Alarm event rule mining method for intelligent manufacturing industry

Abstract

The embodiment of the invention provides an alarm event rule mining method for an intelligent manufacturing industry, which is applied to intelligent production of petroleum refining and comprises the following steps: acquiring multi-source heterogeneous operation data of equipment in an intelligent production workshop of petroleum refining; normalizing the multi-source heterogeneous operation data to form regular mining basic data; obtaining a petroleum refining intelligent production process description file, and analyzing the file to obtain a process sequence and association relationship feature; based on the technological process sorting and association relationship characteristics, constructing an association topological structure between the petroleum refining intelligent production equipment, and adding a production influence weight on an association topological result to generate an equipment association topological structure with the weight; and generating an executable alarm rule base according to the device association topological structure with the weight.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a method for mining alarm event rules in the intelligent manufacturing industry. Background Art

[0002] In the field of intelligent manufacturing, oil refining and petrochemical production processes are being upgraded to intelligent systems based on technologies such as big data and the Internet of Things. Real-time collection and analysis of equipment operating data has become crucial for ensuring production safety. However, oil refining workshops feature a wide variety of equipment and complex processes, and the amount of heterogeneous data generated by sensors, control systems, and business systems is exploding. Extracting effective rules related to production safety from this massive amount of data has become a core challenge in improving early warning capabilities for alarm events.

[0003] Existing alarm rule mining solutions are typically based solely on statistical analysis of historical alarm data, identifying frequently occurring alarm patterns through association rule algorithms. While these methods can discover co-occurrence relationships at the data level, they lack deep integration with the characteristics of petroleum refining processes and are unable to effectively utilize key information such as the physical connections between equipment and process dependencies. For example, traditional solutions struggle to accurately characterize the spatial relationship between the heating furnace and reactor, material flow constraints, and fault propagation paths. As a result, the mined rules may contain logical gaps, failing to truly reflect the cause-and-effect relationships within the production process. This results in low alarm accuracy and reliability under complex operating conditions.

[0004] Furthermore, existing technologies lack the ability to dynamically adapt to production baselines, making it difficult to adjust rule thresholds based on different operating conditions (e.g., normal production, start-up and shutdown phases). Furthermore, they lack a scientific evaluation system for rule priorities, resulting in high-value rules being buried among a large number of low-value rules, making it impossible to provide accurate decision support for safe production. Therefore, an alarm rule mining method that can integrate process topology features with multi-source data features is urgently needed to improve the prediction accuracy and response efficiency of alarm events in intelligent production in petroleum refining. Summary of the Invention

[0005] Based on the above problems, an embodiment of the present application provides an alarm event rule mining method for the intelligent manufacturing industry to solve the problems existing in the above-mentioned prior art.

[0006] The present application embodiment provides a method for mining alarm event rules in the intelligent manufacturing industry, which is applied to intelligent production in petroleum refining and petrochemicals, and includes:

[0007] Acquire multi-source heterogeneous operating data of equipment in intelligent production workshops of petroleum refining;

[0008] Normalize multi-source heterogeneous operation data to form basic data for rule mining;

[0009] Obtain the petroleum refining intelligent production process description file and parse it to obtain the process flow sorting and correlation relationship characteristics;

[0010] Based on the process sequencing and association relationship characteristics, the association topology structure between the intelligent production equipment of petroleum refining is constructed, and the production impact weight is added to the association topology result to generate a weighted equipment association topology structure;

[0011] Generate an executable alarm rule base based on the weighted device association topology.

[0012] This application provides a method for mining alarm event rules in the intelligent manufacturing industry, which has the following technical advantages:

[0013] 1. This application obtains and parses the intelligent production process description files for petroleum refining to derive process flow ranking and correlation features. This step extracts the complex process characteristics of petroleum refining in a digitized form, overcoming the shortcomings of traditional solutions that rely solely on statistical analysis of historical alarm data and lack deep integration of process characteristics.

[0014] In this application, the association topology between the intelligent production equipment of petroleum refining is constructed based on the process flow sorting and association relationship characteristics, and the production impact weight is added to it. This makes the association between the equipment no longer a simple co-occurrence relationship at the data level, but integrates process characteristics, such as the physical connection relationship between the equipment, process dependency, etc. For example, it can accurately characterize the spatial position association, material flow constraints and fault propagation path between the heating furnace and the reactor, so that the rules mined are more in line with the cause and effect relationship in the production process, avoiding logical faults and improving the accuracy and reliability of alarms under complex working conditions.

[0015] 2. In this application, the weighted device association topology takes into account the importance and impact range of different devices in the production process, which provides a basis for adjusting the rule threshold according to different working conditions. Under different working conditions (such as normal production and start-up and shutdown stages), the threshold of the alarm rule can be dynamically adjusted according to the weight of each device in the device association topology. For example, during the start-up and shutdown stages, the weight of certain key devices may change, and their alarm rule thresholds should be adjusted accordingly, so that the rules can better adapt to different working conditions and improve the adaptability of the alarm rules.

[0016] 3. In this application, production impact weights are added to the associated topology results. These weights can reflect the importance of the equipment in the production process and the impact on the overall production. When constructing alarm event rule mining labels based on weighted device association topology structures, the weight information of the equipment can be incorporated into the rules. In this way, among the many alarm event rules mined, the priority of the rules can be scientifically evaluated based on the weights of the equipment involved in the rules. The priority of rules corresponding to devices with high weights will also be increased accordingly, avoiding high-value rules from being submerged in a large number of low-value rules, and providing accurate decision support for safe production.

[0017] 4. This application uses multi-source heterogeneous operational data, normalizes it, and constructs a weighted equipment association topology based on process flow sequencing and association characteristics. This method integrates multi-source data and process characteristics during the production process. This makes the mined alarm event rules more comprehensive and accurate, better reflecting the actual production process, and thus improving the accuracy of alarm event prediction.

[0018] Alarm event rule mining tags, built based on weighted device association topology, enable scientific evaluation and priority sorting of rules, providing accurate decision support for safe production. When an alarm event occurs, a rapid response can be made based on the priority of the rules, improving response efficiency, reducing the occurrence of production accidents, and ensuring the safe and stable operation of intelligent production in petroleum refining. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0020] Figure 1 This is an embodiment of the present application, a method for mining alarm event rules in the intelligent manufacturing industry. DETAILED DESCRIPTION

[0021] The implementation of any technical solution in the embodiments of the present application does not necessarily require achieving all of the above advantages at the same time.

