Intelligent decision framework construction method and device based on dynamic ontology

Abstract

The invention relates to the technical field of intelligent decision rule base construction, and discloses an intelligent decision framework construction method and device based on a dynamic ontology. The method comprises the following steps: acquiring original data streams from a plurality of heterogeneous data sources in real time, and constructing an initial dynamic ontology structure through a dynamic ontology modeling unit; performing semantic annotation and relation extraction on the original data flow by using the structure to generate semantic enhanced data; generating a candidate decision rule set based on the semantic enhancement data, and obtaining a verified rule set through consistency verification and conflict detection; calculating the adaptability score of the verified rule set in combination with the real-time environment data, and screening out an optimal decision rule subset according to the score and a preset threshold value; the method is integrated into an intelligent decision rule base, and an optimization loop is triggered based on an update state of the base. According to the method, the data utilization efficiency and the rule quality are improved, the rule base can dynamically adapt to the environment, and decision effectiveness and reliability are enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision rule base construction technology, specifically to a method and apparatus for constructing an intelligent decision framework based on dynamic ontology. Background Technology

[0002] In various scenarios that rely on intelligent decision-making, whether it's process optimization in industrial production, risk control in the financial sector, or resource allocation in public services, the accuracy and timeliness of decisions depend heavily on the quality of the intelligent decision-making rule base (or intelligent decision-making framework). Existing technologies face numerous insurmountable bottlenecks in constructing intelligent decision-making rule bases.

[0003] In the data acquisition and utilization phase, data sources across various application scenarios often exhibit significant heterogeneity, encompassing multiple types such as structured, semi-structured, and unstructured data. The data formats and semantic definitions of different data sources vary greatly. Traditional methods typically only collect and process data from a single type of data source, making it difficult to achieve real-time and comprehensive acquisition from multiple heterogeneous data sources. This results in a one-sided data foundation upon which subsequent decision rule construction relies, failing to fully reflect the complexities of real-world scenarios. Even when some methods attempt to integrate multi-source data, the lack of effective semantic fusion techniques often leads to a "semantic gap" between data from different sources, making it difficult to form an effective dataset with unified semantics, thus affecting the rationality of decision rules.

[0004] In ontology modeling and rule generation, existing technologies mostly use static ontology structures as the basis for rule construction. Once a static ontology is built, its concept definitions, relationships between concepts, and other content remain fixed and cannot be adjusted in real time to adapt to dynamic changes in external environmental data. However, in real-world applications, environmental information is often in constant flux, such as fluctuations in market demand, aging of production equipment, and changes in user behavior. Static ontology struggles to effectively characterize this dynamic information, leading to a gradual decrease in the adaptability of the generated decision rules to real-world application scenarios. Furthermore, traditional rule generation processes are often disconnected from data semantics. Rule generation units often rely on simple data statistics or manual experience input, failing to fully explore the semantic relationships hidden behind the data. This results in candidate decision rule sets that are often redundant, semantically ambiguous, and have chaotic logical relationships between rules, making them difficult to directly apply to actual decision-making.

[0005] In the rule verification and optimization stages, existing technologies handle candidate decision rule sets rather crudely. Most methods only perform simple syntax error checks, neglecting potential logical consistency issues and conflicts between rules. When contradictory or duplicate rules exist in the rule set, it can lead to decision confusion during actual operation, preventing the system from outputting definite decisions. Even if some rules pass preliminary verification, existing technologies lack a scientific and effective mechanism to evaluate the applicability of rules in real-time environments. The applicability of rules is often greatly affected by real-time environmental data; a rule may be effective in a specific environment but may no longer be applicable after the environment changes. Once a traditional rule base is built, it lacks the ability to dynamically select the optimal rules based on the real-time environment, and can only use a fixed set of rules. This results in decision results that cannot respond to environmental changes in a timely manner, reducing the effectiveness of intelligent decision-making. Finally, existing rule bases generally lack a continuous optimization cycle mechanism. Rule base updates mostly rely on regular manual maintenance, which is time-consuming, costly, and difficult to precisely optimize based on the actual operating status and environmental changes of the rule base. This makes it difficult for the performance of the rule base to continuously improve with the development of application scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for constructing an intelligent decision-making framework based on dynamic ontology, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides a method for constructing an intelligent decision-making framework based on dynamic ontology, the method comprising:

[0008] Real-time acquisition of raw data streams from multiple heterogeneous data sources;

[0009] Based on the original data stream, an initial dynamic ontology structure is constructed using a dynamic ontology modeling unit;

[0010] Using the initial dynamic ontology structure, semantic annotation and relation extraction are performed on the original data stream to generate semantically enhanced data;

[0011] Based on the semantically enhanced data, a set of candidate decision rules is generated by the rule generation unit;

[0012] The candidate decision rule set is subjected to consistency verification and conflict detection to generate a verified rule set;

[0013] Based on real-time environmental data, the applicability score of the verified rule set is calculated by the rule evaluation unit;

[0014] Based on the applicability score and the preset threshold, the optimal subset of decision rules is selected;

[0015] The optimal decision rule subset is integrated into the intelligent decision rule base;

[0016] Based on the update status of the intelligent decision rule base, a rule base optimization loop is triggered.

[0017] Preferably, raw data streams are collected in real time from multiple heterogeneous data sources, including:

[0018] For each of the heterogeneous data sources, an independent data acquisition agent is configured, and the data acquisition agent continuously monitors the output of the data source.

