Intelligent data query generation method based on large model and knowledge graph

Abstract

The invention relates to the technical field of data processing, in particular to an intelligent data query generation method based on a large model and a knowledge graph, and the method comprises the steps: obtaining data, and judging a fault event; determining a correlation coefficient; screening temporary nodes; determining risk nodes; connecting map edges; determining an abnormal node; adjusting a threshold value and generating a report. According to the method, multiple key data are acquired, equipment fault events are identified, correlation coefficients of data sets are determined according to a large model, temporary nodes are screened on the basis, then risk nodes are determined according to a prediction deviation rate, a correlation network is constructed by using data correlation to determine abnormal nodes, and meanwhile, a threshold value is dynamically optimized. Finally, a query report is generated, it is ensured that the core data set reflecting the equipment fault can be accurately positioned, and the problems that due to the fact that a static threshold value and a changeable working environment are excessively relied on, the fault data query accuracy is low, and query response lags behind are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent data query and generation method based on large models and knowledge graphs. Background Technology

[0002] In industrial production, when equipment experiences a continuous decline in output efficiency without a clear cause, traditional methods require maintenance personnel to manually check massive amounts of sensor data, process parameters, and quality records. This is time-consuming and prone to missing crucial fault clues. Therefore, automatically triggering intelligent data queries to quickly locate the core data sets causing equipment malfunctions is crucial for subsequent work.

[0003] Chinese Patent Application Publication No. CN120687475A discloses a knowledge graph query method, apparatus, electronic device, and storage medium. The method includes: responding to a natural query statement entered by a query user on a query interface, determining second query information corresponding to first query information in the natural query statement from a pre-constructed knowledge graph, and determining semantic information of the second query information based on the knowledge graph; generating an initial database query statement based on the second query information and guided by the semantic information; optimizing the initial database query statement based on all triples in the knowledge graph to obtain a target database query statement; and querying data in the knowledge graph based on the target database query statement.

[0004] Therefore, the knowledge graph query method has the following problems: it passively responds to user input and cannot proactively discover hidden abnormal relationships in the data before the user realizes the problem; it relies heavily on the quality of the natural query statement entered by the user, and if the user misjudges or describes the root cause of the fault inaccurately, the generated query statement will fail to hit the core problem; the query logic and thresholds are fixed, resulting in poor environmental adaptability. Summary of the Invention

[0005] To address this, the present invention provides an intelligent data query generation method based on large models and knowledge graphs, which overcomes the problems of low accuracy and delayed query response in existing technologies due to over-reliance on static thresholds and changing operating environments through multi-dimensional data analysis and dynamic adjustment mechanisms.

[0006] To achieve the above objectives, this invention provides an intelligent data query and generation method based on large models and knowledge graphs, comprising: Real-time acquisition of the output counting interval of production equipment in the industrial production database, the prediction deviation rate of each monitoring node in the knowledge graph constructed with each set of data to be measured during the operation of production equipment as monitoring nodes, the process criticality, and the data correlation between any two monitoring nodes. A device malfunction event is determined based on the output counting interval and the preset interval threshold. Based on the equipment failure event, the correlation coefficient of each monitoring node is determined according to the preset large model and each monitoring node; Several temporary nodes are selected based on the correlation coefficients and preset correlation thresholds; Several risk nodes are determined based on the prediction deviation rate of each of the temporary nodes; Connect two risk nodes based on the data correlation between any two risk nodes to obtain several graph edges; Several abnormal nodes are determined based on the risk determination frequency of each risk node within a preset determination time and the number of times each risk node is connected by the graph edge. The preset correlation threshold or the preset interval threshold is adjusted based on the prediction deviation rate, process criticality, and number of occurrences of each abnormal node. A fault association data query report is generated based on the correlation coefficients of all the abnormal nodes obtained after adjusting the preset correlation threshold or the preset interval threshold.

[0007] Furthermore, the process of selecting several temporary nodes based on the aforementioned correlation coefficients includes: Based on the comparison between the correlation coefficient and the preset correlation threshold, the monitoring node is determined to be a temporary node, so as to filter out a number of temporary nodes.

[0008] Furthermore, the process of determining a number of risk nodes based on the prediction deviation rate of each of the temporary nodes includes: Based on the comparison between the predicted deviation rate and the preset deviation rate threshold, the temporary node is determined to be a marked node, and several marked nodes are obtained; By comparing the prediction deviation rates of all the marked nodes, the marked node with the largest prediction deviation rate is determined as the node of interest; Several risk nodes are determined based on the prediction deviation rate of the nodes of interest and any of the marked nodes within a preset time period.

[0009] Furthermore, the process of determining several risk nodes based on the prediction deviation rate of the nodes of interest and any of the marked nodes within a preset time period includes: Calculate several first deviation rate fluctuation values ​​based on the prediction deviation rate of the node of interest within the preset time period; Calculate several second deviation rate fluctuation values ​​based on the prediction deviation rate of the marked nodes within the preset time period; The deviation pattern matching degree is calculated based on all the first deviation rate fluctuation values ​​and all the second deviation rate fluctuation values; Based on the comparison result between the deviation pattern matching degree and the preset matching degree threshold, the nodes of interest and the marked nodes are determined as risk nodes, thereby identifying several risk nodes.

[0010] Furthermore, the process of connecting two risk nodes based on the data relevance of any two risk nodes to obtain several graph edges includes: The correlation deviation is calculated based on the data correlation degree and the preset data correlation threshold. Based on the comparison between the correlation deviation and the preset correlation deviation threshold, two risk nodes are connected to obtain several graph edges.

[0011] Furthermore, the process of determining several abnormal nodes based on the risk determination frequency of each risk node within a preset determination time period and the number of times each risk node is connected by the graph edge includes: Based on the comparison between the risk determination frequency and the preset frequency threshold, the risk node is determined to be a high-frequency risk node; Based on the number of times the high-frequency risk node is connected by the graph edge at each risk determination time within the preset determination period, a number of abnormal nodes are determined. The risk determination time refers to the moment when each of the risk nodes is determined to be a risk node within the preset determination time.

