Hydropower engineering carbon emission intelligent supervision method and system fusing rule engine

Abstract

The application relates to the technical field of digital supervision of carbon emission of hydropower engineering, and particularly discloses a hydropower engineering carbon emission intelligent supervision method and system fusing a rule engine, which first performs timestamp alignment and vectorization encapsulation on multi-source carbon monitoring data, and matches device numbers with a carbon supervision space-time knowledge graph; then, taking the focus entity obtained through matching as the center, the graph is traversed and pruned according to space-time conditions to generate a construction context subgraph and an event trigger load; then, the node attributes and edge relations of the subgraph are analyzed, converted and compiled into a working condition dynamic rule set; the event trigger load is injected into a rule engine session and mode matching reasoning is performed to obtain carbon emission abnormality assertion results, on the basis of which closed-loop rectification instructions are generated, and the dynamic rules are destroyed and the session memory is released after reasoning is completed. In this way, the accuracy of abnormality research and judgment under complex construction conditions can be improved, and rule redundancy and the occupation of computing resources can be reduced.

Description

Technical Field

[0001] This application relates to the field of digital monitoring technology for carbon emissions from hydropower projects, and more specifically, to a method and system for intelligent monitoring of carbon emissions from hydropower projects that integrates a rule engine. Background Technology

[0002] Hydropower projects involve long construction periods and wide construction areas, with numerous emission sources such as diesel generators, heavy transport vehicles, and concrete pouring equipment, whose operating conditions change frequently. Carbon emission levels are also affected by factors such as the construction stage, the degree of equipment concentration, the work area, and ambient temperature. Traditional manual ledger accounting and post-event statistics methods are insufficient to identify local emission anomalies in a timely manner, failing to meet the requirements for real-time monitoring, accurate judgment, and closed-loop management during construction. Therefore, automated processing and anomaly analysis of multi-source carbon monitoring data are needed to achieve intelligent supervision of carbon emissions from hydropower projects.

[0003] Existing carbon emission monitoring schemes for hydropower projects typically collect data such as equipment fuel consumption, vehicle trajectories, and environmental parameters, using fixed thresholds or manually configured rules for anomaly detection. Some schemes also utilize knowledge graphs to store the relationships between equipment, processes, regions, regulatory indicators, and rectification measures. However, knowledge graphs primarily describe entity relationships using nodes and edges, while rule engines rely on defined logical conditions to perform matching reasoning. The two differ in data structure and computational methods, making it difficult to automatically convert node attributes and edge relationships in knowledge graphs into executable rules adapted to the current construction conditions, thus hindering their use for rule-based judgment. Therefore, existing rule engines often directly input monitored values ​​into static rules, judging solely based on whether the values ​​exceed limits. This fails to fully perceive the construction stage, spatial region, associated equipment, and environmental status of abnormal equipment, easily leading to false alarms. Pre-combining various equipment, regions, processes, and environmental conditions into rules would result in an excessive number of rules, increasing rule maintenance difficulty, matching computation overhead, and memory usage.

[0004] Furthermore, the sampling frequency of multi-source monitoring data is not consistent. Sensors in high-altitude environments may also generate data or interpolated data that have not been updated for a long time due to freezing or communication interruption. Existing pattern matching networks usually use equal-weighted binary judgment for data of different freshness, making it difficult to distinguish the reliability of real-time sampled data from old filled data. When low-reliability data slightly exceeds the limit, it may still trigger rule activation, thereby increasing invalid conflict arbitration and causing rectification instructions to be triggered erroneously.

[0005] Therefore, an optimized intelligent monitoring solution for carbon emissions from hydropower projects is desired. Summary of the Invention

[0006] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for intelligent monitoring of carbon emissions from hydropower projects that integrates a rule engine.

[0007] According to one aspect of this application, a method for intelligent monitoring of carbon emissions from hydropower projects integrating a rule engine is provided, comprising: Step 1: Timestamp alignment and vectorization encapsulation of multi-source carbon monitoring data to obtain carbon emission feature vectors, and matching device numbers with nodes of a pre-built carbon regulatory spatiotemporal knowledge graph to obtain focus entity identifiers; Step 2: Using the spatiotemporal coordinates in the carbon emission feature vector as pruning conditions, and centering on the focal entity identifier, perform constrained breadth traversal and dynamic subgraph pruning in the carbon regulatory spatiotemporal knowledge graph to obtain the construction context subgraph, and serialize the numerical attributes in the carbon emission feature vector into event triggering loads. Step 3: Use the domain ontology converter to perform topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph to obtain the dynamic rule set of the working condition. Step 4: Load the dynamic rule set of operating conditions into the isolated rule engine session, inject the event trigger payload as a fact object into the working memory, and perform deductive reasoning through the pattern matching network to obtain the carbon emission anomaly assertion result. Step 5: Based on the abnormal cause code in the carbon emission anomaly assertion result, retrieve the rectification measures library to generate a closed-loop rectification instruction. After the reasoning is completed, destroy the dynamic rule set of the operating condition and release the rule engine session memory.

[0008] According to another aspect of this application, a smart monitoring system for carbon emissions from hydropower projects that integrates a rule engine is provided, comprising: The carbon feature encapsulation and entity matching module is used to perform timestamp alignment and vectorization encapsulation on multi-source carbon monitoring data to obtain carbon emission feature vectors, and to match the device number with the nodes of the pre-built carbon regulatory spatiotemporal knowledge graph to obtain the focus entity identifier. The subgraph pruning and load generation module is used to perform constrained breadth traversal and dynamic subgraph pruning in the carbon regulatory spatiotemporal knowledge graph with the spatiotemporal coordinates in the carbon emission feature vector as pruning conditions and the focus entity identifier as the center to obtain the construction context subgraph, and serialize the numerical attributes in the carbon emission feature vector into event-triggered loads. The graph semantic conversion and rule compilation module is used to perform topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph through the domain ontology converter to obtain the dynamic rule set of the working condition. The rule engine matching and reasoning module is used to load the dynamic rule set of working conditions into the isolated rule engine session, inject the event trigger payload as a fact object into the working memory, and perform deductive reasoning through the pattern matching network to obtain the carbon emission anomaly assertion result. The rectification loop and session destruction module is used to retrieve the rectification measure library based on the abnormal cause code in the carbon emission anomaly assertion result, generate a closed-loop rectification instruction, destroy the dynamic rule set of the working condition and release the rule engine session memory after the inference is completed.

[0009] Compared with existing technologies, this application provides a method and system for intelligent carbon emission monitoring of hydropower projects that integrates a rule engine. Driven by multi-source carbon monitoring events, it performs time-series alignment and feature encapsulation of equipment fuel consumption, vehicle trajectories, ambient temperature, and carbon emission factors in the ledger. It also associates monitoring equipment with a carbon regulatory spatiotemporal knowledge graph, dynamically extracting the current construction context based on event entities and their spatiotemporal conditions. The entity attributes and relationships in the context are converted into dynamic rules adapted to the current working conditions. Anomaly reasoning is completed in isolated sessions, combining monitoring facts to bridge the gap between graph semantics and executable rules, avoiding static thresholds detached from construction conditions and excessive expansion of pre-set rules. Simultaneously, it assesses the freshness and confidence of data in each dimension, suppressing and filtering low-confidence rule activation to reduce misjudgments caused by outdated or interpolated data. After reasoning, it generates rectification instructions and releases temporary rules and session resources, achieving accurate judgment and closed-loop monitoring of carbon emission anomalies. Attached Figure Description

[0010] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0011] Figure 1 This is a flowchart of a method for intelligent monitoring of carbon emissions from hydropower projects using a fusion rule engine, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in the intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine according to an embodiment of this application. Figure 3 This is a flowchart of step four of the intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine according to an embodiment of this application. Figure 4 The flowchart illustrates the intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine according to embodiments of this application. It describes how event-triggered loads are deserialized into fact objects and injected into the root node of an identification and matching network structure for layer-by-layer pattern propagation matching to obtain a conflict resolution agenda queue. Figure 5 This is a block diagram of a smart monitoring system for carbon emissions from hydropower projects based on an embodiment of this application using a fusion rule engine. Detailed Implementation

[0012] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0013] To address the challenges in existing carbon emission monitoring for hydropower projects, such as the difficulty in converting graph context into executable rules, the inability of static rules to adapt to changes in construction conditions, rule combination expansion, and the susceptibility of outdated or interpolated data to anomalies and misjudgments, this solution proposes an intelligent carbon emission monitoring method for hydropower projects that integrates a rule engine. First, it performs time alignment and vectorization processing on equipment fuel consumption, vehicle trajectory, ambient temperature, and carbon emission factors in the ledger, and determines the focal entities in the graph based on equipment codes. Then, using event time and spatial location as constraints, it trims the carbon monitoring spatiotemporal knowledge graph to form a structure encompassing equipment, processes, regions, and environmental relationships. The system extracts the construction context subgraph and extracts monitoring values ​​to form event trigger loads. Then, it parses the node attributes and edge relationships of the subgraph, transforms and compiles them into dynamic rules adapted to the current working conditions, injects event facts into the isolated rule engine session, and assigns confidence based on the sampling freshness of each data dimension. Anomaly inference is completed through soft gating propagation, trusted threshold filtering, and priority arbitration. Finally, it generates rectification instructions based on the cause of the anomaly and the degree of exceedance, and unloads the dynamic rules and releases session memory after the inference is completed, thereby reducing false alarms and invalid rule matching, controlling computation and storage overhead, and realizing closed-loop supervision of carbon emission anomalies.

[0014] Figure 1 This is a flowchart of a method for intelligent monitoring of carbon emissions from hydropower projects using a fusion rule engine, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a smart monitoring method for carbon emissions from hydropower projects based on an embodiment of this application using a fusion rule engine. Figure 1 and Figure 2As shown, the intelligent monitoring method for carbon emissions in hydropower projects using a fusion rule engine according to an embodiment of this application includes: Step 1, performing timestamp alignment and vectorization encapsulation on multi-source carbon monitoring data to obtain carbon emission feature vectors, and matching equipment numbers with nodes in a pre-built carbon monitoring spatiotemporal knowledge graph to obtain focus entity identifiers; Step 2, using the spatiotemporal coordinates in the carbon emission feature vectors as pruning conditions, performing constrained breadth-first traversal and dynamic subgraph pruning in the carbon monitoring spatiotemporal knowledge graph centered on the focus entity identifiers to obtain construction context subgraphs, and serializing the numerical attributes in the carbon emission feature vectors into... The process involves several steps: Step 3: Using a domain ontology converter, topological semantic parsing and logical antecedent mapping are performed on the node attributes and edge relationships of the construction context subgraph to obtain a dynamic rule set for the working conditions. Step 4: The dynamic rule set for the working conditions is loaded into an isolated rule engine session, and the event-triggered payload is injected into the working memory as a fact object. Deductive reasoning is then performed through a pattern matching network to obtain carbon emission anomaly assertion results. Step 5: Based on the anomaly cause code in the carbon emission anomaly assertion results, the rectification measure library is retrieved to generate closed-loop rectification instructions. After the reasoning is completed, the dynamic rule set for the working conditions is destroyed and the rule engine session memory is released.