[0022] like Figure 1 As shown, the embodiment of the present application provides a method for mining alarm event rules in the intelligent manufacturing industry, which is applied to intelligent production in petroleum refining, and includes:

[0023] Acquire multi-source heterogeneous operating data of equipment in intelligent production workshops of petroleum refining;

[0024] Normalize multi-source heterogeneous operation data to form basic data for rule mining;

[0025] Obtain the petroleum refining intelligent production process description file and parse it to obtain the process flow sorting and correlation relationship characteristics;

[0026] Based on the process sequencing and association relationship characteristics, the association topology structure between the intelligent production equipment of petroleum refining is constructed, and the production impact weight is added to the association topology result to generate a weighted equipment association topology structure;

[0027] Generate an executable alarm rule base based on the weighted device association topology.

[0028] Optionally, obtain multi-source heterogeneous operating data of equipment in the petroleum refining intelligent production workshop, including:

[0029] Based on the deployed edge data collection nodes, the operating parameter data generated by the equipment operation in the petroleum refining intelligent production workshop and the log data of the control system are collected from different data sources;

[0030] Based on the deployed centralized data collection platform, the adaptive communication protocol adapter component is called to build a data sampling channel to sample operation parameter data and log data from the edge data collection node and generate multi-source heterogeneous operation data based on this.

[0031] Optionally, based on the deployed centralized data collection platform, the adaptive communication protocol adapter component is called to build a data sampling channel to sample operating parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operating data based on the data, including:

[0032] Send a preset general detection data packet to the node edge data collection node to receive the response data packet returned by the node and extract the protocol characterization features;

[0033] Match the extracted protocol characterization features with the constructed communication protocol multi-dimensional feature library to determine the matching protocol type;

[0034] Extracting protocol features of the matching protocol type to generate protocol conversion rules. Protocol features include data frame format, data encoding method, command set, and communication parameters.

[0035] A data sampling channel is constructed based on protocol conversion rules to sample operation parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operation data accordingly.

[0036] In summary, the above solution for acquiring multi-source heterogeneous operating data of equipment in the intelligent production workshop of petroleum refining has the following technical advantages:

[0037] 1. This application collects equipment operating parameter data and control system log data from different types of data sources by deploying edge data collection nodes. In the intelligent production workshop of petroleum refining, there are various types of equipment, and different equipment may use different data collection methods and interfaces. Different types of data sources can cover information at all levels during the operation of the equipment. For example, the real-time operating parameters of the equipment (temperature, pressure, flow, etc.) reflect the physical state of the equipment, while the control system log data records the control instructions, abnormal events and other information during the operation of the equipment. This multi-source data collection method ensures that the full picture of the equipment operation can be obtained, providing a rich data foundation for subsequent rule mining, and avoiding inaccurate judgments on the equipment operation status and alarm events due to missing data.

[0038] This application further ensures data integrity by integrating edge data collection nodes and a centralized data collection platform. Edge data collection nodes can quickly acquire data close to the data source, reducing the risk of data transmission delays and loss. The centralized data collection platform can centrally manage and integrate edge-collected data to ensure that all important data is collected. For example, during the operation of some critical equipment, some short-lived abnormal parameters may be lost during transmission due to network fluctuations, but this multi-level collection architecture can maximize the guarantee that this data is fully recorded.

[0039] 2. This application calls the adaptive communication protocol adapter component to build a data sampling channel. This design enables the system to flexibly respond to various communication protocols that may exist in the intelligent production workshop of petroleum refining. Different devices may use different communication protocols for data transmission. The traditional fixed protocol acquisition method requires the development of a specific acquisition module for each protocol, which is costly and has poor scalability. The adaptive communication protocol adapter component interacts with the edge data acquisition node and extracts protocol characterization features through a preset general detection data packet, and then matches it with the constructed communication protocol multi-dimensional feature library to automatically identify the protocol type used by the device. This method greatly improves the compatibility of the system. No matter what protocol equipment is added to the workshop, the system can quickly adapt and collect data without the need for large-scale system modifications.

[0040] In this application, after determining the matching protocol type, the protocol features are extracted to generate protocol conversion rules, and data sampling channels are constructed based on these rules. This enables the system to convert data of different protocols into a unified format for transmission and processing, avoiding data processing difficulties caused by protocol differences. At the same time, the data sampling channel constructed based on the protocol conversion rules can be optimized for specific protocols to improve the efficiency of data sampling. For example, for some devices with large data volumes and high transmission frequencies, the sampling frequency and data packet size can be optimized through protocol conversion rules, reducing network bandwidth usage while ensuring data integrity, and improving the performance of the entire data acquisition system.

[0041] 3. This application uses an adaptive communication protocol adapter component to avoid developing separate acquisition modules for each communication protocol, reducing hardware investment and software development workload. In intelligent production workshops of petroleum refineries, there are numerous devices and complex communication protocols. If traditional acquisition methods were used, it would be necessary to purchase corresponding acquisition equipment for each protocol and devote a significant amount of manpower to software development and debugging. The introduction of the adaptive communication protocol adapter component greatly simplifies the architecture of the data acquisition system and reduces hardware procurement and software development costs.

[0042] In this application, since the adaptive communication protocol adapter component can automatically identify and process multiple communication protocols, when equipment is replaced or the protocol is upgraded in the workshop, there is no need to make large-scale modifications and debugging to the data acquisition system. Maintenance personnel only need to focus on updating the protocol feature library and configuring the components, which greatly simplifies maintenance work. For example, when a device is replaced with a model that uses a new communication protocol, it is only necessary to add the features of the new protocol to the feature library and perform simple configuration of the components. The system can then automatically adapt to the new protocol for data collection, reducing system downtime and maintenance costs caused by equipment replacement or protocol upgrades.

[0043] 4. This application accurately determines the protocol type used by a device by precisely matching the extracted protocol characterization features with a constructed multi-dimensional feature library of communication protocols. This precise protocol matching ensures the correct parsing and transmission of data during the acquisition process, avoiding issues such as data garbled or lost due to protocol identification errors, thereby improving data quality. High-quality data is the foundation for subsequent rule mining and alarm event analysis, ensuring the accuracy and reliability of mining results.

