Pollution source data quality problem intelligent diagnosis method based on knowledge graph
By constructing quadruple data based on knowledge graphs and tracing the root causes of anomalies, the difficulties in semantic understanding and root cause tracing in pollution source data quality diagnosis are solved, achieving high-precision, dynamically adaptable data quality diagnosis and risk warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-26
AI Technical Summary
Existing methods for diagnosing pollution source data quality suffer from weak semantic understanding, incomplete anomaly identification, lack of cross-source data correlation verification, and difficulty in tracing the root causes, leading to false alarms, missed alarms, and blind rectification.
Based on knowledge graphs, quadruple data is constructed for anomaly type diagnosis and root cause labeling. Through specific weight optimization, confidence detection, source tracing report generation, and critical feature database, multi-dimensional and dynamically adapted data quality diagnosis is achieved.
This improves the accuracy of anomaly identification, precisely pinpoints key root causes, establishes a full-cycle data quality assurance system, and reduces resource waste and blind rectification.
Smart Images
Figure CN122089335A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data governance and environmental monitoring technology, specifically to an intelligent diagnostic method for pollution source data quality problems based on knowledge graphs. Background Technology
[0002] With the advancement of digital transformation in environmental regulation, pollution source monitoring data (including online monitoring, manual sampling, and production conditions) has become a core basis for environmental law enforcement, pollution prevention and control decisions, and corporate compliance assessments. However, existing pollution source data quality diagnostic methods mainly rely on traditional threshold detection and simple rule verification, which have significant limitations: First, they have weak semantic understanding capabilities, only able to identify numerical anomalies in a single dimension, and unable to capture implicit semantic relationships and complex logical conflicts between data, such as the difficulty in detecting issues like "production shutdown and non-zero pollutant emission data" or "contradictory pollutant emission ratios at different discharge outlets of the same enterprise." Second, anomaly identification is one-sided, lacking cross-source data correlation verification, and unable to comprehensively judge based on multi-dimensional information such as production conditions, meteorological conditions, and equipment status, easily leading to false alarms and missed alarms. Third, root cause tracing is difficult, only able to locate the abnormal data itself, unable to clarify the cause of the anomaly (such as equipment failure, human interference, and fluctuations in operating conditions), which is not conducive to subsequent repair and problem rectification.
[0003] Therefore, there is an urgent need for an intelligent diagnostic method for pollution source data quality issues based on knowledge graphs, which can achieve multi-dimensional, high-precision, interpretable, and dynamically adaptable pollution source data quality diagnosis. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent diagnostic method for pollution source data quality problems based on knowledge graphs, comprising the following steps:
[0005] Step s1: Convert multi-source data of several historical pollution discharge units into quadruple data, construct a knowledge graph based on the quadruple data, perform anomaly type diagnosis and anomaly root source labeling on the knowledge graph, optimize the weights of the diagnosed anomaly types in the knowledge graph, and obtain the specific weights of each anomaly type.
[0006] Step s2: Convert the multi-source data stored in the database at the current moment into quadruple data and access the knowledge graph to execute the intelligent diagnosis process. If an anomaly type is generated, perform confidence detection on the anomaly type based on the specific weight of the anomaly type and filter the core anomaly type.
[0007] Step s3: Construct anomaly nodes and potential root cause nodes for core anomaly types, trace the root causes of anomalies based on the anomaly nodes and potential root cause nodes, and generate a standardized traceability report based on the traceability results and feed it back to the sewage discharge unit.
[0008] Step s4: If no anomaly type is generated, construct a critical feature database, perform retrospective testing on the quadruple data, and determine whether to generate a predicted anomaly root cause based on the retrospective testing results and feed it back to the sewage discharge unit.
[0009] Furthermore, a cloud management center is constructed, which communicates with data transmission terminals of several pollution discharge units within a preset range. The data transmission terminals are used to upload multi-source data from the pollution discharge units to the cloud management center. The cloud management center executes steps s1 to s4. The cloud management center is equipped with a database, which contains several blockchain nodes. The blockchain nodes are interconnected to form a blockchain network. Each blockchain node is linked to a data uplink terminal, which is used to store the multi-source data of the pollution discharge units to the blockchain node and mark the uplink cycle.
[0010] Furthermore, the process of constructing a knowledge graph based on quadruple data includes:
[0011] Entities, relationships, attributes, and axiom constraints in the pollution source domain are predefined. Based on these entities, relationships, attributes, and axiom constraints, multi-source data from several historical chain cycles of each pollution discharge unit in the database are transformed to generate several historical chain cycles of quadruple data. Entities in all quadruple data are used as nodes, and relationships in the quadruple data are used as connections between nodes to construct a knowledge graph. The attributes and axiom constraints in each quadruple data are then added to each node.