[0015] Specifically, in step one, the multi-source carbon monitoring data is timestamped and vectorized to obtain carbon emission feature vectors, and the device number is matched with nodes in a pre-built carbon regulatory spatiotemporal knowledge graph to obtain focal entity identifiers. It is worth noting that the multi-source carbon monitoring data here includes real-time fuel consumption of diesel generators, location trajectory coordinates of heavy trucks, temperature in the dam construction area, and carbon emission factors in the ledger. It is understandable that, since the carbon emission monitoring data at hydropower construction sites comes from different types of sensors and management systems, the collection frequency, physical dimensions, and update cycles of each data source differ significantly. If unaligned and unstandardized heterogeneous data is directly fed into the subsequent inference process, it will lead to temporal misalignment causing event matching failures, dimensional confusion causing rule threshold comparison failures, and the lack of clear graph entity anchor points will prevent subsequent subgraph pruning from being performed. Therefore, in the technical solution of this application, carbon emission feature vectors are obtained by timestamping and vectorizing multi-source carbon monitoring data, and the device number is matched with the nodes of a pre-constructed carbon regulatory spatiotemporal knowledge graph to obtain the focus entity identifier. This unifies the heterogeneous and frequency-varying original monitoring data into a time-consistent, dimensionless structured feature representation, and determines the location node of the current carbon emission event in the knowledge graph. In this way, a unified input data foundation with a clear graph entry point can be provided for subsequent subgraph pruning and dynamic rule compilation, ensuring the continuity and effectiveness of the inference link.

[0016] Specifically, in the embodiments of this application, step one includes: performing heterogeneous frequency resampling and timestamp alignment on multi-source carbon monitoring data to obtain a time-aligned carbon monitoring matrix; performing dimension-wise standardization and tensor flattening encapsulation on the time-aligned carbon monitoring matrix based on the mean and standard deviation of the sliding time window to obtain a carbon emission feature vector; and performing entity addressing matching between the device identifier field in the carbon emission feature vector and the candidate node encoding in the carbon regulatory spatiotemporal knowledge graph based on device encoding to obtain a focal entity identifier.

[0017] More specifically, in a particular example of this application, multi-source carbon monitoring data is first collected at the hydropower project construction site. This multi-source carbon monitoring data includes real-time fuel consumption of diesel generators, positioning trajectory coordinates of heavy trucks, temperature of the dam pouring area, and carbon emission factors in the ledger. The real-time fuel consumption of diesel generators is collected once per second by a turbine flow meter installed on the fuel line of the generator set. The positioning trajectory coordinates of heavy trucks are reported in longitude, latitude, and altitude once every 5 seconds by an on-board Beidou positioning terminal. The temperature of the dam pouring area is transmitted back once every 30 seconds by a temperature sensor embedded in the concrete surface. The carbon emission factors in the ledger are maintained by the project carbon emission management system according to construction batches and updated daily.

[0018] Next, using the sampling timestamp sequence of real-time fuel consumption of diesel generators as the global reference time series, for time points in the heavy truck positioning trajectory coordinates and dam pouring area temperature where the sampling interval is greater than the reference interval, if there are valid sampling values ​​before and after the time point, weighted interpolation is performed on the two adjacent valid sampling values ​​based on the reciprocal of the time distance to generate resampling results aligned to the reference time axis. If there are no valid sampling values ​​after the time point due to high altitude freezing or communication link interruption, the calculation is switched to maintaining or extrapolating based on the changing trend of the last valid sampling value and its neighboring historical sampling values. While completing the above resampling and filling, the time of the actual back-transmission value of the sensor corresponding to each dimension is uniformly registered in the sampling metadata table as the last true sampling timestamp of that dimension. For dimensions that are switched to maintaining or extrapolating due to high altitude freezing or communication link interruption, the last true sampling timestamp is still the time of the last actual back-transmission of the sensor. At this time, the deviation between the filled value of that dimension and the actual physical state continues to expand as the lag time increases. A zero-order preservation strategy is adopted for the carbon emission factors in the ledger to copy and fill the current effective factor values ​​into all corresponding time sections in the baseline time series. After splicing, a time-aligned carbon monitoring matrix is ​​obtained.

[0019] Subsequently, a sliding time window covering recent sampling data is set, and the mean and standard deviation of each feature dimension in the time-aligned carbon monitoring matrix are calculated within this window. Z-score standardization is performed on numerical columns such as fuel consumption, dam construction area temperature, and carbon emission factor to eliminate dimensional differences between liters per hour, degrees Celsius, and kilograms of carbon dioxide per liter of diesel. For dimensions such as the carbon emission factor that have constant values ​​and zero standard deviation within the window, a preset smoothing constant is used as the lower limit of the denominator to avoid division by zero. At the same time, the physical quantity values ​​and their unit identifiers of the above dimensions are retained in parallel in the carbon emission feature vector in the form of original value fields. The longitude, latitude and altitude, event timestamp, and equipment identification fields obtained from the heavy truck positioning trajectory coordinate analysis are also retained. The above original value fields and spatiotemporal coordinate fields do not participate in the standardization process. Next, the standardized numerical sequences of each dimension are flattened and concatenated with the original quantity field, spatiotemporal coordinate field, and equipment identification field according to a predefined tensor structure to encapsulate a carbon emission feature vector. This vector possesses both dimensionless feature representation and retains physical quantities directly usable for subsequent spatiotemporal pruning condition calculations and rule threshold comparisons. Finally, the unique equipment code recorded in the equipment identification field is parsed from the carbon emission feature vector. All candidate entity nodes storing equipment code attributes in the carbon regulatory spatiotemporal knowledge graph are traversed. The unique equipment code is compared with the code attribute of each candidate node using hash values. Nodes with completely identical hash values ​​are selected as the hit results, and the unique identifier of this node in the graph is extracted as the focus entity identifier. This focus entity identifier is the starting anchor point for subsequent subgraph traversal and pruning centered on this equipment.

[0020] Specifically, in step two, using the spatiotemporal coordinates in the carbon emission feature vector as pruning conditions, and centering on the focal entity identifier, a constrained breadth-first traversal and dynamic subgraph pruning are performed in the carbon regulatory spatiotemporal knowledge graph to obtain the construction context subgraph. The numerical attributes in the carbon emission feature vector are then serialized into event-triggered payloads. It should be understood that since the carbon regulatory spatiotemporal knowledge graph contains tens of thousands of equipment nodes, process nodes, regional nodes, and environmental status nodes throughout the entire lifecycle of a hydropower project, traversing the entire graph and passing all information into the rule compilation stage would not only introduce a large number of redundant nodes unrelated to the current carbon emission event, increasing the number of rules and matching computation overhead, but also dilute the contextual relevance, leading to a decrease in the accuracy of subsequent rule judgments. Therefore, in the technical solution of this application, the spatiotemporal coordinates in the carbon emission feature vector are used as pruning conditions, and constrained breadth-first traversal and dynamic subgraph pruning are performed in the carbon regulatory spatiotemporal knowledge graph centered on the focal entity identifier. The numerical attributes in the carbon emission feature vector are serialized into event trigger payloads. This allows for the accurate extraction of spatially adjacent and temporally valid local contexts from the global graph that are relevant to the current carbon emission event. Simultaneously, the monitoring values ​​to be determined are independently encapsulated as fact payloads that the rule engine can accept. This limits the scope of subsequent rule compilation and inference to the local subgraphs closely related to the abnormal event, reducing the interference of irrelevant nodes on the rule generation and matching process, while ensuring that event values ​​enter the inference chain in a form independent of the graph structure.

[0021] Specifically, in this embodiment, step two includes: separating and extracting the spatiotemporal and numerical dimensions of the carbon emission feature vector to obtain spatiotemporal pruning conditions and event triggering loads; performing breadth-first traversal and weight pruning on the adjacent nodes under the spatiotemporal pruning condition constraints in the carbon regulatory spatiotemporal knowledge graph, centered on the focal entity identifier, to obtain candidate nodes and edge sets; and performing graph topology mapping and context subgraph instantiation on the candidate nodes and edge sets to obtain the construction context subgraph.

[0022] More specifically, in a concrete example of this application, the carbon emission feature vector is first segmented and extracted according to a predefined dimension index. The dimension segments recording the three-dimensional coordinates of the equipment space and the event timestamp are extracted and encapsulated as spatiotemporal pruning conditions. For example, when a diesel generator triggers a carbon emission event on the right bank of the dam's pouring surface, the spatiotemporal pruning conditions include the longitude, latitude, altitude of the generator, and the precise timestamp of the event. Simultaneously, the carbon emission feature vector also includes records of fuel consumption, heavy truck positioning trajectory coordinates, truck speed obtained by the difference between the trajectory coordinates of two adjacent time sections, and pouring area temperature. The carbon emission factor values ​​are extracted and converted into a carbon emission rate in kilograms per hour by multiplying the fuel consumption value at the current time point by the carbon emission factor value in the ledger. All dimensions, along with this converted value, are serialized into a key-value pair format event trigger payload. The last true sampling timestamp for the truck speed value is taken from the reporting time of the latest positioning trajectory point participating in the differential calculation; the last true sampling timestamp for the fuel consumption-converted carbon emission rate is taken from the last true sampling timestamp of the fuel consumption value dimension. The ledger carbon emission factor, as a static parameter maintained in batches, has physical validity that does not decay over time within the same construction batch. Next, using the node corresponding to the focus entity identifier as the traversal starting point, a breadth-first search queue with a maximum depth of K hops is initialized in the carbon regulatory spatiotemporal knowledge graph. During the queue's layer-by-layer expansion, for each newly arrived adjacent node, the spatial coordinates and time effective interval recorded in the node's attributes are extracted, and the spatiotemporal association weight score between it and the spatiotemporal pruning conditions is calculated. This score is calculated using a negative exponential decay function.

[0023] in, Indicates the k-th adjacent node The spatiotemporal correlation weight score; This represents the spatial distance attenuation coefficient, which controls the penalty intensity of spatial span on the weight. Represents a node The three-dimensional spatial coordinates recorded in the attribute; This represents the coordinates of the carbon emission event center extracted from the spatiotemporal pruning conditions; This represents the square of the Euclidean distance between the node coordinates and the event center coordinates; This represents the time difference decay coefficient, which controls the penalty intensity of the weight based on the time span. Represents a node The status update timestamp is recorded in the attribute; This represents the timestamp of the event extracted from the spatiotemporal pruning conditions; This represents the absolute difference between the node timestamp and the event timestamp. The formula uses a negative exponential function to jointly map spatial distance and time difference into decay weights. Nodes closer to the location of the carbon emission event and whose state update time is closer to the current event time receive higher weight scores. Nodes spatially far from the construction area or no longer in the current construction phase have their weights rapidly decay to near zero. For example, heavy truck nodes and mixing plant nodes located in the same right bank pouring area as the generator and currently in the concrete pouring process will receive higher scores and be retained. Drilling equipment nodes located on the left bank diversion tunnel that have not yet started construction will be eliminated due to their large spatial distance and ineffective time interval. Subsequently, nodes and their associated edges with spatiotemporal correlation weight scores below a preset critical threshold are eliminated, while nodes and their connected edges with scores above the threshold are retained and grouped into a candidate node and edge set.