[0044] Furthermore, by generating multi-source, heterogeneous operational data in a unified format based on protocol conversion rules, data from different devices and protocols can be processed and analyzed on the same platform. This significantly improves data availability and facilitates subsequent data storage, querying, mining, and application. For example, when mining alarm event rules, unified data formats facilitate feature extraction and correlation analysis, helping to quickly and accurately identify potential issues and alarm rules in device operation.

[0045] Alternatively, in a specific scenario, based on the deployed centralized data collection platform, the adaptive communication protocol adapter component is called to build a data sampling channel to sample operating parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operating data based on this, including:

[0046] Conduct multi-level business semantic mining on industrial communication protocols used in intelligent production in petroleum refining to build a semantic model at the communication protocol layer;

[0047] Based on the constructed communication protocol layer semantic model, a semantic rule base for different industrial communication protocols is generated to construct a semantic association graph of protocol semantic elements;

[0048] Generate protocol conversion rules based on the semantic association graph of protocol semantic elements;

[0049] A data sampling channel is constructed based on protocol conversion rules to sample operation parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operation data accordingly.

[0050] To this end, the above solution has the following technical advantages:

[0051] Considering that in the intelligent production scenario of petroleum refining, industrial communication protocols carry key information about equipment operation. Multi-level business semantic mining can deeply analyze the protocol, not just staying at the format level of the data frame, but understanding the specific meaning of each field in the actual production business. For example, a certain protocol field may correspond to a specific operating state of the equipment (such as high temperature alarm, pressure abnormality, etc.), and the communication protocol layer semantic model constructed through semantic mining can accurately identify these meanings. This makes the operating parameter data and log data obtained from the edge data collection node more accurate when parsing, avoids data misinterpretation due to insufficient understanding of the protocol semantics, and provides a reliable data foundation for subsequent alarm event rule mining.

[0052] Furthermore, industrial communication protocols in the petroleum refining industry are often complex and diverse, encompassing numerous business rules and special agreements. Multi-level semantic mining can comprehensively cover these complexities and build a comprehensive semantic model. For example, some protocols may contain nested data structures or implicit business logic relationships. Deeper mining can accurately parse these complexities, ensuring that all data is fully and correctly interpreted, without missing any critical information.

[0053] Furthermore, the semantic rule base generated based on the semantic model provides clear specifications for conversions between different industrial communication protocols. For example, when converting data from one protocol to another, the rules in the rule base provide guidance on how to accurately map the semantic elements of the source protocol to the target protocol. Furthermore, the constructed semantic association diagram of protocol semantic elements clearly demonstrates the relationships between semantic elements of different protocols, facilitating the generation of more rational protocol conversion rules. This semantic-based conversion approach is more intelligent and flexible than traditional format-based conversion, better adapting to differences between protocols and improving the accuracy and efficiency of protocol conversion.

[0054] Furthermore, thanks to its in-depth semantic understanding and conversion capabilities for various industrial communication protocols, this solution significantly enhances the centralized data collection platform's compatibility with diverse devices. Regardless of the type of new equipment using different protocols added to the workshop, the system can quickly adapt and collect data using semantic models and conversion rules. This reduces the difficulty and cost of integrating devices into the system, eliminating the need for enterprises to replace or significantly renovate equipment to adapt to the system, thereby improving the flexibility and scalability of production systems.

[0055] Furthermore, with the continuous advancement of technology and changes in production requirements, smart oil refining workshops may introduce new equipment or adopt new communication protocols. Because this solution already has a comprehensive semantic model and rule base, for new protocols, simply perform multi-level business semantic mining, update the semantic rule base and semantic association graph, and generate the corresponding protocol conversion rules. This enables the system to quickly adapt to protocol changes without requiring large-scale redevelopment and redeployment of the entire data acquisition system, greatly improving system scalability.

[0056] When protocols change or new protocols require support, maintenance personnel only need to update the semantic model and rule base without having to make large-scale modifications to the underlying code of the data collection platform. This simplifies system maintenance, reducing maintenance workload and time. Furthermore, due to the improved system stability and reliability, system failures and data errors caused by protocol issues are reduced, further reducing maintenance costs.

[0057] In this application, the data sampling channel constructed based on accurate protocol conversion rules can sample high-quality multi-source heterogeneous operation data from edge data collection nodes. These data are semantically consistent and can accurately reflect the operating status and business information of the equipment. In the process of mining alarm event rules, high-quality data can provide richer features and more accurate correlations, which helps to mine more accurate and valuable alarm event rules. For example, by analyzing a large amount of accurate multi-source heterogeneous data, the potential laws and abnormal patterns between the equipment operating parameters can be discovered, thereby predicting and warning possible alarm events in advance.

[0058] Finally, the multi-source heterogeneous operational data generated in this application can not only be used for alarm event rule mining, but also support a variety of other production management application scenarios. For example, by analyzing equipment operation data, production processes can be optimized and equipment utilization can be improved; by mining log data, potential problems and safety hazards in the system can be discovered. The high-quality data provided by this solution provides strong support for these application scenarios, helping to improve the management efficiency and decision-making level of the entire petroleum refining intelligent production.

[0059] Optionally, normalize the multi-source heterogeneous operation data to form basic data for rule mining, including:

[0060] Perform spatiotemporal alignment of multi-source heterogeneous operation data and correlate the spatial positions of corresponding equipment in the production process to build a process correlation map;

[0061] Based on the historical production process data of petroleum refining, the process correlation map is calibrated to extract the statistical characteristics, time series characteristics and correlation characteristics of multi-source heterogeneous operation data;

[0062] Generate rules to mine basic data based on statistical features, time series features and association features.

[0063] Optionally, based on historical production process data of petroleum refining, the process correlation map is calibrated to extract statistical features, time series features, and correlation features of multi-source heterogeneous operation data, including:

[0064] Based on the source equipment, collection time, and process flow relationship of the historical production process data of petroleum refining, the corresponding equipment nodes and process relationship edges are determined to establish the association index between the generated process data and the graph, so as to calibrate the process association graph and obtain the association topology index of the refining process data;

[0065] Based on the equipment node attribute values ​​in the topological index of the refining process data, descriptive statistical calculations are performed on the multi-source heterogeneous operation data to extract the statistical characteristics of the multi-source heterogeneous operation data;

[0066] Based on the time series edge weights in the topological index of the refining process data, the state transition evaluation is performed on the multi-source heterogeneous operation data to extract the time series characteristics of the multi-source heterogeneous operation data;

[0067] Based on the weighted directed edges in the association topology index of refining process data, causal association mining is performed on multi-source heterogeneous operation data to extract the association features of multi-source heterogeneous operation data.