[0012] Furthermore, the process of anomaly type diagnosis and anomaly root cause annotation for knowledge graphs includes:
[0013] Extract multiple attributes of the same entity from the knowledge graph and the axiom constraints corresponding to each attribute. Determine whether each attribute is within the threshold range of the corresponding axiom constraints. If it is, extract multiple attributes of related entities in the knowledge graph that have a connection relationship with the same entity. Based on historical chain cycles, compare the attributes of the same entity and the attributes of related entities within the same historical chain cycle. If the attributes of the same entity and the attributes of related entities conform to the logical and semantic constraints in the corresponding axiom constraints, mark the attributes of the same entity as normal. If the attributes of the same entity and the attributes of related entities do not conform to the logical constraints in the corresponding axiom constraints, mark the attributes of the same entity as logical conflict anomalies. Mark the attributes of related entities in the same historical chain cycle as the attributes marked as logical conflict anomalies as potential anomalous nodes of the same entity. If the attributes of the same entity and the attributes of related entities do not conform to the semantic constraints in the corresponding axiom constraints, mark the attributes of the same entity as semantic mismatch anomalies. Mark the attributes of related entities in the same historical chain cycle as the attributes marked as semantic mismatch anomalies as potential anomalous nodes of the same entity.
[0014] If the attributes of the same entity are not within the threshold range of the corresponding axiom constraint, then the attributes of the same entity are marked as numerical out-of-limit anomalies, the attributes of related entities that have a connection relationship with the same entity are extracted, and the attributes of related entities that are in the same historical chain cycle as the attributes marked as numerical out-of-limit anomalies are marked as potential anomalous nodes of the same entity.
[0015] Anomaly root source nodes are constructed based on preset anomaly root source types. Anomaly root source annotations are performed on entities that have completed anomaly type marking, and the anomaly root source nodes corresponding to the entities are connected to the entities.
[0016] The exception types include numerical out-of-limit exceptions, logical conflict exceptions, and semantic mismatch exceptions.
[0017] Furthermore, the process of obtaining the specific weights for each anomaly type includes:
[0018] The anomaly types and their associated root causes in the knowledge graph are categorized to obtain an anomaly type set and an anomaly root cause set. The total number of anomaly root causes and the total number of different types of anomaly root causes associated with each anomaly type are extracted from the anomaly type set and the anomaly root cause set. Based on the total number of anomaly root causes and the total number of different types of anomaly root causes associated with each anomaly type, the specific weight of each anomaly type is obtained.
[0019] Furthermore, the process of executing the intelligent diagnostic workflow and screening for core abnormality types includes:
[0020] The multi-source data stored in the database at the current moment is transformed to generate four-tuple data. The four-tuple data is then connected to the knowledge graph for anomaly type diagnosis. If an anomaly type is generated corresponding to the four-tuple data, the specific weight of each anomaly type, the number of anomaly root sources associated with each anomaly type, and the entity corresponding to each anomaly type are extracted. It is determined whether the entity corresponding to the anomaly type has generated the same anomaly type in the historical chain cycle. If it has, the attributes of the entity in the historical chain cycle are extracted. The similarity of the attributes of the entity in the historical chain cycle with the attributes connected to the entity at the current moment is matched to obtain the historical data matching degree. At the same time, the anomaly type corresponding to the entity is obtained. Based on the specific weight of each anomaly type, the number of anomaly root sources associated with each anomaly type, the historical data matching degree of the entity corresponding to each anomaly type, and the anomaly type, the confidence of each anomaly type is obtained.
[0021] A pre-set confidence threshold is used to filter out anomaly types with a confidence level greater than the threshold and mark them as core anomaly types.
[0022] Furthermore, the process of constructing the anomaly nodes and potential root cause nodes of the core anomaly types includes:
[0023] Extract the entities corresponding to the core anomaly types in the knowledge graph and other nodes that are connected to the entities. Mark the entities corresponding to the core anomaly types as anomaly nodes. Extract the anomaly root source nodes that are connected to the anomaly nodes and other nodes. Mark the anomaly root source nodes as potential root source nodes.
[0024] Furthermore, the process of tracing the root causes of anomalies based on abnormal nodes and potential root cause nodes, and generating a standardized source tracing report based on the tracing results and feeding it back to the discharging unit includes:
[0025] Based on the path length between each potential root source node and anomaly node in the knowledge graph, the initial correlation coefficient between each node in the knowledge graph is preset, the call records of each relationship between each node in the knowledge graph are counted, the initial correlation coefficient between each node is updated, the real-time correlation coefficient is generated, and the anomaly root source annotation and historical tracing results of the knowledge graph are statistically analyzed to obtain the anomaly type matching degree of each potential root source node.