[0024] Next, the corresponding local adjacency submatrix is ​​extracted from the global adjacency matrix of the carbon regulatory spatiotemporal knowledge graph based on the node indices contained in the candidate node and edge set. Based on this submatrix, an independent graph data structure is instantiated in memory and the equipment type, process stage, regional affiliation and environmental constraint attributes of each retained node are bound to generate a construction context subgraph. This construction context subgraph only contains local topological structures closely related to the current carbon emission event, such as equipment, process and environmental nodes closely related in spatiotemporal space and their interrelationships. Subsequent steps will perform topological parsing on its node attributes and edge relationships and compile them into dynamic rules.

[0025] Specifically, in step three, the domain ontology converter performs topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph to obtain a dynamic rule set for the work conditions. It should be understood that since the construction context subgraph stores the topological relationships and node attributes between equipment, processes, regions, and the environment in the form of a graph data structure, while the rule engine requires input as a conditional judgment script conforming to a specific syntax, there are fundamental differences in their data expression forms and computational paradigms. The semantic relationships and numerical constraints in the graph cannot be directly parsed and executed by the rule engine. If rules are pre-written manually for various work condition combinations, the number of rules increases dramatically with the combination relationships of equipment, processes, and regions, leading to a simultaneous increase in maintenance costs and matching overhead. Therefore, in the technical solution of this application, the domain ontology converter performs topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph to automatically convert the descriptive topological structure of the graph into a conditional logic script that the rule engine can load and execute, achieving isomorphic mapping between graph semantics and rule logic. In this way, the rule engine can obtain dynamic rules that are precisely matched with the current construction conditions during each inference, without the need to pre-set and maintain a large number of static rules, thus ensuring the relevance of the rule context while suppressing the expansion of the number of rules.

[0026] Specifically, in this embodiment, step three includes: performing topological feature parsing and entity constraint extraction on the directed edges and node attributes of the construction context subgraph to obtain a set of context triples and an entity static constraint matrix; performing isomorphic transformation from semantics to logical antecedents on the set of context triples and the entity static constraint matrix based on the domain mapping dictionary to obtain an abstract logic syntax tree; and performing node-by-node code mapping and just-in-time bytecode instantiation on the abstract logic syntax tree to obtain a dynamic rule set for the working conditions.

[0027] More specifically, in a specific example of this application, firstly, all directed edges in the construction context subgraph are traversed, and the source node, target node, and semantic relation labels marked on each edge are extracted. These are then combined in the form of head entity, relation, and tail entity to generate a set of context triples. For example, triples are extracted such as diesel generator node connected to the right bank dam pouring work surface node via the "belongs" relation, heavy truck node connected to the concrete pouring process node via the "serves" relation, and right bank dam pouring work surface node connected to the pouring stage node via the "being" relation. Simultaneously, each independent node in the subgraph is traversed, and the static business constraints recorded in the node attributes are extracted, including the upper limit of the carbon emission rate threshold for the diesel generator in the pouring stage being 45 kg / h, the upper limit of the number of heavy machinery allowed to operate in parallel on the work surface being 6 units, and the upper limit of the allowable temperature range for the pouring area being 28 degrees Celsius. The numerical constraint attributes of each node are vectorized and then concatenated row by row to construct an entity static constraint matrix. Next, a pre-defined domain mapping dictionary is invoked to perform a semantic-to-logical antecedent isomorphic transformation on the context triple set and the entity static constraint matrix. For each triple, the logical operator corresponding to its relation label is found, mapping the "belonging" relation to an existence predicate, the "serving" relation to a variable association binding predicate, and the "being in" relation to a process state matching predicate. For each row of constraint vectors in the entity static constraint matrix, it is mapped to a comparison operation predicate. For example, the upper limit of the carbon emission rate threshold is mapped to a greater than predicate, the upper limit of the number of parallel machines is mapped to a count exceeding predicate, and the upper limit of the allowable temperature range of the pouring area is mapped to a range exceeding predicate. Using relation predicates as internal branch nodes and comparison operation predicates as leaf nodes, an abstract logic syntax tree is hierarchically assembled according to logical conjunction relations. The first-order predicate logic expression expressed by this abstract logic syntax tree is:

[0028] in, This represents the predicate logic expression corresponding to the constructed abstract logic syntax tree; This represents the logical conjunction operation, used to connect independent conditional predicates into a compound antecedent; Represents the set of context triples Middle of the entity ,relation Tail-end entity The triplet formed; This represents the transformation function that maps the topological relation of the triple to the relational predicate of the rule engine; This indicates the number of rows in the entity static constraint matrix, i.e., the number of entities with numerical constraints; Represents the first element in the static constraint matrix of the entity. Row constraint vector; This represents the transformation function that maps the constraint vector to a threshold comparison predicate. The first part of the expression, the conjunction, converts the topological connectivity of the subgraph into entity existence and association conditions in the rule antecedent, enabling the rule to identify whether the current carbon emission event occurs under a specific combination of equipment and work surfaces. The second part, the conjunction, converts the carbon emission threshold and upper limit of the number of machines in the node attributes into numerical comparison conditions in the rule antecedent. These two parts, through logical conjunction, together constitute the complete rule antecedent, thus giving the generated rule both working condition identification and numerical judgment capabilities.

[0029] Then, the domain-specific language generator is activated, and a depth-first traversal algorithm is used to sequentially access each level of the abstract logical syntax tree. For each logical predicate node, code text mapping is performed according to the script syntax template of the target rule engine. Existence judgment operators are converted into pattern matching statement blocks, variable association binding operators are converted into variable binding statements, and comparison operation predicates are converted into conditional constraint expressions. These are then assembled into a complete rule antecedent script text according to the syntax tree hierarchy. Further, the rule engine's built-in just-in-time compiler is invoked to perform lexical analysis, syntax verification, and bytecode compilation on the script text. The text-based rule code is instantiated into a dynamic rule set that can be directly loaded and executed. The judgment conditions of each rule in this rule set are derived from the actual topology and node constraint attributes of the construction context subgraph, corresponding to the equipment combination, process stage, and regional environment of the current carbon emission event. This set is used to load the rules into an isolated rule engine session for anomaly reasoning in subsequent steps.

[0030] Specifically, in step four, the dynamic rule set for the operating conditions is loaded into an isolated rule engine session, and the event trigger payload is injected into the working memory as a fact object. The result is then obtained through deductive reasoning via a pattern matching network to derive anomaly assertions about carbon emissions. It should be understood that since the dynamic rule set for the operating conditions is merely a logical set describing the judgment conditions, it lacks the ability to judge real-time monitoring values. Furthermore, this rule set is generated specifically for the operating condition of the current carbon emission event and is only valid in this inference. Loading it into the globally shared rule engine runtime environment would cause it to mix with rules from other concurrent operating conditions, polluting the working memory state. Simultaneously, if each rule is sequentially compared with all monitoring values, the computational load increases with the product of the number of rules and the number of facts, making it difficult to meet the real-time judgment requirements of the construction site. Therefore, in the technical solution of this application, the dynamic rule set of operating conditions is further loaded into an isolated rule engine session, and the event trigger payload is injected into the working memory as a fact object. Deductive reasoning is then performed through a pattern matching network to obtain carbon emission anomaly assertion results. This allows for the matching and determination of dynamic rules and real-time monitoring values ​​within independent memory spaces that do not interfere with each other, and outputs anomaly conclusions with quantified levels of exceedance. This avoids state pollution between concurrent operating condition inferences, reduces matching computation overhead through condition node sharing and layer-by-layer filtering, and ensures that the anomaly conclusions carry quantified indicators that can be directly used for subsequent rectification decisions.

[0031] Figure 3 This is a flowchart of step four of the intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine, according to an embodiment of this application. Figure 3 As shown, step four includes: S41, starting the network compiler in the isolated sandbox session, performing syntactic dimensionality reduction parsing and conditional topology expansion on the dynamic rule set of the working conditions to obtain the identification matching network structure; S42, deserializing the event triggering payload into a fact object and injecting it into the root node of the identification matching network structure to perform layer-by-layer pattern propagation matching to obtain the conflict resolution agenda queue; S43, based on the dynamic priority arbitration strategy, selecting and executing the activation rules in the conflict resolution agenda queue and calculating the carbon emission exceedance degradation deviation to obtain the carbon emission anomaly assertion result.

[0032] Accordingly, in step S41, a network compiler is started in the isolated sandbox session to perform syntactic dimensionality reduction parsing and conditional topology expansion on the dynamic rule set of the working conditions to obtain the identification matching network structure. It should be understood that since the dynamic rule set of the working conditions is generated in real time by the construction context subgraph corresponding to the current carbon emission event, its rule content varies with different construction conditions. In hydropower projects, multiple work surfaces will generate carbon emission events in parallel during the same period and generate their own rule sets. If each set of rule sets is loaded into the shared rule engine runtime environment, the fact objects and rule activation states in their working memory will be visible to each other, which will cause cross-working condition state pollution and rule misactivation. Moreover, the residual rule objects after the inference is difficult to clear in a targeted manner. At the same time, the antecedents of many rules in the rule set contain a large number of repetitive conditions. If the antecedents of each rule are repeatedly evaluated during inference, the number of matching operations increases with the product of the number of rules and the number of facts. Therefore, in the technical solution of this application, a network compiler is started in an isolated sandbox session to perform syntactic dimensionality reduction parsing and conditional topology expansion on the dynamic rule set of the working conditions to obtain the identification matching network structure. In this way, the rule antecedents are pre-converted into a networked matching structure that can be shared by condition nodes and filtered layer by layer in an independent memory space. In this way, the memory state between concurrent working condition inferences can be isolated, avoiding mutual interference between rules and facts. By reusing nodes with repeated conditions, the number of evaluations of the same condition is compressed from accumulating according to the number of rules to a single evaluation, reducing the computational overhead of subsequent fact matching. At the same time, the memory boundary occupied by this inference is clear, which facilitates the overall release after the inference is completed.