[0019] The data acquisition agent extracts multimodal data packets from the heterogeneous data source;

[0020] The multimodal data packets are timestamped and formatted to generate a standardized data stream;

[0021] All the standardized data streams are aggregated to form the original data stream.

[0022] Preferably, based on the original data stream, an initial dynamic ontology structure is constructed through a dynamic ontology modeling unit, including:

[0023] Parse the entity and attribute information in the original data stream;

[0024] Using ontology learning algorithms, concept hierarchy and relational networks are identified from the entity and attribute information;

[0025] Based on the concept hierarchy and relation network, the initial dynamic ontology structure is generated, wherein the initial dynamic ontology structure includes concept nodes, relation edges, and attribute constraints.

[0026] Preferably, the initial dynamic ontology structure is used to perform semantic annotation and relation extraction on the original data stream to generate semantically enhanced data, including:

[0027] Map the original data stream onto the concept nodes of the initial dynamic ontology structure;

[0028] For each data point, add a semantic label according to the attribute constraints;

[0029] Extract the implicit relationships between data points and generate a relationship graph based on the relationship edges;

[0030] The semantically enhanced data is output by combining the semantic tags and the relationship graph.

[0031] Preferably, based on the semantically enhanced data, a set of candidate decision rules is generated by a rule generation unit, including:

[0032] Extract frequent pattern association rules from the semantically enhanced data;

[0033] The frequent pattern association rules are transformed into formal decision rules using a rule template library.

[0034] The formal decision rules are parameterized to generate the candidate decision rule set.

[0035] Preferably, the candidate decision rule set is subjected to consistency verification and conflict detection to generate a verified rule set, including:

[0036] Load historical rules from the existing intelligent decision-making rule base;

[0037] Compare the logical consistency between the candidate decision rule set and the historical rules;

[0038] Detect rule conflicts within the candidate decision rule set;

[0039] The detected rule conflicts are analyzed using a conflict resolution algorithm;

[0040] Output the verified rule set without any conflicts.

[0041] Preferably, based on real-time environmental data, the applicability score of the verified rule set is calculated by the rule evaluation unit, including:

[0042] Collect current environmental status data;

[0043] The current environment status data is matched with the condition portion of the verified rule set;

[0044] Calculate the condition matching degree and execution effect prediction value for each rule;

[0045] By combining the condition matching degree and the predicted execution effect value, an applicability score is generated for each rule.

[0046] Preferably, based on the applicability score and a preset threshold, an optimal subset of decision rules is selected, including:

[0047] Set the applicability rating threshold and priority weight;

[0048] The rules in the verified rule set are sorted according to their applicability scores;

[0049] Select rules whose applicability score is higher than the applicability score threshold;

[0050] The selection results are adjusted by applying priority weights to generate the optimal decision rule subset.

[0051] Preferably, based on the update status of the intelligent decision rule base, a rule base optimization loop is triggered, including:

[0052] Monitor the frequency of rule usage and effect feedback of the intelligent decision-making rule base;

[0053] When the frequency of rule usage is lower than a preset frequency threshold or the effect feedback is lower than a preset feedback threshold, the rule base optimization loop is started.

[0054] In the rule base optimization loop, raw data streams are collected again in real time from multiple heterogeneous data sources to update the intelligent decision rule base.

[0055] Preferably, the present invention also includes an intelligent decision rule base construction device based on dynamic ontology, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described intelligent decision framework construction method based on dynamic ontology.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] From the perspective of data processing origin, this method can collect raw data streams in real time from multiple heterogeneous data sources, breaking the dependence of traditional methods on a single data source. It achieves comprehensive capture of various types of structured, semi-structured, and unstructured data, ensuring that the data upon which decision rules are based fully covers various information within the application scenario, thus avoiding decision bias caused by data partiality. Simultaneously, the real-time acquisition characteristic ensures that the data promptly reflects the latest state of the environment, laying a solid data foundation for subsequently building decision rules that fit actual needs.

[0058] In the ontology modeling stage, the initial dynamic ontology structure constructed by the dynamic ontology modeling unit overcomes the limitations of traditional static ontology. Dynamic ontology can flexibly adjust its conceptual framework and inter-concept relationships based on new data features and semantic associations discovered during subsequent semantic annotation and relation extraction, as well as changes in real-time environmental data. This ensures that the ontology remains synchronized with the ever-changing application scenarios. This dynamic adaptability provides the decision rules generated based on the ontology with the foundation to cope with dynamic environmental changes, avoiding the problem of rules becoming disconnected from the actual scenario due to ontology rigidity.

[0059] By semantically annotating and extracting relationships from the original data stream using an initial dynamic ontology structure, the generated semantically enhanced data possesses immense value. Semantic annotation endows data, which previously lacked clear semantic definitions, with a unified semantic interpretation, eliminating the "semantic gap" between different heterogeneous data sources. This allows data to be understood and utilized within a unified semantic framework. Relationship extraction uncovers hidden intrinsic connections between data points, transforming previously isolated data into a logically linked organic whole. This semantically enhanced data provides rich and precise semantic support for subsequent rule generation units, enabling the generated candidate decision rule set to closely align with the actual business logic behind the data. This reduces rule redundancy and semantic ambiguity, thereby improving the overall quality of candidate rules.

