An intelligent decision support system and method based on cognitive logic and contextualized semantics

Abstract

This invention relates to the field of enterprise management technology, and in particular to an intelligent decision support system and method based on cognitive logic and contextualized semantics. The method acquires enterprise decision-making data through multi-source data monitoring and preprocesses it to generate cross-modal enterprise scenario cognitive information; analyzes cross-modal correlations between data, integrates cognitive logic reasoning rules and semantic understanding models, and constructs contextualized semantic decision fusion features; trains an enterprise decision quality assessment model based on historical cases to perceive the quality of the current decision-making scenario and outputs decision quality information; judges the execution effect in conjunction with expected goals, and if deviations from the goals are detected, an optimization mechanism is initiated, using a historical knowledge base and multi-objective optimization algorithms to generate candidate solutions and select the optimal solution, thereby achieving intelligent adjustment of the decision-making plan. This method integrates multi-source heterogeneous data and cognitive logic, enhancing the semantic understanding, dynamic evaluation, and adaptive optimization capabilities of the decision-making system, and improving the scientific rigor and real-time performance of enterprise decisions.

Description

Technical Field

[0001] This invention relates to the field of enterprise management technology, and in particular to an intelligent decision support system and method based on cognitive logic and contextualized semantics. Background Technology

[0002] As the business environment becomes increasingly complex and dynamic, the decision-making process faces a growing number of information dimensions. Traditional decision support systems are gradually showing their limitations in dealing with complex and ever-changing business decision-making scenarios. Most existing systems rely on statistical analysis of structured data and simple matching of empirical rules, lacking the ability to deeply integrate and semantically understand multi-source heterogeneous data across departments and systems (such as internal operational data, market environment data, and user interaction behavior). This makes it difficult to achieve a comprehensive perception and dynamic understanding of business decision-making scenarios. Especially when facing unstructured text, scenarios with significant temporal fluctuations, or incomplete information, traditional methods often fail to effectively extract key decision-making elements, and are even less capable of modeling the deep relationships between multimodal data.

[0003] Furthermore, most current intelligent decision-making models remain at the data-driven, superficial prediction level, lacking effective mechanisms to integrate the cognitive logic of human experts (such as causal reasoning, prioritization, and contextual dependence) into the machine decision-making process. This results in poor model interpretability and weak generalization ability, making it difficult to support high-risk, high-complexity enterprise-level decision-making tasks. Although some studies have attempted to introduce knowledge graphs or deep learning models to improve semantic expression capabilities, a semantic gap still exists in integrating the semantics of the external environment with the internal operational status. Moreover, they rarely consider dynamic evaluation and real-time feedback optimization mechanisms for decision quality, failing to proactively adjust decision-making schemes based on actual performance.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent decision support system and method based on cognitive logic and contextualized semantics. This aims to solve the technical problems of existing intelligent decision support systems, which are unable to effectively integrate multi-source heterogeneous data and incorporate cognitive logic for deep semantic understanding and dynamic quality assessment, resulting in incomplete perception, unreliable reasoning, and untimely optimization in complex enterprise decision-making scenarios.

[0006] To achieve the above objectives, this invention provides an intelligent decision support method based on cognitive logic and contextual semantics, the method comprising:

[0007] Multi-source data monitoring is conducted on the target enterprise's decision-making scenarios to obtain enterprise decision-making monitoring data. The obtained enterprise decision-making monitoring data is preprocessed to obtain cross-modal enterprise scenario cognitive information.

[0008] Based on the cross-modal enterprise scenario cognitive information analysis, the cross-modal correlation of each monitoring data is analyzed, and the cognitive logic reasoning rules are integrated to construct semantic fusion features, thereby obtaining scenario-based semantic decision fusion features;

[0009] A corporate decision-making quality assessment model is constructed. The scenario-based semantic decision fusion features are input into the trained corporate decision-making quality assessment model to perform quality perception of the current decision-making scenario and obtain corporate decision-making quality perception information.

[0010] Based on the enterprise decision quality perception information, it is determined whether the execution effect of the current decision scenario meets the expected goal. If not, the current decision-making scheme is optimized to provide intelligent decision support for the enterprise.

[0011] Furthermore, to achieve the above objectives, the present invention also provides an intelligent decision support system based on cognitive logic and contextualized semantics, the system comprising:

[0012] The data perception module is used to monitor the decision-making scenarios of target enterprises from multiple sources to obtain enterprise decision-making monitoring data, and to preprocess the obtained enterprise decision-making monitoring data to obtain cross-modal enterprise scenario cognitive information.

[0013] The semantic fusion module is used to analyze the cross-modal correlation of each monitoring data based on the cross-modal enterprise scenario cognitive information, and to integrate cognitive logic reasoning rules to construct semantic fusion features, thereby obtaining scenario-based semantic decision fusion features;

[0014] The quality assessment module is used to construct an enterprise decision-making quality assessment model. The scenario-based semantic decision fusion features are input into the trained enterprise decision-making quality assessment model to perform quality perception of the current decision-making scenario and obtain enterprise decision-making quality perception information.

[0015] The decision optimization module is used to determine whether the execution effect of the current decision scenario meets the expected goals based on the enterprise decision quality perception information. If it does not meet the expected goals, the module optimizes the current decision scheme to provide intelligent decision support for the enterprise.

[0016] Furthermore, to achieve the above objectives, the present invention also provides an intelligent decision support device based on cognitive logic and contextual semantics. The device includes: a memory, a processor, and an intelligent decision support program based on cognitive logic and contextual semantics stored in the memory and executable on the processor. The intelligent decision support program based on cognitive logic and contextual semantics is configured to implement the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described above.

[0017] Furthermore, to achieve the above objectives, the present invention also provides a medium storing an intelligent decision support program based on cognitive logic and contextual semantics, wherein when the intelligent decision support program based on cognitive logic and contextual semantics is executed by a processor, it implements the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described above.

[0018] Furthermore, to achieve the above objectives, the present invention also provides a computer program product, including an intelligent decision support program based on cognitive logic and contextual semantics, wherein when the intelligent decision support program based on cognitive logic and contextual semantics is executed by a processor, it implements the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described above.

[0019] This invention provides an intelligent decision support method based on cognitive logic and contextualized semantics. The method achieves comprehensive and accurate cognitive perception of enterprise decision-making scenarios through multi-source data monitoring and cross-modal preprocessing. By integrating cognitive logic reasoning rules with contextualized semantic analysis, it enhances the understanding of correlations and semantic expression capabilities between multimodal data, improving the interpretability and contextual adaptability of decision features. By constructing an enterprise decision quality assessment model, it achieves dynamic perception and quantitative evaluation of the execution quality of the current decision scenario. Furthermore, it introduces an optimization mechanism when the decision effect deviates from expectations, supporting intelligent feedback and adaptive solution adjustments. This method significantly improves the scientific rigor, real-time performance, and robustness of enterprise decision-making in complex and dynamic environments, achieving a leap from data-driven to cognitively driven and semantically guided intelligent decision-making. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of the intelligent decision support method based on cognitive logic and contextualized semantics of the present invention.

[0021] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the intelligent decision support method based on cognitive logic and contextual semantics of the present invention. An embodiment of the intelligent decision support method based on cognitive logic and contextual semantics of the present invention is proposed.

[0024] In one embodiment, the intelligent decision support method based on cognitive logic and contextual semantics includes:

[0025] In one embodiment, an intelligent decision support method based on cognitive logic and contextual semantics is provided, the intelligent decision support method based on cognitive logic and contextual semantics includes:

[0026] Step S100: Multi-source data monitoring is performed on the target enterprise decision-making scenario to obtain enterprise decision-making monitoring data. The obtained enterprise decision-making monitoring data is preprocessed to obtain cross-modal enterprise scenario cognitive information.

[0027] The target enterprise decision-making scenario can be a specific business context within the enterprise that requires decision-making, encompassing specific objectives, constraints, and contextual environment. It can serve as the input boundary and starting point for intelligent decision support methods, defining the scope of data collection and semantic modeling. In an exemplary embodiment, the target enterprise decision-making scenario may include, but is not limited to, one or more scenarios such as supply chain disruption emergency response, new product market launch strategy formulation, and customer churn risk intervention. Multi-source data can be a heterogeneous collection of data from different systems, departments, or external environments, including structured, semi-structured, and unstructured forms. It can be used to provide original observational evidence of enterprise operational status and changes in the external environment. For example, multi-source data may include internal ERP system log data, social media sentiment text, and IoT device time-series sensor data.

[0028] Enterprise decision monitoring data can be a set of raw observation data obtained from multi-source data monitoring of the target enterprise's decision-making scenarios. Furthermore, enterprise decision monitoring data can be acquired in real-time or in batches from multi-source data through API integration, log collection, web crawling, or sensor streams, and can be used as input for the preprocessing stage, carrying raw but unaligned multimodal information. Cross-modal enterprise scenario cognitive information can be a structured information set that has been preprocessed, possesses a unified semantic representation, and retains multimodal characteristics. In a specific embodiment, cross-modal enterprise scenario cognitive information can be generated through cross-modal preprocessing techniques such as modality alignment, missing value imputation, noise filtering, and semantic normalization, and can be used to provide a consistent and computable input foundation for subsequent correlation analysis and semantic fusion.

[0029] Step S200: Analyze the cross-modal correlation of each monitoring data based on cross-modal enterprise scenario cognitive information, and integrate cognitive logic reasoning rules to construct semantic fusion features, thereby obtaining scenario-based semantic decision fusion features.

[0030] Cross-modal relationships can be implicit connections between different modalities of data (such as text, numerical values, images, and time series) at the semantic or causal level. Furthermore, cross-modal relationships can be mined using graph neural networks, attention mechanisms, or multi-view learning models to uncover co-occurrence patterns or dependency structures between modalities. This can be used to reveal the deep coupling logic between multi-source heterogeneous data, supporting the rationality of semantic fusion. In a specific embodiment, cross-modal relationships can be a process that works in conjunction with cognitive logic reasoning rules to construct semantic fusion features. Cognitive logic reasoning rules can be formally expressed causal, priority, or context-dependent reasoning logic used by human experts in specific decision-making scenarios. For example, cognitive logic reasoning rules can be extracted from expert interviews, historical decision logs, or domain ontologies using knowledge engineering methods and encoded into a rule base. This can be used to endow machine decision-making with human-understandable reasoning paths, enhancing model interpretability.

[0031] Semantic fusion features can be intermediate representation vectors generated by fusing cross-modal relationships and cognitive logical reasoning rules. In an exemplary embodiment, semantic fusion features can be generated through joint embedding space mapping, rule-guided attention weighting, or graph structure enhancement, and can serve as a direct basis for constructing contextualized semantic decision fusion features. Contextualized semantic decision fusion features can be high-order feature representations that combine specific decision scenario context, multimodal semantics, and cognitive logic. Furthermore, contextualized semantic decision fusion features can be dynamically modulated by introducing scenario meta-information (such as time windows, business types, and risk levels) on top of semantic fusion features, which can be used to improve the adaptability and semantic richness of decision features to the current context, supporting high-quality evaluation. In a specific embodiment, contextualized semantic decision fusion features can serve as input features for an enterprise decision quality assessment model.

[0032] Step S300: Construct an enterprise decision quality assessment model. Input the scenario-based semantic decision fusion features into the trained enterprise decision quality assessment model to perform quality perception of the current decision scenario and obtain enterprise decision quality perception information.

[0033] The enterprise decision-making quality assessment model can be a machine learning or rule-driven model used to quantitatively evaluate whether the execution effect of the current decision-making plan meets the expected goals. Furthermore, the enterprise decision-making quality assessment model can be obtained through supervised learning or reinforcement learning training based on historical decision result labels (success / failure, degree of deviation, etc.), which can be used to achieve dynamic and quantitative perception of decision execution quality. In an exemplary embodiment, the enterprise decision-making quality assessment model can include, but is not limited to, one or more of the following: regression-based quality assessment models, categorical quality assessment models, and multi-objective Pareto assessment models.