[0033] More specifically, in a concrete example of this application, the rule engine session lifecycle manager first creates an isolated sandbox session, allocating an independent working memory area, an independent class loader, and a preset upper limit for heap memory quota to this session, and assigning a globally unique sandbox session identifier. This ensures that the dynamic rule set generated by the carbon emission event of the right bank dam pouring face resides in a memory space that is invisible to each other, while the rule sets concurrently generated by other work faces during the same time period reside in memory spaces that are not visible to each other. This sandbox session identifier is also used for memory release verification after inference ends. Next, the network compiler is started within the isolated sandbox session and the dynamic rule set of the working conditions is loaded into the compilation input buffer. Syntax dimensionality reduction parsing is performed on the antecedent of each rule, and the composite antecedent is broken down layer by layer into the smallest indivisible decision units using the logical conjunction operator. Decision units involving only a single fact attribute are marked as single-condition types, including equipment type equal to diesel generator, fuel consumption converted carbon emission rate greater than the upper limit of the carbon emission rate threshold during the pouring stage, temperature in the dam pouring area exceeding the allowable range, and heavy truck speed lower than the idle speed discrimination threshold, etc. Decision units involving two or more fact attributes are marked as... The multi-condition association types include the joint determination between the carbon emission rate converted from diesel generator fuel consumption and the temperature of the dam pouring area, and the joint determination between the carbon emission rate converted from diesel generator fuel consumption and the number of parallel-operating machines on the right bank pouring face of the dam to which the generator belongs. The number of parallel-operating machines is obtained by counting the equipment nodes under the node of the working face in the construction context subgraph. The numerical attributes of the context fact object participate in the association determination. The last true sampling timestamp is taken from the time when the construction context subgraph was pruned, so that it participates in the subsequent confidence assessment and soft-gating propagation together with the numerical dimensions of the load.

[0034] Then, all single-condition decision units are normalized and hashed according to the decision attribute name, comparison operator, and comparison threshold. Decision units with consistent encoding are merged into the same single-condition filter node, so that the equipment type decision and work surface attribution decision that occur repeatedly in multiple rules are retained in only one node instance. Each rule that references the condition shares its matching result through the node's outgoing edge. Furthermore, with the root node as the entry point, the merged single-condition filter nodes are connected in series as a filter link in order of passing ratio from low to high, so that fact objects that do not meet the conditions are intercepted at a shallow level. Then, decision units of multi-condition association type are instantiated as multi-condition association nodes, and their input ends are connected to the output ends of the single-condition filter nodes participating in the association decision, and left and right memory areas are configured for them to cache the partially passed matching results. Subsequently, a terminal node is attached to the convergence point of all its predecessor conditions for each rule, and the rule identifier and priority weight metadata are bound to the terminal node. The root node, single-condition filtering node, multi-condition association node, terminal node and the directed edges for data transmission between them together form a directed acyclic graph, which is instantiated as an identification matching network structure in the working memory of the isolated sandbox session. This structure is then used to inject the event trigger payload into the fact object after it is deserialized into a fact object and propagate the matching layer by layer from the root node.

[0035] Accordingly, in step S42, the event trigger payload is deserialized into a fact object and injected into the root node of the authentication matching network structure for layer-by-layer pattern propagation matching to obtain the conflict resolution agenda queue. It should be understood that since the authentication matching network structure is merely a static decision framework composed of condition nodes and directed edges for data transmission, it does not contain any real-time monitoring values. Furthermore, the event trigger payload is serialized data in key-value pair format and lacks the object structure required for the rule engine to perform attribute addressing and condition evaluation. If a data access and propagation relationship is not established between the two, the condition nodes in the network cannot obtain decision input, and the satisfaction of rule antecedents cannot be determined. Simultaneously, a single carbon emission event may simultaneously establish the antecedents of multiple rules. If the established rule instances, along with their bound fact data, are not uniformly collected, subsequent arbitration and selection among multiple candidate rules cannot be implemented. Therefore, in the technical solution of this application, the event-triggered load is further deserialized into a fact object, and the number of parallel operating machines obtained by counting the equipment nodes under the work surface node in the construction context subgraph is instantiated into a context fact object. This fact object is written into the working memory of the isolated sandbox session and registered at the root node. The last real sampling timestamp of this dimension is taken from the time when the construction context subgraph is pruned, which is the root node of the injection identification matching network structure (Rete) to perform layer-by-layer pattern propagation matching to obtain the conflict resolution agenda queue. This allows the real-time monitoring values ​​to enter the network in the form of objects addressable by the rule engine and complete the judgment layer by layer along the condition nodes. The rule instances with fully valid antecedents are collected into the arbitration queue. In this way, fact objects that do not meet the conditions can be intercepted at shallow nodes and no longer participate in subsequent judgments, reducing the number of invalid condition evaluations. At the same time, it provides a structurally unified and directly selectable set of activation rules for subsequent priority arbitration and deviation calculation.

[0036] More specifically, in a specific example of this application, each key-value pair in the event trigger payload is first read, and deserialization is performed on it according to the predefined fact object template. The fuel consumption value, truck speed value, pouring area temperature value, and carbon emission factor value are bound to the corresponding numerical fields of the fact object, respectively. The equipment identifier and equipment type are bound to the metadata fields of the fact object, and data type validation and value range validation are performed on each field. After the validation is passed, the fact object is written to the working memory of the isolated sandbox session and registration is completed at the root node. Next, the root node assigns the fact object to the corresponding single-condition filter node link based on the type label of the fact object. Each single-condition filter node addresses the single numerical field it is judging from the fact object, evaluates the value of the field with the node's preset comparison operator and comparison threshold, and outputs the judgment result using a hard binarized indicator function. When the field value meets the node condition, it outputs an activation signal 1 and continues to pass the fact object to the lower layer. When it does not meet the condition, it outputs 0 and terminates the subsequent propagation on the link. For example, the single-condition filter node that judges whether the carbon emission rate converted from the fuel consumption of the diesel generator exceeds the upper limit of the carbon emission rate threshold during the pouring stage outputs 1 when the rate exceeds the limit, while the single-condition filter node that judges whether the temperature in the dam pouring area exceeds the allowable range outputs 0 when the temperature does not exceed the limit. Then, the fact objects of the single-condition filtering nodes are written into the left and right memory areas of the downstream multi-condition association nodes. The multi-condition association nodes perform Cartesian joins on the partial matching results in the left and right memory areas and evaluate them according to the cross-variable constraints. Similarly, they output 0 or 1 with hard-binarized indicator functions. For example, when the joint judgment condition between the fuel consumption-converted carbon emission rate and the number of parallel-operating machines on the right bank of the dam is met, the output is 1. Finally, the outputs of all the single-condition filtering nodes and multi-condition association nodes associated with this rule are multiplied together to obtain the activation judgment result of the rule.

[0037] in, Indicates the first Rules The activation determination result is 1, which means that the antecedent of the rule is fully met and activated, and 0 means that the rule is not activated. Indicates the first The total number of single-condition filtering (Alpha) nodes associated with each rule; Indicates the first The total number of multi-condition association (Beta) nodes associated with each rule; This indicates a Boolean logic indicator function that returns 1 if the condition within the parentheses is true and 0 if it is false. This represents the monitoring numerical data carried by the fact object obtained by deserializing the event-triggered payload; This indicates that the logical relationship is satisfied, meaning that the value of the fact object satisfies the predefined judgment condition of the node on the right. Indicates the first The first rule associated with A single-condition filtering node; Indicates the first The first rule associated with This formula involves multiple condition-related nodes. The indicator function in the formula compresses the judgment result of each node into two discrete values, 0 or 1, so that the judgments such as whether the fuel consumption-converted carbon emission rate exceeds the limit and whether the temperature of the pouring area exceeds the limit are uniformly expressed as binary signals. The first multiplication term gathers all the single-variable judgment results involved in the rule, corresponding to the item-by-item verification of independent attributes such as equipment type, fuel consumption-converted carbon emission rate and pouring area temperature. The second multiplication term gathers the cross-variable related judgment results, corresponding to the joint verification between fuel consumption-converted carbon emission rate and the number of parallel operating machines on the work surface. The multiplication of the two parts means that if the output of any node is 0, the overall product will be zero. The antecedent of the rule is judged to be invalid and will not enter the subsequent process. Only when all node outputs are 1 will the product value be 1 and the rule will be judged to be activated. This constitutes a judgment paradigm in which all conditions are passed and the rule is activated. The contribution weight of each node to the rule activation result is 1, which is independent of the sampling source of the numerical dimension judged by the node. Subsequently, rules with an activation determination result of 1 are encapsulated as activation items at their terminal nodes. These activation items include rule identifiers, references to the bound fact objects, and pre-set priority weight metadata on the terminal nodes. All activation items generated during this propagation are collected and written into the cache area to obtain a conflict resolution agenda queue, which is then used to select and execute and calculate the carbon emission exceedance and degradation deviation degree based on the dynamic priority arbitration strategy.

[0038] Specifically, in the intelligent carbon emission monitoring system for hydropower projects, the core reasoning process of the rule engine relies on the hierarchical pattern propagation and matching determination of the injected fact payload by the Rete identification matching network structure. In the first embodiment, after the event trigger payload is uniformly deserialized into a standard business fact object, it is directly injected into the root node of the Rete network. Subsequently, in the single-condition filtering node and the multi-condition association node, a hard-binarized indicator function is used to perform a Boolean decision of all pass or all fail. As long as a certain numerical dimension of the fact object meets the node's preset matching condition, the node outputs an activation signal of 1; otherwise, it outputs 0. Finally, the outputs of each node are multiplied together to determine whether the rule is activated as a whole.

[0039] However, the event-triggered load is not a set of homogeneous and equally weighted values. This load originates from the upstream steps of timestamp interpolation and alignment of multi-source heterogeneous sensor data. During this alignment process, some dimensions represent the sensor's actual high-frequency sampling points (corresponding to high data confidence), while others represent linear interpolation fillers from times farther from the actual sampling (corresponding to low data confidence). Especially in high-altitude hydropower stations (such as the frigid plateau region where the Yaxia Hydropower Project is located), temperature sensors are highly susceptible to freezing or communication link interruptions, leading to a sudden drop in sampling frequency or even prolonged periods without updates. In such cases, the system can only rely on outdated historical data to calculate the current temperature value using interpolation algorithms. The deviation between this calculated value and the actual physical state continues to widen over time.

[0040] The Rete network propagation mechanism in the first embodiment is completely unaware of the aforementioned dimensional data quality differences. A dimension based on outdated interpolated temperature data that is tens or even minutes behind and only slightly exceeds the threshold, and a dimension based on real-time high-frequency fuel consumption sampling from a diesel generator that significantly exceeds the limit, are assigned the exact same activation contribution weight of 1 in the matching determination of single-condition filtering nodes and multi-condition association nodes. This leads to a situation where, in a typical scenario where multiple heavy machines operate in parallel during the peak of dam construction, the cross-variable association determination of multi-condition association nodes involving outdated temperature interpolated data can easily erroneously activate a rule in form, triggering entry into the conflict resolution agenda queue and ultimately generating unnecessary rectification control instructions. This systematic false rule activation not only consumes limited computational arbitration resources but may also lead to frequent accidental shutdowns of equipment at the construction site, affecting project progress and management efficiency.