[0060] Consistency verification and conflict detection of candidate decision rule sets effectively eliminate logical contradictions, duplicates, and redundancies, ensuring that each rule in the verified rule set is logically sound and unique. This process avoids the problem of decision system chaos caused by rule conflicts, allowing the rule set to stably output consistent decision-making criteria in practical applications and enhancing the reliability of the decision rule base.

[0061] The applicability score of the rule set, calculated and verified based on real-time environmental data, provides a scientific and objective basis for selecting the optimal subset of decision rules. Different rules have varying applicability in different environments. The applicability score accurately quantifies the degree of matching between each rule and the current real-time environment. Combined with the optimal subset of decision rules selected using preset thresholds, this ensures that the rules integrated into the intelligent decision rule base are those best suited to the actual needs of the current environment, resulting in highly targeted and effective decision outputs from the decision system.

[0062] By integrating a subset of optimal decision rules into the intelligent decision rule base, an optimization loop is triggered based on the rule base's update status, forming a continuously iterative and self-improving closed loop. Whenever the rule base is updated due to the integration of a new subset of optimal rules, the optimization loop is initiated, re-optimizing the rule base from stages such as data collection, ontology modeling, semantic enhancement, rule generation, verification, and scoring selection. This cyclical optimization mechanism allows the intelligent decision rule base to continuously adjust and improve itself in response to environmental changes and data updates, eliminating the need for large-scale, regular manual maintenance, reducing maintenance costs, and ensuring that the rule base's performance continuously improves with the development of application scenarios, enabling the intelligent decision system to maintain efficient and accurate decision-making capabilities in the long term. Whether it's helping enterprises optimize production processes in real time in industrial production, assisting institutions in accurately controlling risks in the financial sector, or assisting departments in efficiently allocating resources in public services, the intelligent decision rule base constructed using this method can play a significant role in promoting higher-quality operation in various intelligent decision-making scenarios. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent decision-making framework construction method based on dynamic ontology described in this invention.

[0064] Figure 2 A flowchart for collecting raw data streams from multiple heterogeneous data sources;

[0065] Figure 3 A flowchart for generating semantically enhanced data. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Please see Figure 1This invention provides a method and apparatus for constructing an intelligent decision-making framework based on dynamic ontology. The method includes: collecting raw data streams in real time from multiple heterogeneous data sources; constructing an initial dynamic ontology structure based on the data streams; performing semantic annotation and relation extraction to generate semantically enhanced data; generating a candidate decision rule set using the semantically enhanced data; forming a validated rule set after consistency verification and conflict detection; calculating applicability scores based on real-time environmental data and selecting the optimal subset of decision rules; and finally integrating the subset into an intelligent decision rule base and triggering an optimization loop based on the update status. This method relies on dynamic ontology technology to achieve autonomous evolution and continuous optimization of the rule base. Through deep fusion and semantic enhancement of multi-source heterogeneous data, it significantly improves data utilization and knowledge discovery capabilities. A multi-level rule verification mechanism ensures the consistency and reliability of the decision-making system. Combined with real-time environmental perception and rule applicability evaluation, the decision-making process has high context adaptability and accuracy. The resulting closed-loop optimization architecture enables the intelligent decision-making system to continuously adapt to complex and changing environmental conditions, maintaining decision accuracy and execution efficiency in dynamic scenarios, providing efficient and reliable technical support for complex decision-making scenarios in fields such as intelligent manufacturing, the Internet of Things, and smart cities.

[0068] Example 1: See Figure 2 This process involves real-time acquisition of raw data streams from multiple heterogeneous data sources and the construction of an initial dynamic ontology structure. This is achieved through the configuration of independent data acquisition agents, each customized for the specific characteristics of its data source. For example, the agent for relational databases uses a JDBC connection pool to listen for data table change events, capturing incremental and full data updates. The agent for message queues continuously consumes message streams from topics using a publish-subscribe pattern. For IoT sensor networks, a lightweight MQTT protocol is used to establish long-lived connections to receive telemetry data packets. During operation, the data acquisition agents maintain heartbeat mechanisms and retry logic to ensure the continuity and reliability of data transmission, preventing data loss due to network fluctuations or source-end anomalies. The acquired multimodal data packets retain their original format and metadata, including binary image streams, structured records, and text log fragments.

[0069] The timestamp alignment operation for multimodal data packets uses the Network Time Protocol to synchronize the local clocks of each data source, eliminating time offsets caused by the distribution of acquisition devices in different time zones or differences in system time settings. The alignment process is based on Coordinated Universal Time (UTC), and a globally consistent time stamp is attached to each data packet, accurate to the millisecond level to meet real-time processing requirements. Format standardization converts heterogeneous data packets into a unified intermediate representation, such as parsing XML documents into JSON objects or decoding binary image streams into Base64 encoded strings. At the same time, key metadata such as data source identifiers, sampling rates, and field descriptors are extracted. The standardized data stream is buffered and transmitted through a distributed message pipeline to ensure a balance between data ordering and throughput.