[0034] The current decision-making scenario can be a real-time or near-real-time enterprise decision instance being evaluated and supported, and can be used as the object of quality assessment and feedback optimization. The expected goal can be the business indicators or states that the enterprise hopes to achieve in a specific decision-making scenario, and can be used as a benchmark to judge whether the decision effect deviates from the target. The current decision-making scheme can be a set of specific action strategies that have been generated or are being implemented for the current decision-making scenario, and can be used as a direct modification object for optimization operations. Enterprise decision quality perception information can be a quantitative or qualitative feedback signal about the current decision effect output by the enterprise decision quality assessment model. In a specific embodiment, the enterprise decision quality perception information can be used by the model to infer the contextualized semantic decision fusion features to generate scores, confidence levels, or deviation indicators, which can be used as the basis for judging whether to trigger the optimization mechanism. Furthermore, the enterprise decision quality perception information can be the initiation condition driving the optimization of the current decision-making scheme.

[0035] Step S400: Based on the enterprise decision quality perception information, determine whether the execution effect of the current decision scenario meets the expected goal. If not, optimize the current decision plan to provide intelligent decision support for the enterprise.

[0036] Enterprise intelligent decision support can be provided by the system to decision-makers with cognitively enhanced and semantically guided optimization suggestions or automatic adjustment instructions, which can be used to assist or replace manual completion of highly complex and high-risk decision-making tasks. Judging whether the execution effect of the current decision-making scenario meets the expected goals based on enterprise decision quality perception information can be achieved by comparing the quality perception information with preset expected goal thresholds or rules to determine whether there are significant deviations. In an exemplary embodiment, this operation can set a dynamic threshold that is automatically adjusted according to the business cycle, or use fuzzy matching rules to handle the uncertainty of the goal description, thereby triggering or inhibiting the activation of the optimization mechanism.

[0037] If the decision does not meet the requirements, the current decision-making scheme is optimized to provide intelligent decision support for the enterprise. This can be achieved by calling an optimization algorithm to adjust the parameters or structure of the current decision-making scheme when the execution effect deviates from expectations. Furthermore, this operation can fine-tune the strategy parameters based on gradient backpropagation, or retrieve and replace a better strategy implementation from the alternative scheme library, thereby realizing closed-loop adaptive decision adjustment.

[0038] Taking the dynamic adjustment of retail enterprise promotion strategies as an example, the intelligent decision support method based on cognitive logic and contextual semantics in this embodiment can be as follows: During a large-scale e-commerce promotion, the system monitors multiple signals such as a decrease in inventory turnover, a surge in negative comments on social media, and a sudden drop in competitor prices; after cross-modal preprocessing, unified cognitive information is formed; through analysis, it is found that there is a strong correlation between user emotional fluctuations and inventory backlog, and combined with the cognitive rule of "prioritizing clearance of high inventory", a contextual semantic decision fusion feature emphasizing clearance priority is constructed; this feature is input into a trained quality assessment model, and the output shows that the quality score of the current promotion plan is lower than the threshold; the system determines that the effect has not met expectations, automatically triggers the optimization mechanism, adjusts the original full reduction strategy to a combination of limited-time discount and gifts, and pushes it to the operations team for execution.

[0039] In one embodiment, multi-source data monitoring is performed on the target enterprise decision-making scenario to obtain enterprise decision-making monitoring data. The obtained enterprise decision-making monitoring data is then preprocessed to obtain cross-modal enterprise scenario cognitive information, specifically including:

[0040] In the target enterprise decision-making scenario, the integrated data interface is used to monitor the business system in real time to obtain enterprise decision-making monitoring data, which includes internal operation monitoring data, external environment monitoring data and interactive behavior monitoring data.

[0041] The integrated data interface can be a standardized communication middleware that unifies access to data sources from multiple business systems, enabling integrated collection of multi-dimensional monitoring data on internal operations, external environment, and interactive behaviors. In an exemplary embodiment, the integrated data interface can aggregate real-time or batch data streams from different systems through API gateways, message queues, or ETL tools. For example, the integrated data interface may include, but is not limited to, one or more of the following: RESTful API integration interface, Kafka message bus interface, and database CDC synchronization interface. Business systems can be a collection of information application systems used by an enterprise to support daily operations, serving as the primary source of enterprise decision-making monitoring data. Further, business systems may include, but are not limited to, one or more of the following: ERP system, CRM system, and SCM system. Internal operational monitoring data can be structured or semi-structured data reflecting the internal processes, resources, and performance status of an enterprise, providing objective evidence of the organization's internal operational status. In a specific embodiment, internal operational monitoring data may include, but is not limited to, one or more of the following: production work order completion rate, inventory turnover days, and financial cash flow indicators.

[0042] External environment monitoring data can be macro-level or competitive information originating outside the enterprise's boundaries and influencing business decisions. It can be used to capture the impact of changes in the market, policy, or technological environment on the enterprise. For example, external environment monitoring data may include, but is not limited to, one or more of the following: industry price indices, policy and regulatory texts, and competitor marketing activity logs. Interaction behavior monitoring data can be behavioral trajectory data recording the interaction process between users, customers, or partners and the enterprise's systems. It can be used to reveal the micro-dynamics of demand preferences and experience feedback. Furthermore, interaction behavior monitoring data may include, but is not limited to, one or more of the following: app clickstream logs, customer service dialogue texts, and webpage dwell time sequences.

[0043] The enterprise decision monitoring data is decomposed into wavelet packets using preset wavelet basis functions to obtain different sub-bands and calculate the Shannon entropy of each sub-band. Based on the calculated Shannon entropy, hard threshold filtering is performed to extract key semantic frequency bands.

[0044] The preset wavelet basis function can be a pre-selected mother wavelet function used for wavelet transform, possessing specific time-frequency localization characteristics. It can be used to support multi-resolution decomposition of non-stationary time-series signals, preserving key semantic frequency bands. In an exemplary embodiment, the preset wavelet basis function can be selected from standard wavelet libraries (such as Daubechies and Symlets) based on signal characteristics. For example, the preset wavelet basis function can include, but is not limited to, one or more of the Daubechies wavelet basis, Coiflet wavelet basis, and Morlet complex wavelet basis. Wavelet packet decomposition can be a fully binary tree-like time-frequency decomposition method that recursively subdivides the signal into high- and low-frequency subbands, which can be used to characterize the signal's frequency domain structure more comprehensively than traditional wavelet transform. Furthermore, wavelet packet decomposition can decompose the original signal into multiple fine subbands by iteratively applying wavelet filter banks. In a specific embodiment, wavelet packet decomposition can provide the energy distribution basis for each subband for Shannon entropy calculation. Different subbands can be signal components with different frequency ranges and time resolutions obtained after wavelet packet decomposition, which can be used as basic units for information content evaluation and frequency band selection.

[0045] Shannon entropy can be an information-theoretic metric for measuring the uncertainty of signal sub-band information. It can be used to quantify the richness of effective semantic information contained in each sub-band, guiding the extraction of key frequency bands. In an exemplary embodiment, Shannon entropy can be used to calculate the information entropy value based on the probability distribution of sub-band coefficients. Key semantic frequency bands can be a set of sub-bands that are determined to have high information content after Shannon entropy evaluation. They can be used to retain signal components containing decision-related semantics and eliminate redundant or noisy frequency bands. Furthermore, key semantic frequency bands can be filtered out by setting an entropy threshold to select sub-bands with values ​​higher than the threshold. Hard threshold filtering can be a denoising method that truncates wavelet coefficients by amplitude truncation, setting those below the threshold to zero. It can be used to achieve frequency domain sparsity reconstruction oriented towards semantic preservation. In a specific embodiment, hard threshold filtering can set a threshold based on the Shannon entropy ranking result, retaining only the wavelet coefficients of key semantic frequency bands.

[0046] After filtering, the enterprise decision monitoring data is synchronized with time series. Bilateral filtering is used to eliminate data noise and moving average smoothing is used to eliminate data fluctuations. A dynamic window algorithm is introduced to slide based on a preset time window and calculate the confidence interval and coefficient of variation of the data within the window. Data repair is performed in combination with spline interpolation algorithm.

[0047] Among these, time-series synchronization processing can be an operation to align the time axis of time series acquired asynchronously from multiple sources, ensuring comparability and consistency of cross-modal data in the time dimension. In an exemplary embodiment, time-series synchronization processing can unify the timestamp reference of each data stream through interpolation, resampling, or event alignment mechanisms. Bilateral filtering can be a nonlinear smoothing filter that simultaneously considers spatial proximity and numerical similarity, used to suppress noise while preserving abrupt changes (such as inflection points and jumps) in the signal. Furthermore, bilateral filtering can weight the data points within a sliding window, with the weights determined by both time and numerical distance. Moving average smoothing processing can be a linear filtering method that takes the average within a local window of the time series to weaken short-term fluctuations, used to eliminate high-frequency random fluctuations and highlight long-term trends. In a specific embodiment, moving average smoothing processing can use a fixed or adaptive window length to calculate the moving average sequence. Dynamic window algorithms can be time-series processing strategies that automatically adjust the size or position of the analysis window based on local data characteristics, used to achieve local quality perception and adaptive processing of non-stationary sequences. Furthermore, dynamic window algorithms can slide based on a preset time window and calculate statistical indicators within each window to evaluate data stability. For example, dynamic window algorithms may include, but are not limited to, one or more of the following: sliding fixed window algorithm, adaptive variable length window algorithm, event-triggered window algorithm, etc.

[0048] A preset time window can be a fixed or configurable time span used for local statistical calculations in a dynamic window algorithm, which can be used to define the observation range for local data quality assessment. A confidence interval can be the range of values ​​within which the estimated true value of a parameter might fall at a given confidence level, which can be used to quantify the reliability range of local data and assist in anomaly detection. In an exemplary embodiment, the confidence interval can be calculated based on the sample mean and standard deviation within the window to determine the statistical confidence boundary. The coefficient of variation can be the ratio of the standard deviation to the mean, a dimensionless indicator used to measure the relative dispersion of data, which can be used to assess the severity of local data fluctuations relative to its average level. Further, the coefficient of variation can be calculated within the dynamic window as the ratio of the standard deviation to the mean. A spline interpolation algorithm can be a method for smoothly reconstructing missing or outlier data points using a piecewise polynomial function, which can be used to repair incomplete data while maintaining the continuity of the overall trend. In a specific embodiment, the spline interpolation algorithm can fit a cubic spline curve based on nearby valid points to estimate missing values.

[0049] Data repair can be the process of reasonably filling in or correcting missing, anomalous, or unreliable observations. It can improve data integrity and usability, and enhance the robustness of the system in scenarios with missing information. Furthermore, data repair can combine dynamic window quality assessment results with spline interpolation to generate alternative values. Time-series synchronization processing of enterprise decision-making monitoring data after filtering can involve resampling or interpolating and aligning multi-source data streams according to a unified time benchmark. This operation can ensure strict alignment of cross-modal data in the time dimension. Using bilateral filtering to eliminate data noise can involve weighted averaging of data points within the time neighborhood, with weights determined by both time distance and numerical differences. This operation can smooth noise while preserving key abrupt changes. Eliminating data fluctuations through moving average smoothing can involve applying a sliding window mean to time-series data to weaken short-term random fluctuations. This operation can highlight long-term trends and improve data stability.

[0050] After the enterprise decision monitoring data is repaired, Z-score standardization is performed to convert continuous data into a distribution with a mean of 0 and a standard deviation of 1. Semantic bucketing is used to perform contextual encoding on discrete data to generate cross-modal enterprise scenario cognitive information.