[0041] To address the aforementioned shortcomings, a second embodiment is proposed. Specifically, in another specific example of this application, Figure 4 The flowchart illustrates the process of a method for intelligent monitoring of carbon emissions from hydropower projects using a fusion rule engine, based on embodiments of this application. This method involves deserializing event-triggered payloads into fact objects, injecting them into the root node of an identification and matching network structure, and performing layer-by-layer pattern propagation matching to obtain a conflict resolution agenda queue. Figure 4 As shown, step S42 includes: S421, performing time-series freshness decay assessment and confidence score dimension-by-dimensional labeling on each numerical dimension of the event-triggered payload to obtain a set of confidence-labeled fact objects; S422, injecting the set of confidence-labeled fact objects into the root node of the discrimination matching network structure for soft-gated mode propagation to obtain a confidence-weighted activation score vector; S423, based on a preset minimum confidence activation threshold, performing low-confidence rule filtering and discarding and high-confidence rule priority descending order encapsulation on the comprehensive activation scores of each rule in the confidence-weighted activation score vector to obtain a conflict resolution agenda queue.

[0042] In step S421, time-series freshness decay assessment and confidence score dimension-by-dimensional labeling are performed on each numerical dimension of the event-triggered payload to obtain a set of confidence-labeled fact objects. It should be understood that in the real-time inference scenario of carbon emission monitoring in hydropower projects, the reliability of each dimension of the payload data directly determines the accuracy of downstream rule activation judgments. Different sensors exhibit drastically different data update frequencies and latency characteristics due to differences in deployment environments (such as high-altitude freezing risks and communication obstruction). Therefore, in order to establish data freshness profiles for each dimension before the payload enters the identification and matching network structure, providing a quantitative basis for subsequent soft-gating propagation, the technical solution of this application further performs time-series freshness decay assessment and confidence score dimension-by-dimensional labeling on each numerical dimension of the event-triggered payload. This allows the reliability differences of each dimension of the payload data to be explicitly quantified and enter the network along with the numerical attributes. Low-quality dimensions carry self-inhibiting signal markers upon entering the network, thereby providing a data foundation for suppressing spurious rule activation driven by outdated interpolated data.

[0043] In a specific example of this application, firstly, each numerical dimension in the event-triggered load and the numerical dimension of the context fact object obtained by counting the construction context subgraph are traversed. The last true sampling timestamp of this dimension is queried from the sampling metadata table of the sensor data access layer. Here, the last true sampling timestamp only records the moment when the sensor actually transmits the data, and does not include the virtual time point filled by the interpolation algorithm during the upstream timestamp alignment process, so that the timestamp can truly reflect the physical acquisition time of the data in this dimension. Next, the current system clock timestamp is read, and the time difference between it and the last true sampling timestamp of this dimension is calculated to obtain the lag time experienced by the data in this dimension since the last true acquisition. Then, a time-series decay rate constant is configured for each dimension. This constant is calibrated according to the nominal data update frequency of the specific sensor type. Sensors with a high nominal update frequency are configured with a larger decay rate constant, and sensors with a low nominal update frequency are configured with a smaller decay rate constant, so that the same lag time produces a score decay amplitude that is adapted to its normal sampling rhythm on different types of sensors. Furthermore, the above time difference is mapped to a dimension confidence score between 0 and 1 using a negative exponential decay function.

[0044] in, Indicates the first The time-series decay confidence score is calculated for each dimension, with a value range of (0,1]. The time-series decay rate constant of the sensor corresponding to the d-th dimension is calibrated by the nominal data update frequency of the specific sensor type. The higher the nominal update frequency, the larger the value. It is worth mentioning that for static parameter dimensions such as carbon emission factors in the ledger that are maintained in batches and context value dimensions obtained by trimming the construction context subgraph, the time-series decay rate constant is calibrated to a value close to zero, so that its confidence score remains close to 1 within the current inference time scale, thus not exerting additional suppression on the joint confidence coefficient. Indicates the current system clock timestamp; Indicates the first Each dimension corresponds to the last true sampling timestamp of the sensor. It should be noted that when the last true sampling time of a certain dimension completely coincides with the current time, the confidence score reaches its maximum value of 1, indicating that the data for that dimension is the freshest data just collected. As the time difference continues to increase, the score approaches 0 exponentially, reflecting the nonlinear weakening effect of the sensor sampling interval on the physical reliability of the data in harsh high-altitude environments. That is, the interpolated data farther from the true sampling point exhibits an exponential decline in its credible contribution to the assessment of carbon emission anomalies. The time difference term in this formula... In a physical sense, it measures the degree of disconnect between the numerical value of this dimension and the actual physical state. The temperature of the pouring trough surface continues to change under the influence of hydration heat and ambient temperature during the sensor freezing period. The longer the lag time, the greater the cumulative deviation between the interpolated value and the actual temperature of the trough surface; decay rate constant This determines the rate at which the degree of disconnection translates into credibility loss. For a diesel generator fuel consumption flow meter that is nominally transmitted once per second, a lag of several seconds indicates a significant anomaly, and its decay rate constant takes a larger value. For a pouring area temperature sensor that is nominally transmitted once every 30 seconds, the same lag time is within the normal sampling rhythm, and its decay rate constant takes a smaller value. The negative exponential function constrains the above relationship to a monotonically decreasing continuous mapping between 0 and 1 without touching 0. This ensures that fresh data receives a full credibility contribution close to 1, while ensuring that the credibility contribution of old data is compressed to close to 0 but not directly judged as invalid. This allows this dimension to participate in subsequent judgments with a very low weight without interrupting the inference link.

[0045] Furthermore, the calculated confidence scores are bound dimension-by-dimensionally to the corresponding numerical attributes, generating a set of confidence-labeled fact objects where each attribute carries a time-series freshness label. In the high-altitude hydropower station scenario, temperature data that has not been updated for a long time due to freezing or communication interruption will automatically receive an extremely low confidence score, such as 0.12, while the real-time fuel consumption dimension of the diesel generator, which is transmitted normally in real time, receives a high confidence score, such as 0.97. The carbon emission rate calculated from fuel consumption using this dimension uses the same confidence score, establishing a quantifiable credibility classification profile for subsequent network propagation from the source. This ensures that low-quality dimensions carry a self-suppressing signal marker when entering the identification and matching network structure. The binding relationship between the above scores and numerical attributes is stored within the fact objects as an attribute-accompanying field, allowing the confidence score to flow synchronously with the propagation path of the numerical attribute in the network. It can be directly read by condition nodes at all levels, providing a basis for subsequent processing steps.

[0046] In step S422, the set of confidence-labeled fact objects is injected into the root node of the discriminative matching network structure for soft-gated mode propagation to obtain a confidence-weighted activation score vector. It should be understood that after the aforementioned dimension-level temporal decay confidence assessment and dimensional labeling, although the payload data already possesses dimensional quality information, the confidence score at this point only exists as an accompanying label within the set of confidence-labeled fact objects. If the original hard binarization propagation logic is still used, each node will only read the numerical attribute itself and not its associated confidence score during evaluation. Therefore, the labeling information will not play a practical role in the matching process, and the outdated interpolation dimension and the real-time sampling dimension will still output the same activation signal 1 at the single-condition filtering node and the multi-condition association node. Therefore, in the technical solution of this application, the set of confidence-labeled fact objects is further injected into the root node of the discriminative matching network structure for soft-gated mode propagation to obtain a confidence-weighted activation score vector. This structurally modifies the step-by-step propagation mechanism of the discriminative matching network structure, enabling the confidence score to be deeply embedded in the matching output of each node. In this way, the activation conclusion of each rule can simultaneously reflect the validity of its judgment conditions and the credibility of the data it relies on, enabling the network to have data quality awareness capabilities without modifying the logic of the rules themselves.

[0047] In a specific example of this application, the set of confidence-labeled fact objects is first written into the working memory of the isolated sandbox session and injected into the root node of the self-identifying matching network structure. Each numerical attribute in this set of fact objects carries the aforementioned time-decayed confidence score bound dimension by dimension. This score flows synchronously along the network propagation path with the corresponding attribute and can be directly addressed and read by nodes at each level along the way. Next, at each single-condition filtering node, the original hard binary propagation is transformed into soft-gated propagation. Instead of simply outputting a Boolean pass signal, the binary result of attribute matching is multiplied by the dimension confidence score bound to the attribute to output a weighted matching score. For example, the single-condition filtering node that determines whether the carbon emission rate converted from diesel generator fuel consumption exceeds the upper limit of the carbon emission rate threshold during the pouring stage outputs 0.97 instead of 1 when the condition is met; for another rule with temperature exceeding the limit as a single variable condition, the single-condition filtering node that determines whether the temperature in the dam pouring area exceeds the allowable range outputs 0.12 instead of 1 when the condition is met. Then, at the multi-condition association node, since it involves cross-variable association judgments across multiple dimensions, the confidence scores of each dimension involved in the judgment are read from its left and right memory areas. The geometric mean of the confidence scores of all dimensions involved in the judgment is taken as the joint confidence coefficient. For example, in the cross-variable association judgment involving both the fuel consumption-to-carbon emission rate of diesel generators and the temperature of the dam construction area, the confidence scores of the two dimensions involved in the judgment are 0.97 and 0.12, respectively, and the joint confidence coefficient is taken as the geometric mean of the two, which is approximately 0.34. Furthermore, when the fact object propagates to the terminal node, the weighted scores output by all single-condition filtering nodes and multi-condition association nodes are multiplied and aggregated to generate a comprehensive confidence activation score for each rule. For a rule whose activation path consists of the aforementioned fuel consumption single-condition filtering node and the multi-condition association node, and whose temperature dimension only participates in the joint judgment through the multi-condition association node, its comprehensive confidence activation score is the product of 0.97 and 0.34, which is approximately 0.33. Then, the comprehensive confidence activation scores of each rule in this propagation are organized sequentially by rule identifier index into a confidence-weighted activation score vector, wherein the comprehensive confidence activation score is calculated as follows:

[0048] in, Indicates the first The combined activation score of the rules after confidence-based soft-gating modulation; Indicates the first The total number of single-condition filter nodes associated with each rule; Indicates the first The total number of Beta nodes associated with each rule; This indicates a Boolean logic indicator function that returns 1 if the condition is true, otherwise returns 0. The injected fact data, in this embodiment, is the monitoring numerical data carried by the set of confidence-labeled fact objects. Its numerical content is consistent with the fact objects obtained by deserialization of event-triggered payloads in the first embodiment. The difference is that each numerical attribute is attached with a time-decayed confidence score bound dimension by dimension. Symbols representing logical satisfaction relationships; Indicates the first The first rule associated with A single-condition filtering node; Indicates the first The first rule associated with Multiple condition-related nodes; Indicates the first The time-decay confidence score corresponding to the data dimension determined by each single-condition filtering node; Indicates the first The set of all data dimension indexes involved when performing cross-variable association determination for a Beta node; Represents a set The number of dimensions in; The geometric mean of the confidence scores of the associated dimensions of the multi-condition associated nodes is used as the joint confidence coefficient. The indicator function term in this formula fully retains the judgment logic of the rule's antecedents. Its value depends only on whether attributes such as fuel consumption-converted carbon emission rate and pouring zone temperature meet the node's preset comparison conditions. Therefore, the logical content of the rule is not modified. The time-decayed confidence score term multiplied by it acts on the binary result in the form of multiplicative gain. In a physical sense, it represents the degree to which the current value of this dimension can represent the real physical state. This causes the same condition to produce different levels of activation contribution due to the different freshness of the data source. The fuel consumption-converted carbon emission rate dimension, which is transmitted in real time and frequently, contributes almost the full amount, while the temperature dimension, which has not been updated for a long time due to freezing or communication interruption, contributes only a very small share. At multi-condition association nodes, the geometric mean, rather than the arithmetic mean, is used as the joint confidence coefficient. The physical basis for this is that the geometric mean is extremely sensitive to the lowest confidence level among the participating dimensions; if the confidence level of any dimension approaches 0, the joint confidence coefficient will simultaneously approach 0. In contrast, the arithmetic mean allows the high-confidence dimension of fuel consumption converted to carbon emission rate to mask the outdated temperature dimension. This aligns with the physical essence of cross-variable association decisions, where the conclusion of a multi-condition association node requires all participating dimensions to accurately reflect the current physical state, and its reliability is limited by the least reliable data source. The full-path multiplication aggregation term causes the confidence loss to accumulate progressively along the entire activation path of the rule. The more low-quality dimensions involved in the activation path, and the more outdated they are, the greater the reduction in the overall confidence activation score. Simultaneously, this multiplication structure retains the interception characteristics of the original hard decision; a zero indicator function at any node still results in a zero overall product. Rules whose conditions themselves are invalid will not be activated simply because the data they rely on has high confidence.

[0049] It should be noted that the overall activation score is no longer a discrete state of either 0 or 1, but a gradient value that continuously varies between 0 and 1. When the activation path of a carbon emission anomaly rule includes outdated temperature data with a confidence level of only 0.12 as one of the joint judgment conditions of multi-condition association nodes, the overall activation score of the rule will be significantly reduced, for example, from 1.0 under the hard binarization propagation of the first embodiment to the order of 0.34. In the above example, it is the product of 0.97 and 0.34, which is approximately 0.33. This achieves soft suppression of low-quality data participating in rule activation at the data structure level of network propagation. The above modification transforms the original hard decision paradigm of "all pass equals activation" into a flexible decision paradigm of confidence-weighted gradient activation. Without modifying the logic of the rule itself, it enables the network to have data quality awareness capabilities, avoiding outdated data generated by high-altitude frozen sensors from obtaining the same activation contribution weight as high-quality real-time data in the discrimination and matching network structure. In the scenario of multiple heavy machines operating in parallel during the peak period of dam construction, the rules that use outdated interpolated temperature data to determine cross-variable associations have their scores suppressed to a low level, while the rules that are driven by real-time fuel consumption sampling and significantly exceed the standard have their scores maintained at a high level. This creates a comparable gradient difference between the scores, providing quantitative input for subsequent filtering and priority ranking based on the lowest credible activation threshold.

[0050] In step S423, based on a preset minimum confidence activation threshold, the comprehensive activation scores of each rule in the confidence-weighted activation score vector are filtered and discarded for low-confidence rules, and high-confidence rules are sorted and encapsulated in descending order of priority to obtain a conflict resolution agenda queue. It should be understood that after propagation through the soft-gating mode, each rule in the confidence-weighted activation score vector has obtained a comprehensive activation score reflecting the overall data confidence level of its activation path. However, this score is only attached to the activation item as a numerical label and does not participate in the admission determination. If no admission threshold is set, rules with low confidence scores will still enter the subsequent arbitration process along with rules with high confidence scores, consuming unnecessary computational resources and increasing the risk of misjudgment. Furthermore, the dynamic priority arbitration strategy cannot distinguish the reliability of the data on which each rule depends when selecting execution rules. Rules that participate in activation using outdated interpolation dimensions may still be selected for execution based on their preset priority weights. Therefore, in the technical solution of this application, based on a preset minimum trusted activation threshold, the comprehensive activation scores of each rule in the trusted weighted activation score vector are further processed by filtering and discarding low-confidence rules and prioritizing and encapsulating high-confidence rules in descending order to obtain a conflict resolution agenda queue. This adds an access barrier based on data quality before rules enter the agenda queue. This prevents low-confidence activation rules from entering the subsequent conflict arbitration and rectification instruction generation stages, avoiding their consumption of limited arbitration computing resources, while making data trustworthiness an explicit ranking basis for the rule execution order.

[0051] In a specific example of this application, a minimum trusted activation threshold, such as 0.6, is first set. This threshold is calibrated based on the nominal data update frequency of various sensors in the field and the historical false alarm statistics. After calibration, it is written into the inference configuration item of the isolated sandbox session. Next, the comprehensive activation score of each rule in the confidence-weighted activation score vector is traversed, and each rule is compared with the minimum trusted activation threshold. Rules with a comprehensive activation score lower than the threshold are judged as low-confidence activations and directly filtered and discarded. At the same time, the reference of the activation item to the fact object is removed, and the space occupied by the activation item in the working memory is reclaimed. This is equivalent to establishing a data quality firewall at the rule activation determination level. Then, for rule instances with scores higher than the threshold, the comprehensive trusted activation score is written as an additional priority weight factor into the rule metadata. This, together with the priority weights pre-set on the terminal node based on the proportion of equipment carbon emission contribution and the sensitivity of process control, constitutes the arbitration basis for the rule. For example, the two rules with overall activation scores of approximately 0.33 and 0.12 are both below the threshold and are therefore filtered out. However, the rules whose activation paths consist only of a single-condition filtering node for fuel consumption-converted carbon emission rate and a multi-condition associated node consisting of the number of parallel-operating machines at that rate, have weighted scores of 0.97 and approximately 0.98, respectively, and an overall activation score of approximately 0.96, which is above the threshold and is therefore retained.

[0052] Furthermore, each rule instance that has passed the filtering process is sorted in descending order using the additional priority weight factor as the sorting key. After being sorted in descending order of score, the rule instances are encapsulated into the conflict resolution agenda queue. The first instance of the queue is the rule instance with the highest comprehensive confidence activation score. This queue is then output to the selection and execution stage based on the dynamic priority arbitration strategy to calculate the carbon emission exceedance and degradation deviation and generate carbon emission anomaly assertion results.

[0053] In high-carbon emission scenarios where multiple diesel generators operate in parallel during dam construction, even if a rule passes the matching conditions of all nodes in the identification and matching network structure due to a slight numerical deviation from the limit, its overall score will be suppressed below the threshold and intercepted if its activation path involves outdated or interpolated low-quality dimensional data. This suppresses the systemic risk of false activation rules driven by low-quality data entering the subsequent conflict arbitration stage and causing misjudgment and rectification orders. At the same time, genuine carbon emission anomaly rules activated based on high-quality real-time data naturally obtain higher arbitration execution priority because their overall scores far exceed the threshold, ensuring that truly reliable carbon emission anomaly signals are responded to and handled first. Thus, threshold comparison physically constitutes an access gate based on data freshness, while descending order determines the traversal order of subsequent dynamic priority arbitration and allows data credibility to directly enter the arbitration weighting through an additional priority weighting factor. This ensures that genuine over-emission events driven by real-time fuel consumption sampling on the pouring surface are handled before suspicious events linked to freezing temperature sensor interpolation data.

[0054] In summary, the second embodiment described above achieves the following technical effects: In the harsh environment of high-altitude hydropower projects where sensor data updates are delayed due to freezing or communication interruptions, the system can automatically identify and quantify the temporal freshness differences of data in various dimensions. This upgrades the rule activation determination for identifying and matching network structures from the original hard binary equal-weighted voting mode to a gradient-based selection mode that considers data quality weights. This effectively reduces the false rule activation rate caused by outdated interpolated data, reduces unnecessary conflict arbitration computation overhead and the frequency of misjudgment rectification instructions, and improves the inference accuracy and computational resource utilization efficiency of the rule engine in complex multi-condition concurrent scenarios. Simultaneously, high-confidence rules are given a natural priority response position in the agenda queue, which helps to improve the ability of real-time reliable data-driven carbon emission anomaly signals to penetrate conflict arbitration and trigger closed-loop rectification at the first moment. This strengthens the robust judgment capability and rapid response time of the intelligent carbon emission monitoring system for hydropower projects under extreme conditions.

[0055] Accordingly, in step S43, based on the dynamic priority arbitration strategy, the activated rules in the conflict resolution agenda queue are selected and executed, and the carbon emission exceedance and degradation deviation is calculated to obtain the carbon emission anomaly assertion result. It should be understood that since multiple activated rules may be encapsulated in the conflict resolution agenda queue at the same time, when multiple diesel generators and heavy trucks are operating in parallel on the right bank of the dam, the carbon emission rate exceedance rule, the number of machines on the working face exceedance rule, and the temperature anomaly rule in the pouring area may be established at the same time, and give overlapping or even opposite disposal requirements for the same equipment. If all activated rules are executed one by one according to the order of writing in the queue, it will result in repeated issuance or mutual cancellation of rectification actions and occupy limited arbitration calculation resources. At the same time, the activation of a rule only indicates that its antecedent is established, which is a qualitative conclusion on whether there is an anomaly, without quantitative information on the extent of exceedance. Subsequent retrieval of rectification measures library and calculation of equipment parameter adjustment amount lack directly usable quantitative input. Therefore, the technical solution of this application further utilizes a dynamic priority arbitration strategy to select and execute activation rules in the conflict resolution agenda queue and calculate the carbon emission exceedance and degradation deviation. This determines a unique execution target among multiple candidate activation rules and transforms qualitative rule activation conclusions into structured anomaly conclusions with quantitative exceedance levels. This avoids conflicts in rectification instructions and waste of arbitration resources caused by concurrent execution of multiple rules, ensures that the output anomaly conclusions directly carry the quantitative indicators required for subsequent closed-loop rectification, and prioritizes disposal resources for anomalies that contribute significantly to emissions, are highly sensitive to control, and have reliable data.