[0012] Furthermore, the process of determining several abnormal nodes based on the number of times the high-frequency risk node is connected by the graph edge at each risk determination time within the preset determination period includes: The average number of times a high-frequency risk node is connected by a graph edge is calculated based on the number of times the risk is determined at all risk determination moments within the preset determination time period. Based on the comparison between the mean frequency and the preset mean frequency threshold, the frequency fluctuation value is calculated according to the number of times the high-frequency risk node is connected by the graph edge at all the risk determination times within the preset determination time. Based on the comparison between the frequency fluctuation value and the preset frequency fluctuation threshold, the high-frequency risk node is determined as the abnormal node, thereby identifying several abnormal nodes.

[0013] Furthermore, the process of adjusting the preset relevance threshold or the preset interval threshold based on the prediction deviation rate, process criticality, and number of occurrences of all the abnormal nodes includes: Calculate the mean prediction deviation based on all the prediction deviation rates described; The relative prediction deviation is calculated based on the mean prediction deviation and the preset deviation rate threshold. The deviation index is calculated based on the predicted relative deviation, the preset predicted relative deviation weight, the process criticality, and the preset criticality weight. The preset relevant threshold or the preset interval threshold is adjusted based on the deviation index, the preset deviation index range, and the number of times.

[0014] Further, the process of adjusting the preset correlation threshold or the preset interval threshold based on the deviation index, the preset deviation index range, and the number of times includes: Based on the comparison between the deviation index and the maximum value of the preset deviation index range, the preset relevant threshold is adjusted according to the deviation index and the maximum value of the preset deviation index range. Based on the comparison between the deviation index and the minimum value of the preset deviation index range, the index deviation is calculated according to the deviation index and the minimum value of the preset deviation index range. The deviation of the number of times is calculated based on the number of times and the preset average threshold of the number of times. The adjustment factor is calculated based on the frequency deviation, the preset frequency deviation weight, the exponential deviation, and the preset exponential deviation weight; The preset interval threshold is adjusted according to the adjustment factor.

[0015] Furthermore, the process of determining the occurrence of a device malfunction event based on the output counting interval and the preset interval threshold includes: Calculate the interval average based on all output count intervals within the preset counting period; A device malfunction event is determined based on the comparison between the average interval and the preset interval threshold.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: by acquiring a variety of key data, it first identifies equipment failure events; then, a large model understands the semantic correlation of each data group to obtain correlation coefficients; and based on this, relevant temporary nodes are screened out. Subsequently, the degree of data anomaly is quantified by combining the prediction deviation rate to determine risk nodes. Then, the association network of risk nodes is constructed using the data correlation in the knowledge graph, thereby identifying abnormal nodes in key positions in the topology. At the same time, the screening criteria are dynamically optimized through an adaptive threshold mechanism. Finally, a fault association data query report is generated, ensuring that the core data groups reflecting equipment failures can be continuously and accurately located in massive industrial data, greatly improving operation and maintenance efficiency and diagnostic accuracy. It effectively solves the problems of low accuracy and delayed query response caused by over-reliance on static thresholds and changing operating environments.

[0017] Furthermore, by monitoring macroscopic performance trends to capture fundamental changes in equipment operating status rather than responding to instantaneous fluctuations, the output counting interval serves as a direct quantitative indicator of overall equipment efficiency. By calculating the average interval over a preset counting period, random interference in the production process can be effectively filtered out, thereby capturing continuous performance degradation trends. When the average interval exceeds a preset threshold, it indicates that the equipment has transitioned from occasional delays to a systemic efficiency decline, providing an accurate fault trigger signal for subsequent multi-dimensional data correlation analysis. This effectively avoids false triggers caused by instantaneous anomalies, ensuring that the entire intelligent diagnostic process is launched both timely and reliably.

[0018] Furthermore, by pre-setting relevant thresholds to screen correlation coefficients, a semantic association field centered on fault events is essentially established in the multidimensional data space. Only nodes with association strength exceeding the critical value are retained as temporary nodes. This effectively overcomes the limitations of traditional keyword matching or rule engines, captures implicit associations that are difficult for humans to discover from the underlying data logic, and builds a high-quality data foundation for subsequent accurate analysis, thereby significantly improving the accuracy and efficiency of the entire intelligent diagnostic process.

[0019] Furthermore, by comparing the prediction deviation rate with the preset deviation rate threshold, the system can filter out the actual abnormal marker nodes from the temporary nodes obtained from semantic correlation. If the prediction deviation rate is greater than the preset deviation rate threshold, it indicates that the actual state of the data group deviates from its expected normal state by more than the threshold, which means that there is an anomaly in the data group. Using the node of interest with the highest prediction deviation rate as the core reference point, by analyzing the correlation between other marker nodes and this node of interest, it is possible to effectively identify the associated data groups that have synchronous abnormal behavior with the core fault, thereby discovering potential fault propagation paths or collaborative abnormal phenomena, which greatly improves the systematicness and comprehensiveness of fault diagnosis.

[0020] Furthermore, by matching temporal fluctuation patterns, node groups with coordinated abnormal behavior can be identified. Since real faults often trigger synchronous abnormal fluctuation patterns in multiple related data groups, by calculating the standard deviation sequence of the prediction deviation rate of the nodes of interest and the marked nodes within a preset time period, and comparing the cosine similarity of these two fluctuation sequences, the coordinated change patterns exhibited by different data groups during the fault development process can be effectively captured. This not only verifies the abnormal correlation between nodes, but also discovers potential fault propagation paths, thereby integrating seemingly scattered anomalies into risk groups with intrinsic correlation, significantly improving the systematicness and accuracy of fault diagnosis.