[0068] Optionally, based on statistical features, time series features, and association features, basic data for rule mining is generated, including:

[0069] Encode statistical features, temporal features and correlation features and fuse them to obtain a fused feature vector;

[0070] The fused feature vector is normalized by topological centrality, causal strength, and association rule confidence to generate basic data for rule mining.

[0071] To this end, the above solution has the following technical advantages:

[0072] In this application, the time-space alignment of multi-source heterogeneous operation data and the spatial position association of corresponding equipment in the production process can eliminate the differences in time and space between different data sources. In the intelligent production of petroleum refining, the data collected by different equipment may have deviations in time or lack correlation in spatial position, which will cause the data to be unable to accurately reflect the actual production situation. By constructing a process association map, the data is organized in time and space order, making the association between the data clearer, and providing an accurate data basis for subsequent feature extraction and rule mining. For example, when analyzing the anomalies of a certain production link, the process association map can be used to quickly locate the operation data of the relevant equipment at a specific time point, thereby improving the accuracy of problem troubleshooting.

[0073] Furthermore, based on the source equipment, collection time, and process relationships of historical oil refining production process data, we identify equipment nodes and process relationship edges, establish an association index between the generated process data and the graph, and calibrate the process association graph to obtain an association topology index for the refining process data. This calibration method accurately maps historical data to nodes and edges in the process association graph, ensuring that the graph truly reflects the actual production process. For example, when analyzing data changes for a particular piece of equipment at different process stages, the association topology index can quickly locate relevant historical data, providing strong support for data analysis and rule mining.

[0074] Furthermore, based on the associative topological index of refining process data, multi-source heterogeneous operating data is extracted from three dimensions: statistical features, time series features, and correlation features. Statistical features, such as mean and variance, reflect the overall distribution of data, helping to understand the stability and fluctuation range of equipment operation. Time series features capture the patterns of data change over time, such as trends and periodicity, and are important for predicting future equipment operating status and alarm events. Correlation features reveal the causal relationships and mutual influences between different equipment and data, helping to identify potential problems and optimization points in the production process. By extracting features from these three dimensions, the characteristics of multi-source heterogeneous operating data can be comprehensively characterized, providing rich feature information for rule mining.

[0075] Furthermore, descriptive statistical calculations based on device node attribute values ​​are used to extract statistical features, enabling rapid and accurate data statistical information. State transition assessments based on time series edge weights are used to extract time series features, effectively capturing temporal trends in data. Causal association mining based on weighted directed edges is used to extract correlation features, enabling in-depth analysis of causal relationships between data. These targeted computational methods improve the efficiency and accuracy of feature extraction, enabling the extracted features to better reflect the actual characteristics of the data.

[0076] Furthermore, statistical, temporal, and correlation features are encoded and fused to create a fused feature vector, integrating features from different dimensions to form a comprehensive feature representation. This fused feature vector more comprehensively reflects the characteristics of the data, providing richer information for rule mining. For example, when mining alarm event rules, the fused feature vector can comprehensively consider the device's operating status, temporal trends, and relationships with other devices, thereby improving the accuracy and effectiveness of rule mining.

[0077] In addition, normalizing the fused feature vectors by topological centrality, causal strength, and association rule confidence can eliminate dimensional differences and inconsistent influence weights between different features. This normalization ensures that all features have the same weight and comparability during the rule mining process, improving the quality and stability of rule mining. For example, when calculating association rules, normalization can prevent certain features from having an excessive impact on the rule mining results due to excessively large or small values, ensuring that the mined rules are more reasonable and reliable.

[0078] Furthermore, the process correlation map is calibrated based on historical production process data from oil refining, making the solution adaptable to different production scenarios and process flows. Historical production process data contains a wealth of production experience and patterns. By analyzing and utilizing this data, a process correlation map and feature extraction method that aligns with actual production conditions can be constructed. When production processes change or new equipment is introduced, the map and feature extraction method can be adjusted accordingly based on the updated historical data to quickly adapt to the new production scenario, improving the system's adaptability and flexibility.

[0079] Furthermore, the feature extraction method provided by this solution is highly scalable. As production technology evolves and data analysis needs increase, new feature extraction methods and dimensions can be continuously introduced, such as feature extraction based on machine learning algorithms and feature representation based on deep learning. Furthermore, newly added devices or data sources can be integrated and processed according to the existing feature extraction framework, eliminating the need for large-scale system modifications and supporting continuous system upgrades and optimization.

[0080] Optionally, obtain a petroleum refining intelligent production process description file and parse it to obtain process flow sorting and correlation relationship features, including:

[0081] Obtain the petroleum refining intelligent production process description file and parse it to generate a set of structured entity relationship triples;

[0082] Based on the structured entity relationship triple set, the process flow sorting and association relationship features are generated.

[0083] Optionally, a petroleum refining intelligent production process description file is obtained and parsed to generate a set of structured entity relationship triples, including:

[0084] Obtain piping and instrumentation diagrams and process flow diagrams and infer pipeline connection relationships based on symbol recognition models;

[0085] Extract operating procedure documents and capture historical alarm logs and maintenance record texts in the DCS control system;

[0086] Multimodal semantic parsing and entity recognition are performed on operating procedure documents, historical alarm logs, and maintenance record texts to obtain a set of structured entity relationship triples.

[0087] Optionally, based on the structured entity relationship triple set, process flow sorting and association relationship features are generated, including:

[0088] Perform temporal logic mining and causal relationship modeling on the structured entity relationship triple set to obtain a causal relationship graph with temporal constraints;

[0089] Based on the construction of the petroleum refining word vector model, the development process is identified in the causal relationship diagram with time constraints, and the process chain and process segment division results are obtained;

[0090] Based on the pipeline connection relationship reasoning and the process chain and process segment division results, the process flow sorting and association relationship characteristics are generated.