[0070] The process aggregates all standardized data streams to form a raw data stream using a windowing mechanism within a stream processing framework. Data packets are batch-combined by dividing the data stream into windows based on time or quantity. The window size is dynamically adjusted according to the data rate; under high load, the window shrinks to reduce latency, while under low load, the window expands to improve processing efficiency. The raw data stream enters subsequent processing modules as a continuous event stream, preserving complete temporal characteristics and source tracing information. When constructing the initial dynamic ontology structure, entity and attribute information in the raw data stream is parsed. Entity recognition algorithms scan structured fields and unstructured text in the data stream, using named entity recognition technology to identify entity types such as people, locations, and organizations. Attribute extraction retrieves field constraints, such as data type, value range, and optionality, from database schemas or data dictionaries. The parsing process simultaneously handles semantic ambiguity and synonym mapping, normalizing similar concepts from different data sources to a unified identifier. Ontology learning algorithms are used to identify concept hierarchy and relationship network from entity and attribute information. The concept hierarchy is constructed using a hierarchical clustering algorithm. The similarity of entities is quantified by calculating the cosine similarity between entity attribute vectors (applicable to numerical attributes) or the Jaccard coefficient (applicable to categorical attributes), and a similarity matrix is ​​constructed. A bottom-up aggregation strategy is adopted, initially treating each entity as an independent cluster, and gradually merging the clusters with the highest similarity until a parent-child relationship tree that meets the preset hierarchical depth is formed (e.g., the optimal number of clusters is determined by the Silhouette Score). Relationship network discovery is based on association rule mining technology, employing the FP-Growth algorithm (an efficient algorithm for mining frequent patterns in large-scale data). Entity pairs are used as "itemsets," and co-occurrence frequency is used as "support." An FP-tree is constructed to mine association rules that meet minimum support (e.g., 5%) and confidence (e.g., 80%). For the rule "Entity A → Entity B," the lift (Lift) is calculated to quantify the relationship strength (Lift > 1 indicates a positive correlation between A and B; the larger the Lift, the more significant the relationship). Finally, rules with high confidence and high lift are retained as ontology relation edges (e.g., "located in," "belongs to," etc.). The algorithm iteratively optimizes the hierarchical structure and relation weights, gradually improving the accuracy and coverage of the ontology model. An initial dynamic ontology structure is generated based on the concept hierarchy and relation network. This structure is stored in a graph database in the form of a directed graph. Nodes represent concept categories and are subject to attribute constraints. Edges represent semantic relations and are labeled with relation type and strength. Attribute constraints define the value range and data type of concept instances. For example, for numeric types, upper and lower bounds need to be specified, and for character types, regular expression patterns need to be defined. The initial dynamic ontology structure is output as a machine-readable OWL file or a set of RDF triples for use in subsequent semantic enhancement steps.

[0071] Example 2: See Figure 3This process utilizes an initial dynamic ontology structure to perform semantic annotation and relation extraction on the raw data stream to generate semantically enhanced data. First, the raw data stream is mapped to concept nodes in the initial dynamic ontology structure. This mapping operation is based on a combination of semantic similarity calculation and rule matching: for structured data fields, an exact matching algorithm is used to compare data values ​​with the ontology's concept identifiers; for unstructured text content, word embedding techniques and cosine similarity calculation are used to find the most matching concept node. During the mapping process, a dynamic cache layer is maintained to store frequently accessed concept mapping relationships, improving the efficiency of large-scale data processing. Each successfully mapped data point records its corresponding concept node URI as metadata tags. For each data point, a semantic tag is added based on attribute constraints. These constraints come from the concept attribute specifications defined in the initial dynamic ontology structure, including data type constraints, value ranges, and business rule restrictions: numeric data points need to be verified to be within a defined numerical range, character data points need to match regular expression patterns, and enumerated data points need to be checked to see if they belong to a predefined set of values. Data points that pass the constraint validation are annotated with semantic annotations in OWL format, including concept type, attribute value, and confidence score. Data points that fail the validation are entered into the anomaly handling process for manual review or rule correction.

[0072] The implicit relationships between data points are extracted based on the relation edges of the initial dynamic ontology structure. These relation edges define the allowed semantic relationship types between concepts and their constraints. A graph traversal algorithm is used to explore potential connection paths starting from labeled data points and following relation edges. Temporal relationships are identified by analyzing the timestamp sequences of data points to discover causal associations, and spatial relationships are determined by calculating the distance between geographic coordinates and topological relationships. The relationship extraction process employs a multi-strategy fusion approach: a rule-based method executes predefined relationship derivation rules; a statistical method analyzes the co-occurrence patterns and correlation coefficients of data points; and a machine learning method uses a neural network model to predict potential relationship types. The generated relationship graph is stored and maintained using an attribute graph model. Nodes represent data point instances with semantic labels, and edges represent extracted semantic relationship instances, labeled with relationship type, direction, and temporal validity. The graph update mechanism supports incremental expansion and dynamic weight adjustment, and relationship strength is dynamically decayed or enhanced based on freshness and confidence.

[0073] The system combines semantic tags and relationship graphs to output semantically enhanced data. This data is serialized using an extended JSON-LD format, preserving the original data values ​​while embedding structured annotations: the `@context` field defines the namespace mapping of ontology concepts, the `@graph` field contains a graph structure representation of data points and their relationships, and the `@timestamp` field records the time stamp of the data processing. The semantically enhanced data is distributed to downstream processing modules via a message queue and persistently stored in a distributed file system for historical tracing and offline analysis. In practical applications of smart grid monitoring systems, the implementation of Example 2 demonstrates its ability to handle complex data streams. This system needs to process multi-source heterogeneous data from generator sets, transmission lines, distribution transformers, and user meters in real time. This data includes equipment operating parameters, power quality indicators, environmental monitoring data, and user electricity consumption records. The initial dynamic ontology structure has established a power grid domain model containing concept nodes such as "generator," "transmission line," and "electricity load," as well as relational edges such as "supplying to," "connected to," and "affecting."