[0026] Obtain the node paths between abnormal nodes and each potential root cause node in the knowledge graph, obtain the average real-time correlation coefficient of the node paths between abnormal nodes and each potential root cause node, obtain the influence coefficient of each potential root cause node based on the path length, anomaly type matching degree and average real-time correlation coefficient, select the potential root cause node with the largest influence coefficient and mark it as the key root cause node, and mark the node paths between abnormal nodes and key root cause nodes in the knowledge graph as the source tracing paths, generate a standardized source tracing report based on abnormal nodes, key root cause nodes and source tracing paths and feed it back to the pollution discharge unit.
[0027] Furthermore, the process of constructing a critical feature database includes:
[0028] Extract the historical chain cycles of entities that have completed anomaly type labeling and anomaly root source labeling from the knowledge graph, as well as the potential anomaly nodes of the entities. Extract multiple attributes of entities and potential anomaly nodes in the historical chain cycles of the entities. Compare the multiple attributes of entities and potential anomaly nodes in the historical chain cycles of the entities to obtain critical features. Associate the critical features with the anomaly root sources of the entities to construct a critical feature database. Store the critical features that have completed the anomaly root source association into the critical feature database.
[0029] Furthermore, the process of retrospectively testing the quadruple data includes:
[0030] If no anomaly type is generated when the quadruple data of the current on-chain cycle is accessed to the knowledge graph for anomaly type diagnosis, the end timestamp of the current on-chain cycle is used as the review detection time point. At the review detection time point, critical features are extracted from the attributes of each entity accessed to the knowledge graph in the current on-chain cycle. The critical features are then matched with the critical feature database to obtain the cosine similarity of each critical feature in the critical feature database. A preset cosine similarity threshold is set. If there is a critical feature in the critical feature database with a cosine similarity greater than the cosine similarity threshold, the anomaly root cause associated with the critical feature is extracted, the anomaly root cause is marked as the estimated anomaly root cause, and fed back to the sewage discharge unit.
[0031] Compared with the prior art, the beneficial effects of the present invention are:
[0032] 1. By assigning specific weights to anomaly types based on historical data and root cause correlation analysis, and combining multi-dimensional factors to calculate confidence levels to screen core anomaly types, we can effectively focus on key data quality issues, reduce interference from irrelevant anomalies, improve the accuracy of anomaly identification, and avoid wasting resources on handling low-value anomalies.
[0033] 2. By building a network of connections between abnormal nodes and potential root cause nodes based on knowledge graphs, and calculating the impact coefficient through multiple dimensions such as path length, real-time correlation coefficient, and anomaly type matching degree, the key root cause can be accurately identified, making the location of anomaly causes more evidence-based, providing a clear direction for the rapid rectification of pollution discharge units, and reducing the blindness of problem investigation.
[0034] 3. Construct a critical feature database to conduct retrospective testing on quadruple data that does not directly generate anomaly types, identify potential data quality risks and predict the root causes of anomalies in advance, and extend from "managing existing anomalies" to "early warning of potential anomalies" to form a full-cycle data quality assurance system.
[0035] 4. Enhance the dynamic adaptability of the diagnostic system. Continuously optimize key parameters such as anomaly type specific weights and node correlation coefficients based on historical chain cycle data, so that the diagnostic model can be continuously iterated and upgraded with the accumulation of data, dynamically adapting to changes in the production conditions and environmental conditions of the sewage discharge unit, and continuously improving the adaptability and effectiveness of the diagnostic method. Attached Figure Description
[0036] Figure 1 This is a flowchart of a knowledge graph-based intelligent diagnostic method for pollution source data quality problems, as described in an embodiment of this application. Detailed Implementation
[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0038] like Figure 1 As shown, the intelligent diagnostic method for pollution source data quality problems based on knowledge graphs includes the following steps:
[0039] Step s1: Convert multi-source data of several historical pollution discharge units into quadruple data, construct a knowledge graph based on the quadruple data, perform anomaly type diagnosis and anomaly root source labeling on the knowledge graph, optimize the weights of the diagnosed anomaly types in the knowledge graph, and obtain the specific weights of each anomaly type.
[0040] Step s2: Convert the multi-source data stored in the database at the current moment into quadruple data and access the knowledge graph to execute the intelligent diagnosis process. If an anomaly type is generated, perform confidence detection on the anomaly type based on the specific weight of the anomaly type and filter the core anomaly type.
[0041] Step s3: Construct abnormal nodes and potential root source nodes for core abnormal types, trace the root causes of abnormalities based on abnormal nodes and potential root source nodes, and generate a standardized traceability report based on the traceability results (abnormal nodes, key root source nodes, and traceability paths) and feed it back to the pollution discharge unit.