[0091] In summary, the above solution, from acquiring process description files to generating process flow sequencing and correlation features, uses technical means such as symbol recognition models, multimodal semantic analysis, temporal logic mining, and causal relationship modeling. The system can automatically complete data parsing, feature extraction, and relationship modeling without extensive manual intervention, providing the following technical benefits:

[0092] In this application, when obtaining the description file of the intelligent production process of petroleum refining, not only the pipeline instrumentation diagram and process flow chart are used, but also the operating procedure documents are extracted and the historical alarm logs and maintenance record texts in the DCS control system are captured. This multi-source data fusion method can comprehensively obtain various types of knowledge about the intelligent production process of petroleum refining. For example, the pipeline instrumentation diagram and process flow chart intuitively show the connection relationship of the equipment and the direction of the process flow; the operating procedure document contains the standard steps and precautions for equipment operation; the historical alarm logs and maintenance record texts record the problems and solutions encountered in the actual production process. By integrating these multi-source data, we can have a more complete understanding of the overall picture of the production process and provide a rich knowledge base for subsequent feature extraction and rule mining.

[0093] Secondly, multimodal semantic parsing and entity recognition of operating procedure documents, historical alarm logs, and maintenance records can deeply explore the potential information within these texts. Traditional text processing methods may only extract surface textual information, but multimodal semantic parsing can combine natural language processing technology and domain knowledge to identify entities in the text (such as equipment names, operating procedures, etc.) and the relationships between them. For example, semantic parsing can identify the causal relationship between a device failure mentioned in an alarm log and the corresponding maintenance operation, providing more accurate data for subsequent causal relationship modeling.

[0094] Furthermore, obtaining piping and instrumentation diagrams (PIDs) and process flow diagrams (PFDs) and reasoning about pipeline connections based on the symbol recognition model can accurately reflect the physical connections between equipment. In petroleum refining and chemical production, equipment connections directly impact the progress of the process and the transfer of materials. By analyzing PIDs and PFDs using the symbol recognition model, it is possible to automatically identify the pipeline routes, connected equipment, and the locations of control elements such as valves, thereby constructing an accurate equipment connection diagram. This provides a crucial basis for subsequent process sequencing and the generation of association features, avoiding errors and omissions that can occur with manual identification.

[0095] Furthermore, temporal logic mining and causal relationship modeling are performed on sets of structured entity relationship triples to generate a causal relationship graph with temporal constraints, which can reveal the inherent logical relationships within petroleum refining production processes. In actual production processes, the operating status and operation steps of equipment often have temporal and causal characteristics. Temporal logic mining can determine the chronological order of events, while causal relationship modeling can analyze the causal relationships between events. For example, it can be determined that the startup of one piece of equipment is a prerequisite for the operation of another, or that an error in a certain operation step can lead to a series of subsequent failures. The causal relationship graph with temporal constraints can intuitively display these logical relationships, providing strong support for subsequent process chain and process segmentation.

[0096] Furthermore, based on the constructed petroleum refining word vector model, we can identify development processes within a causal relationship graph with timing constraints, enabling precise segmentation of process chains and process segments. The word vector model converts process vocabulary into vector representations and identifies similar operational steps and processes by calculating the similarity between these vectors. Combined with the causal relationship graph with timing constraints, we can accurately segment different process chains and process segments, clarifying the inputs, outputs, and operational requirements of each process. This helps us gain a deeper understanding of the flow and structure of the production process, providing a more refined basis for process sequencing and generating correlation features.

[0097] Finally, based on the reasoning of pipeline connection relationships and the division of process chains and process segments, process flow ranking and association features are generated, providing more targeted features for alarm event rule mining. These features not only capture the connection relationships between equipment and the order of process flows, but also reflect the associations and dependencies between processes. During the alarm event rule mining process, these features can be used to accurately locate the links and equipment where alarms may occur and analyze the association patterns between alarm events. For example, when a piece of equipment fails, the process flow ranking and association features can be used to quickly determine the subsequent processes and equipment that may be affected, thereby mining more accurate and valuable alarm event rules.

[0098] In particular, in a specific application scenario, the generation and implementation process of the above process flow sorting and association relationship features are as follows:

[0099] 1. Multimodal document parsing and entity relationship extraction

[0100] Symbol recognition model based on graph neural network:

[0101]

[0102] The edge attention coefficient is:

[0103]

[0104] in: The subscript v represents the vth node in the diagram (such as an instrument symbol), and the superscript l represents the lth layer of the network. That is, the feature vector of node v in layer l, The set of neighbor nodes of node v (such as adjacent pipes and valve symbols), c v,u : Normalization constant to ensure the stability of message passing, W (l) ,b (l) : The weight matrix and bias vector of the lth layer, α v,u : The attention coefficient from node v to node u, reflecting the importance of the connection, a: The trainable parameter vector of the attention mechanism.

[0105] As can be seen, this application uses a graph attention network (GAT) to perform symbol recognition and relational reasoning on piping and instrumentation diagrams (P&IDs). In the P&ID diagram of an ethylene plant, the model can identify the connection between "Pump P-101" and "Control Valve FV-102" and assign weights to different connections through an attention mechanism (e.g., a weight of 0.85 for material flow edges and 0.6 for control signal edges). The model achieved a symbol recognition accuracy of 97.2% on 1,000 refining drawings, and a relational reasoning F1 score of 0.93.

[0106] 2. Multimodal semantic parsing and entity recognition

[0107] Multimodal entity relationship triple generation based on BERT:

[0108] Triple(e i ,r,e j )=Softmax(W3·tanh(W2·[e i ;e j ;r]+b2)+b3)

[0109] Where entities and relations are embedded:

[0110] e i =BERT [CLS] (D i )+γ·CNN(I i )

[0111] r=TransE(r)+β·GCN(r topo )

[0112] Among them: Triple(e i ,r,e j ): represents entity e i With e j The probability distribution of the relationship r between them, Di : Entity e i Text description (such as "reactor temperature is too high"), I i : Entity e i Image features (such as DCS interface screenshots), r topo : topological structure representation of relation r (such as pipeline connection), γ=0.3, β=0.4: multimodal fusion weight.

[0113] This application uses BERT to capture text semantics, combining image features extracted by CNN and topological features extracted by GCN to generate structured triples. For example, the triple "[Pump P-101][Outlet pressure too high][Control valve FV-102]" was extracted from an operating procedure document with a confidence level of 0.92. This model achieved an 89.5% accuracy rate in triple extraction on a refining and chemical industry corpus, a 14 percentage point improvement over a single-modality approach.