[0021] Furthermore, a dynamic tolerance connection mechanism is used to accurately construct the knowledge graph network. Strongly correlated nodes should exhibit stable correlation consistency in the data dimension. By calculating the relative deviation between the data correlation and a preset threshold, and setting the correlation deviation threshold as the connection tolerance range, it is ensured that there is a sufficient correlation strength between connected nodes, while allowing the correlation to fluctuate within a reasonable range. This effectively overcomes the connection instability caused by data noise and normal fluctuations, and constructs a graph topology structure that reflects the real data relationship and has good robustness, laying a reliable foundation for subsequent abnormal node identification based on the network structure.

[0022] Furthermore, through a dual mechanism of temporal persistence verification and network influence assessment, it is ensured that the selected abnormal nodes are not only accidental anomalies, but also core fault sources with sustained influence capabilities. The risk determination frequency verifies the persistence of node anomalies from a temporal perspective, eliminating instantaneous interference, while the number of connections at each risk determination moment proves the stability of the node's influence in the knowledge graph from a spatial perspective. This can accurately identify key nodes that are continuously active and always at the center of the associated network throughout the entire fault evolution process, thereby effectively distinguishing core faults from derivative phenomena and significantly improving the accuracy and reliability of root cause diagnosis.

[0023] Furthermore, by calculating the average number of connections, we ensure that the node has sufficient network centrality. At the same time, we verify the stability of this influence by analyzing the fluctuation value of the number of connections. The true root cause node not only needs to be continuously abnormal, but also needs to maintain a stable topological influence in the knowledge graph. The lower the fluctuation value, the more stable the connection between the node and other nodes is, rather than accidental or variable. This can effectively filter out those interfering nodes with unstable connection relationships, ensuring that the abnormal node finally located is a reliable root cause that maintains its core hub status throughout the entire failure period.

[0024] Furthermore, by quantitatively assessing the overall quality and business importance of identified abnormal nodes, and dynamically optimizing screening criteria based on the discovery scale, the system generates a comprehensive deviation index by weighting and fusing the mean prediction deviation and process criticality according to the statistical characteristics and business process attributes of the abnormal node group. This index reflects both the reliability of the current anomaly detection results and the actual impact on production and operations, enabling the system to intelligently determine whether the current threshold setting is too strict or too lenient, and make precise adjustments accordingly. This ensures that the entire intelligent diagnostic system can maintain optimal detection performance and business relevance during continuous operation.

[0025] Furthermore, when the deviation index exceeds the normal range, the system analyzes the synergistic relationship between the quality characteristics and frequency characteristics of the abnormal node group and adopts targeted adjustment strategies accordingly. A deviation index greater than the upper limit indicates that the currently captured abnormal nodes are not only connected by graph edges frequently but also have large prediction biases and high process criticality, strongly suggesting that the system is in an over-alarm state. At this point, the preset interval threshold is increased, raising the threshold for fault event judgment to achieve strategic contraction. Only a more severe decline in output efficiency will trigger subsequent complex analysis processes, directly reducing unnecessary analysis tasks, lowering system load, and fundamentally suppressing oversensitivity. False alarms are generated; if the deviation index is less than the lower limit, it means that the abnormal nodes found by the system are connected by the graph edges too few times or the quality is not high, indicating that the system may have missed important information and is too conservative. At this time, by weighting the comprehensive deviation degree of quality and frequency, an adjustment factor is obtained, and the preset correlation threshold is reduced to lower the screening criteria for temporary nodes. The system allows more data groups with slightly weaker semantic correlation to enter the subsequent analysis, expands the basis of analysis, increases the chance of discovering potential abnormal associations, thereby improving the detection rate of the system, avoiding false alarms, effectively solving the defect of fixed sensitivity in traditional systems, and enabling the entire diagnostic system to have a self-optimizing dynamic balance capability. Attached Figure Description

[0026] Figure 1 This is a flowchart of the intelligent data query and generation method based on large models and knowledge graphs in this embodiment; Figure 2 This is a logic diagram for determining the occurrence of a device malfunction event in this embodiment; Figure 3 This is a logic diagram for selecting temporary nodes in this embodiment; Figure 4 The logic diagram for determining risk nodes in this embodiment is shown. Detailed Implementation

[0027] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0028] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0029] Please see Figure 1 The flowchart shown is a process for the intelligent data query generation method based on large models and knowledge graphs in this embodiment. This embodiment provides an intelligent data query generation method based on large models and knowledge graphs, including: Real-time acquisition of the output counting interval of production equipment in the industrial production database, the prediction deviation rate of each monitoring node in the knowledge graph constructed with each set of data to be measured during the operation of production equipment as monitoring nodes, the process criticality, and the data correlation between any two monitoring nodes. A device malfunction event is determined based on the output counting interval and the preset interval threshold. Based on the equipment failure event, the correlation coefficient of each monitoring node is determined according to the preset large model and each monitoring node; Several temporary nodes are selected based on the correlation coefficients and preset correlation thresholds; Several risk nodes are determined based on the prediction deviation rate of each of the temporary nodes; Connect two risk nodes based on the data correlation between any two risk nodes to obtain several graph edges; Several abnormal nodes are determined based on the risk determination frequency of each risk node within a preset determination time and the number of times each risk node is connected by the graph edge. The preset correlation threshold or the preset interval threshold is adjusted based on the prediction deviation rate, process criticality, and number of occurrences of each abnormal node. A fault association data query report is generated based on the correlation coefficients of all the abnormal nodes obtained after adjusting the preset correlation threshold or the preset interval threshold.

[0030] In this embodiment, the industrial production database is an integrated data storage system used to store and manage real-time and historical data of the entire production process, including but not limited to output counting intervals. Each data set to be measured refers to a collection of various monitored data related to the operation of the production equipment. The output counting interval is the time interval taken for the production equipment to complete the production of one unit of product; it is a core performance indicator for measuring equipment operating efficiency and health status. It is triggered by photoelectric sensors installed at the end of the production line, and the time difference between two consecutive output signals is the output counting interval. The prediction deviation rate measures the degree of deviation between the actual observed value of a data set and its historical values, calculated by acquiring the relative deviation between current data and historical operating data in real time. Process criticality is used to assess the importance and influence level of the equipment, parameter, or link represented by the data set in the entire production process. Based on the topology of a knowledge graph, using PageRank or center centrality isograph algorithms, the centrality score of the data set's nodes in the network is calculated as its process criticality. Data relevance quantifies the strength of a linear or non-linear association between any two data sets in their historical behavior. When constructing a knowledge graph, based on process flow diagrams, equipment connection relationships, and control logic, inherent relationships between data sets are predefined, and the degree of association is assigned to these relationships. Risk determination frequency is the cumulative frequency at which each risk node is determined as a risk node within a preset determination period.