[0042] Step s4: If no anomaly type is generated, construct a critical feature database, perform retrospective testing on the quadruple data, and determine whether to generate a predicted anomaly root cause based on the retrospective testing results and feed it back to the sewage discharge unit.
[0043] It should be further explained that, in the specific implementation process, a cloud management center is constructed. The cloud management center communicates with the data transmission terminals of several pollution discharge units within a preset range. The data transmission terminals are used to upload multi-source data of the pollution discharge units to the cloud management center. The cloud management center executes steps s1 to s4. The cloud management center is equipped with a database, and the database is equipped with several blockchain nodes. The blockchain nodes are linked to each other to form a blockchain network. Each blockchain node is linked to a data uplink terminal, which is used to store the multi-source data of the pollution discharge units to the blockchain node and mark the uplink period.
[0044] It should be further explained that, in the specific implementation process, the multi-source data of several historical pollution discharge units across multiple cycles is transformed into quadruple data. The process of constructing a knowledge graph based on the quadruple data includes:
[0045] The entities, relationships, attributes, and axiomatic constraints in the pollution source domain are predefined. Entities include pollution discharge units (e.g., factories, enterprises), discharge outlets (e.g., wastewater discharge outlets, exhaust gas discharge outlets), pollutants (e.g., COD, ammonia nitrogen, particulate matter), monitoring equipment (e.g., online monitoring instruments, sampling equipment), standard limits (e.g., national emission standards, local emission standards), production conditions (e.g., production load, operating time), and meteorological factors (e.g., temperature, humidity, wind speed). Core relationships include emission (pollution discharge unit - pollutant), monitoring (monitoring equipment - discharge outlet), association (pollution discharge unit - discharge outlet), compliance (pollutant - standard limit), and impact (meteorological factors - pollutant emission, production conditions - pollutant emission). Core attributes include pollutant concentration and emission flow rate. Monitoring parameters include time, equipment range, equipment calibration time, production load, temperature, and humidity. Axiomatic constraints are categorized into three types: numerical constraints (e.g., pollutant concentration ≥ 0, wastewater pH 6-9, exhaust gas particulate matter concentration ≤ 10 mg / m³), logical constraints (e.g., the error between online monitoring data and manual sampling data for the same pollutant from the same discharge unit at the same time period ≤ 15%, monitoring equipment range ≥ 1.2 times the corresponding pollutant standard limit), and semantic constraints (e.g., no exhaust gas pollutants are emitted from industrial wastewater discharge outlets, pollutant emission concentration should be close to 0 when production load is 0). Based on the entities, relationships, attributes, and axiomatic constraints within the pollution source domain, the multi-source data from several historical chain cycles of various discharge units in the database are transformed to generate quadruple data from several historical chain cycles. ;
[0046] ;in, This represents a quadruple describing multi-source data, where j represents a specific entity object. Represents a set of entity objects, where u describes the attributes of the entity objects. Used to describe the relationships between entity objects, d and g represent the set of attributes and the set of relationships used to describe the entity objects in s, respectively, and h represents the set of literal values for attributes and relationships. The axioms and constraints used to describe entity objects are used to construct a knowledge graph by treating all entities in the quadruple data as nodes (where an entity alignment algorithm based on the fusion of attribute similarity and semantic similarity is used to eliminate semantic conflicts between cross-source data and align entities with the same meaning as the same entity, such as aligning "Factory A" and "Enterprise A" as the same entity). The relations in the quadruple data are used as the connection relations between nodes. The attributes and axioms and constraints in each quadruple data are then added to each node.
[0047] It should be further explained that, in the specific implementation process, the process of diagnosing anomalies and labeling root causes of anomalies in the knowledge graph includes:
[0048] Extract multiple attributes of the same entity from the knowledge graph and the axiom constraints corresponding to each attribute. Determine whether each attribute falls within the threshold range of the corresponding axiom constraints. If it does, extract multiple attributes of related entities in the knowledge graph that have connections with the same entity. Based on historical chain cycles, compare the attributes of the same entity within the same historical chain cycle with the attributes of related entities. If the attributes of the same entity and the attributes of related entities conform to the logical and semantic constraints in the corresponding axiom constraints, then the attributes of the same entity are marked as normal. If the attributes of the same entity and the attributes of related entities do not conform to the logical constraints in the corresponding axiom constraints (e.g., the production load of the sewage discharge unit is 0 (attribute 1), but the COD concentration at the sewage outlet is 12 mg / L (attribute 2), then it is considered a violation). If the attribute of the same entity is marked as a logical conflict anomaly (the attribute of the same entity is marked as a logical conflict anomaly), and the attribute of the related entity that is in the same historical chain period as the attribute marked as a logical conflict anomaly is marked as a potential anomaly node of the same entity, if the attribute of the same entity and the attribute of the related entity do not conform to the semantic constraints in the corresponding axiom constraints (the pollutant attribute of wastewater discharge outlet B is "particulate matter" (air pollutant), which violates "wastewater discharge outlet does not discharge air" → judged as "semantic mismatch anomaly"), then the attribute of the same entity is marked as a semantic mismatch anomaly, and the attribute of the related entity that is in the same historical chain period as the attribute marked as a semantic mismatch anomaly is marked as a potential anomaly node of the same entity.