[0074] When the raw data stream enters the semantic processing module, the system first maps real-time data packets to concept nodes in the ontology structure: voltage and current readings are mapped to the "electrical parameters" concept, equipment status codes are mapped to the "operating status" concept, and geographic coordinates are mapped to the "equipment location" concept. The mapping process employs a multi-level caching strategy, residing frequently accessed concept nodes in memory to accelerate matching. For fuzzy concepts, the nearest neighbor algorithm is used to find the closest concept node in the vector space. When adding semantic labels to each data point, the system verifies them based on attribute constraints in the ontology: voltage values ​​are checked to ensure they are within a reasonable range (0-500kV), frequency values ​​are verified to ensure they are within the standard power frequency range (49-51Hz), and equipment status codes are confirmed to belong to a predefined status set. Data points that pass verification receive semantic annotations, including data source, credibility indicators, and timeliness markers. Data points that fail verification are transferred to the exception handling process, and a quality alarm is triggered.

[0075] When extracting implicit relationships between data points, the system discovers potential connections based on spatiotemporal correlation analysis: equipment within the same substation automatically establishes a "co-location" relationship, monitoring points with the same timestamp establish a "synchronous monitoring" relationship, and voltage fluctuations and subsequent load changes establish a "potential impact" relationship. The relationship extraction algorithm uses a sliding time window to analyze time series patterns and discovers statistical dependencies between variables through mutual information calculation. The generated relationship graph adopts a dynamic update mechanism, with node attributes reflecting the latest equipment status in real time, and edge weights dynamically adjusted according to relationship strength: the weight of frequently occurring equipment relationships increases, while the weight of equipment relationships that have not interacted for a long time decreases. The graph is stored using an in-memory graph database, supporting millisecond-level multi-hop relationship queries and path analysis. The output semantically enhanced data is encapsulated in an extended JSON format, with raw data values ​​retained in the "rawValue" field, semantic annotations stored in the "semanticTags" object, and relationship information recorded in the form of a "relationships" array. Each data packet is appended with a processing timestamp and quality score for downstream systems to judge data freshness and reliability. In power grid control scenarios, this semantic enhancement enables the system to understand complex semantic relationships such as "a sudden voltage drop at a substation may cause a power outage for downstream users," providing rich contextual information for intelligent decision-making.

[0076] Example 3: The process of generating candidate decision rule sets from semantically enhanced data and performing consistency verification and conflict detection on these sets. This process first extracts frequent pattern association rules from the semantically enhanced data. Frequent pattern mining uses an improved FP-Growth algorithm to process the graph structure information in the semantically enhanced data. The algorithm extends the traditional frequent itemset concept to adapt to complex data patterns containing semantic relationships and attribute constraints: by traversing the relation graph in the semantically enhanced data, it identifies frequently co-occurring concept node combinations and their associations, forming a weighted pattern set. Pattern support calculation considers not only frequency of occurrence but also incorporates semantic similarity factors, allowing patterns that are semantically similar but have different expressions to be reasonably aggregated. A rule template library is applied to transform frequent pattern association rules into formal decision rules. The rule template library is organized in an object-oriented manner, containing two main categories: condition templates and action templates. Condition templates define various logical expression structures, such as comparison expressions, set operations, and temporal constraints; action templates define executable operation types, such as state setting, message sending, and process triggering. The template instantiation process maps frequent patterns to the most matching template structure, automatically filling in parameters and logical operators to generate formal decision rules that conform to standard syntax specifications.

[0077] Formal decision rules are parameterized using an optimization algorithm based on historical data. The algorithm analyzes the statistical correlation between the antecedent and consequent of the rule in historical instances, and automatically calibrates threshold parameters and confidence indices. For numerical parameters, a sliding window statistical method is used to calculate the optimal value range; for categorical parameters, information gain analysis is used to determine the most discriminative value combination. A weighted support index is introduced during the parameter adjustment process to guide the optimization direction.

[0078]

[0079] in: Indicates weighted support. The semantic weight representing the i-th pattern. It is a traditional support metric. This represents the similarity between the pattern vector and the ideal direction. The adjusted rule set is then output as candidate decision rules to the verification stage.

[0080] When performing consistency checks on the candidate decision rule set, historical rules from the existing intelligent decision rule library are loaded to establish a verification benchmark. These historical rules are represented using a logic programming language and loaded into the working memory of the inference engine. The consistency check employs descriptive logic reasoning techniques to detect sub-sumption relationships between new and existing rules: by calculating the logical inclusion relationship of the rule's condition part and the compatibility of its action part, potential logical contradictions are identified. The check process pays particular attention to handling rule exceptions, ensuring that new rules do not disrupt the exception handling mechanism of the existing rule system. A multi-dimensional conflict detection strategy is used to detect rule conflicts within the candidate decision rule set: syntactic conflict detection checks whether the rule structure conforms to syntactic rules; semantic conflict analysis examines the behavioral interactions of rules in specific scenarios; and temporal conflict detection focuses on race conditions that may result from the order of rule execution. The conflict detection algorithm establishes a rule influence propagation model, simulating the path of system state changes after rule execution, and identifying rule pairs that may produce contradictory state transitions.