[0051] Z-score standardization can be a linear transformation that converts continuous variables into a standard normal distribution with a mean of 0 and a standard deviation of 1. It can be used to eliminate dimensional differences, making different indicators comparable during the fusion stage. In an exemplary embodiment, Z-score standardization can normalize each feature according to its historical mean and standard deviation. Continuous data can be numerical observation variables that can take any value within the real number range, and can be used as the object of Z-score standardization. A distribution with a mean of 0 and a standard deviation of 1 can be a dimensionless standard normal distribution formed after Z-score standardization, and can be used to provide scale-consistent input for subsequent multimodal fusion. Discrete data can be non-continuous variables with values ​​in a finite number of categories or integers, and can be used as the object of semantic binning.

[0052] Context encoding can map discrete categories to vector representations carrying semantic association information, enabling semantic computability of categorical variables. In one specific embodiment, context encoding can construct embedding or bucketing rules based on business context (such as time, scenario, co-occurrence relationships). Semantic bucketing can cluster discrete values ​​into semantically consistent buckets based on semantic similarity or business logic and encode them, enhancing the semantic expressiveness of discrete features and supporting cross-modal semantic alignment. Furthermore, semantic bucketing can utilize domain knowledge or clustering algorithms to group discrete values, assigning a unique embedding vector to each group. For example, semantic bucketing can include, but is not limited to, one or more of the following: semantic bucketing based on business rules, semantic bucketing based on co-occurrence graphs, and semantic bucketing based on pre-trained language models. Performing Z-score standardization on the repaired enterprise decision monitoring data can normalize each continuous feature according to its historical mean and standard deviation. Furthermore, this operation can eliminate the influence of dimensions and unify the numerical scale. Converting continuous data into a distribution with a mean of 0 and a standard deviation of 1 can be achieved by applying the Z-score formula to perform a linear transformation. Furthermore, this operation can generate standardized inputs, which facilitates subsequent model processing.

[0053] For example, in the scenario of equipment fault early warning and maintenance decision-making in manufacturing, the intelligent decision support method based on cognitive logic and scenario-based semantics in this embodiment can be as follows: In a smart factory, the system collects PLC sensor time-series data (internal operations), supply chain delivery delay early warning (external environment), and maintenance work order text (interactive behavior) in real time through an integrated data interface; the vibration signal is decomposed into wavelet packets using the Symlets wavelet basis, and the high-entropy sub-band is retained after calculating the Shannon entropy of each sub-band to extract the fault-sensitive frequency band; then, the multi-source data is synchronized in time, and bilateral filtering is used to retain the impact characteristics and moving average is used to smooth the high-frequency jitter; a sudden increase in the coefficient of variation and the expansion of the confidence interval are detected by the dynamic window algorithm, and it is determined that there is a data anomaly, and spline interpolation is called to repair the missing points; finally, the Z-score of continuous variables such as temperature and speed is standardized, and the fault code is mapped to the semantic vector of "bearing-high frequency vibration" through semantic bucketing, and finally cross-modal enterprise scenario cognitive information is generated to support subsequent fault root cause reasoning and maintenance strategy generation.

[0054] In one embodiment, cross-modal correlations of various monitoring data are analyzed based on cross-modal enterprise scenario cognitive information, and semantic fusion features are constructed by incorporating cognitive logic reasoning rules to obtain scenario-based semantic decision fusion features, specifically including:

[0055] Acquire cross-modal enterprise scenario cognitive information, extract interactive behavior monitoring data and parse enterprise interactive text monitoring data through cross-modal enterprise scenario cognitive information, and input enterprise interactive text monitoring data into BERT pre-trained model to obtain context embedding vectors;

[0056] The cross-modal enterprise scenario cognitive information can be an aligned multimodal dataset, serving as a unified input source for this refined process. Interaction behavior monitoring data can be raw records of inter-subject interactions collected from internal enterprise systems or user touchpoints, such as clickstreams, conversation logs, and approval process trajectories, reflecting the actual behavioral patterns of decision-making stakeholders and providing behavioral context for intent recognition. For example, interaction behavior monitoring data may include, but is not limited to, one or more of customer service dialogue logs, internal approval operation sequences, and user page browsing paths. Enterprise interaction text monitoring data can be unstructured text interaction content generated during enterprise operations, including customer service dialogues, meeting minutes, and email correspondence, carrying decision intent and semantic information, and serving as the basic input for constructing the first semantic feature. In an exemplary embodiment, enterprise interaction text monitoring data may include, but is not limited to, customer complaint ticket texts, cross-departmental collaborative emails, and transcribed sales negotiation recordings.

[0057] The BERT pre-trained model is a bidirectional encoder language model based on the Transformer architecture. It has been pre-trained on large-scale corpora to generate context embedding vectors that preserve context dependencies, improving the accuracy of semantic role recognition. Furthermore, the BERT pre-trained model can load publicly available or domain-fine-tuned BERT weights to generate context-sensitive vector representations of the input text. Context embedding vectors can be dense vectors output by the pre-trained language model, representing the semantics of words in a specific context, and used as the numerical basis for subsequent semantic role labeling and relation density calculation.

[0058] The decision intent index is constructed by weighted fusion of key semantic role annotations and entity relationship density using context embedding vectors. Subsequently, a knowledge entity network of enterprise interactive text is constructed using cognitive graph algorithm. The ratio of network node degree to edge weight is calculated as the semantic association strength index. Combined with the sentiment features of enterprise operational text, the first semantic feature information is generated.

[0059] Key semantic role labeling can be the process and result of identifying components in text that perform specific semantic functions (such as agent, patient, target, cause, etc.), used to structurally express the logical relationships between decision-making subjects, actions, and objects. In a specific embodiment, key semantic role labeling can be implemented using a sequence labeling model (such as BiLSTM-CRF) or a prompt-based method based on context embedding vectors. Entity relation density can be the number of semantic relations between entities identified per unit text length, reflecting the density of text information, and is used as a weighting factor in the construction of the decision intent index to enhance the contribution of high-information text. Furthermore, entity relation density can be calculated and normalized using a relation extraction model to count the number of effective triples. The decision intent index can be a quantitative indicator formed by fusing key semantic role labeling and entity relation density, characterizing the strength of the implicit decision tendency in the text, and used to provide semantic signals guiding intent for the first semantic feature. In an exemplary embodiment, the decision intent index can be achieved by weighted fusion (such as weighted summation or product) of role label confidence and relation density.

[0060] Cognitive graph algorithms can be methods that combine domain knowledge and semantic parsing techniques to automatically construct structured knowledge networks from unstructured text, transforming enterprise interaction text into a computable knowledge entity network. Furthermore, cognitive graph algorithms can generate graph structures with types and attributes by integrating named entity recognition, relation extraction, and ontology alignment modules. The knowledge entity network can be a directed graph structure composed of entities extracted from enterprise interaction text and their semantic relationships, serving as the topological basis for calculating the strength of semantic associations. The network node degree can be the number of edges connected to a node in the knowledge entity network, reflecting the centrality or activity of that entity, used to measure the importance of the entity in the semantic network. The edge weight can be the semantic strength or confidence value carried by the edge connecting two nodes in the knowledge entity network, used to characterize the reliability or closeness of the relationship between entities.

[0061] The semantic association strength index can be a normalized index obtained by calculating the ratio of network node degree to edge weight, used to quantify the semantic influence of an entity in the network. Further, the semantic association strength index can be standardized by calculating the degree of each node and dividing it by the average edge weight (or a weighted average). Enterprise operational text sentiment features can be emotional tendencies, attitude polarities, or satisfaction indicators extracted from internal enterprise texts (such as work orders and reports), used to enhance the perception of subjective states by the first semantic feature. In a specific embodiment, enterprise operational text sentiment features can output sentiment scores or categories through sentiment analysis models (such as FinBERT or domain-adjusted RoBERTa). For example, enterprise operational text sentiment features may include, but are not limited to, customer satisfaction sentiment scores, employee emotional fluctuation indices, and supplier cooperation willingness sentiment tags. The first semantic feature information can be an internal semantic representation formed by fusing the decision intention index, semantic association strength index, and enterprise operational text sentiment features, used to represent the decision semantic state in the internal interactive context of the enterprise, serving as a query vector for the cognitive attention mechanism. Further, the first semantic feature information can integrate multi-dimensional semantic signals through vector concatenation, weighted fusion, or gating mechanisms.

[0062] By acquiring external environment monitoring data, market sentiment data is extracted and TF-IDF is applied to generate a sentiment keyword matrix. The sentiment keyword matrix is ​​then input into a pre-trained cognitive reasoning engine to complete the knowledge graph and generate a semantic ontology feature vector for the enterprise scenario, thus obtaining the second semantic feature information.

[0063] External environment monitoring data can be observational data from outside the enterprise, reflecting market, policy, or social dynamics, used to provide external semantic input for the construction of second semantic features. Market sentiment data can be public discussion content about the enterprise, industry, or competitors collected from news, social media, forums, etc., used to reflect the potential impact of the external environment on enterprise decision-making. For example, market sentiment data can include, but is not limited to, industry policy interpretation texts, competitor product evaluation posts, and macroeconomic news summaries. TF-IDF can be a statistical method used to assess the importance of a word in a document set, used to extract a distinctive keyword matrix from sentiment text. Furthermore, TF-IDF can generate keyword weights by calculating the product of term frequency (TF) and inverse document frequency (IDF). The sentiment keyword matrix can be a sparse numerical matrix generated by TF-IDF, with rows representing documents and columns representing keywords, used as structured input for the cognitive reasoning engine, supporting knowledge graph completion.

[0064] The pre-trained cognitive reasoning engine can be a neural symbolic system pre-trained on a large-scale knowledge graph, possessing relational reasoning and missing link prediction capabilities. It is used to map the keyword matrix to the enterprise scenario ontology space, generating structured semantic vectors. Furthermore, the pre-trained cognitive reasoning engine can be trained based on TransE, ComplEx, or neural logic programming frameworks, supporting graph completion under ontology constraints. Knowledge graph completion can be a process of predicting missing relations or entities based on existing triples, used to enrich external semantic representations and bridge the semantic gap between keywords and ontology concepts. The semantic ontology feature vector of the enterprise scenario can be a semantic embedding vector generated after knowledge graph completion, conforming to the enterprise domain ontology structure, used as a core component of the second semantic feature. The second semantic feature information can be an external environment semantic representation composed of the semantic ontology feature vector of the enterprise scenario, used to represent the external market and social context, serving as a key / value vector for the cognitive attention mechanism.

[0065] A cognitive attention mechanism is introduced, using the first semantic feature information as the query vector and the second semantic feature information as the key vector and value vector. The attention weight distribution is constrained by cognitive logic rules, and the rule-weighted attention score is calculated to perform feature fusion and generate semantically enhanced feature information.

[0066] The cognitive attention mechanism can be an improved attention calculation module that introduces cognitive logic rule constraints on top of the standard attention mechanism, ensuring that attention focuses on semantically aligned regions that conform to expert reasoning logic. Furthermore, the cognitive attention mechanism can apply a rule-guided bias term or mask to the original attention score before Softmax. The query vector can be a vector used to retrieve relevant information in the attention mechanism; in this scheme, it is provided by the first semantic feature information and used to drive directional attention to external semantic information. The key vector can be an index vector used to match the query vector in the attention mechanism; it is provided by the second semantic feature information and used to support semantic similarity calculation. The value vector can be a vector carrying actual information content in the attention mechanism; it is provided by the second semantic feature information and used to form the output features after weighting.