[0049] If the attributes of the same entity are not within the threshold range of the corresponding axiom constraint, then the attributes of the same entity are marked as numerical out-of-limit anomalies, the attributes of related entities that have a connection relationship with the same entity are extracted, and the attributes of related entities that are in the same historical chain cycle as the attributes marked as numerical out-of-limit anomalies are marked as potential anomalous nodes of the same entity.
[0050] Anomaly root source nodes are constructed based on preset anomaly root source types (including equipment failure, transmission interruption, human interference, etc.). Entities that have completed anomaly type marking are manually annotated with anomaly root sources, and the anomaly root source nodes corresponding to the entities are connected to the entities.
[0051] The exception types include numerical out-of-limit exceptions, logical conflict exceptions, and semantic mismatch exceptions.
[0052] It should be further explained that, in the specific implementation process, the process of optimizing the weights of the diagnosed anomaly types in the knowledge graph and obtaining the specific weights for each anomaly type includes:
[0053] In existing anomaly root cause diagnosis, each anomaly root cause in the ontology knowledge base is associated with a certain number of anomaly types. The assumption of this static association is that all anomaly types associated with a root cause collectively affect that root cause. That is, if the detected anomaly type set A completely covers all anomaly types associated with a certain anomaly root cause, with no redundant or missing anomaly types, then the core anomaly root cause of that anomaly type problem can be considered to be ri. Under this assumption, if two anomaly root causes r1 and r2 both contain all or the same anomaly types in the detected anomaly type set A, it cannot be assumed that the correlation (influence weight) between these two root causes and the data quality problem is the same. The influence of other anomaly types (undetected anomaly types) associated with each of these two root causes also needs to be considered.
[0054] To address the aforementioned issues, this invention categorizes the anomaly types and associated root causes within the knowledge graph, obtaining an anomaly type set and an anomaly root cause set. From these sets, the total number of root causes M and the total number of different types of root causes Ma associated with each anomaly type are extracted. For example, if "constant data value" is only associated with "equipment failure" in the anomaly type and root cause set, then Ma=1; if "data missing" is associated with both "transmission interruption" and "human error," then Ma=2. Based on the total number of root causes and the total number of different types of root causes associated with each anomaly type, the specific weight of each anomaly type is obtained.
[0055] It should be further explained that, in the specific implementation process, the process of obtaining the specific weight for each anomaly type includes:
[0056] ,in, Here, M represents the specific weights, and M represents the total number of outlier sources. The number of different types of root causes associated with an anomaly type is the total number of root causes (Ma). The smaller the number of root causes (Ma) associated with a certain anomaly type, the higher the weight (Ma). The higher the value, the more targeted the exception type is at pinpointing a specific root cause.
[0057] It should be further explained that, in the specific implementation process, the multi-source data currently stored in the database is transformed into four-tuple data and accessed through the knowledge graph to execute the intelligent diagnostic process. If an anomaly type is generated, a confidence test is performed on the anomaly type based on its specific weight. The process of screening core anomaly types includes:
[0058] Convert the multi-source data stored at the current moment in the database to generate quadruple data, and connect the quadruple data to the knowledge graph for abnormal type diagnosis. If the abnormal types corresponding to the generated quadruple data (a quadruple data may not only be diagnosed as one abnormal type, but may violate multiple axiom constraints in the knowledge graph at the same time: numerical, logical, and semantic constraints, and multiple groups of abnormal types may be diagnosed), extract the specific weights of each abnormal type, the number of abnormal roots associated with each abnormal type, and the entities corresponding to each abnormal type. Determine whether the entity corresponding to the abnormal type generated the same abnormal type in the previous chain cycle. If so, extract the attributes of the entity in the previous chain cycle, and perform a similarity match (cosine similarity match) between the attributes of the entity in the previous chain cycle and the attributes accessed by the entity at the current moment to obtain the historical data matching degree. At the same time, obtain the types of abnormal types corresponding to the entity (such as being marked by both numerical overlimit abnormality and logical conflict abnormality, then the confidence level is increased). According to the specific weights of each abnormal type, the number of abnormal roots associated with each abnormal type, the historical data matching degree of the entities corresponding to each abnormal type, and the types of abnormal types, obtain the confidence levels of each abnormal type;
[0059] ;
[0060] Among them, is the confidence level of the abnormal type (value range: 0 < C ≤ 1, the closer to 1, the higher the credibility);
[0061] : The specific weight of the abnormal type (reflecting the uniqueness of the abnormal type determination rule, value range: 0.5 ≤ ≤ 1);
[0062] : The number of abnormal roots associated with the abnormal type (that is, how many roots this abnormal type can correspond to, value range: ≥ 1);
[0063] : The historical data matching degree of the entity corresponding to the abnormal type (similarity of features with the historically confirmed abnormality, value range: 0 < ≤ 1);
[0064] : The weight of the abnormal type category (assigned based on the priority of the abnormal type, value range: 0.5 ≤ ≤ 1);
[0065] , , , : Weight coefficient (satisfying + + + =1, default value: =0.35, =0.25, =0.25, =0.15).