[0081] Detected rule conflicts are analyzed using a conflict resolution algorithm. This algorithm employs a priority-based conflict resolution framework: each rule is assigned a dynamic priority score, determined by the rule's source authority, historical execution success rate, and business urgency. For conflicting rules with overlapping conditions, a rule merging technique is used to generate a comprehensive rule, eliminating conflicts by adjusting the condition scope and adding exception clauses. The resolution process retains all conflict resolution traces, forming an auditable decision log. The final verified rule set guarantees logical consistency and execution compatibility, providing reliable input for subsequent evaluation.

[0082] The semantically enhanced data collected by the power grid monitoring system contains rich information such as equipment status, power parameters, and environmental factors. After ontology annotation, this data forms a structured information set with clear semantic relationships. The rule generation unit first uses an improved Apriori algorithm to mine frequent patterns from this data. The algorithm is particularly adapted to the characteristics of time-series data and can discover time-series association patterns such as "a sudden voltage drop event is often accompanied by a sudden load change within 2 minutes". The minimum support is set to 0.1 and the minimum confidence is 0.7. The generated association rules cover various types such as equipment anomalies, network faults, and environmental impacts. When applying the rule template library, the system calls predefined fault diagnosis rule templates. These templates adopt an IF-THEN structure and contain power grid-specific condition expressions: the IF part supports condition types such as equipment status judgment, electrical parameter threshold comparison, and time-series relationship description, while the THEN part contains action types such as fault classification, warning level, and handling suggestions. The template instantiation process converts frequent patterns into formal rules that conform to IEEE standards. For example, the frequent pattern "transformer oil temperature > 85℃ and load rate > 90%" is instantiated into the formal rule "IF transformer_oil_temperature > 85 AND loading_rate > 90 THEN issue_overheat_alert".

[0083] The parameterization adjustment phase involves fine-tuning the generated formal rules. The system accesses the historical fault record database to analyze the correlation between rule condition parameters and fault occurrence: for over-temperature alarm rules, the threshold is optimized from the initial 85℃ to 82℃ by statistically analyzing the actual fault occurrence rate under different temperature thresholds; for voltage sag rules, the duration threshold is adjusted from 1 second to 0.8 seconds based on equipment tolerance data analysis. Seasonal variations are considered during the adjustment process, with the over-temperature threshold appropriately lowered in summer and the triggering conditions of the anti-icing flashover rule adjusted in winter. The candidate rule set with completed parameter adjustments enters the verification phase. The system first loads 278 historical rules from the existing fault diagnosis rule library to establish a verification benchmark. Consistency verification employs a description-based inference engine to examine the logical relationship between new and existing rules. A new rule, "IFvoltage_dip<0.8puTHENclassify_voltage_sag", was found to have overlapping conditions but inconsistent classification results with the existing rule "IFvoltage<0.75puTHENclassify_voltage_sag", thus marking it as a logical conflict. Another new rule, "IFline_temperature>100℃THENissue_emergency_alert", had a threshold conflict with the existing rule "IFline_temperature>120℃THENissue_emergency_alert".

[0084] The conflict detection employs a multi-dimensional analysis strategy, identifying three rules within the candidate rule set that offer different handling suggestions for the same fault phenomenon: Rule A suggests immediately disconnecting the load, Rule B suggests activating backup power, and Rule C suggests adjusting the network topology. The system's impact propagation simulation demonstrates that simultaneous triggering of these three rules could lead to operational conflicts and equipment damage. The conflict resolution algorithm assigns a priority score to each rule, considering the reliability of the rule source, historical execution success rate, and fault severity: rules confirmed by the power grid dispatch center are given the highest priority, automatically generated rules have moderate priorities, and rules based on field experience are used as a reference. The resolution process generates detailed audit logs, recording the basis for resolving each conflict and the adjustments made. The final verified rule set contains 152 conflict-free fault diagnosis rules, categorized by voltage level and equipment type. Each rule includes a version identifier and verification certificate, ready for subsequent evaluation and deployment into the real-time fault diagnosis system.

[0085] Example 4: The process of calculating and validating the applicability score of the rule set after real-time environmental data verification, and selecting the optimal decision rule subset based on the score and threshold, is implemented through a rule evaluation unit to achieve a dynamic rule optimization mechanism. In a specific application scenario of an intelligent temperature control system, the system continuously monitors parameters such as ambient temperature, humidity, and personnel activity status. The verified rule set contains multiple decision rules for adjusting air conditioning and ventilation equipment. The rule evaluation unit collects current environmental state data through a distributed sensor network. Temperature sensors report current readings once per minute, humidity sensors detect air humidity, and infrared sensors capture motion signals within the area. The collected data is preprocessed to form a unified format environmental state snapshot, including fields such as timestamp, sensor ID, numerical type, and measured value.

[0086] When matching environmental state data with the conditional components of the validated rule set, a multi-pattern matching algorithm is used to process different types of conditional expressions. Numerical conditions, such as "temperature > 28℃", are matched through range comparison; Boolean conditions, such as "personnel presence = true", are validated through exact matching; and composite conditions are parsed using an abstract syntax tree and then recursively evaluated. The matching process generates a conditional matching degree value for each rule, which ranges from 0 to 1, representing the degree to which the current environment satisfies the rule conditions. The predicted execution effect value for each rule is calculated based on a prediction model built from historical execution records. The model analyzes the historical execution results of the rule under similar environmental conditions, including factors such as energy consumption changes, comfort indices, and equipment wear. The predicted value calculation comprehensively considers short-term benefits and long-term impacts, and a weighted average method is used to generate a standardized prediction score between 0 and 1.