[0031] In this embodiment, when the equipment experiences a continuous decline in output efficiency but the specific cause is unclear, intelligent data query is automatically triggered to quickly locate the core data group reflecting the equipment abnormality, thereby accurately identifying the root cause of the fault, significantly shortening the diagnosis time, and effectively reducing production losses.

[0032] In this embodiment, the fault association data query report uses the correlation coefficient, a unified quantitative indicator calculated by a large model, to classify the importance of all abnormal nodes. This ultimately converges the multidimensional and complex abnormal features to the core dimension of semantic association strength with the fault event. By arranging them in descending order, a clear spectrum from the root cause to the associated phenomena is formed. This allows the generated query report to not only present a list of abnormal data but also reveal the core nature and impact path of each abnormal node in the fault network. This provides maintenance personnel with a basis for decision-making by following the map, significantly improving the efficiency and accuracy of fault diagnosis.

[0033] The pre-trained large-scale model is based on a Transformer architecture-based pre-trained language model. It understands semantic knowledge within the industrial domain and quantifies the potential correlation strength between any monitoring node and specific equipment failure events. This model is built upon a large-scale model trained on a general corpus (such as web pages, books, and encyclopedias), with initial parameters being weights published by this general model. Further pre-training is performed using industrial text corpora comprised of historical work orders, maintenance reports, equipment manuals, and process documents. Simultaneously, a supervised training dataset is constructed, consisting of failure event descriptions, relevant data nodes, and correlation coefficients. These correlation coefficients are labeled by domain experts based on historical cases. During training, a fine-tuning method is employed, allowing the model to learn to generate correlation coefficients for corresponding monitoring nodes based on failure events, ensuring the model's output closely approximates the expert-annotated correlation coefficients. Mean squared error is used as the loss function, ultimately resulting in a large-scale model specifically designed for industrial fault diagnosis. This model retains general language understanding capabilities but is specifically optimized for understanding the correlation between failures and data in industrial scenarios. When the system is running, the model takes the device failure event and the description of each monitoring node as input and passes them to the trained large model. The model will output a value between 0 and 1, which is the correlation coefficient. This coefficient represents the semantic and historical correlation strength between the monitoring node and the current failure event.

[0034] The preset interval threshold is the critical value used to determine whether a device has malfunctioned. It depends on the device's designed capacity, historical operating data, and process standards, and is typically set between 1.5 and 3 times the baseline interval (the baseline interval refers to the standard time interval taken for a unit of product output under normal and stable operating conditions, which depends on the device's theoretical design cycle, historical stable operating data, and process standard requirements; in this embodiment, it is set to 15 seconds). In this embodiment, it is set to twice the baseline interval, which allows for normal production fluctuations while effectively identifying significant efficiency declines, avoiding excessive false alarms due to an overly strict threshold or missed alarms due to an overly lenient threshold.

[0035] The preset relevance threshold is a critical value used to filter data groups related to equipment failure events. It depends on the statistical learning results of the large model on the correlation distribution of normal and failure modes in historical data, and is usually set between 0.5 and 0.8. In this embodiment, it is set to 0.7, which can effectively focus on the core data groups that are strongly correlated with failures, avoid introducing too many weakly correlated nodes to interfere with subsequent analysis, and improve the overall quality of temporary nodes while ensuring recall.

[0036] By acquiring various key data, the system first identifies equipment failure events. Then, a large model understands the semantic correlation of each data group to obtain correlation coefficients. Based on this, relevant temporary nodes are screened out. Subsequently, the prediction deviation rate is combined to quantify the degree of data anomaly and determine risk nodes. Next, the data correlation in the knowledge graph is used to construct the association network of risk nodes, thereby identifying abnormal nodes in key positions in the topology. At the same time, the screening criteria are dynamically optimized through an adaptive threshold mechanism. Finally, a fault association data query report is generated, ensuring that the core data groups reflecting equipment failures can be continuously and accurately located in massive industrial data. This significantly improves operation and maintenance efficiency and diagnostic accuracy, effectively solving the problems of low accuracy and delayed query response caused by over-reliance on static thresholds and changing operating environments.

[0037] Please see Figure 2 As shown, this is a logic diagram for determining the occurrence of a equipment failure event in this embodiment. In this embodiment, the process of determining the occurrence of a equipment failure event based on the output counting interval and the preset interval threshold includes: Calculate the average value of all output counting intervals within the preset counting time to obtain the interval mean; A device malfunction event is determined to have occurred when the average interval is greater than the preset interval threshold.

[0038] The preset counting duration refers to the time length used to calculate the average output counting interval. It depends on the operating cycle of the production equipment and the real-time requirements of fault detection, and is usually set between 1 minute and 24 hours. In this embodiment, it is set to 30 minutes, which can ensure the stability of data statistics while responding promptly to changes in production rhythm and avoiding misjudgments caused by instantaneous fluctuations.

[0039] By monitoring macroscopic performance trends, the system captures fundamental changes in equipment operating status, rather than responding to instantaneous fluctuations. The output counting interval is a direct quantitative indicator of overall equipment efficiency. By calculating the average interval over a preset counting period, random interference in the production process can be effectively filtered out, thereby capturing continuous performance degradation trends. When the average interval exceeds a preset threshold, it indicates that the equipment has transitioned from occasional delays to a systemic efficiency decline, providing an accurate fault trigger signal for subsequent multi-dimensional data correlation analysis. This effectively avoids false triggers caused by instantaneous anomalies, ensuring that the entire intelligent diagnostic process is initiated both timely and reliably.