[0066] A pre-set confidence threshold (90%) is used to filter out anomaly types with a confidence level greater than the confidence threshold and mark them as core anomaly types.
[0067] It should be further explained that, in the specific implementation process, the process of constructing the anomaly nodes and potential root source nodes of the core anomaly types includes:
[0068] Extract the entities corresponding to the core anomaly types in the knowledge graph and other nodes that are connected to the entities. Mark the entities corresponding to the core anomaly types as anomaly nodes. Extract the anomaly root source nodes that are connected to the anomaly nodes and other nodes. Mark the anomaly root source nodes as potential root source nodes.
[0069] It should be further explained that, in the specific implementation process, the process of tracing the root causes of anomalies based on abnormal nodes and potential root cause nodes, and generating a standardized source tracing report based on the tracing results (abnormal nodes, key root cause nodes, and tracing paths) and feeding it back to the discharging unit includes:
[0070] Based on the path length between each potential root source node and anomaly node in the knowledge graph (path length is the number of nodes between the potential root source node and the anomaly node; the shorter the path, the higher the weight; direct association paths have higher weight than multi-hop association paths), an initial association coefficient is preset between each node in the knowledge graph. The call records of each relationship between nodes in the knowledge graph are statistically analyzed, and the initial association coefficients between nodes are updated to generate a real-time association coefficient. The real-time association coefficient = (the number of times the relationship actually triggers an anomaly / the total number of times the relationship is called). For example, if the call records of each relationship in the knowledge graph are statistically analyzed, such as the relationship "sensor drift → COD concentration exceeding the limit" being called 100 times, of which 92 times actually caused an anomaly, then the real-time association coefficient = 92 / 100 = 0.92. Furthermore, statistical analysis is performed on the anomaly root source annotation and historical tracing results of the knowledge graph to obtain the anomaly type matching degree of each potential root source node. The anomaly type matching degree = (the historical number of times a certain root source type caused this anomaly type / the total historical number of times this root source type caused all anomaly types).
[0071] Obtain the node paths between abnormal nodes and each potential root cause node in the knowledge graph. Obtain the average real-time correlation coefficient (the average of the real-time correlation coefficients between each node in the node path) of the node paths between abnormal nodes and each potential root cause node. Perform weighted averaging based on the path length, anomaly type matching degree, and average real-time correlation coefficient of each potential root cause node to obtain the influence coefficient of each potential root cause node. Select the potential root cause node with the largest influence coefficient and mark it as the key root cause node. Mark the node paths between abnormal nodes and key root cause nodes in the knowledge graph as source tracing paths. Generate a standardized source tracing report based on abnormal nodes, key root cause nodes, and source tracing paths and feed it back to the pollution discharge unit.
[0072] It should be further explained that, in the specific implementation process, the process of constructing the critical feature database includes:
[0073] Extract the historical chain cycles of entities and potential anomalous nodes of entities that have completed anomaly type labeling and anomaly root source labeling in the knowledge graph. Extract multiple attributes of entities and potential anomalous nodes in the historical chain cycles of entities. Compare the multiple attributes of entities and potential anomalous nodes in the historical chain cycles of entities to obtain critical features. Critical features include the Pearson correlation coefficient and the average standard deviation of multiple attributes of entities and potential anomalous nodes in the historical chain cycles. Associate critical features with the anomaly root sources of entities to construct a critical feature database. Store the critical features that have completed the anomaly root source association in the critical feature database.