[0114] 3. Temporal Logic Mining and Causal Relationship Modeling

[0115] Temporal Causal Discovery Based on Dynamic Time Warping:

[0116]

[0117] The mutual information is calculated as:

[0118]

[0119] Among them: Causal(X,Y): causal strength of variable X to Y, MI(X t ,Y t+τ ):X t With Y t+τ The mutual information between DTW(X 1:t ,Y 1:t+τ ):dynamic time warping distance, measuring time series similarity, τ:time lag parameter (typical value in oil refining scenario is 5-15 minutes), σ=0.2:time window width parameter.

[0120] For example, in a hydrocracking unit, the model found that a change in the "recycle hydrogen compressor inlet pressure" occurred 8 ± 3 minutes after the "reactor temperature" increased, with a causal strength of 0.78. By processing non-uniformly sampled data using dynamic time warping, the model resolved the issue of inconsistent sampling frequencies across different equipment in refining operations (e.g., a 30-second temperature sampling interval and a 1-minute pressure sampling interval).

[0121] 4. Process identification and process segmentation

[0122] Process chain identification based on hidden Markov model:

[0123] π=argmaxπ P(O|π)P(π)

[0124] The observation probability is:

[0125]

[0126] State transition probability:

[0127]

[0128] Where: π: optimal process chain sequence, O: observation sequence (such as temperature, pressure and other sensor data), S t : The process status at time t (such as "feeding stage", "reaction stage"), Process status t The vector representation of σ s =0.5: state transition smoothing parameter.

[0129] For example, in a catalytic reforming unit, the model divides the production process into four stages: pretreatment, reaction, separation, and refining, with an accuracy rate of 94.3%. When the reactor temperature rises from 200°C to 480°C while the feed flow rate remains stable at 95% of the design value, the model accurately identifies this as the reaction stage and predicts with a probability of 0.87 that the next stage will be the separation stage.

[0130] 5. Process flow sorting and relationship generation

[0131] Process modeling based on Petri net:

[0132]

[0133] The dynamic correlation calculation is:

[0134]

[0135] Where: P: set of places (such as equipment and materials), T: set of transitions (such as reactions and transports), F: set of flow relations (connecting places and transitions), W: arc weight function, M0: initial identifier (initial state), Rel k (p i ,p j ): p under the kth association type i With p j The correlation degree, The timestamp of the k-th association, σ t =10: time decay parameter (minutes).

[0136] For example, in a constant-pressure distillation unit, Petri net modeling revealed the process sequence of "crude oil → primary distillation tower → atmospheric tower → vacuum tower." Dynamic correlation calculations revealed that when the "primary distillation tower top temperature" increased, the "atmospheric tower feed temperature" changed after 15 ± 5 minutes, and the correlation dynamically adjusted from 0.68 to 0.85, accurately reflecting the temporal relationship between the process stages.

[0137] Process flow sequencing and correlation characteristics

[0138]

[0139] V proc ={v1,v2,...,v n}: process node set, v i Represents the i-th process

[0140] Process relationship set with time constraints

[0141] Process feature embedding function

[0142] Ω:E proc →[0,1]:process association strength function

[0143] Time lag distribution function

[0144] For example, in a specific oil refining scenario application example, the process flow of the oil hydrogenation unit is extracted:

[0145] 1. Document parsing: Identify 127 equipment symbols and 213 connection relationships from P&ID diagrams, and extract 386 operating instructions from operating procedures

[0146] 2. Entity recognition: Generate triples such as "[Crude oil buffer tank V-101][Feed to][Hydrogenation reactor R-102]" (confidence 0.94)

[0147] 3. Causal modeling: The causal strength between "reactor temperature" and "hydrogen-to-oil ratio" was found to be 0.82, with a time lag of 10 ± 2 minutes.

[0148] 4. Process division: The production process is divided into three process sections: "raw material pretreatment", "hydrogenation reaction" and "product separation"

[0149] 5. Process sequencing: Construct a Petri net model to determine the process flow of "V-101→P-101→R-102→E-103→V-104"

[0150] Optionally, based on the process flow sorting and association relationship characteristics, an association topology structure between the intelligent production equipment of the petroleum refining is constructed, and production impact weights are added to the association topology result to generate a weighted equipment association topology structure, including:

[0151] Perform equipment entity recognition and spatial coordinate mapping on the process flow sequencing and association relationship features to generate the equipment space topology base map;

[0152] The process logic associated edges are constructed on the equipment space topology base graph to generate a process associated enhanced graph;

[0153] Perform multi-dimensional weight quantization and fusion processing on the process correlation enhancement graph to generate a topological graph with initial weights;

[0154] Industrial constraint optimization and dynamic adjustment are performed on the topology graph with initial weights to generate a device association topology structure with weights.

[0155] Optionally, the process flow sequence and association relationship features are processed for equipment entity recognition and spatial coordinate mapping to generate an equipment space topology base map, including:

[0156] Perform equipment entity semantic analysis on process flow sequencing and association relationship features to generate a standardized equipment entity set;

[0157] Performing spatial coordinate mapping processing on the standardized device entity set to generate a device coordinate mapping table;

[0158] The device coordinate mapping table is processed by constructing an adjacency matrix to generate a device space topology base graph.

[0159] Optionally, the equipment space topology base graph is processed with process logic association edges to generate a process association enhancement graph, including:

[0160] Perform process edge extraction on the equipment space topology base graph to generate a process edge set;

[0161] Perform parameter causal edge construction processing on the device space topology base graph to generate a parameter causal edge set;

[0162] Perform fault propagation edge reasoning on the device space topology base graph to generate a set of fault propagation edges;

[0163] The equipment space topology base graph and the process flow edge set, parameter causal edge set, and fault propagation edge set are fused to generate a process association enhanced graph.

[0164] Optionally, multi-dimensional weight quantization and fusion processing are performed on the process association enhancement graph to generate a topology graph with initial weights, including:

[0165] Construct a production impact weight indicator system for the process correlation enhancement diagram to generate a weight dimension calculation engine;

[0166] Based on the weight dimension calculation engine, single-dimensional weight is calculated to generate a multi-dimensional weight edge set;

[0167] The multi-dimensional weighted edge sets are fused to generate a topological graph with initial weights.