[0040] Please see Figure 3 As shown, this is the decision logic diagram for filtering temporary nodes in this embodiment. In this embodiment, the process of filtering several temporary nodes according to the correlation coefficients includes: When the correlation coefficient is greater than the preset correlation threshold, the monitoring node is determined to be a temporary node, so as to filter out a number of temporary nodes.

[0041] By pre-setting relevant thresholds to screen correlation coefficients, it is essentially to establish a semantic association field centered on fault events in a multi-dimensional data space. Only nodes with association strength exceeding the critical value are retained as temporary nodes. This can effectively overcome the limitations of traditional keyword matching or rule engines, capture implicit associations that are difficult for humans to discover from the underlying data logic, build a high-quality data foundation for subsequent accurate analysis, and thus significantly improve the accuracy and efficiency of the entire intelligent diagnosis process.

[0042] Specifically, the process of determining a number of risk nodes based on the prediction deviation rate of each of the temporary nodes includes: When the prediction deviation rate is greater than a preset deviation rate threshold, the temporary node is determined to be a marked node, and several marked nodes are obtained; By comparing the prediction deviation rates of all the marked nodes, the marked node with the largest prediction deviation rate is determined as the node of interest; Several risk nodes are determined based on the prediction deviation rate of the nodes of interest and any of the marked nodes within a preset time period.

[0043] The preset deviation rate threshold is a critical value used to determine whether a data set exhibits significant abnormal behavior. It depends on the equipment's operational accuracy requirements and the statistical range of normal fluctuations in historical data, and is typically set between 0.1 and 0.3. In this embodiment, it is set to 0.2, which effectively controls the false alarm rate while ensuring the recall of genuine anomalies, providing a reliable foundation of abnormal data for determining subsequent nodes of interest.

[0044] By comparing the prediction deviation rate with the preset deviation rate threshold, the system can filter out the actual abnormal marker nodes from the temporary nodes obtained from semantic correlation. If the prediction deviation rate is greater than the preset deviation rate threshold, it indicates that the actual state of the data group deviates from its expected normal state by more than the threshold, which means that there is an anomaly in the data group. Using the node of interest with the highest prediction deviation rate as the core reference point, by analyzing the correlation between other marker nodes and this node of interest, it is possible to effectively identify the associated data groups with synchronous abnormal behavior with the core fault, thereby discovering potential fault propagation paths or collaborative abnormal phenomena, which greatly improves the systematicness and comprehensiveness of fault diagnosis.

[0045] Please see Figure 4 As shown, this is the logic diagram for determining risk nodes in this embodiment. In this embodiment, the process of determining several risk nodes based on the prediction deviation rate of the nodes of interest and any of the marked nodes within a preset time period includes: Calculate the standard deviation of all prediction deviation rates of the nodes of interest from the initial time to each time within the preset time period to obtain several first deviation rate fluctuation values; Calculate the standard deviation of all prediction deviation rates of the marked nodes from the initial time to each time within the preset time period to obtain several second deviation rate fluctuation values; The first deviation rate fluctuation values ​​are subjected to maximum-minimum normalization to obtain several first deviation rate normalization values. Max-min normalization is performed on all the second deviation rate fluctuation values ​​to obtain several second deviation rate normalization values; Calculate the cosine similarity between all normalized values ​​of the first deviation rate and all normalized values ​​of the second deviation rate to obtain the deviation pattern matching degree; When the deviation pattern matching degree is greater than a preset matching degree threshold, the node of interest and the marked node are determined as risk nodes, thereby identifying several risk nodes.

[0046] The preset time interval refers to the length of time used to calculate the similarity of the predicted deviation rate fluctuation pattern. It depends on the time characteristics of the equipment failure development and the data acquisition frequency, and is usually set between 10 minutes and 2 hours. In this embodiment, it is set to 30 minutes, which allows for sufficient observation of the evolution trend of abnormal fluctuation patterns to ensure the timeliness and accuracy of pattern matching.

[0047] The preset matching degree threshold is a critical value for judging whether the fluctuation patterns of different nodes are similar. It depends on the analysis results of abnormal propagation patterns in historical failure cases and is usually set between 0.7 and 0.9. In this embodiment, setting it to 0.85 can effectively identify strongly correlated cooperative abnormal nodes, while excluding random fluctuation matching, ensuring the reliability of the inherent correlation of the risk node group.

[0048] By matching temporal fluctuation patterns, we can identify clusters of nodes with coordinated abnormal behavior. Since real faults often trigger synchronous abnormal fluctuation patterns in multiple related data groups, we can effectively capture the coordinated change patterns of different data groups during fault development by calculating the standard deviation sequence of the prediction deviation rate of the nodes of interest and the marked nodes within a preset time period and comparing the cosine similarity of these two fluctuation sequences. This not only verifies the abnormal correlation between nodes but also discovers potential fault propagation paths, thereby integrating seemingly scattered anomalies into risk groups with intrinsic connections, and significantly improving the systematicness and accuracy of fault diagnosis.

[0049] Specifically, the process of connecting two risk nodes based on their data correlation to obtain several graph edges includes: The relative deviation between the data correlation degree and the preset data correlation threshold is calculated to obtain the correlation deviation; When the correlation deviation is less than a preset correlation deviation threshold, two risk nodes are connected to obtain several graph edges.

[0050] The preset data correlation threshold is a benchmark value for judging whether there is a significant data correlation between two monitoring nodes. It depends on the statistical distribution of historical data correlation in the knowledge graph and the requirements of the business scenario for the correlation strength, and is usually set between 0.5 and 0.8. In this embodiment, it is set to 0.6, which can effectively filter out correlations with actual business significance, provide a reliable data foundation for the subsequent construction of graph edges, and at the same time avoid introducing too many weak correlation connections that affect the analysis effect.