[0067] Cognitive logic rules can be formally expressed expert reasoning constraints used to limit the legitimate distribution range of attention weights, ensuring that the output of the attention mechanism conforms to human cognitive logic such as causality, priority, or contextual dependence. For example, cognitive logic rules can include, but are not limited to, causal direction constraint rules, business compliance filtering rules, and risk avoidance priority rules. Attention weight distribution can be a normalized matching score vector of each key-query pair in the attention mechanism, used to determine the fusion ratio of the value vector. Rule-weighted attention score can be the attention score modified under the constraints of cognitive logic rules, used for feature fusion. Furthermore, the rule-weighted attention score can be normalized by adding the original attention score to a rule mask or bias term and then adding the result. Semantic enhancement feature information can be an enhanced semantic representation generated by fusing the first and second semantic features through the cognitive attention mechanism, used as a high-level semantic input for cross-modal fusion, bridging the semantic gap between internal and external systems. Furthermore, semantic enhancement feature information can be input into the cognitive graph neural network along with internal operational monitoring data and external environmental monitoring data.

[0068] By leveraging cross-modal enterprise scenario cognitive information to extract internal operational monitoring data and external environmental monitoring data, and combining semantic enhancement feature information, the data is input into a cognitive graph neural network for cross-modal semantic fusion, resulting in scenario-based semantic decision fusion features.

[0069] Internal operational monitoring data can be structured business operation data collected from enterprise ERP, CRM, SCM, and other systems, used to provide a quantitative baseline for decision-making scenarios. The cognitive graph neural network (GNN) can be a neural network model integrating cognitive logic rules and graph structure learning capabilities, supporting semantic propagation of multimodal nodes and edges. It is used to achieve deep cross-modal fusion of semantically enhanced features and original monitoring data, outputting scenario-based semantic decision fusion features. Furthermore, the GNN can embed rule-guided aggregation functions or node update strategies during message passing. Examples include, but are not limited to, rule-constrained graph attention networks, causal perceptive graph convolutional networks, and context-aware heterogeneous graph neural networks.

[0070] Utilizing cross-modal enterprise scenario cognitive information to extract internal operational monitoring data and external environmental monitoring data can separate structured operational data from external non-textual monitoring data (such as prices and indices) from cognitive information. Furthermore, this operation can be achieved through data pattern recognition and field mapping, thus providing raw modal input to graph neural networks. Combining semantically enhanced feature information into the cognitive graph neural network for cross-modal semantic fusion can involve using semantically enhanced features as node attributes or edge features to construct a heterogeneous graph with the raw monitoring data and perform message passing. Further, this operation can employ heterogeneous graph neural networks to process multiple types of nodes or insert rule verification modules between GNN layers, thereby achieving deep fusion of semantic, numerical, and temporal modalities. Obtaining scenario-based semantic decision fusion features can be achieved by reading the fused node or graph-level representation from the final layer of the cognitive graph neural network. Further, this operation can be achieved through global pooling or node embedding projection, thus outputting high-order decision features with context adaptability and cognitive interpretability.

[0071] For example, in scenarios involving dynamic optimization of financial risk control strategies, the intelligent decision support method based on cognitive logic and contextual semantics in this embodiment can be used by banks to monitor customer repayment delays, surges in customer service complaints, and negative public opinion on social media during post-loan management. The system extracts phrases such as "customers complain about high interest rates" from interactive text, encodes them using BERT, identifies the semantic roles of "customer-complaint-interest rate," and calculates high relation density. Simultaneously, a knowledge network is constructed to discover that the "interest rate" node has high degree but low edge weight, indicating concentrated controversy but weak consensus, which, combined with negative sentiment, generates the first semantic feature. External public opinion is extracted using TF-IDF with keywords such as "interest rate hike" and "default wave," which are then supplemented into the macroeconomic ontology through a cognitive reasoning engine to generate the second semantic feature. Under the constraint of the "risk aversion priority" rule, the cognitive attention mechanism strengthens the focus on external signals related to the "default wave," generating semantically enhanced features. Finally, this feature, along with operational data such as delinquency rates and customer assets, is input into a cognitive graph neural network, outputting a contextual semantic decision fusion feature emphasizing liquidity risk, triggering early collection and interest rate negotiation strategies.

[0072] In one embodiment, internal operational monitoring data and external environmental monitoring data are extracted using cross-modal enterprise scenario cognitive information, and then combined with semantic enhancement feature information and input into a cognitive graph neural network for cross-modal semantic fusion to obtain scenario-based semantic decision fusion features, specifically including:

[0073] In the time dimension, the data input to the cognitive graph neural network is transformed into dynamic temporal cognitive nodes. Each dynamic temporal cognitive node contains the enterprise's operational indicators, public opinion hotspots, business interaction events, and environmental risk parameters at the current moment.

[0074] In this context, dynamic temporal cognitive nodes can be structured time-step units organized along the time dimension, containing multimodal enterprise state elements. They can serve as basic processing units in temporal modeling within cognitive graph neural networks, carrying cross-modal dynamic information. In an exemplary embodiment, dynamic temporal cognitive nodes may include, but are not limited to, one or more of the following: operational indicator-driven cognitive nodes, public opinion event-triggered cognitive nodes, and risk parameter mutation-driven cognitive nodes. Enterprise operational indicators can be quantitative metrics reflecting the internal business operation status of an enterprise, such as inventory turnover rate, order fulfillment rate, and cash flow. These can be used to provide a quantitative baseline for decision-making scenarios and to assess changes in business status.

[0075] Public opinion hotspots can be topics or events that are the focus of public or media discussion about a company, industry, or competitors within a specific time period. They can be used to characterize potential pressures or opportunities from the external environment on a company's decision-making. Business interaction events can be discrete interactions with business significance that occur between internal and external entities, such as contract signing, complaint submission, and approval. They can be used to provide key event anchors in the decision-making context and enhance the granularity of temporal causal modeling.

[0076] Environmental risk parameters can be quantitative or categorical variables that describe external uncertainties, such as policy change indices, supply chain disruption probabilities, and exchange rate volatility. They can be used to characterize the macro or meso-level risk context in which decisions are made.

[0077] The forward cognition layer captures the instantaneous impact of decision adjustments on the business state through confidence gating, while the backward cognition layer uses causal strength gating to filter the long-term impact of decision adjustments on the business state, generating time-series decision causal features.

[0078] The forward cognition layer can be the forward propagation module in a cognitive graph neural network responsible for modeling the immediate impact of decision adjustments on the business state. It can be used to capture short-term, high-confidence responsiveness. In one specific embodiment, the forward cognition layer can introduce a confidence gating mechanism during the time-series message passing process to adjust the contribution of the current decision disturbance to the state update.

[0079] Confidence gating is a gating mechanism that dynamically adjusts the amount of information passing through based on current observation consistency or model uncertainty. It can be used to suppress transient misjudgments under noise interference and enhance the reliability of short-term responses. For example, confidence gating can calculate the confidence score of input features (such as the inverse of prediction variance or consistency index) to scale the state update signal.

[0080] The backward cognition layer can be a reverse modeling module in a cognitive graph neural network used to backtrack and filter decision paths with long-term causal influence, and can be used to identify decision factors that truly drive long-term business trends. Furthermore, the backward cognition layer can introduce a causal strength assessment mechanism in temporal backpropagation or attention backtracking. The causal strength gate can be a gating unit used to quantify and filter the strength of causal effects between variables, and can be used to filter out spurious correlations and retain causal paths with long-term explanatory power. In an exemplary embodiment, the causal strength gate can estimate the magnitude of causal effects based on Granger causality, intervention-response simulation, or structural causality models, and weight the information flow accordingly.

[0081] The causal features of time-series decisions can be represented by a time-series feature representation that integrates forward instantaneous response and backward long-term causal effects. This can be used to provide a time-dimensional representation of the final decision features that is both timely and causally interpretable. For example, the causal features of time-series decisions can be achieved by concatenating or weighting the outputs of the forward and backward cognitive layers. The forward cognitive layer captures the instantaneous impact of decision adjustments on the business state through confidence gating. This can be achieved by introducing information flow regulated by confidence gating into the time-series state update, allowing only high-confidence perturbations to pass. Furthermore, this operation can be implemented by dynamically adjusting the gating opening based on the variance of the prediction residuals or by generating confidence using multi-model consensus voting, thereby improving robustness to short-term response signals. The backward cognitive layer uses causal strength gating to filter the long-term impact of decision adjustments on the business state. This can be achieved by filtering weak causal paths based on causal strength gating in backpropagation or time-series attention. Furthermore, this operation can be implemented by estimating the strength of causal effects based on intervention simulations or by using prior constraints on gating parameters in structural causal models, thereby enhancing the causal interpretability of long-term effect modeling. Generating causal features for time-series decisions can be achieved by merging the output feature vectors of the forward and backward cognitive layers, thereby forming a time-series representation that integrates short- and long-term effects.

[0082] In the spatial dimension, the data input to the cognitive graph neural network is aggregated into topological relationships using graph convolutional layers to obtain graph feature representations. The graph feature representations are then subjected to global weighted pooling based on node importance to generate semantically saliency weight vectors.

[0083] In this context, the graph convolutional layer can be a foundational layer in a cognitive graph neural network, performing neighborhood information aggregation. It can capture topological dependencies between nodes and generate feature representations that are aware of local structure. Topological relationship aggregation can be a process of weighting and summarizing neighboring node features based on edge weights or types, enabling the local diffusion and integration of spatial semantic information. The graph feature representation can be a vector representation reflecting the semantic position of a node in the graph structure, obtained after processing by the graph convolutional layer. It can be used as the basic input for subsequent spatial weighting and pooling. Global weighted pooling of node importance can be an operation that weights and aggregates the features of a node into a global vector based on its semantic saliency throughout the graph. It can be used to generate a compact vector that represents the semantic focus of the entire graph. In one specific embodiment, global weighted pooling of node importance can calculate the importance score of each node (such as PageRank, centrality, or task-related attention weights) for weighted summation. The semantic saliency weight vector can be a weight distribution representing the semantic saliency of each node, generated by global weighted pooling of node importance. It can be used to guide subsequent adaptive weighting of graph features.

[0084] The temporal hidden states of the cognitive graph neural network are obtained and fused with the semantic saliency weight vector using tensors. The importance score of each cognitive node in the decision context is calculated, and the original graph feature representation is weighted by nodes to obtain spatial semantic enhancement features.

[0085] In this context, the temporal hidden state of the cognitive graph neural network can be the internal memory state retained by the network after processing dynamic temporal cognitive node sequences. It can be used to encode historical evolution trajectories and for cross-dimensional fusion with spatial semantics. The importance score can be a scalar value reflecting the comprehensive importance of each node in the current decision context, calculated by combining the temporal hidden state and the semantic saliency weight vector. It can be used to drive node-level reweighting of the original graph features. In an exemplary embodiment, the importance score can be generated by tensor fusion (such as outer product, bilinear mapping) followed by MLP or similarity calculation. The spatial semantic enhancement feature can be a graph feature representation weighted by the importance score, highlighting key semantic nodes in the current decision context and providing context-sensitive spatial structure representations.

[0086] The temporal decision causal features and spatial semantic enhancement features are input into a multilayer perceptron for feature fusion to generate contextualized semantic decision fusion features for the target enterprise's decision-making scenario.

[0087] The multilayer perceptron (MLP) can be a feedforward neural network composed of stacked fully connected layers, used for high-dimensional feature fusion. It can integrate temporal causal features and spatial semantic features to generate a unified decision fusion representation. Inputting temporal causal features and spatial semantic enhancement features into the MLP for feature fusion can be achieved by concatenating the two types of features and then feeding them into the MLP for nonlinear combination. Furthermore, this operation can be achieved by using residual connections to preserve the original feature information or by introducing feature cross-layers to enhance interactive modeling, thereby generating a unified, high-order fusion representation. Generating scenario-based semantic decision fusion features for the target enterprise's decision-making scenario can be achieved by outputting a fixed-dimensional vector from the final layer of the MLP as the decision features for that scenario, thus providing the final semantic input that can be used for quality assessment and optimization.