[0056] More specifically, in a specific example of this application, the rule metadata of each active item in the conflict resolution agenda queue is first read, and the priority weights preset by the terminal node are extracted. The preset priority weights are calibrated based on the carbon emission contribution ratio of the equipment pointed to by the rule and the control sensitivity of the process in which it is located. For active items that have been written with additional priority weight factors after being filtered by the confidence threshold according to the second embodiment, the additional priority weight factors are also extracted. The factor is the comprehensive confidence activation score of the rule. For active items obtained by hard binarization propagation according to the first embodiment, the additional priority weight factor takes the default value of 1. Next, the dynamic arbitration score of each activation rule is obtained by multiplying the two weights according to the dynamic priority arbitration strategy. Then, the activation rule with the highest dynamic arbitration score is selected from the head of the queue in descending order of the established conflict resolution agenda queue and enters the execution stack. If there are more than two activation rules with the same dynamic arbitration score, the one with the higher comprehensive confidence activation score is selected according to the queue order. If the comprehensive confidence activation scores are also equal, the rule with the latest last real sampling timestamp in the fact object it is bound to is selected to ensure that the monitoring value on which the selected rule is based is closest to the current physical state.

[0057] Then, the consequent action of the selected activation rule is triggered. The actual carbon emission monitoring value that triggered this activation is retrieved via the fact object reference stored in the activation item, such as the fuel consumption-converted carbon emission rate of the diesel generator at the current time section. The current operating condition carbon emission dynamic threshold, determined by the construction context subgraph node attributes, is read from the antecedent of the rule, such as the upper limit of the carbon emission rate threshold for this type of generator during the pouring stage. Furthermore, the carbon emission exceedance and degradation deviation is calculated using the following formula:

[0058] in, This indicates the calculated deviation of carbon emission exceedance from the standard. This indicates the actual monitored carbon emissions that triggered the activation of this rule; This represents the current operating condition's dynamic carbon emission threshold, determined by the compilation of the construction context subgraph. This represents the absolute difference between the actual monitored value and the dynamic threshold. This indicates the proportion of the absolute difference exceeding the dynamic threshold; Represents a non-negative truncation function; This represents the pre-set severity penalty weighting coefficient in the activated rule. The relative excess ratio term in this formula converts the absolute emission difference into a dimensionless ratio, which physically represents the extent to which the current emission level exceeds the allowable emission level under this operating condition. This makes the degree of exceedance comparable between large diesel generators with rated power differences of several times and light transport vehicles, preventing small fluctuations of high-power units from being judged as more urgent than serious emission overruns of low-power equipment due to larger absolute differences. Since the denominator is taken from a dynamic threshold rather than a fixed limit, the same equipment will have different deviations during the diversion and pouring phases due to different thresholds, allowing the quantitative results to be adaptively adjusted according to the construction conditions. The non-negative cutoff term defines the boundary between abnormal and normal. When the actual monitored value does not exceed the dynamic threshold, the deviation is taken as 0, so that normal emission fluctuations on the pouring surface caused by intermittent mechanical start-stop are not included in the degree of abnormality, and only when the limit is actually exceeded will a non-zero abnormality measure be formed. The severity penalty weighting coefficient amplifies the deviation according to equipment and process. Units in key processes such as dam pouring or those that make the main contribution to emissions have a larger coefficient, so that key equipment will have a higher abnormality level under the same excess ratio, thereby obtaining a greater pressure reduction in subsequent parameter adjustment calculations.

[0059] Next, the calculated carbon emission exceedance and degradation deviation is assembled with the anomaly cause code, anomaly occurrence timestamp, equipment identifier, and equipment spatial coordinates mapped to the executed activation rule identifier, and a carbon emission anomaly assertion result is generated and written to the inference result buffer of the isolated sandbox session. The anomaly cause code is used to locate the corresponding control strategy template in the remediation measure library, and the carbon emission exceedance and degradation deviation is used to calculate parameter adjustment amounts in conjunction with the equipment's carbon emission reduction response weights. Afterward, the executed activation rule is removed from the conflict resolution agenda queue and its activation item is marked as consumed to prevent the same fact object from triggering the rule repeatedly in this inference. The remaining unselected activation rules in the queue are released when this inference ends, thus completing this deductive reasoning and outputting the carbon emission anomaly assertion result to the closed-loop remediation instruction generation stage.

[0060] Specifically, in step five, the rectification measure library is retrieved based on the anomaly cause code in the carbon emission anomaly assertion result to generate a closed-loop rectification instruction. After the inference is completed, the dynamic rule set for operating conditions is destroyed and the rule engine session memory is released. It should be understood that since the carbon emission anomaly assertion result is only structured data describing the abnormal facts, it does not have the ability to control the equipment at the construction site. If it is not converted into control messages that can be recognized and executed by the front-end IoT devices, the supervision process will stop at alarm display and cannot form a closed-loop handling. At the same time, the dynamic rule set for operating conditions is generated in real time by the construction context subgraph corresponding to the current carbon emission event and is only valid in this inference. If it remains in the isolated sandbox session after the inference is completed, as the carbon emission events at the construction site continue to be triggered, the rule sets generated in previous times and their identification and matching network structures will continue to accumulate in memory, the heap memory occupied will continue to grow and eventually exhaust the session quota, and the remaining expired rules may be mistakenly loaded by subsequent inference and make judgments based on the expired operating condition thresholds. Therefore, in the technical solution of this application, a closed-loop rectification instruction is generated by retrieving the rectification measure library based on the anomaly cause code in the carbon emission anomaly assertion result. After the inference is completed, the dynamic rule set of the operating condition is destroyed and the rule engine session memory is released. This converts the quantified anomaly conclusion into a device control message that can be directly issued and executed. After a single inference closure, the rule objects and memory occupied by this inference are completely cleared. In this way, a complete closed loop is formed from the identification to the handling of carbon emission anomalies, while ensuring that memory occupation does not accumulate and increase with the number of event triggers, thus guaranteeing the long-term stability of session memory and the timeliness and accuracy of inference results in high-frequency regulatory scenarios.

[0061] Specifically, in this embodiment, step five includes: retrieving the rectification measures library based on the anomaly cause code and deviation degree of exceeding the standard in the carbon emission anomaly assertion result; calculating the parameter adjustment amount and assembling the communication protocol message in combination with the equipment carbon emission reduction response weight to obtain a closed-loop rectification instruction; performing forced unregistration and heap memory release on the dynamic rule set of the operating condition through the rule engine session lifecycle manager, and verifying the memory clearing status through the garbage collection status monitoring interface to obtain the memory destruction identifier; and concatenating and hashing the sandbox session identifier, destruction timestamp, and memory destruction identifier to obtain the rule engine memory release confirmation receipt.

[0062] More specifically, in a concrete example of this application, firstly, the anomaly cause code and the deviation degree of carbon emission exceedance and degradation are parsed from the carbon emission anomaly assertion results. Using the anomaly cause code as the search key, key-value matching is performed in the rectification measure library to retrieve the control strategy template mapped to that cause code. For example, a cause code for exceeding the carbon emission rate limit is mapped to a strategy template for reducing the operating power of the diesel generator, and a cause code for exceeding the limit for the number of machines on the work surface is mapped to a strategy template for evacuating idle machines on the work surface. Next, based on the equipment identifier in the carbon emission anomaly assertion results, the carbon emission reduction response weight of the equipment is retrieved. This weight is pre-calibrated based on the equipment's rated power, carbon emission intensity per unit power, and its adjustable margin in the current process. A larger calibration value indicates that the equipment bears more pressure reduction responsibility. Then, combining the deviation degree of exceedance and degradation and the equipment's carbon emission reduction response weight, the target adjustment amount of the equipment's operating parameters is calculated using the following formula:

[0063] in, This indicates the calculated operating parameter settings that the target device should execute; This indicates the current actual operating parameter values ​​of the target device when this anomaly was triggered; This represents the preset global penalty adjustment coefficient, used to control the aggressiveness of emission reduction adjustments; This represents the deviation from carbon emission exceedance and degradation extracted from the carbon emission anomaly assertion results. This represents the carbon emission reduction response weighting coefficient of the target equipment; This represents the pressure drop calculated based on the degree of exceedance and the equipment's emission reduction responsibility. This formula uses the equipment's current actual operating parameters as a benchmark for reduction correction. Physically, it ensures that the control command is incrementally adjusted from the equipment's current operating condition rather than abruptly switching to a fixed level. This avoids power outages or instability in concrete delivery rhythm caused by sudden parameter drops in the diesel generator continuously supplying power to the pouring area. The deviation from the exceedance standard in the pressure drop is positively correlated with the degree of exceedance; minor exceedances result in only small adjustments, while severe exceedances result in significant pressure drops, achieving on-demand handling. The carbon emission reduction response weighting item ensures that units with high emission intensity and large adjustable margins bear a larger share of the pressure drop for the same degree of exceedance, while equipment with limited load capacity or in critical continuous operation phases bears only a smaller share, thus ensuring the continuity of pouring operations while achieving emission reduction targets. The global penalty adjustment coefficient uniformly constrains the overall control intensity; a larger value results in rapid emission reduction response but increased equipment parameter fluctuations, while a smaller value ensures stable equipment operation but slower emission reduction. It can be calibrated overall according to the control requirements of each stage of the project. Furthermore, the aforementioned target adjustment amount is assembled with the message header, device address field, and check field of the device's communication protocol to form a closed-loop rectification instruction that can be recognized and executed by the front-end IoT device and then sent to the field control bus.

[0064] Then, a status notification indicating the completion of this inference is submitted to the rule engine session lifecycle manager. The lifecycle manager then performs a forced unregistration of the dynamic rule set mounted in the isolated sandbox session, removing the rule set from the session's rule base registry and detaching it from the binding relationships between it and the terminal nodes of the authentication and matching network structure. Next, the single-condition filtering nodes, multi-condition association nodes, and some matching results cached in their left and right memory areas in the authentication and matching network structure are destroyed in sequence. The fact objects or confidence annotation fact object sets residing in the working memory, the context fact objects generated by the construction context subgraph counting, and the residual activation items in the conflict resolution agenda queue are cleared, and the heap memory space occupied by the above objects is released. At the same time, the independent class loader used when compiling the rule set is unloaded to ensure that the dynamically generated rule bytecode is also invalidated. Subsequently, the garbage collection status monitoring interface is invoked to verify the memory clearing status. The count of live object instances in the dynamic rule set of the sandbox session and the number of bytes of heap memory currently occupied by the session are read. When the instance count reaches zero and the occupied memory falls below the baseline memory level when the session is idle, the memory destruction flag is set to true; otherwise, it is set to false, triggering a second memory cleanup and re-verification. Further, the sandbox session identifier assigned during the creation of the isolated sandbox session, the destruction timestamp at the completion of the memory release operation, and the verified memory destruction flag are extracted. These three pieces of information are concatenated and concatenated in a predetermined order. A hash digest is calculated on the concatenated result to generate a fixed-length, tamper-proof checksum. This checksum is then encapsulated as a rule engine memory release confirmation receipt and written to the regulatory audit log.