[0087] The applicability score is generated by combining the condition matching degree and the execution effect prediction value in a linear combination method. The weight of the condition matching degree focuses on the compliance with the current environment, while the weight of the execution effect prediction value reflects the historical performance of the rule. The scoring formula parameters can be adjusted according to the application scenario. In high-temperature environments, the weight of energy-saving rules may be increased, while in office scenarios, comfort indicators are emphasized.

[0088] Table 1: Rule Evaluation Results

[0089] Rule Number Temperature conditions Humidity conditions Personnel activities Condition matching degree Predicted effect value Applicability rating R202 >28℃ <70% any 0.92 0.85 0.89 R305 >30℃ <65% exist 0.76 0.92 0.83 R418 >26℃ <75% Does not exist 0.88 0.78 0.84 R527 >29℃ <60% any 0.95 0.72 0.85 R639 >27℃ <80% exist 0.82 0.88 0.84

[0090] Referring to Table 1, when setting the applicability score threshold and priority weight, the system administrator configures the baseline parameters according to business needs. The score threshold is usually set between 0.7 and 0.9 to avoid selecting inefficient rules. The priority weight allocation takes into account the differences in importance of rule types. The weight coefficient for safety-related rules is set to 1.2, the weight for comfort rules is 1.0, and the weight for energy-saving rules is 0.9. The weight coefficient is multiplied by the applicability score to generate the adjusted final score.

[0091] The rules in the validated rule set are sorted by their applicability scores using a quicksort algorithm. The algorithm arranges the rule list in descending order of score value, and rules with the same score are sorted in ascending order by rule number to maintain determinism. During the sorting process, the integrity of rule metadata is maintained to ensure traceability. Rules with applicability scores above a threshold are selected by scanning the sorted rule list using a traversal algorithm. All rules with scores exceeding the threshold are collected to form an initial candidate set. The selection rationale and score details for each rule are recorded for subsequent auditing. Boundary cases are specially handled during the selection process; rules with scores exactly equal to the threshold are selected using an upward-taking principle.

[0092] When adjusting the selection results using priority weights, the system recalculates the weighted scores of the rules in the initial candidate set according to their weight coefficients. A second round of sorting and filtering is then performed based on these weighted scores to ensure that high-priority rules receive an appropriate advantage. Weight adjustments may alter the selection status of rules, potentially allowing some low-scoring but high-weight rules to enter the final subset. The generated optimal decision rule subset is output as an XML configuration file, containing information such as rule ID, execution parameters, and effective time. A selection report explaining the filtering logic and weight allocation basis is also generated. The subset is distributed to the rule execution engine via a message bus and takes effect immediately. The entire evaluation and filtering process is executed automatically and periodically to ensure the rule base continuously adapts to environmental changes.

[0093] Example 5: The process of triggering rule base optimization loops based on the update status of the intelligent decision rule base is demonstrated in the actual operation of a smart home temperature control system. After system deployment, the actual performance of each rule in the rule base is continuously monitored. An embedded monitoring agent collects execution logs and feedback data for each rule. Monitoring data includes multiple indicators such as rule trigger count, execution time, resource consumption, and user adjustment records. Rule usage frequency statistics employ a sliding time window algorithm, calculating the call frequency of each rule across three dimensions: 1 hour, 24 hours, and 7 days, generating time-series data for trend analysis. Effect feedback is collected through multiple channels: sensors directly measure changes in environmental parameters, user manual adjustments reflect preference deviations, and device status logs record execution success or failure. This feedback data is normalized and converted into an effect score within the range of 0-1.

[0094] When a rule's usage frequency falls below a preset threshold, the system initiates an anomaly detection process. The frequency threshold is dynamically adjusted based on the rule type: critical safety rules are triggered at least once daily, comfort rules at least three times weekly, and energy-saving rules at least five times monthly. When feedback is below a preset feedback threshold, the system analyzes the specific reasons. Feedback thresholds are differentiated by rule category: temperature control rules require a feedback score of at least 0.8, humidity control rules at least 0.7, and fresh air control rules at least 0.75. Threshold judgment employs a multi-condition triggering mechanism; if any condition is not met, the rule is marked as needing optimization. Dependencies between rules are also considered to avoid misjudgments due to the failure of associated rules.