[0088] Taking emergency decision-making for supply chain disruptions in the manufacturing industry as an example, the intelligent decision support method based on cognitive logic and contextualized semantics in this embodiment can be as follows: When a manufacturing enterprise encounters a port worker incident that causes a delay in the arrival of raw materials, the system organizes daily inventory levels (enterprise operation indicators), social media discussions about alternative suppliers (public opinion hotspots), emergency order adjustment records (business interaction events), and geopolitical risk indices (environmental risk parameters) into dynamic temporal cognitive nodes. The forward cognitive layer identifies that the "temporary air freight" decision significantly alleviates production line shutdowns on the same day through confidence gating (high-confidence instantaneous effect), while the backward cognitive layer discovers through causal strength gating that although the "switching to a second-tier supplier" has no obvious effect in the initial stage, it significantly reduces the material shortage rate after three weeks (strong long-term causality). At the same time, in the spatial dimension, the graph convolutional layer aggregates the supplier-material-factory topology relationship, and node importance pooling identifies "key chip suppliers" as the semantic focus; combined with the temporal hidden state, its importance score is calculated to be 0.92, and spatial semantic enhancement features are generated after weighting the graph features. Ultimately, the multilayer perceptron integrates temporal decision-making causal features with spatial semantic enhancement features to output scenario-based semantic decision-making fusion features that emphasize "long-term supplier diversification," supporting high-level management in formulating structural response strategies.

[0089] In one embodiment, an enterprise decision-making quality assessment model is constructed. Contextualized semantic decision fusion features are input into the trained enterprise decision-making quality assessment model to perform quality perception of the current decision-making scenario, thereby obtaining enterprise decision-making quality perception information. Specifically, this includes:

[0090] Historical decision cases of different quality are retrieved from the historical decision database. Based on the retrieved historical decision cases, the historical decision characteristics at the time of execution of each historical decision case are extracted to obtain historical decision characteristic information.

[0091] The historical decision repository can be a structured database storing past decision cases of an enterprise, their execution results, features, and quality labels. It can provide labeled supervision signals for model training. Historical decision cases of different quality levels can be complete decision instances in the historical decision repository, categorized by actual business effectiveness (such as achievement rate, deviation rate, and ROI). These can serve as the basis for positive and negative samples in building the evaluation model. Historical decision features at the time of execution can be a set of multimodal input features corresponding to each historical decision case at the execution moment. These can reflect the contextual state at the time of decision-making and be used to establish a mapping relationship with the results. Historical decision feature information can be a structured feature set extracted and integrated from multiple historical decision cases. This can be used as the original input for cognitive partitioning and graph modeling.

[0092] Retrieving historical decision cases of varying quality from a historical decision database can be achieved by selecting representative cases based on quality labels (e.g., high / medium / low). Further, this operation can be implemented by ensuring a balanced quality distribution through quantile sampling or by selecting the case with the most information based on an active learning strategy, thereby constructing a training sample set covering the quality spectrum. Historical decision features are then extracted from the acquired historical decision cases to obtain historical decision feature information. This can be achieved by parsing structured feature vectors from case metadata and associated logs, thus reconstructing the multimodal input state at the time of decision-making.

[0093] Extract the business processes corresponding to each historical decision case, perform cognitive partitioning of the historical decision feature information, and generate several decision feature subsets of different historical decision cases. Each decision feature subset corresponds to a different semantic label of the decision scenario.

[0094] The business processes corresponding to each historical decision case can describe the specific business operation path or workflow stage to which the historical decision case belongs. These processes can serve as key metadata for distinguishing decision semantic scenarios and guiding cognitive partitioning. Cognitive partitioning can be a process of semantic clustering or grouping of historical decision feature information based on business processes. It can be used to align feature subsets with specific decision contexts, improving the semantic consistency of subsequent modeling. In a specific embodiment, cognitive partitioning can achieve feature subset division through clustering algorithms or rule matching guided by process labels. Decision feature subsets can be feature subgroups belonging to the same business process or semantic scenario formed after cognitive partitioning. These can be used as basic units for constructing cognitive nodes in the semantic topology graph. Decision scenario semantic labels can be symbols or vectors abstracted from business processes to identify the semantic category to which the decision feature subset belongs. These labels can be used to support cross-case semantic alignment and generalization reasoning.

[0095] Extracting the business processes corresponding to each historical decision case can be done by extracting the business line, stage, or workflow ID from the case records, thus providing a semantic grouping basis for cognitive partitioning. Cognitively partitioning the historical decision feature information generates several decision feature subsets for different historical decision cases. This can be done by dividing features into mutually exclusive or overlapping subsets based on business process labels. For example, this operation can be achieved by using process labels as hard partitioning conditions or by using soft clustering to achieve cross-process feature sharing, thereby enabling feature organization aligned with semantic scenarios.

[0096] Map the historical decision feature information after cognitive partitioning to the corresponding historical decision quality, and use decision features as cognitive nodes and decision quality as evaluation nodes. Define business processes as scenario identifiers for cognitive nodes and establish a semantic topology graph.

[0097] In this context, cognitive nodes can be graph nodes representing a subset of decision features in the semantic topology graph, used to represent decision states within a specific semantic scenario. Evaluation nodes can be graph nodes representing historical decision quality scores, used as supervisory signals embedded in the graph structure to guide cognitive nodes in learning quality-related representations. Scenario identifiers can be meta-attributes used to mark the business process or decision context to which a cognitive node belongs, preserving the original semantic context and preventing information ambiguity within the graph structure. The semantic topology graph can be a heterogeneous graph structure composed of cognitive nodes, evaluation nodes, and their associated edges, defined by business logic or statistical co-occurrence, used to explicitly model the structured semantic relationships between decision features and quality. In an exemplary embodiment, the semantic topology graph can be explicitly connected to nodes in the business process based on mapping relationships, forming a semantically labeled graph. Furthermore, the semantic topology graph can provide an input graph structure for the graph attention network and support relation matrix generation.

[0098] Mapping historical decision features after cognitive partitioning to corresponding historical decision quality can establish a pairing relationship between each subset of decision features and its actual business performance, thus forming the input-output pairs required for supervised learning. Using decision features as cognitive nodes, decision quality as evaluation nodes, and defining business processes as scenario identifiers for cognitive nodes, a semantic topology graph can be built. This can be done by using feature subsets and quality scores as nodes, and establishing edge connections based on business logic or statistical significance. For example, this operation can be achieved by explicitly adding intra-process node connections or automatically constructing cross-process edges based on mutual information thresholds, thereby constructing a heterogeneous graph with semantic structure and supervision signals.

[0099] A corporate decision quality assessment model is constructed based on cognitive logic network and graph attention network. The historical decision feature information after cognitive partitioning is input into the cognitive logic network to extract multi-granular time representations of different decision qualities.

[0100] The cognitive logic network can be a neural network module specifically designed to extract multi-granular dynamic representations that conform to human cognitive patterns from temporal decision features, enhancing the cognitive rationality and interpretability of temporal representations. In one specific embodiment, the cognitive logic network can be implemented by embedding causal masks, priority gating, or contextual condition modules into the network structure. Furthermore, the cognitive logic network can process inputs in parallel with a graph attention network, outputting temporal and spatial features respectively. The graph attention network can be a variant of a graph neural network that performs adaptive weighted message passing on a semantic topology graph, enabling focus on key semantic relationships and improving the discriminative power of spatial decision features. In an exemplary embodiment, the graph attention network can dynamically allocate the influence of neighboring nodes using learnable attention coefficients. Furthermore, the graph attention network can be combined with a cognitive message passing mechanism to achieve rule-guided attention computation. The multi-granular temporal representation can be dynamic features extracted by the cognitive logic network from historical decision features, covering different time scales (such as short-term fluctuations and long-term trends), used to characterize temporal dependencies and rhythmic changes in the decision evolution process. In one specific embodiment, multi-granularity temporal representations can be generated through multi-scale convolution, hierarchical RNNs, or Transformer temporal window mechanisms.

[0101] Building an enterprise decision-making quality assessment model based on cognitive logic networks and graph attention networks can be achieved by connecting the two networks in parallel or in series to form the backbone of feature extraction, thereby enabling joint modeling of temporal cognition and spatial semantics. By inputting historical decision-making feature information after cognitive partitioning into the cognitive logic network, multi-granular temporal representations of different decision qualities can be extracted. A multi-scale temporal encoder can capture short-term response and long-term trend features, thus obtaining dynamic representations with cognitive rationality.

[0102] The relation matrix is ​​obtained by establishing a semantic topology graph. The state of nodes is aggregated and updated based on the cognitive message passing mechanism, and the updated semantic node representation is obtained. The updated semantic node representation of the whole graph is semantically pooled to obtain spatial decision features. Multi-granular temporal representation and spatial decision features are fused across modally to generate cognitive-spatial fusion features.

[0103] The relation matrix can be a mathematical representation describing the strength or type of adjacency between nodes in a semantic topology graph, and can be used as the topological basis for message passing in a graph neural network. In one specific embodiment, the relation matrix can be extracted from the edge set of the semantic topology graph and encoded as a sparse or dense matrix. The cognitive message passing mechanism can be an information propagation strategy guided by cognitive logic rules integrated into the graph structure, which can be used to ensure that the node aggregation process conforms to expert reasoning logic and enhance interpretability. In an exemplary embodiment, the cognitive message passing mechanism can be implemented by introducing rule constraints or attention biases on the basis of standard graph message passing. The node state aggregation update can be an iterative process of a graph neural network that calculates new representations based on neighbor node information and its own state, which can be used to gradually integrate local semantic context to form higher-order node representations.

[0104] The updated semantic node representation can be a node embedding vector obtained after one or more rounds of message passing, which can be used to reflect the node's position and role in the global semantic structure. Semantic pooling can be an operation that aggregates all updated semantic node representations of the entire graph to generate a global graph-level representation, which can be used to generate a fixed-dimensional vector representing the overall decision semantic space. In a specific embodiment, semantic pooling can adopt scene-identified weighted average pooling or attention-guided graph readout mechanism. Spatial decision features can be global feature vectors obtained after semantic pooling, representing the current decision semantic space structure, which can be used to capture collaborative semantic patterns among multiple nodes as one input for cross-modal fusion. Cross-modal fusion can be a process of jointly modeling multi-granular representations in the temporal dimension and semantic structural features in the spatial dimension, which can be used to generate a unified representation that simultaneously includes temporal cognitive depth and spatial semantic breadth. In an exemplary embodiment, cross-modal fusion can adopt bilinear fusion or gated cross-attention mechanism. Cognitive-spatial fusion features can be high-dimensional joint representations formed by integrating multi-granular temporal representations and spatial decision-making features. They can be used as the final input of the quality assessment layer to comprehensively reflect the dynamics and structure of decision-making.

[0105] Obtaining the relation matrix by establishing a semantic topology graph can convert the graph's adjacency relationships into a numerical form suitable for matrix operations, thus providing topological input for graph neural networks. Aggregating and updating node states based on a cognitive message passing mechanism to obtain updated semantic node representations can introduce cognitive rules to constrain aggregation weights during message passing. Furthermore, this operation can be achieved by adding causal priors to attention weights or imposing penalties on edges that violate rules, ensuring that the graph propagation process conforms to expert logic. Semantic pooling of the updated semantic node representations across the entire graph yields spatial decision features, which can be achieved by aggregating all node embeddings to generate graph-level representations, thereby compressing local semantics into a global decision context vector. Cross-modal fusion of multi-granular temporal representations and spatial decision features generates cognitive-spatial fusion features, which can be achieved by jointly encoding temporal and graph structural information through a fusion module. For example, this operation can be implemented using cross-attention to align spatiotemporal dimensions or by modeling higher-order interactions through tensor fusion networks, thus forming a unified decision representation that combines dynamism and structure.