[0051] The preset association deviation threshold is a standard value for controlling the accuracy of risk node connections. It depends on the system's requirements for association consistency and the fluctuation characteristics of data correlation, and is usually set between 0.1 and 0.3. In this embodiment, it is set to 0.2, which provides appropriate flexibility while ensuring connection quality, allowing data correlation to fluctuate within a reasonable range, and ensuring that the constructed graph network is both stable and reflects the true association relationships.

[0052] The knowledge graph network is accurately constructed through a dynamic tolerance connection mechanism. Strongly correlated nodes should exhibit stable correlation consistency in the data dimension. By calculating the relative deviation between the data correlation and a preset threshold, and setting the correlation deviation threshold as the connection tolerance range, it is ensured that there is a sufficient correlation strength between connected nodes, while allowing the correlation to fluctuate within a reasonable range. This effectively overcomes the connection instability caused by data noise and normal fluctuations, and constructs a graph topology structure that reflects the real data relationship and has good robustness, laying a reliable foundation for subsequent abnormal node identification based on the network structure.

[0053] Specifically, the process of determining several abnormal nodes based on the risk determination frequency of each risk node within a preset determination time period and the number of times each risk node is connected by the graph edge includes: When the risk determination frequency is greater than a preset frequency threshold, the risk node is determined to be a high-frequency risk node; Based on the number of times the high-frequency risk node is connected by the graph edge at each risk determination time within the preset determination period, a number of abnormal nodes are determined. The risk determination time refers to the moment when each of the risk nodes is determined to be a risk node within the preset determination time.

[0054] The preset judgment duration refers to the length of time used to statistically analyze the frequency and number of connections of risk nodes. It depends on the characteristics of fault evolution and the system's trade-off between stability and real-time performance, and is usually set between 10 minutes and 4 hours. In this embodiment, it is set to 1 hour, which can ensure that continuous abnormal patterns are captured while responding promptly to changes in production status and effectively distinguishing between transient interference and real faults.

[0055] The preset frequency threshold is a standard value for judging whether a risk node is continuously abnormal. It depends on the density of the analysis cycle and the requirement for the persistence of the anomaly, and is usually set between 0.6 and 0.9. In this embodiment, it is set to 0.7, which means that the node is identified as a risk node within 70% of the preset judgment time. This can effectively filter out nodes with continuous abnormal characteristics, avoid misjudging occasional fluctuations as critical anomalies, and ensure that the real source of failure is not missed.

[0056] Through a dual mechanism of time-series persistence verification and network influence assessment, it is ensured that the selected abnormal nodes are not only accidental anomalies, but also core fault sources with sustained influence capabilities. The risk determination frequency verifies the persistence of node anomalies from a time dimension, eliminating instantaneous interference, while the number of connections at each risk determination moment proves the stability of the node's influence in the knowledge graph from a spatial dimension. It can accurately identify key nodes that are continuously active and always at the center of the associated network throughout the entire fault evolution process, thereby effectively distinguishing core faults from derivative phenomena and significantly improving the accuracy and reliability of root cause diagnosis.

[0057] Specifically, the process of determining several abnormal nodes based on the number of times the high-frequency risk node is connected by the graph edge at each risk determination time within the preset determination period includes: Calculate the average number of times the high-frequency risk node is connected by the graph edge at all risk determination moments within the preset determination time period to obtain the average number of times; When the average number of occurrences is greater than a preset average number of occurrences threshold, the standard deviation of the number of times the high-frequency risk node is connected by the graph edge at all risk determination times within the preset determination time is calculated to obtain the number fluctuation value; When the fluctuation value of the number of times is less than the preset fluctuation threshold, the high-frequency risk node is determined to be the abnormal node, so as to identify a number of abnormal nodes.

[0058] The preset average threshold is the minimum number of graph connections required to identify key anomalous nodes. It depends on the network density of the knowledge graph and the requirements for identifying core nodes in fault diagnosis, and is typically set between 2 and 5. In this embodiment, it is set to 3, which effectively filters out core anomalous nodes with significant influence in the network, avoids selecting overly isolated anomalous points, and ensures that the final identified anomalous nodes are systematic and representative.

[0059] The preset fluctuation threshold is a standard value for judging the stability of the number of connections of high-frequency risk nodes in the network graph. It depends on the system's requirements for the stability of node associations and the dynamic characteristics of the network structure, and is usually set between 0.1 and 0.3. In this embodiment, it is set to 0.15, which can effectively identify core nodes that maintain stable influence in the network and ensure that the finally determined abnormal nodes have reliable structural importance.

[0060] By calculating the average number of connections, we ensure that the node has sufficient network centrality. At the same time, we verify the stability of this influence by analyzing the fluctuation value of the number of connections. The true root cause node not only needs to be continuously abnormal, but also needs to maintain a stable topological influence in the knowledge graph. The lower the fluctuation value, the more stable the connection between the node and other nodes is, rather than accidental or variable. This can effectively filter out those interfering nodes with unstable connection relationships, ensuring that the abnormal node finally located is a reliable root cause that maintains its core hub status throughout the entire failure.

[0061] Specifically, the process of adjusting the preset relevance threshold or the preset interval threshold based on the prediction deviation rate of all the abnormal nodes, the process criticality, and the number of occurrences includes: Calculate the average of all the prediction deviation rates to obtain the mean prediction deviation; The relative deviation between the mean prediction deviation and the preset deviation rate threshold is calculated to obtain the prediction relative deviation; The deviation index is obtained by weighting and summing the prediction relative deviation, the preset prediction relative deviation weight, the process criticality, and the preset criticality weight. The preset relevant threshold or the preset interval threshold is adjusted based on the deviation index, the preset deviation index range, and the number of times.

[0062] The preset prediction relative deviation weight is a weighting coefficient used to measure the importance of prediction relative deviation in the deviation index calculation. It depends on the system's trade-off strategy between the degree of data anomaly and the importance of business processes, and is usually set between 0.4 and 0.7. In this embodiment, it is set to 0.6, which allows the system to pay more attention to the degree of anomaly of the data itself, while maintaining a reasonable balance with the criticality of the process in the deviation index calculation.