[0168] Optionally, the topology graph with initial weights is subjected to industrial constraint optimization and dynamic adjustment processing to generate a weighted device association topology structure, including:

[0169] Injecting process expert knowledge into the topological graph with initial weights to generate a rule-constraint enhanced graph;

[0170] Performing dynamic weight adjustment processing on the rule constraint enhancement graph to generate a dynamic weight topology graph;

[0171] Perform topological structure optimization on the dynamic weight topological graph to generate a community division topological graph;

[0172] The community division topology graph is weight-encoded to generate a weighted device association topology structure.

[0173] In summary, the above solution, which accurately depicts equipment relationships, reasonably quantifies weights, dynamically adapts to production changes, and optimizes decision-making, has the following technical benefits:

[0174] First, the equipment space topology base map is generated by identifying and mapping equipment entities to spatial coordinates based on the process flow sequence and association relationship characteristics. Semantic analysis of equipment entities generates a standardized set of equipment entities to ensure accurate equipment identification. Spatial coordinate mapping generates an equipment coordinate mapping table to clearly define the equipment's location in space. An adjacency matrix is ​​constructed to mathematically express the spatial relationships between equipment, generating the equipment space topology base map. This series of steps comprehensively and accurately depicts the spatial distribution of equipment and the relationships between equipment entities, laying a solid foundation for the subsequent construction of more complex association topologies and providing a clearer understanding of the layout and interactions of equipment in a production environment.

[0175] Secondly, based on the equipment spatial topology base graph, we construct process logic association edges to generate a process association enhancement graph. We extract process flow edges, construct parameter causal edges, and infer fault propagation edges, and then fuse these edge sets together. This approach not only considers the spatial location of the equipment but also deeply explores the logical associations between the equipment in terms of process flow, parameter causality, and fault propagation. This makes the associations between equipment more comprehensive and rich, and more accurately reflects the complex interactions between equipment in actual production processes.

[0176] Furthermore, the process-related enhancement graph undergoes multi-dimensional weight quantification and fusion processing to generate a topology graph with initial weights. First, a production impact weight indicator system is constructed to form a weight dimension calculation engine, which comprehensively considers the impact of equipment relationships on production from multiple dimensions. This engine calculates single-dimensional weights, generates a multi-dimensional weighted edge set, and finally fuses them. This multi-dimensional weight quantification and fusion method comprehensively considers various factors related to equipment relationships, such as the importance of the process flow, the strength of the correlation between parameters, and the potential impact of fault propagation. This ensures that the weights in the generated topology graph with initial weights more reasonably reflect the impact of equipment relationships on production, providing a more accurate basis for subsequent analysis and decision-making.

[0177] Furthermore, the topology with initial weights undergoes industrial constraint optimization and dynamic adjustment to generate a weighted device-association topology. First, process expert knowledge is injected to generate a rule-constraint enhancement diagram, incorporating the experience and knowledge of process experts into the topology to enhance its rationality and practicality. Next, dynamic operating condition weight adjustment is performed to dynamically adjust weights based on changes in actual production conditions, enabling the topology to reflect the actual production process in real time. Topology optimization is then performed to generate a community-partitioned topology, dividing closely related devices into communities and further optimizing the topology. Finally, weight encoding is performed to generate a weighted device-association topology. This series of steps enables the generated topology to dynamically adapt to complex and changing production environments, providing accurate and effective device association information under varying production conditions, facilitating the timely detection and resolution of potential alarm events.

[0178] Finally, the resulting weighted device association topology provides strong support for alarm event rule mining and decision-making in the intelligent manufacturing industry. In terms of alarm event rule mining, the weighted topology clearly demonstrates the associations and impacts between devices, helping to mine more accurate and valuable alarm event rules. For example, the weight can be used to determine the scope and extent of the impact of a device failure on other devices, thereby determining the propagation path and impact range of the alarm event. Furthermore, this topology can provide a reference for decision-making in production scheduling, equipment maintenance, and other areas. For example, based on the associations and weights between devices, equipment maintenance plans can be rationally arranged to avoid interruptions to the entire production process due to equipment failures, thereby improving production efficiency and stability.

[0179] In particular, in a specific scenario, the above weighted device association topology is implemented as follows:

[0180] 1. Construction of device space topology base map

[0181] Multimodal embedding for device entity semantic parsing:

[0182]

[0183] Among them: e i :Subscript i represents the i-th device, e i That is, the semantic embedding vector of device i (dimension 2048). i :Subscript i represents the i-th device, D i is a text description of the device i (e.g., “hydrogenation reactor R 101 ”). The subscript i represents the i-th device, is the neighborhood topology of device i, that is, the local topology composed of other devices directly connected to device i and their connection relationships. i :Subscript i represents the i-th device, I i is the CAD drawing image feature of device i.

[0184] α=0.5,β=0.3,γ=0.2: modality fusion weights.

[0185] In this application, considering the numerous different types and naming schemes of equipment in petroleum refining enterprises, their descriptions include both textual information and topological connectivity and spatial layout information. Processing equipment text descriptions using the BERT model can uncover the semantic associations underlying different textual expressions, for example, semantically aligning different expressions such as "reforming reactor" and "core reaction equipment of a certain model of reforming unit." A graph convolutional network (GCN) is used to process the neighborhood topology of the equipment and analyze the connectivity between devices, such as determining whether the devices are connected in series, in parallel, or in other complex configurations. A convolutional neural network (CNN) extracts features from the equipment's CAD drawing images to identify information such as the device's spatial location and shape. By fusing these three modalities according to specific weights (α = 0.5, β = 0.3, γ = 0.2), a semantic embedding vector for the device is derived. This uniformly maps heterogeneous device descriptions into a 2048-dimensional space, enabling accurate identification of entities representing the same device despite different representations. Entities with a cosine similarity greater than 0.85 are confirmed as the same device. This multimodal fusion approach is applicable to all equipment in refining and chemical enterprises, solving the identification problem caused by the diversity of equipment naming.