[0106] The cognitive-spatial fusion features are input into the quality assessment layer to score the decision quality. The decision quality assessment results are calibrated and the parameters are fine-tuned through a preset validation set. After iterative training, an enterprise decision quality assessment model that meets the preset standards is obtained.

[0107] The quality assessment layer can be a learnable scoring module at the end of the enterprise decision-making quality assessment model, used to map fused features into interpretable quantitative indicators of decision quality. In one specific embodiment, the quality assessment layer can be implemented using a fully connected neural network or a regression head. The decision quality score can be a numerical value or level output by the quality assessment layer, representing the expected effectiveness of the current decision-making scheme, and can directly constitute the core content of the enterprise's perceived decision quality information. The preset validation set can be a set of high-quality labeled cases reserved from the historical decision database that have not participated in training, used to calibrate the model output distribution and guide parameter fine-tuning. Parameter fine-tuning can be a process of making small adjustments to the model parameters based on the feedback from the validation set to optimize generalization performance, and can be used to improve the model's evaluation accuracy in unseen scenarios. Iterative training can be a model optimization loop that repeatedly executes forward inference, loss calculation, backpropagation, and parameter updates, used to gradually approach the optimal evaluation capability under preset standards. Preset standards can be performance thresholds or convergence conditions used to determine whether the enterprise decision-making quality assessment model has completed training, and can be used to control the timing of training termination to ensure that the model reaches a usable level.

[0108] By inputting cognitive-spatial fusion features into the quality assessment layer to score decision quality, quality predictions can be output through regression or classification heads, thereby generating quantifiable estimates of decision effectiveness. The decision quality assessment results are calibrated and parameters are fine-tuned using a pre-set validation set. This can involve calculating the assessment error on the validation set and adjusting the model parameters in reverse, thus improving the accuracy and stability of the model output. Through iterative training, a corporate decision quality assessment model conforming to pre-set standards is obtained. This can be achieved by repeating the training-validation cycle until convergence is met, resulting in a reliable model that can be used for real-time quality perception.

[0109] Obtain contextualized semantic decision fusion features, input these features into the trained enterprise decision quality assessment model to perform quality perception of the current decision scenario, and obtain enterprise decision quality perception information.

[0110] Obtaining contextualized semantic decision fusion features can be achieved by receiving fusion features of the current decision scenario from previous steps, thus providing real-time input for quality assessment. These contextualized semantic decision fusion features are then input into a trained enterprise decision quality assessment model to perform quality perception on the current decision scenario, obtaining enterprise decision quality perception information. This can be achieved by executing forward inference of the model and parsing the output into structured perception signals, thereby enabling dynamic quantitative evaluation of the current decision execution effect.

[0111] Taking the decision-making assessment of supply chain disruption response in the manufacturing industry as an example, the intelligent decision support method based on cognitive logic and scenario-based semantics in this embodiment can be as follows: The system retrieves supply chain disruption cases caused by port strikes and other factors in the past three years from the historical decision database, and classifies them into three categories of high / medium / low quality according to the recovery timeliness and cost overrun. It extracts features such as inventory level, alternative supplier quotations, and logistics delay days at the time of each decision, and performs cognitive partitioning according to its "raw material procurement emergency process". It constructs a semantic topology graph, where cognitive nodes represent feature combinations of each process stage, and evaluation nodes are the actual recovery scores. It extracts multi-granularity time representations of "short-term material shortage impact" and "long-term supplier switching" through the cognitive logic network, and at the same time, the graph attention network aggregates cross-node semantics on the topology graph and obtains spatial features through pooling. After the two are fused, they are input into the quality assessment layer, and the output is a quality score of 62 points (below the threshold of 75 points) for the emergency plan formulated under the current new port congestion event, triggering plan optimization.

[0112] In one embodiment, based on enterprise decision quality perception information, it is determined whether the execution effect of the current decision-making scenario meets the expected goals. If not, the current decision-making scheme is optimized to provide intelligent decision support for the enterprise, specifically including:

[0113] Obtain enterprise decision quality perception information, and generate decision effect evolution curve for the current decision scenario based on the enterprise decision quality perception information. The decision effect evolution curve represents the decision quality index of the current decision scenario at different time points.

[0114] The decision performance evolution curve can be a time-series function or data sequence representing the trend of the decision quality index at different timestamps in the current decision-making scenario. It can be used to dynamically depict the evolution of decision execution effectiveness, supporting deviation identification and trend prediction. In this embodiment, the decision performance evolution curve can be generated by aggregating quality index sequences by timestamps based on continuously acquired enterprise decision quality perception information and fitting them into a curve. Furthermore, the decision performance evolution curve can include, but is not limited to, one or more of the following: monotonically decreasing evolution curve, fluctuating convergent evolution curve, and abrupt transition evolution curve. The decision quality index can be a standardized numerical indicator output by the enterprise decision quality assessment model, used to quantify the current decision performance, and can be used to provide comparable and traceable decision performance metrics. In an exemplary embodiment, the decision quality index can constitute the basic data points of the decision performance evolution curve.

[0115] Generating an evolution curve of decision-making effectiveness in the current decision-making scenario based on enterprise decision-making quality perception information can be achieved by aligning time-series quality perception information by timestamps and converting it into a continuous or discrete quality index sequence to form an evolution curve. Furthermore, this operation can be implemented by using spline interpolation to smooth discrete points to generate a continuous curve or by employing time series clustering to identify typical evolution patterns, thereby enabling dynamic visualization and trend modeling of decision-making effectiveness.

[0116] Obtain the expected target threshold of the current decision-making scenario, calculate the deviation between the decision quality index corresponding to different timestamps and the expected target threshold by combining the decision effect evolution curve, and compare the calculated deviation with the preset fault tolerance range.

[0117] The expected target threshold can be the lowest or best acceptable boundary value of the decision quality index set by the enterprise in the current decision-making scenario. It can be used as a quantitative benchmark to judge whether the decision deviates from the expectation. For example, the expected target threshold can be a static fixed threshold, a dynamic sliding threshold, or a multi-level segmented threshold. The deviation can be the absolute or relative difference between the decision quality index and the expected target threshold at a specific time stamp. It can be used to quantify the degree of deviation between the current decision effect and the target. In a specific embodiment, the deviation can be calculated by numerical subtraction or normalized distance. The preset fault tolerance range can be the maximum acceptable deviation range that allows the decision quality index to fluctuate around the expected target threshold. It can be used to avoid unnecessary optimization operations triggered by small fluctuations and improve system robustness. Furthermore, the preset fault tolerance range can include, but is not limited to, one or more of the following: symmetric fault tolerance range, asymmetric fault tolerance range, and adaptive fault tolerance band.

[0118] The deviation between the decision quality index and the expected target threshold at different time points is calculated by combining the decision effect evolution curve. This can be achieved by performing difference or relative error calculations on the quality index and the corresponding threshold for each time point. Furthermore, this operation can be implemented by calculating the absolute deviation |Q(t)-T| or the normalized deviation (Q(t)-T) / T, thereby quantifying the degree of decision deviation at each time point. Comparing the calculated deviation with a preset tolerance range can determine whether the deviation exceeds the upper or lower tolerance limits. Further, this operation can be implemented through hard threshold judgment (triggered if deviation > upper tolerance limit) or soft probability judgment (trigger probability calculated based on deviation distribution), thereby determining whether to trigger the optimization process.

[0119] If the error exceeds the preset tolerance range, it means that the current decision control parameters are not suitable for the expected goal of the current decision scenario. In this case, the preset decision scheme of the current decision scenario is obtained and the scheme is optimized.

[0120] Among them, the current decision control parameters can be adjustable strategy variables or configuration items in the current decision scheme, which directly affect the execution effect and can be used as the object of adjustment for the optimization algorithm to determine the specific form of the decision behavior. The preset decision scheme can be a set of default or initial decision strategies pre-configured for the current decision scenario, which can be used as the starting point for optimization or as a replacement benchmark when deviation from the effect is detected.

[0121] Retrieve historical decision-making cases that meet the current decision quality requirements and business domain from the preset historical decision knowledge base, and extract the scenario-based historical decision control schemes corresponding to each historical decision-making case.

[0122] The pre-defined historical decision knowledge base can be a structured, semantically encoded database containing historical successful / failed decision cases, their contexts, parameters, and effects. This database can provide transferable decision priors and control scheme templates for new scenarios. In one exemplary embodiment, the pre-defined historical decision knowledge base can be organized using knowledge graphs or vector indexing techniques, supporting retrieval based on semantic similarity. Furthermore, the pre-defined historical decision knowledge base can include, but is not limited to, one or more of the following: an industry-wide general decision case library, an enterprise-specific historical solution library, and a cross-domain migration decision knowledge base. Historical decision cases can be complete records of decision instances executed in similar business domains and quality requirements in the past, serving as a cognitive source for initializing optimization algorithms. Scenario-based historical decision control schemes can be reusable strategy templates extracted from historical decision cases, containing specific control parameters and execution logic, which can be used as initial population individuals for multi-objective cognitive bee colony optimization algorithms.

[0123] Retrieving historical decision-making cases that meet the current decision quality requirements and business domain from a pre-defined historical decision-making knowledge base can be done by performing a similarity search within the knowledge base based on the semantic tags, quality requirements, and business type of the current scenario. Furthermore, this operation can be achieved by using vector similarity (such as cosine similarity) to retrieve nearest neighbor cases in the embedding space or by finding isomorphic scenario cases through knowledge graph path matching, thereby obtaining historical experience in context matching as optimization priors.

[0124] A multi-objective cognitive bee colony optimization algorithm is introduced. The cognitive population is initialized according to the scenario-based historical decision control scheme corresponding to each historical decision case. The objective function is preset and business constraints are set. The objective function value of each individual in the initial cognitive bee colony is calculated through the objective function. Non-dominated sorting is performed to divide all individuals in the population into different cognitive frontier levels.

[0125] The multi-objective cognitive bee colony optimization algorithm can be a multi-objective optimization algorithm that integrates swarm intelligence search mechanisms and cognitive logic guidance, simulating bee colony foraging behavior and embedding semantic reasoning. It can be used to efficiently search for Pareto optimal solutions in high-dimensional, complex decision spaces. In one specific embodiment, the multi-objective cognitive bee colony optimization algorithm can introduce cognitive frontier, density evaluation, and semantic shift mechanisms based on traditional multi-objective artificial bee colony algorithms. Furthermore, the multi-objective cognitive bee colony optimization algorithm can include, but is not limited to, one or more of rule-guided cognitive bee colony algorithms, graph neural network-enhanced cognitive bee colony algorithms, and context-based cognitive bee colony algorithms. The cognitive population can be an initial solution set composed of multiple candidate decision schemes (individuals), with each individual encoded as a decision control parameter vector, which can be used as the starting point for the optimization algorithm's search. In this embodiment, the cognitive population can be semantically mapped and parameter instantiated based on contextualized historical decision control schemes.

[0126] The objective function can be a multidimensional mathematical expression used to evaluate the merits of candidate decision schemes. It typically contains multiple conflicting optimization objectives and can drive the search direction of the optimization algorithm. For example, the objective function can be a cost-benefit joint objective function, a risk-return trade-off objective function, or a time-efficiency-quality balance objective function. Business constraints can be boundary conditions of the feasible solution space, consisting of enterprise operating rules, resource limitations, or compliance requirements. These can be used to ensure that the optimization results are executable in actual business operations. The objective function value can be a multidimensional evaluation score vector of a single candidate decision scheme under the objective function, which can be used for non-dominated ranking and individual merit comparison. The cognitive frontier hierarchy can be a Pareto hierarchy after non-dominated ranking, where individuals within the same hierarchy are not mutually dominant. This can be used to achieve hierarchical organization of multi-objective solutions, prioritizing the retention of frontier solutions. In an exemplary embodiment, the cognitive frontier hierarchy can be stratified using a fast non-dominated ranking algorithm (such as the method in NSGA-II).