[0065] In summary, the intelligent carbon emission monitoring method for hydropower projects based on the fusion rule engine of this application is explained. Driven by multi-source carbon monitoring events, it performs temporal alignment and feature encapsulation of equipment fuel consumption, vehicle trajectory, ambient temperature, and carbon emission factors in the ledger. It also associates monitoring equipment with a carbon regulatory spatiotemporal knowledge graph, dynamically extracting the current construction context based on event entities and their spatiotemporal conditions. The entity attributes and relationships in the context are converted into dynamic rules adapted to the current working conditions. Anomaly reasoning is completed in isolated sessions by combining monitoring facts, thereby bridging the gap between graph semantics and executable rules, avoiding static thresholds from being detached from construction conditions and excessive expansion of pre-set rules. Simultaneously, the freshness and confidence of data in each dimension are evaluated, suppressing and filtering the activation of low-confidence rules to reduce misjudgments caused by outdated or interpolated data. After reasoning, rectification instructions are generated and temporary rules and session resources are released, achieving accurate judgment and closed-loop monitoring of carbon emission anomalies.

[0066] Furthermore, a smart monitoring system for carbon emissions from hydropower projects that integrates a rules engine is also provided.

[0067] Figure 5 This is a block diagram of a smart monitoring system for carbon emissions from hydropower projects based on an embodiment of this application using a fusion rule engine. Figure 5 As shown, the intelligent monitoring system 100 for carbon emissions in hydropower projects according to the fusion rule engine of this application includes: a carbon feature encapsulation and entity matching module 110, used to perform timestamp alignment and vectorization encapsulation on multi-source carbon monitoring data to obtain carbon emission feature vectors, and to match the device number with a pre-built carbon monitoring spatiotemporal knowledge graph node to obtain a focus entity identifier; and a subgraph pruning and load generation module 120, used to perform constrained breadth traversal and dynamic subgraph pruning in the carbon monitoring spatiotemporal knowledge graph with the focus entity identifier as the pruning condition, using the spatiotemporal coordinates in the carbon emission feature vector as the pruning condition, and to obtain a construction context subgraph by performing constrained breadth traversal and dynamic subgraph pruning in the carbon monitoring spatiotemporal knowledge graph with the focus entity identifier as the center, and to serialize the numerical attributes in the carbon emission feature vector into event-triggered loads. The graph semantic conversion and rule compilation module 130 is used to perform topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph through the domain ontology converter to obtain the dynamic rule set of the working condition; the rule engine matching and reasoning module 140 is used to load the dynamic rule set of the working condition into the isolated rule engine session, inject the event triggering payload as a fact object into the working memory, and perform deductive reasoning through the pattern matching network to obtain the carbon emission anomaly assertion result; the rectification closure and session destruction module 150 is used to retrieve the rectification measure library based on the anomaly cause code in the carbon emission anomaly assertion result to generate a closed-loop rectification instruction, destroy the dynamic rule set of the working condition and release the rule engine session memory after the reasoning is completed.

[0068] As described above, the intelligent monitoring system 100 for carbon emissions from hydropower projects using a fusion rule engine according to embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent monitoring algorithms for carbon emissions from hydropower projects using a fusion rule engine. In one possible implementation, the intelligent monitoring system 100 for carbon emissions from hydropower projects using a fusion rule engine according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the intelligent monitoring system 100 for carbon emissions from hydropower projects using a fusion rule engine can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent monitoring system 100 for carbon emissions from hydropower projects using a fusion rule engine can also be one of many hardware modules of the wireless terminal.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent monitoring of carbon emissions from hydropower projects integrating a rule engine, characterized in that, include: Step 1: Timestamp alignment and vectorization encapsulation of multi-source carbon monitoring data to obtain carbon emission feature vectors, and matching device numbers with nodes of a pre-built carbon regulatory spatiotemporal knowledge graph to obtain focus entity identifiers; Step 2: Using the spatiotemporal coordinates in the carbon emission feature vector as pruning conditions, and centering on the focal entity identifier, perform constrained breadth traversal and dynamic subgraph pruning in the carbon regulatory spatiotemporal knowledge graph to obtain the construction context subgraph, and serialize the numerical attributes in the carbon emission feature vector into event triggering loads. Step 3: Use the domain ontology converter to perform topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph to obtain the dynamic rule set of the working condition. Step 4: Load the dynamic rule set of operating conditions into the isolated rule engine session, inject the event trigger payload as a fact object into the working memory, and perform deductive reasoning through the pattern matching network to obtain the carbon emission anomaly assertion result. Step 5: Based on the abnormal cause code in the carbon emission anomaly assertion result, retrieve the rectification measures library to generate a closed-loop rectification instruction. After the reasoning is completed, destroy the dynamic rule set of the operating condition and release the rule engine session memory.

2. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 1, characterized in that, Multi-source carbon monitoring data includes real-time fuel consumption of diesel generators, location trajectory coordinates of heavy trucks, temperature of the dam construction area, and carbon emission factors in the ledger.

3. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 2, characterized in that, Step one includes: Heterogeneous frequency resampling and timestamp alignment of multi-source carbon monitoring data are performed to obtain a time-aligned carbon monitoring matrix; Based on the mean and standard deviation of the sliding time window, the time-aligned carbon monitoring matrix is ​​subjected to dimension-wise standardization and tensor flattening to obtain the carbon emission feature vector. Based on equipment coding, entity addressing matching is performed between the equipment identification field in the carbon emission feature vector and the candidate node coding in the carbon regulatory spatiotemporal knowledge graph to obtain the focus entity identifier.

4. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 1, characterized in that, Step two includes: The spatiotemporal and numerical dimensions of carbon emission feature vectors are separated and extracted to obtain spatiotemporal pruning conditions and event triggering loads; Using the focal entity identifier as the center, we perform breadth traversal and weight pruning on the adjacent nodes in the spatiotemporal knowledge graph of carbon regulation under spatiotemporal pruning conditions to obtain candidate nodes and edge sets. The candidate node and edge set are mapped to a graph topology and instantiated with a context subgraph to obtain the construction context subgraph.

5. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 1, characterized in that, Step three includes: Topological feature parsing and entity constraint extraction are performed on the directed edges and node attributes of the construction context subgraph to obtain the context triple set and entity static constraint matrix; Based on the domain mapping dictionary, the semantic-to-logical antecedent isomorphic transformation of the context triple set and the entity static constraint matrix is ​​performed to obtain the abstract logic syntax tree. The abstract logic syntax tree is mapped node by node and instantiated with just-in-time bytecode to obtain a dynamic rule set for operating conditions.

6. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 1, characterized in that, Step four includes: In an isolated sandbox session, the network compiler is started to perform syntactic dimensionality reduction parsing and conditional topology expansion on the dynamic rule set of the working conditions to obtain the identification and matching network structure. The event-triggered payload is deserialized into a fact object and injected into the root node of the identification matching network structure for layer-by-layer pattern propagation matching to obtain the conflict resolution agenda queue. Based on a dynamic priority arbitration strategy, activation rules in the conflict resolution agenda queue are selected and executed, and the deviation degree of carbon emission exceedance and degradation is calculated to obtain the carbon emission anomaly assertion result.

7. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 1, characterized in that, Step five includes: Based on the abnormal cause code and the deviation of excessive degradation in the carbon emission anomaly assertion results, the rectification measures library is retrieved. Combined with the carbon emission reduction response weight of the equipment, the parameter adjustment amount is calculated and the communication protocol message is assembled to obtain the closed-loop rectification instruction. Through the rule engine session lifecycle manager, the dynamic rule set of the working condition is forcibly unregistered and the heap memory is released. The memory clearing status is verified by the garbage collection status monitoring interface to obtain the memory destruction mark. The sandbox session identifier, destruction timestamp, and memory destruction identifier are concatenated and hashed to obtain the rule engine memory release confirmation receipt.

8. The intelligent monitoring method for carbon emissions from hydropower projects using a fusion rule engine as described in claim 6, characterized in that, The event-triggered payload is deserialized into a fact object and injected into the root node of the authentication matching network structure for layer-by-layer pattern propagation matching to obtain the conflict resolution agenda queue, including: Each numerical dimension of the event-triggered payload is labeled with temporal freshness decay assessment and confidence score to obtain a set of confidence-labeled fact objects; The set of confidence-labeled fact objects is injected into the root node of the discriminative matching network structure to perform soft-gated pattern propagation in order to obtain the confidence-weighted activation score vector. Based on a preset minimum confidence activation threshold, the comprehensive activation scores of each rule in the confidence-weighted activation score vector are filtered and discarded for low confidence rules and encapsulated in descending order of priority for high confidence rules to obtain a conflict resolution agenda queue.

9. A smart monitoring system for carbon emissions from hydropower projects integrating a rules engine, characterized in that, include: The carbon feature encapsulation and entity matching module is used to perform timestamp alignment and vectorization encapsulation on multi-source carbon monitoring data to obtain carbon emission feature vectors, and to match the device number with the nodes of the pre-built carbon regulatory spatiotemporal knowledge graph to obtain the focus entity identifier. The subgraph pruning and load generation module is used to perform constrained breadth traversal and dynamic subgraph pruning in the carbon regulatory spatiotemporal knowledge graph with the spatiotemporal coordinates in the carbon emission feature vector as pruning conditions and the focus entity identifier as the center to obtain the construction context subgraph, and serialize the numerical attributes in the carbon emission feature vector into event-triggered loads. The graph semantic conversion and rule compilation module is used to perform topological semantic parsing and logical antecedent mapping on the node attributes and edge relationships of the construction context subgraph through the domain ontology converter to obtain the dynamic rule set of the working condition. The rule engine matching and reasoning module is used to load the dynamic rule set of working conditions into the isolated rule engine session, inject the event trigger payload as a fact object into the working memory, and perform deductive reasoning through the pattern matching network to obtain the carbon emission anomaly assertion result. The rectification loop and session destruction module is used to retrieve the rectification measure library based on the abnormal cause code in the carbon emission anomaly assertion result, generate a closed-loop rectification instruction, destroy the dynamic rule set of the working condition and release the rule engine session memory after the inference is completed.

10. The intelligent monitoring system for carbon emissions from hydropower projects based on the fusion rule engine according to claim 9, characterized in that, The rule engine matching and reasoning module includes: The identification matching network compilation unit is used to start the network compiler in the isolated sandbox session to perform grammatical dimensionality reduction parsing and conditional topology expansion on the dynamic rule set of the working conditions to obtain the identification matching network structure. The fact injection and pattern propagation unit is used to deserialize the event-triggered payload into a fact object and inject it into the root node of the identification matching network structure to perform layer-by-layer pattern propagation matching in order to obtain the conflict resolution agenda queue. The rule arbitration and anomaly assertion unit is used to select and execute active rules in the conflict resolution agenda queue based on a dynamic priority arbitration strategy, and calculate the carbon emission exceedance and degradation deviation to obtain the carbon emission anomaly assertion result.

Images