[0074] It should be further explained that, in the specific implementation process, the retrospective testing of the quaternion data, and the process of determining whether to generate a predicted anomaly root cause and feeding it back to the wastewater discharge unit based on the retrospective test results, include:
[0075] If no anomaly type is generated when the quadruple data of the current on-chain cycle is accessed to the knowledge graph for anomaly type diagnosis, the end timestamp of the current on-chain cycle is used as the review detection time point. At the review detection time point, critical features are extracted from the attributes of each entity accessed to the knowledge graph in the current on-chain cycle. Critical features include the Pearson correlation coefficient and the average standard deviation of multiple attributes between entities with connection relationships. The critical features are matched with the critical feature database to obtain the cosine similarity of each critical feature in the critical feature database. The preset cosine similarity threshold is 0.9. If there is a critical feature in the critical feature database with a cosine similarity greater than the cosine similarity threshold, the anomaly root cause associated with the critical feature is extracted, the anomaly root cause is marked as the estimated anomaly root cause and fed back to the sewage discharge unit.
[0076] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A knowledge graph-based intelligent diagnostic method for pollution source data quality issues, characterized in that, Includes the following steps: Step s1: Convert multi-source data of several historical pollution discharge units into quadruple data, construct a knowledge graph based on the quadruple data, perform anomaly type diagnosis and anomaly root source labeling on the knowledge graph, optimize the weights of the diagnosed anomaly types in the knowledge graph, and obtain the specific weights of each anomaly type. Step s2: Convert the multi-source data stored in the database at the current moment into quadruple data and access the knowledge graph to execute the intelligent diagnosis process. If an anomaly type is generated, perform confidence detection on the anomaly type based on the specific weight of the anomaly type and filter the core anomaly type. Step s3: Construct anomaly nodes and potential root cause nodes for core anomaly types, trace the root causes of anomalies based on the anomaly nodes and potential root cause nodes, and generate a standardized traceability report based on the traceability results and feed it back to the sewage discharge unit. Step s4: If no anomaly type is generated, construct a critical feature database, perform retrospective testing on the quadruple data, and determine whether to generate a predicted anomaly root cause based on the retrospective testing results and feed it back to the sewage discharge unit.
2. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 1, characterized in that, A cloud management center is constructed, which communicates with data transmission terminals of several pollution discharge units within a preset range. The data transmission terminals are used to upload multi-source data of the pollution discharge units to the cloud management center. The cloud management center executes steps s1 to s4. The cloud management center is equipped with a database, which contains several blockchain nodes. The blockchain nodes are interconnected to form a blockchain network. Each blockchain node is linked to a data uplink terminal, which is used to store the multi-source data of the pollution discharge units to the blockchain node and mark the uplink period.
3. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 2, characterized in that, The process of constructing a knowledge graph based on quadruplet data includes: Entities, relationships, attributes, and axiom constraints in the pollution source domain are predefined. Based on these entities, relationships, attributes, and axiom constraints, multi-source data from several historical chain cycles of each pollution discharge unit in the database are transformed to generate several historical chain cycles of quadruple data. Entities in all quadruple data are used as nodes, and relationships in the quadruple data are used as connections between nodes to construct a knowledge graph. The attributes and axiom constraints in each quadruple data are then added to each node.
4. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 3, characterized in that, The process of anomaly type diagnosis and anomaly root cause annotation for knowledge graphs includes: Extract multiple attributes of the same entity from the knowledge graph and the axiom constraints corresponding to each attribute. Determine whether each attribute is within the threshold range of the corresponding axiom constraints. If it is, extract multiple attributes of related entities in the knowledge graph that have a connection relationship with the same entity. Based on historical chain cycles, compare the attributes of the same entity and the attributes of related entities within the same historical chain cycle. If the attributes of the same entity and the attributes of related entities conform to the logical and semantic constraints in the corresponding axiom constraints, mark the attributes of the same entity as normal. If the attributes of the same entity and the attributes of related entities do not conform to the logical constraints in the corresponding axiom constraints, mark the attributes of the same entity as logical conflict anomalies. Mark the attributes of related entities in the same historical chain cycle as the attributes marked as logical conflict anomalies as potential anomalous nodes of the same entity. If the attributes of the same entity and the attributes of related entities do not conform to the semantic constraints in the corresponding axiom constraints, mark the attributes of the same entity as semantic mismatch anomalies. Mark the attributes of related entities in the same historical chain cycle as the attributes marked as semantic mismatch anomalies as potential anomalous nodes of the same entity. If the attributes of the same entity are not within the threshold range of the corresponding axiom constraint, then the attributes of the same entity are marked as numerical out-of-limit anomalies, the attributes of related entities that have a connection relationship with the same entity are extracted, and the attributes of related entities that are in the same historical chain cycle as the attributes marked as numerical out-of-limit anomalies are marked as potential anomalous nodes of the same entity. Anomaly root source nodes are constructed based on preset anomaly root source types. Anomaly root source annotations are performed on entities that have completed anomaly type marking, and the anomaly root source nodes corresponding to the entities are connected to the entities. Exception types include numerical out-of-limit exceptions, logical conflict exceptions, and semantic mismatch exceptions.
5. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 4, characterized in that, The process of obtaining the specific weights for each anomaly type includes: The anomaly types and their associated root causes in the knowledge graph are categorized to obtain an anomaly type set and an anomaly root cause set. The total number of anomaly root causes and the total number of different types of anomaly root causes associated with each anomaly type are extracted from the anomaly type set and the anomaly root cause set. Based on the total number of anomaly root causes and the total number of different types of anomaly root causes associated with each anomaly type, the specific weight of each anomaly type is obtained.
6. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 5, characterized in that, The process of executing the intelligent diagnostic workflow and screening for core abnormality types includes: The multi-source data stored in the database at the current moment is transformed to generate four-tuple data. The four-tuple data is then connected to the knowledge graph for anomaly type diagnosis. If an anomaly type is generated corresponding to the four-tuple data, the specific weight of each anomaly type, the number of anomaly root sources associated with each anomaly type, and the entity corresponding to each anomaly type are extracted. It is determined whether the entity corresponding to the anomaly type has generated the same anomaly type in the historical chain cycle. If it has, the attributes of the entity in the historical chain cycle are extracted. The similarity of the attributes of the entity in the historical chain cycle with the attributes connected to the entity at the current moment is matched to obtain the historical data matching degree. At the same time, the anomaly type corresponding to the entity is obtained. Based on the specific weight of each anomaly type, the number of anomaly root sources associated with each anomaly type, the historical data matching degree of the entity corresponding to each anomaly type, and the anomaly type, the confidence of each anomaly type is obtained. A pre-set confidence threshold is used to filter out anomaly types with a confidence level greater than the threshold and mark them as core anomaly types.
7. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 6, characterized in that, The process of constructing the exception nodes and potential root source nodes for the core exception types includes: Extract the entities corresponding to the core anomaly types in the knowledge graph and other nodes that are connected to the entities. Mark the entities corresponding to the core anomaly types as anomaly nodes. Extract the anomaly root source nodes that are connected to the anomaly nodes and other nodes. Mark the anomaly root source nodes as potential root source nodes.
8. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 7, characterized in that, The process of tracing the root causes of anomalies based on anomalous nodes and potential root cause nodes, and generating a standardized source tracing report based on the tracing results and feeding it back to the discharging unit, includes: Based on the path length between each potential root source node and anomaly node in the knowledge graph, the initial correlation coefficient between each node in the knowledge graph is preset, the call records of each relationship between each node in the knowledge graph are counted, the initial correlation coefficient between each node is updated, the real-time correlation coefficient is generated, and the anomaly root source annotation and historical tracing results of the knowledge graph are statistically analyzed to obtain the anomaly type matching degree of each potential root source node. Obtain the node paths between abnormal nodes and each potential root cause node in the knowledge graph, obtain the average real-time correlation coefficient of the node paths between abnormal nodes and each potential root cause node, obtain the influence coefficient of each potential root cause node based on the path length, anomaly type matching degree and average real-time correlation coefficient, select the potential root cause node with the largest influence coefficient and mark it as the key root cause node, and mark the node paths between abnormal nodes and key root cause nodes in the knowledge graph as the source tracing paths, generate a standardized source tracing report based on abnormal nodes, key root cause nodes and source tracing paths and feed it back to the pollution discharge unit.
9. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 8, characterized in that, The process of constructing a critical feature database includes: Extract the historical chain cycles of entities that have completed anomaly type labeling and anomaly root source labeling from the knowledge graph, as well as the potential anomaly nodes of the entities. Extract multiple attributes of entities and potential anomaly nodes in the historical chain cycles of the entities. Compare the multiple attributes of entities and potential anomaly nodes in the historical chain cycles of the entities to obtain critical features. Associate the critical features with the anomaly root sources of the entities to construct a critical feature database. Store the critical features that have completed the anomaly root source association into the critical feature database.
10. The intelligent diagnostic method for pollution source data quality problems based on knowledge graphs according to claim 9, characterized in that, The process of retrospectively testing quadruple data includes: If no anomaly type is generated when the quadruple data of the current on-chain cycle is accessed to the knowledge graph for anomaly type diagnosis, the end timestamp of the current on-chain cycle is used as the review detection time point. At the review detection time point, critical features are extracted from the attributes of each entity accessed to the knowledge graph in the current on-chain cycle. The critical features are then matched with the critical feature database to obtain the cosine similarity of each critical feature in the critical feature database. A preset cosine similarity threshold is set. If there is a critical feature in the critical feature database with a cosine similarity greater than the cosine similarity threshold, the anomaly root cause associated with the critical feature is extracted, the anomaly root cause is marked as the estimated anomaly root cause, and fed back to the sewage discharge unit.