[0127] A multi-objective cognitive bee colony optimization algorithm is introduced. The cognitive population is initialized based on the scenario-specific historical decision control schemes corresponding to each historical decision case. This can be achieved by mapping the extracted control scheme parameters to the initial individuals of the algorithm, thus forming the cognitive population. Furthermore, this operation can be implemented by directly copying parameters as initial individuals or by adding Gaussian perturbations to historical schemes to generate diversity initial values. This allows the search starting point to be closer to the feasible region, accelerating convergence.

[0128] Non-dominated sorting divides all individuals in the population into different cognitive frontier levels. This can be achieved by performing Pareto dominance analysis on individuals based on objective function values ​​and then stratifying them. Furthermore, this operation can be accelerated by using FastNon-dominatedSort or a dominance tree structure, thereby enabling hierarchical organization of multi-objective solutions.

[0129] For each cognitive frontier level, cognitive density is assessed to obtain a crowding index. The individual with the lowest crowding in each level is selected as the leader bee, and the remaining individuals are selected as followers bees. The cognitive offset vector of the leader bee in the decision cognitive space is calculated for position update. The optimal decision solution set is output after repeated iterations until the convergence condition is met.

[0130] Cognitive density assessment can be a quantitative analysis of the density of individuals distributed in the target space at the cognitive frontier level, which can be used to maintain population diversity and prevent local clustering. In a specific embodiment, cognitive density assessment can be achieved by calculating the average distance between an individual and its neighboring individuals in the objective function value space. Crowding index can be a numerical measure characterizing the crowding level of the area where an individual is located in the target space, which can be used to screen leader bees and prioritize individuals in sparse areas to enhance exploration capabilities. Leader bees can be low-crowding individuals selected as search guides at each cognitive frontier level, which can be used to guide the population's exploration direction in the decision cognitive space. Follower bees can be other population individuals besides leader bees, who conduct local development near leader bees, which can be used to enhance local search accuracy and balance exploration and development. Cognitive offset vector can be a direction adjustment vector generated by leader bees in the decision cognitive space based on cognitive logic (such as rules, historical experience), which can be used to make position updates not only dependent on numerical gradients but also incorporate semantic guidance. In an exemplary embodiment, the cognitive offset vector can be combined with semantic similarity, causal reasoning, or contextual matching to calculate the offset direction. The decision-making cognitive space can be a high-dimensional semantic space composed of decision control parameters, where distance reflects the semantic similarity of solutions. This space can serve as the carrier space for the cognitive bee colony algorithm to perform position updates and searches. Convergence conditions can be preset criteria for determining whether an optimization algorithm terminates, such as the maximum number of iterations or the stability of the solution set. These can be used to balance the algorithm's runtime and solution quality. The optimal decision solution set can be the Pareto optimal candidate solution set output after the algorithm converges, which can be used as the basis for generating candidate optimization solutions.

[0131] Crowding indexes are obtained by assessing cognitive density at each cognitive frontier level. This can be achieved by calculating the average distance of each individual within its neighborhood in the target space. Furthermore, this operation can be implemented using the k-nearest neighbor average distance method or the Voronoi unit volume estimation method, thereby identifying sparse regions to maintain population diversity. The cognitive offset vector of the leader bee in the decision-making cognitive space is calculated for position updates. This can be achieved by combining cognitive logic (such as rules and context matching) to generate a direction vector to guide the leader bee's movement. Further, this operation can be implemented by generating offset directions based on knowledge graph path reasoning or by weighting the offset vector using gradient directions from historical success cases, thus incorporating semantic understanding into the search process and surpassing pure numerical optimization.

[0132] Based on the output optimal decision solution set, several candidate decision optimization schemes are generated. The business feasibility of each candidate decision optimization scheme is verified. Based on the verification results, the optimal decision optimization scheme is selected for enterprise intelligent decision support.

[0133] The candidate decision optimization schemes can be several executable decision strategies instantiated from the optimal decision solution set, which can be used for business feasibility verification and screening. Business feasibility verification can be an automated or semi-automated evaluation process to determine whether the candidate decision optimization schemes meet the actual business execution conditions. It can be used to filter out solutions that are technically feasible but not business-feasible, ensuring implementability. In a specific embodiment, business feasibility verification can be completed through a rule engine, simulation, or expert review interface. Furthermore, business feasibility verification can be achieved through automatic verification based on rule engines such as Drools or through simulation of execution effects in a digital twin environment.

[0134] The optimal decision optimization scheme can be the final recommended decision scheme selected after business feasibility verification, which can be used for actual implementation or pushed to decision-makers for confirmation. Several candidate decision optimization schemes are generated based on the output optimal decision solution set, which can be achieved by instantiating the parameter vectors in the Pareto optimal solution into an executable strategy description. Furthermore, this operation can be implemented by generating natural language schemes through template filling or binding parameters to a preset strategy execution engine, thereby connecting the algorithm output with business execution. Business feasibility verification of each candidate decision optimization scheme can be performed by calling a business rule engine or simulation module to evaluate whether the scheme meets resource, compliance, or operational constraints. Furthermore, this operation can be achieved through automatic verification based on rule engines such as Drools or by simulating execution effects in a digital twin environment, thereby ensuring that the recommended scheme can be implemented.

[0135] For example, in a scenario where the execution effect of the current decision-making scenario does not meet the expected goal, the intelligent decision support method based on cognitive logic and scenario-based semantics in this embodiment can be as follows: A factory monitors a decrease in equipment utilization and an increase in delivery delay rate during peak order periods. The system generates a decision effect evolution curve showing that the quality index is continuously lower than the expected target threshold and the deviation exceeds the fault tolerance range. Then, three successful cases with similar high-load scenarios are retrieved from the historical decision knowledge base, and their production scheduling rules, manpower allocation, and outsourcing strategies are extracted as the initial population. A multi-objective cognitive bee colony optimization algorithm is introduced, with "minimizing delay" and "maximizing equipment utilization" as objective functions, and it runs under business constraints such as "daily overtime not exceeding 2 hours". The algorithm divides the frontier level through non-dominated sorting, selects the leading bee based on the crowding degree, and generates an offset vector in combination with the cognitive rule of "prioritizing key equipment". After 10 rounds of iteration, 5 Pareto optimal solutions are output. After generating candidate solutions, they are verified by MES system simulation. Finally, a solution that improves the delivery rate without exceeding the manpower budget is automatically sent to the production scheduling platform.

[0136] Furthermore, to achieve the above objectives, the present invention also provides an intelligent decision support system based on cognitive logic and contextualized semantics, the system comprising:

[0137] The data perception module is used to monitor the decision-making scenarios of target enterprises from multiple sources to obtain enterprise decision-making monitoring data, and to preprocess the obtained enterprise decision-making monitoring data to obtain cross-modal enterprise scenario cognitive information.

[0138] The semantic fusion module is used to analyze the cross-modal correlation of each monitoring data based on the cross-modal enterprise scenario cognitive information, and to integrate cognitive logic reasoning rules to construct semantic fusion features, thereby obtaining scenario-based semantic decision fusion features;

[0139] The quality assessment module is used to construct an enterprise decision-making quality assessment model. The scenario-based semantic decision fusion features are input into the trained enterprise decision-making quality assessment model to perform quality perception of the current decision-making scenario and obtain enterprise decision-making quality perception information.

[0140] The decision optimization module is used to determine whether the execution effect of the current decision scenario meets the expected goals based on the enterprise decision quality perception information. If it does not meet the expected goals, the module optimizes the current decision scheme to provide intelligent decision support for the enterprise.

[0141] Other embodiments or specific implementations of the intelligent decision support system based on cognitive logic and contextual semantics described in this invention can be referred to the above-described method embodiments, and will not be repeated here.

[0142] Furthermore, to achieve the above objectives, the present invention also provides an intelligent decision support device based on cognitive logic and contextual semantics. The device includes: a memory, a processor, and an intelligent decision support program based on cognitive logic and contextual semantics stored in the memory and executable on the processor. The intelligent decision support program based on cognitive logic and contextual semantics is configured to implement the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described above.

[0143] Furthermore, to achieve the above objectives, the present invention also provides a medium storing an intelligent decision support program based on cognitive logic and contextual semantics, wherein when the intelligent decision support program based on cognitive logic and contextual semantics is executed by a processor, it implements the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described above.

[0144] Furthermore, to achieve the above objectives, the present invention also provides a computer program product, including an intelligent decision support program based on cognitive logic and contextual semantics, wherein when the intelligent decision support program based on cognitive logic and contextual semantics is executed by a processor, it implements the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described above.

[0145] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An intelligent decision support method based on cognitive logic and contextualized semantics, characterized in that, The method includes: Multi-source data monitoring is conducted on the target enterprise's decision-making scenarios to obtain enterprise decision-making monitoring data. The obtained enterprise decision-making monitoring data is preprocessed to obtain cross-modal enterprise scenario cognitive information. Based on the cross-modal enterprise scenario cognitive information analysis, the cross-modal correlation of each monitoring data is analyzed, and the cognitive logic reasoning rules are integrated to construct semantic fusion features, thereby obtaining scenario-based semantic decision fusion features; A corporate decision-making quality assessment model is constructed. The scenario-based semantic decision fusion features are input into the trained corporate decision-making quality assessment model to perform quality perception of the current decision-making scenario and obtain corporate decision-making quality perception information. Based on the enterprise decision quality perception information, it is determined whether the execution effect of the current decision scenario meets the expected goal. If not, the current decision-making scheme is optimized to provide intelligent decision support for the enterprise. Specifically, the step of monitoring the target enterprise's decision-making scenario through multi-source data to obtain enterprise decision-making monitoring data, and preprocessing the obtained enterprise decision-making monitoring data to obtain cross-modal enterprise scenario cognitive information, includes: In the target enterprise decision-making scenario, the integrated data interface is used to monitor the business system in real time to obtain enterprise decision-making monitoring data, which includes internal operation monitoring data, external environment monitoring data, and interactive behavior monitoring data. The enterprise decision monitoring data is decomposed into wavelet packets using a preset wavelet basis function to obtain different sub-bands and the Shannon entropy of each sub-band is calculated. Hard threshold filtering is then performed based on the calculated Shannon entropy to extract key semantic frequency bands. After filtering, the enterprise decision monitoring data is synchronized with time series. Bilateral filtering is used to eliminate data noise and moving average smoothing is used to eliminate data fluctuations. A dynamic window algorithm is introduced to slide based on a preset time window and calculate the confidence interval and coefficient of variation of the data within the window. Data repair is performed in combination with spline interpolation algorithm. Z-score standardization is performed on the enterprise decision monitoring data after the repair is completed, transforming continuous data into a distribution with a mean of 0 and a standard deviation of 1. Semantic bucketing is used to perform contextual encoding on discrete data to generate cross-modal enterprise scenario cognitive information. Specifically, the step of analyzing the cross-modal correlations of various monitoring data based on the cross-modal enterprise scenario cognitive information, and integrating cognitive logic reasoning rules to construct semantic fusion features, to obtain scenario-based semantic decision fusion features, includes: Obtain cross-modal enterprise scenario cognitive information, extract interactive behavior monitoring data and parse enterprise interactive text monitoring data through the cross-modal enterprise scenario cognitive information, and input the enterprise interactive text monitoring data into the BERT pre-trained model to obtain context embedding vectors; The decision intent index is constructed by weighted fusion of key semantic role annotation and entity relationship density using context embedding vectors. The key semantic role annotation is the process and result of identifying components in the text that perform specific semantic functions. It is used to structure the logical relationship between decision subjects, actions and objects. Subsequently, a knowledge entity network of enterprise interactive text is constructed using cognitive graph algorithm. The ratio of network node degree to edge weight is calculated as the semantic association strength index. The first semantic feature information is generated by combining the sentiment features of enterprise operation text. Market sentiment data is extracted by acquiring external environment monitoring data and generating a sentiment keyword matrix by applying TF-IDF. The sentiment keyword matrix is ​​then input into a pre-trained cognitive reasoning engine to perform knowledge graph completion and generate semantic ontology feature vectors for enterprise scenarios, thereby obtaining second semantic feature information. A cognitive attention mechanism is introduced, using the first semantic feature information as the query vector and the second semantic feature information as the key vector and value vector. The attention weight distribution is constrained by cognitive logic rules, and the rule-weighted attention score is calculated to perform feature fusion and generate semantically enhanced feature information. By utilizing cross-modal enterprise scenario cognitive information to extract internal operational monitoring data and external environmental monitoring data, and combining the semantic enhancement feature information, the data is input into a cognitive graph neural network for cross-modal semantic fusion to obtain scenario-based semantic decision fusion features.