[0063] The preset criticality weight is a weighting coefficient used to measure the importance of process criticality in the deviation index calculation. It depends on the system's emphasis on the importance of business processes relative to the degree of data anomalies, and is usually set between 0.3 and 0.6. In this embodiment, it is set to 0.4, which can ensure the sensitivity of data anomaly detection while fully considering the criticality of business processes, making the system pay more attention to abnormal nodes that have a greater impact on the production process.

[0064] By quantitatively assessing the overall quality and business importance of identified abnormal nodes, and dynamically optimizing screening criteria based on the discovery scale, the system generates a comprehensive deviation index by weighting and fusing the mean prediction deviation and process criticality according to the statistical characteristics and business process attributes of the abnormal node group. This index reflects both the reliability of the current anomaly detection results and their actual impact on production and operations. As a result, the system can intelligently determine whether the current threshold setting is too strict or too lenient, and make precise adjustments accordingly. This ensures that the entire intelligent diagnostic system can maintain optimal detection performance and business relevance during continuous operation.

[0065] Specifically, the process of adjusting the preset correlation threshold or the preset interval threshold based on the deviation index, the preset deviation index range, and the number of occurrences includes: When the deviation index is greater than the maximum value of the preset deviation index range, the preset interval threshold is increased according to the relative deviation between the deviation index and the maximum value of the preset deviation index range and the preset adjustment coefficient, where T'=T×(1+a×︱Y-Y'︱ / Y'), T' is the adjustment interval threshold, T is the preset interval threshold, a is the preset adjustment coefficient, Y is the deviation index, and Y' is the maximum value of the preset deviation index range; When the deviation index is less than the minimum value of the preset deviation index range, the relative deviation between the deviation index and the minimum value of the preset deviation index range is calculated to obtain the index deviation. Calculate the absolute value of the relative deviation between the number of occurrences and the preset average threshold number of occurrences to obtain the number of occurrences deviation; The adjustment factor is obtained by weighting and summing the frequency deviation, the preset frequency deviation weight, the exponential deviation, and the preset exponential deviation weight. The preset correlation threshold is reduced according to the adjustment factor and the preset adjustment coefficient, where M' = M × (1 - a × F), M' is the adjusted correlation threshold, M is the preset correlation threshold, a is the preset adjustment coefficient, and F is the adjustment factor.

[0066] The preset deviation index range is a benchmark interval for judging whether the system is operating normally. It depends on the statistical distribution of the system's normal state in historical operating data and the stability requirements, and is usually set between [0.5, 0.9]. In this embodiment, it is set to [0.6, 0.8], which can provide a clear boundary standard for the system's adaptive adjustment. When the deviation index exceeds this range, threshold optimization is triggered in time to ensure that the system continues to be in the optimal working state.

[0067] The preset adjustment coefficient is a proportional factor that controls the adjustment range of the interval threshold. It depends on the system's requirements for the threshold change rate and stability considerations, and is usually set between 0.1 and 0.3. In this embodiment, it is set to 0.2, which can ensure the smoothness and controllability of the threshold adjustment and avoid drastic fluctuations in system performance due to excessively large single adjustment ranges.

[0068] The preset frequency deviation weight is a weighted parameter that quantifies the importance of frequency deviation in the adjustment factor calculation. It depends on the system's sensitivity requirements to changes in the number of abnormal nodes and is usually set between 0.3 and 0.6. In this embodiment, it is set to 0.4, which can fully consider the impact of the size of abnormal nodes during the adjustment process and avoid adjustment deviations caused by relying solely on quality indicators.

[0069] The preset index deviation weight is a weighting parameter that measures the weight of index deviation in the adjustment factor calculation. It depends on the system's priority in considering the degree of deviation of the index and is usually set between 0.4 and 0.7. In this embodiment, it is set to 0.6 to ensure that the adjustment factor more accurately reflects the quality status of the system operation, making the threshold adjustment more precise and effective.

[0070] The preset adjustment coefficient is a proportional factor that controls the fineness of the relevant threshold adjustment, depending on the system's requirements for the accuracy of threshold changes and network stability. It is usually set between 0.1 and 0.3. In this embodiment, it is set to 0.1, which enables fine-tuning control of the threshold, ensures the stability of the graph network structure, and avoids large fluctuations in detection results due to sudden threshold changes.

[0071] When the deviation index exceeds the normal range, the system analyzes the synergistic relationship between the quality and frequency characteristics of the abnormal node group and adopts targeted adjustment strategies. A deviation index greater than the upper limit means that the currently captured abnormal nodes are not only connected by graph edges frequently, but also have large prediction biases and high process criticality, strongly suggesting that the system is in an over-alarm state. At this point, the preset interval threshold is increased, raising the threshold for fault event judgment to achieve strategic contraction. Only a more severe decline in output efficiency will trigger subsequent complex analysis processes, directly reducing unnecessary analysis tasks, lowering system load, and fundamentally suppressing the problems caused by oversensitivity. False alarms are reduced; if the deviation index is less than the lower limit, it means that the abnormal nodes found by the system are connected by graph edges too few times or of low quality, indicating that the system may have missed important information and is too conservative. At this time, by weighting the comprehensive deviation degree of quality and frequency, an adjustment factor is obtained, and the preset correlation threshold is reduced to lower the screening criteria for temporary nodes. The system allows more data groups with slightly weaker semantic correlation to enter the subsequent analysis, expands the basis of analysis, increases the chance of discovering potential abnormal associations, thereby improving the detection rate of the system, avoiding false alarms, effectively solving the defect of fixed sensitivity in traditional systems, and enabling the entire diagnostic system to have a self-optimizing dynamic balance capability.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent data query and generation based on large models and knowledge graphs, characterized in that, include: Real-time acquisition of the output counting interval of production equipment in the industrial production database, the prediction deviation rate of each monitoring node in the knowledge graph constructed with each set of data to be measured during the operation of production equipment as monitoring nodes, the process criticality, and the data correlation between any two monitoring nodes. A device malfunction event is determined based on the output counting interval and the preset interval threshold. Based on the equipment failure event, the correlation coefficient of each monitoring node is determined according to the preset large model and each monitoring node; Several temporary nodes are selected based on the correlation coefficients and preset correlation thresholds; Several risk nodes are determined based on the prediction deviation rate of each of the temporary nodes; Connect two risk nodes based on the data correlation between any two risk nodes to obtain several graph edges; Several abnormal nodes are determined based on the risk determination frequency of each risk node within a preset determination time and the number of times each risk node is connected by the graph edge. The preset correlation threshold or the preset interval threshold is adjusted based on the prediction deviation rate, process criticality, and number of occurrences of each abnormal node. A fault association data query report is generated based on the correlation coefficients of all the abnormal nodes obtained after adjusting the preset correlation threshold or the preset interval threshold.