[0095] When initiating the rule base optimization loop, the system first creates a rule base snapshot to save the current state. Then, it suspends the execution of inefficient rules without immediately deleting them, allowing the system to continue running the remaining rules. The optimization loop calls the data acquisition module to re-collect raw data streams from multiple heterogeneous data sources in real time. In the temperature control system example, these data sources include: indoor and outdoor temperature sensor networks, humidity detection devices, human presence sensors, air conditioning equipment operating status interfaces, weather forecast data interfaces, and user-manually set records. During data acquisition, the sampling frequency and range are adjusted, and the acquisition of parameters related to rules with degraded performance is strengthened, such as increasing the sampling density of sensors near heat sources for temperature control rules. The re-acquired data stream undergoes dynamic ontology modeling and semantic enhancement processing to generate updated semantically enhanced data, which specifically contains details of environmental parameter changes during periods of rule performance degradation. The rule generation unit generates candidate rules based on the new semantic data, focusing on the application scenarios of previously inefficient rules and using machine learning to analyze the reasons for rule failure caused by changes in environmental patterns. After verification and evaluation, the newly generated rules are updated in a progressive manner to update the intelligent decision-making rule base: 1. Test environment verification: Offline simulation testing is used, inputting historical environment datasets (covering normal / extreme scenarios), comparing the decision accuracy (e.g., temperature control error ≤ ±0.5℃), response latency (≤100ms), and resource consumption (CPU utilization ≤15%) of the new and old rules. The new rules must meet the requirement of an accuracy improvement of ≥5% and latency / resource consumption no worse than the old rules; 2. Small-scale trial operation: Select typical user groups (e.g., 10% of household users) to deploy the new rules, monitor the user intervention rate (number of manual adjustments by users / number of automatic rule executions, must be ≤10%) and comfort score (user subjective feedback, must be ≥0.85) in real time, and enter the promotion stage after running continuously for 72 hours without any abnormalities; 3. Full system promotion: Adopt a gray release strategy, switch rules in batches by region (20% of users in each batch), monitor the system load (e.g., rule engine QPS fluctuation ≤20%) and global decision consistency (cross-region rule execution result deviation ≤3%) in real time, and complete the full update after ensuring that there are no large-scale abnormalities. The rule base is updated using a version control mechanism. Each optimization cycle generates a new rule base version, while retaining historical versions for easy rollback.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an intelligent decision-making framework based on dynamic ontology, characterized in that, include: Real-time acquisition of raw data streams from multiple heterogeneous data sources; Based on the original data stream, an initial dynamic ontology structure is constructed using a dynamic ontology modeling unit; Using the initial dynamic ontology structure, semantic annotation and relation extraction are performed on the original data stream to generate semantically enhanced data; Based on the semantically enhanced data, a set of candidate decision rules is generated by the rule generation unit; The candidate decision rule set is subjected to consistency verification and conflict detection to generate a verified rule set; Based on real-time environmental data, the applicability score of the verified rule set is calculated by the rule evaluation unit; Based on the applicability score and the preset threshold, the optimal subset of decision rules is selected; The optimal decision rule subset is integrated into the intelligent decision rule base; Based on the update status of the intelligent decision rule base, a rule base optimization loop is triggered.

2. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 1, characterized in that, Raw data streams are collected in real time from multiple heterogeneous data sources, including: For each of the heterogeneous data sources, an independent data acquisition agent is configured, and the data acquisition agent continuously monitors the output of the data source. The data acquisition agent extracts multimodal data packets from the heterogeneous data source; The multimodal data packets are timestamped and formatted to generate a standardized data stream; All the standardized data streams are aggregated to form the original data stream.

3. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 2, characterized in that, Based on the original data stream, an initial dynamic ontology structure is constructed through a dynamic ontology modeling unit, including: Parse the entity and attribute information in the original data stream; Using ontology learning algorithms, concept hierarchy and relational networks are identified from the entity and attribute information; Based on the concept hierarchy and relation network, the initial dynamic ontology structure is generated, wherein the initial dynamic ontology structure includes concept nodes, relation edges, and attribute constraints.

4. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 3, characterized in that, Using the initial dynamic ontology structure, semantic annotation and relation extraction are performed on the original data stream to generate semantically enhanced data, including: Map the original data stream onto the concept nodes of the initial dynamic ontology structure; For each data point, add a semantic label according to the attribute constraints; Extract the implicit relationships between data points and generate a relationship graph based on the relationship edges; The semantically enhanced data is output by combining the semantic tags and the relationship graph.

5. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 4, characterized in that, Based on the semantically enhanced data, a set of candidate decision rules is generated by the rule generation unit, including: Extract frequent pattern association rules from the semantically enhanced data; The frequent pattern association rules are transformed into formal decision rules using a rule template library. The formal decision rules are parameterized to generate the candidate decision rule set.

6. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 5, characterized in that, The candidate decision rule set is subjected to consistency verification and conflict detection to generate a verified rule set, including: Load historical rules from the existing intelligent decision-making rule base; Compare the logical consistency between the candidate decision rule set and the historical rules; Detect rule conflicts within the candidate decision rule set; The detected rule conflicts are analyzed using a conflict resolution algorithm; Output the verified rule set without any conflicts.

7. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 6, characterized in that, Based on real-time environmental data, the applicability score of the verified rule set is calculated by the rule evaluation unit, including: Collect current environmental status data; The current environment status data is matched with the condition portion of the verified rule set; Calculate the condition matching degree and execution effect prediction value for each rule; By combining the condition matching degree and the predicted execution effect value, an applicability score is generated for each rule.

8. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 7, characterized in that, Based on the applicability score and preset threshold, an optimal subset of decision rules is selected, including: Set the applicability rating threshold and priority weight; The rules in the verified rule set are sorted according to their applicability scores; Select rules whose applicability score is higher than the applicability score threshold; The selection results are adjusted by applying priority weights to generate the optimal decision rule subset.

9. The method for constructing an intelligent decision-making framework based on dynamic ontology according to claim 8, characterized in that, Based on the update status of the intelligent decision rule base, a rule base optimization loop is triggered, including: Monitor the frequency of rule usage and effect feedback of the intelligent decision-making rule base; When the frequency of rule usage is lower than a preset frequency threshold or the effect feedback is lower than a preset feedback threshold, the rule base optimization loop is started. In the rule base optimization loop, raw data streams are collected again in real time from multiple heterogeneous data sources to update the intelligent decision rule base.

10. An apparatus for constructing an intelligent decision rule base based on dynamic ontology, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for constructing an intelligent decision-making framework based on dynamic ontology as described in any one of claims 1 to 9.

Images