2. The intelligent decision support method based on cognitive logic and contextual semantics as described in claim 1, characterized in that, The process involves extracting internal operational monitoring data and external environmental monitoring data using cross-modal enterprise scenario cognitive information, combining this with semantic enhancement feature information, and inputting it into a cognitive graph neural network for cross-modal semantic fusion to obtain scenario-based semantic decision fusion features. Specifically, this includes: In the time dimension, the data input to the cognitive graph neural network is transformed into dynamic temporal cognitive nodes, where each dynamic temporal cognitive node contains the enterprise's operational indicators, public opinion hotspots, business interaction events, and environmental risk parameters at the current moment. The forward cognition layer captures the instantaneous impact of decision adjustments on the business state through confidence gating, while the backward cognition layer uses causal strength gating to filter the long-term impact of decision adjustments on the business state, generating time-series decision causal features. In the spatial dimension, the data input to the cognitive graph neural network is aggregated into topological relationships using graph convolutional layers to obtain graph feature representations. The graph feature representations are then subjected to global weighted pooling based on node importance to generate semantic saliency weight vectors. The temporal hidden states of the cognitive graph neural network are obtained and fused with the semantic saliency weight vector using tensors. The importance score of each cognitive node in the decision context is calculated, and the original graph feature representation is weighted by nodes to obtain spatial semantic enhancement features. The temporal decision causal features and spatial semantic enhancement features are input into a multilayer perceptron for feature fusion to generate contextualized semantic decision fusion features for the target enterprise's decision-making scenario.

3. The intelligent decision support method based on cognitive logic and contextual semantics as described in claim 1, characterized in that, The construction of the enterprise decision-making quality assessment model involves inputting the scenario-based semantic decision fusion features into the trained enterprise decision-making quality assessment model to perform quality perception of the current decision-making scenario, thereby obtaining enterprise decision-making quality perception information. Specifically, this includes: Historical decision cases of different quality are retrieved from the historical decision database. Based on the retrieved historical decision cases, the historical decision characteristics at the time of execution of each historical decision case are extracted to obtain historical decision characteristic information. Extract the business processes corresponding to each historical decision case, perform cognitive partitioning on the historical decision feature information, and generate several decision feature subsets of different historical decision cases. Each decision feature subset corresponds to a different decision scenario semantic label. Map the historical decision feature information after cognitive partitioning to the corresponding historical decision quality, and use decision features as cognitive nodes and decision quality as evaluation nodes. Define business processes as scenario identifiers for cognitive nodes and establish a semantic topology graph. A corporate decision quality assessment model is constructed based on cognitive logic network and graph attention network. The historical decision feature information after cognitive partitioning is input into the cognitive logic network to extract multi-granular time representations of different decision qualities. The relation matrix is ​​obtained by establishing a semantic topology graph. The state of the nodes is aggregated and updated based on the cognitive message passing mechanism, and the updated semantic node representation is obtained. The updated semantic node representation of the whole graph is semantically pooled to obtain spatial decision features. The multi-granularity temporal representation and spatial decision features are fused across modally to generate cognitive-spatial fusion features. The cognitive-spatial fusion features are input into the quality assessment layer to score decision quality. The decision quality assessment results are calibrated and the parameters are fine-tuned through a preset validation set. After iterative training, an enterprise decision quality assessment model that meets the preset standards is obtained. Obtain contextualized semantic decision fusion features, input the contextualized semantic decision fusion features into the trained enterprise decision quality assessment model to perform quality perception of the current decision scenario, and obtain enterprise decision quality perception information.

4. The intelligent decision support method based on cognitive logic and contextual semantics as described in claim 1, characterized in that, The step of judging whether the execution effect of the current decision-making scenario meets the expected goals based on the enterprise decision quality perception information, and if not, optimizing the current decision-making scheme to provide intelligent decision support for the enterprise, specifically includes: Obtain enterprise decision quality perception information, and generate a decision effect evolution curve for the current decision scenario based on the enterprise decision quality perception information. The decision effect evolution curve represents the decision quality index of the current decision scenario at different time points. Obtain the expected target threshold of the current decision-making scenario, calculate the deviation between the decision quality index corresponding to different timestamps and the expected target threshold in combination with the decision effect evolution curve, and compare the calculated deviation with the preset fault tolerance range. If the error exceeds the preset tolerance range, it means that the current decision control parameters are not suitable for the expected goal of the current decision scenario. In this case, the preset decision scheme of the current decision scenario is obtained and the scheme is optimized. Retrieve historical decision-making cases that meet the current decision quality requirements and business domain from the preset historical decision knowledge base, and extract the scenario-based historical decision control schemes corresponding to each historical decision-making case. A multi-objective cognitive bee colony optimization algorithm is introduced. The cognitive population is initialized according to the scenario-based historical decision control scheme corresponding to each historical decision case. The objective function is preset and business constraints are set. The objective function value of each individual in the initial cognitive bee colony is calculated through the objective function. Non-dominated sorting is performed to divide all individuals in the population into different cognitive frontier levels. For each cognitive frontier level, cognitive density is assessed to obtain a crowding index. The individual with the lowest crowding in each level is selected as the leader bee, and the remaining individuals are selected as followers bees. The cognitive offset vector of the leader bee in the decision cognitive space is calculated for position update. The optimal decision solution set is output after repeated iterations until the convergence condition is met. Based on the output optimal decision solution set, several candidate decision optimization schemes are generated. The business feasibility of each candidate decision optimization scheme is verified. Based on the verification results, the optimal decision optimization scheme is selected for enterprise intelligent decision support.

5. An intelligent decision support system based on cognitive logic and contextualized semantics, characterized in that, The system includes: The data perception module is used to monitor the decision-making scenarios of target enterprises from multiple sources to obtain enterprise decision-making monitoring data, and to preprocess the obtained enterprise decision-making monitoring data to obtain cross-modal enterprise scenario cognitive information. The semantic fusion module is used to analyze the cross-modal correlation of each monitoring data based on the cross-modal enterprise scenario cognitive information, and to integrate cognitive logic reasoning rules to construct semantic fusion features, thereby obtaining scenario-based semantic decision fusion features; The quality assessment module is used to construct an enterprise decision-making quality assessment model. The scenario-based semantic decision fusion features are input into the trained enterprise decision-making quality assessment model to perform quality perception of the current decision-making scenario and obtain enterprise decision-making quality perception information. The decision optimization module is used to determine whether the execution effect of the current decision scenario meets the expected goal based on the enterprise decision quality perception information. If it does not meet the goal, the current decision scheme is optimized to provide intelligent decision support for the enterprise. Specifically, the step of monitoring the target enterprise's decision-making scenario through multi-source data to obtain enterprise decision-making monitoring data, and preprocessing the obtained enterprise decision-making monitoring data to obtain cross-modal enterprise scenario cognitive information, includes: In the target enterprise decision-making scenario, the integrated data interface is used to monitor the business system in real time to obtain enterprise decision-making monitoring data, which includes internal operation monitoring data, external environment monitoring data, and interactive behavior monitoring data. The enterprise decision monitoring data is decomposed into wavelet packets using a preset wavelet basis function to obtain different sub-bands and the Shannon entropy of each sub-band is calculated. Hard threshold filtering is then performed based on the calculated Shannon entropy to extract key semantic frequency bands. After filtering, the enterprise decision monitoring data is synchronized with time series. Bilateral filtering is used to eliminate data noise and moving average smoothing is used to eliminate data fluctuations. A dynamic window algorithm is introduced to slide based on a preset time window and calculate the confidence interval and coefficient of variation of the data within the window. Data repair is performed in combination with spline interpolation algorithm. Z-score standardization is performed on the enterprise decision monitoring data after the repair is completed, transforming continuous data into a distribution with a mean of 0 and a standard deviation of 1. Semantic bucketing is used to perform contextual encoding on discrete data to generate cross-modal enterprise scenario cognitive information. Specifically, the step of analyzing the cross-modal correlations of various monitoring data based on the cross-modal enterprise scenario cognitive information, and integrating cognitive logic reasoning rules to construct semantic fusion features, to obtain scenario-based semantic decision fusion features, includes: Obtain cross-modal enterprise scenario cognitive information, extract interactive behavior monitoring data and parse enterprise interactive text monitoring data through the cross-modal enterprise scenario cognitive information, and input the enterprise interactive text monitoring data into the BERT pre-trained model to obtain context embedding vectors; The decision intent index is constructed by weighted fusion of key semantic role annotation and entity relationship density using context embedding vectors. The key semantic role annotation is the process and result of identifying components in the text that perform specific semantic functions. It is used to structure the logical relationship between decision subjects, actions and objects. Subsequently, a knowledge entity network of enterprise interactive text is constructed using cognitive graph algorithm. The ratio of network node degree to edge weight is calculated as the semantic association strength index. The first semantic feature information is generated by combining the sentiment features of enterprise operation text. Market sentiment data is extracted by acquiring external environment monitoring data and generating a sentiment keyword matrix by applying TF-IDF. The sentiment keyword matrix is ​​then input into a pre-trained cognitive reasoning engine to perform knowledge graph completion and generate semantic ontology feature vectors for enterprise scenarios, thereby obtaining second semantic feature information. A cognitive attention mechanism is introduced, using the first semantic feature information as the query vector and the second semantic feature information as the key vector and value vector. The attention weight distribution is constrained by cognitive logic rules, and the rule-weighted attention score is calculated to perform feature fusion and generate semantically enhanced feature information. By utilizing cross-modal enterprise scenario cognitive information to extract internal operational monitoring data and external environmental monitoring data, and combining the semantic enhancement feature information, the data is input into a cognitive graph neural network for cross-modal semantic fusion to obtain scenario-based semantic decision fusion features.

6. An intelligent decision support device based on cognitive logic and contextualized semantics, characterized in that, The device includes: a memory, a processor, and an intelligent decision support program based on cognitive logic and contextual semantics stored in the memory and executable on the processor, wherein the intelligent decision support program based on cognitive logic and contextual semantics is configured to implement the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described in any one of claims 1 to 4.

7. A medium, characterized in that, The medium stores an intelligent decision support program based on cognitive logic and contextual semantics. When the intelligent decision support program based on cognitive logic and contextual semantics is executed by the processor, it implements the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described in any one of claims 1 to 4.

8. A computer program product, comprising an intelligent decision support program based on cognitive logic and contextual semantics, characterized in that, When the intelligent decision support program based on cognitive logic and contextual semantics is executed by the processor, it implements the steps of the intelligent decision support method based on cognitive logic and contextual semantics as described in any one of claims 1 to 4.

Images