2. The intelligent data query and generation method based on large models and knowledge graphs according to claim 1, characterized in that, The process of selecting several temporary nodes based on the aforementioned correlation coefficients includes: Based on the comparison between the correlation coefficient and the preset correlation threshold, the monitoring node is determined to be a temporary node, so as to filter out a number of temporary nodes.

3. The intelligent data query and generation method based on large models and knowledge graphs according to claim 2, characterized in that, The process of determining a number of risk nodes based on the prediction deviation rate of each of the aforementioned temporary nodes includes: Based on the comparison between the predicted deviation rate and the preset deviation rate threshold, the temporary node is determined to be a marked node, and several marked nodes are obtained; By comparing the prediction deviation rates of all the marked nodes, the marked node with the largest prediction deviation rate is determined as the node of interest; Several risk nodes are determined based on the prediction deviation rate of the nodes of interest and any of the marked nodes within a preset time period.

4. The intelligent data query and generation method based on large models and knowledge graphs according to claim 3, characterized in that, The process of determining several risk nodes based on the prediction deviation rate between the nodes of interest and any of the marked nodes within a preset time period includes: Calculate several first deviation rate fluctuation values ​​based on the prediction deviation rate of the node of interest within the preset time period; Calculate several second deviation rate fluctuation values ​​based on the prediction deviation rate of the marked nodes within the preset time period; The deviation pattern matching degree is calculated based on all the first deviation rate fluctuation values ​​and all the second deviation rate fluctuation values; Based on the comparison result between the deviation pattern matching degree and the preset matching degree threshold, the nodes of interest and the marked nodes are determined as risk nodes, thereby identifying several risk nodes.

5. The intelligent data query and generation method based on large models and knowledge graphs according to claim 4, characterized in that, The process of connecting two risk nodes based on their data correlation to obtain several graph edges includes: The correlation deviation is calculated based on the data correlation degree and the preset data correlation threshold. Based on the comparison between the correlation deviation and the preset correlation deviation threshold, two risk nodes are connected to obtain several graph edges.

6. The intelligent data query and generation method based on large models and knowledge graphs according to claim 5, characterized in that, The process of determining several abnormal nodes based on the risk determination frequency of each risk node within a preset determination time period and the number of times each risk node is connected by the graph edge includes: Based on the comparison between the risk determination frequency and the preset frequency threshold, the risk node is determined to be a high-frequency risk node; Based on the number of times the high-frequency risk node is connected by the graph edge at each risk determination time within the preset determination period, a number of abnormal nodes are determined. The risk determination time refers to the moment when each of the risk nodes is determined to be a risk node within the preset determination time.

7. The intelligent data query and generation method based on large models and knowledge graphs according to claim 6, characterized in that, The process of determining several abnormal nodes based on the number of times the high-frequency risk node is connected by the graph edge at each risk determination time within the preset determination period includes: The average number of times a high-frequency risk node is connected by a graph edge is calculated based on the number of times the risk is determined at all risk determination moments within the preset determination time period. Based on the comparison between the mean frequency and the preset mean frequency threshold, the frequency fluctuation value is calculated according to the number of times the high-frequency risk node is connected by the graph edge at all the risk determination times within the preset determination time. Based on the comparison between the frequency fluctuation value and the preset frequency fluctuation threshold, the high-frequency risk node is determined as the abnormal node, thereby identifying several abnormal nodes.

8. The intelligent data query and generation method based on large models and knowledge graphs according to claim 7, characterized in that, The process of adjusting the preset relevance threshold or the preset interval threshold based on the prediction deviation rate, process criticality, and number of occurrences of all the abnormal nodes includes: Calculate the mean prediction deviation based on all the prediction deviation rates described; The relative prediction deviation is calculated based on the mean prediction deviation and the preset deviation rate threshold. The deviation index is calculated based on the predicted relative deviation, the preset predicted relative deviation weight, the process criticality, and the preset criticality weight. The preset relevant threshold or the preset interval threshold is adjusted based on the deviation index, the preset deviation index range, and the number of times.

9. The intelligent data query and generation method based on large models and knowledge graphs according to claim 8, characterized in that, The process of adjusting the preset correlation threshold or the preset interval threshold based on the deviation index, the preset deviation index range, and the number of times includes: Based on the comparison between the deviation index and the maximum value of the preset deviation index range, the preset relevant threshold is adjusted according to the deviation index and the maximum value of the preset deviation index range. Based on the comparison between the deviation index and the minimum value of the preset deviation index range, the index deviation is calculated according to the deviation index and the minimum value of the preset deviation index range. The deviation of the number of times is calculated based on the number of times and the preset average threshold of the number of times. The adjustment factor is calculated based on the frequency deviation, the preset frequency deviation weight, the exponential deviation, and the preset exponential deviation weight; The preset interval threshold is adjusted according to the adjustment factor.

10. The intelligent data query and generation method based on large models and knowledge graphs according to claim 9, characterized in that, The process of determining the occurrence of a device failure event based on the output counting interval and the preset interval threshold includes: Calculate the interval average based on all output count intervals within the preset counting period; A device malfunction event is determined based on the comparison between the average interval and the preset interval threshold.

Images