fusion algorithm to perform entity recognition and relation mapping on heterogeneous data from different business systems. Based on the data model, a hybrid intelligent decision engine that integrates causal reasoning and deep learning is used for analysis and processing; Using the analysis results of the hybrid intelligent decision engine, a multi-level causal relationship network among business variables is constructed through a causal relationship discovery algorithm; Based on the aforementioned causal relationship network, an adaptive business scenario analysis model based on reinforcement learning is used to dynamically adjust the analysis strategy according to changes in the business environment. By combining a multi-level causal relationship network and an adaptive business scenario analysis model, a Pareto optimal decision set is generated through a multi-objective optimization algorithm, and operational decision recommendations are output.

2. The intelligent analysis method for operational decision-making based on cross-domain data fusion according to claim 1, characterized in that, The process of performing entity recognition and relation mapping on heterogeneous data from different business systems using a semantic mapping-based multi-source heterogeneous data dynamic fusion algorithm includes: Based on entity definitions in an enterprise multi-domain ontology knowledge base, literal similarity scores are calculated using edit distance and longest common subsequence algorithms to establish an initial candidate set for entity mapping. Utilizing cross-domain relationship mappings in the enterprise multi-domain ontology knowledge base, a pre-trained word vector model is used to calculate the semantic embedding distance of entity concepts, and semantic similarity scores are applied to the initial candidate set. According to semantic constraint rules in the enterprise multi-domain ontology knowledge base, the semantic matching results are verified and constraint consistency scores are generated through data type compatibility checks, numerical range constraint verification, and business rule consistency checks. Based on the weighted fusion results of the three-layer scores, high-confidence entity mapping relationships are confirmed and data transformation rules are established. The verified unified data model is then output to the hybrid decision engine for subsequent analysis and processing.

3. The intelligent analysis method for operational decision-making based on cross-domain data fusion according to claim 1, characterized in that, The analysis and processing using a hybrid intelligent decision engine that integrates causal reasoning and deep learning includes: The symbolic reasoning module receives entity mapping relationships and data transformation rules, and constructs a logical reasoning rule base in conjunction with enterprise business rules. It verifies data consistency and generates business-level logical analysis conclusions through a forward reasoning mechanism. The statistical reasoning module constructs a probability dependency graph between variables based on the logical analysis conclusions and historical data, and uses Bayesian reasoning to calculate the probability distribution and confidence interval of each business indicator. The neural reasoning module receives the probabilistic results from the statistical reasoning module and identifies nonlinear relationship patterns in the data through a deep neural network.

4. The intelligent analysis method for operational decision-making based on cross-domain data fusion according to claim 1, characterized in that, The process of constructing a multi-level causal relationship network among business variables using causal relationship discovery algorithms includes: By using the business logic conclusions of the symbolic reasoning module as prior constraints for causal relationships, a causal network topology that conforms to business logic is established. Based on the probability distribution results of the statistical reasoning module, statistical causal relationships between variables are identified on the basis of the topology through conditional independence testing. Combining the nonlinear relationship patterns identified by the neural reasoning module, Granger causality tests are performed using the temporal characteristics of time series data to determine the directionality of causal relationships. Through causal relationship strength assessment and network stability verification, a complete multi-level causal relationship network is formed.

5. The intelligent analysis method for operational decision-making based on cross-domain data fusion according to claim 1, characterized in that, The process of dynamically adjusting analysis strategies based on changes in the business environment using an adaptive business scenario analysis model includes: By using key causal nodes in the causal relationship network as feature dimensions of the environmental state, a multi-dimensional state space is established to represent the dynamic changes of the business environment. A state transition probability model is designed using relation weights in a causal network, and a sliding window and anomaly detection algorithm are used to monitor environmental change patterns and abnormal events in the state space. The environmental state is used as the observation input of the reinforcement learning agent, different analysis strategies are modeled as action choices, and the Q-learning algorithm is used to learn the state-action value function to optimize the strategy selection. The reward signal is calculated based on the business feedback after the strategy is executed. The Q network parameters are updated through the experience replay mechanism. At the same time, the learned optimal strategy and environment prediction results are output for decision optimization.

6. The intelligent analysis method for operational decision-making based on cross-domain data fusion according to claim 1, characterized in that, The process of generating a Pareto-optimal decision set using a multi-objective optimization algorithm and outputting operational decision recommendations includes: Using causal relationship networks as the decision constraint framework, and combining environmental state prediction and optimal strategy suggestions, a mathematical model for an optimization problem involving multiple conflicting objectives is constructed. The NSGA-II multi-objective genetic algorithm is used to search the decision space under constraints to generate a Pareto optimal solution set covering different trade-off schemes; and a hybrid intelligent decision engine and causal relationship network are invoked to conduct a comprehensive impact assessment of each Pareto solution. The optimal solution is selected from the solution set and output based on the current business priorities and the comprehensive impact assessment results.

7. An intelligent analysis system for operational decision-making based on cross-domain data fusion, characterized in that, include: The data fusion unit is used to perform entity recognition and relationship mapping on heterogeneous data from different business systems based on the enterprise's multi-domain ontology knowledge base and a dynamic fusion algorithm based on semantic mapping, thereby establishing a unified data model. The intelligent decision engine unit is used to perform data analysis based on the data model output by the data fusion unit, using a hybrid intelligent analysis and processing method that integrates causal reasoning and deep learning. The causal relationship unit is used to construct a multi-level causal relationship network between business variables using the analysis results of the intelligent decision engine unit and a causal relationship discovery algorithm. The business scenario analysis unit is used to dynamically adjust the analysis strategy according to changes in the business environment based on the causal relationship network output by the causal relationship unit and the adaptive business scenario analysis model based on reinforcement learning. The decision optimization output unit combines a multi-level causal relationship network and an adaptive business scenario analysis model to generate a Pareto optimal decision set through a multi-objective optimization algorithm, and outputs operational decision recommendations.

Images