[0186] 2. Construction of process logic association edges

[0187] Granger causal strength of parametric causal edges:

[0188]

[0189] Among them: Granger X→Y : Represents the Granger causal strength from event X to event Y, where X and Y represent events related to different devices or parameters. MSE: Mean Squared Error, a measure of the improvement in prediction accuracy, used here to compare predictions of Y based only on historical data of Y t+1 The mean square error is the same as that of predicting Y based on the historical data of Y and X. t+1 The mean square error of τ XY : Actual time lag between X and Y (minutes). μ τ =8,σ τ =2: Historical lag distribution parameter, where μ τ is the historical mean of the time lag between X and Y, σ τ is its standard deviation.

[0190] In this application, it is considered that there are complex causal relationships between various devices and between device parameters in various types of oil refining equipment. Granger causality test is a method used to determine the causal relationship between two time series variables. The formula predicts Y by comparing the historical data of variable Y only. t+1 The mean square error, and the prediction of Y based on the historical data of variables Y and X t+1The mean square error of the model is obtained by using the logarithmic function to obtain the information gain part, which measures whether the addition of X significantly improves the prediction accuracy of Y. At the same time, considering that there is a time lag in the causal relationship in practice, the exponential function is combined with the historical lag distribution parameter (μ τ and σ τ ), constraining time lags. When the calculated Granger causality strength exceeds a certain threshold (e.g., 1.2), a valid causal relationship is assumed to exist from event X to event Y, generating a causal edge. For example, in different refineries, pump vibration may cause pipeline pressure fluctuations, or furnace temperature changes may cause changes in reaction parameters in subsequent reactors. This method can be used to determine causal relationships, providing a basis for fault propagation analysis and early warning.

[0191] 3. Multi-dimensional weight quantization and fusion

[0192] Weight fusion based on analytic hierarchy process:

[0193]

[0194] in:

[0195]

[0196] Where: W ij :Subscripts i and j represent two devices, W ij Represents the weight of the association edge between device i and device j. k : The subscript k represents the sequence number of the weight dimension, ω k is the weight coefficient of the kth dimension in the hierarchical analysis method, where k ranges from 1 to 5 and corresponds to the five dimensions of physical connection strength, process impact, fault propagation speed, maintenance difficulty, and energy consumption correlation. k (e ij ): Subscript k represents the number of the weight dimension, f k is the function for calculating the weight of the kth dimension, e ij represents the set of related attributes of the edge between device i and device j, f k (e ij ) is the weight value of the kth dimension calculated based on the relevant attributes of the edge associated with device i and device j. min(f k ) and max(f k ): They are the minimum and maximum values ​​of the k-th dimension weight among all associated edges.

[0197] In this application, it is considered that the weight of the associated edges between the equipment in the petroleum refining device needs to be determined by integrating multiple dimensions. The hierarchical analysis method decomposes the complex problem of determining the weight of the associated edges into multiple dimensions, namely physical connection strength (such as the connection tightness reflected by the diameter and length of the pipeline), process impact (the degree of influence of the equipment on key indicators such as product quality and output in the process flow), fault propagation speed (the speed at which a device failure causes the failure of the connected equipment obtained from historical data statistics), maintenance difficulty (measurement indicators such as mean repair time MTTR), and energy consumption correlation (the correlation between energy flows between equipment). By assigning a corresponding weight coefficient (ω k ), and normalize the weight value of each dimension (using ), and finally, the weight values ​​of each dimension are linearly combined according to the weight coefficient to obtain the final weight W of the associated edge between device i and device j. ij This method can comprehensively and comprehensively consider the impact of various factors in the refining unit on the equipment association edge weight, and is suitable for quantifying the equipment association relationship in all refining units.

[0198] 4. Industrial Constraint Optimization and Dynamic Adjustment

[0199] Constrained optimization based on Lagrange multipliers:

[0200]

[0201] Where W ij :Subscripts i and j represent two devices, W ij Represents the weight of the association edge between device i and device j, which is a variable in the optimization process. The subscripts i and j represent two devices, Represents the initial weight of the association edge between device i and device j. λ1=0.4,λ2=0.6: Constraint strength coefficient. n: The total number of devices, that is, the number of nodes in the topology graph. C i :Subscript i represents the i-th device, C i is the total connection capacity constraint of device i, such as the maximum flow rate, pressure and other comprehensive constraint values ​​that device i can withstand. ε: represents the set of all associated edges in the topology graph. δ ij :Subscripts i and j represent two devices respectively. When the edge (i, j) does not satisfy the process logic, δ ij =1, when δ ij =0, which is the expert knowledge penalty item.

[0202] In this application, in the actual production of petroleum refining, the weight determination of the equipment association topology structure should not only be based on data-driven analysis, but also consider various constraints in industrial practice. The Lagrange multiplier method is a commonly used constraint optimization method. In the formula, the first term Represents the optimized weight W ij With initial weight We hope to make the optimized weights close to the initial weights as much as possible while satisfying the constraints, and retain the result trend of data-driven analysis. This is a consideration of the device connection capacity constraint, ensuring that the sum of the weights of the edges connected to the device is within the device's capacity limit, such as the device's flow rate, pressure tolerance limit, etc. Expert knowledge penalties are introduced. When edges fail to meet process logic (as determined by expert knowledge), the weights are adjusted using these penalties, ensuring that the topology more closely matches actual process requirements. This approach balances data-driven weight calculations with practical industrial constraints, making it applicable to topology optimization for all refining units.

[0203] 5. Dynamic weight adjustment of working conditions

[0204] Dynamic weights based on Gaussian mixture models:

[0205]

[0206] Where: W ij (t): Subscripts i and j represent two devices, t represents time, W ij (t) represents the dynamic weight of the association edge between device i and device j at time t. The subscripts i and j represent two devices, Represents the baseline weight of the association edge between device i and device j. k (t): Subscript k represents the number of the working mode, t represents time, α k (t) is the weight coefficient of the kth operating condition at time t (predicted by LSTM), where k ranges from 1 to 3, corresponding to low, medium, and high load modes, respectively. Load(t): The production load at time t. μ1 = 0.6, μ2 = 0.85, μ3 = 1.0: The center points of low, medium, and high loads, respectively. The variance of the load distribution under the kth working condition is used to determine the shape of the Gaussian function.

[0207] In this application, considering that the working conditions in the oil refining production process are constantly changing, the importance of the correlation between equipment under different working conditions will also vary. The Gaussian mixture model is a method that can model complex distributions. Here, the production conditions are divided into three modes: low, medium, and high load. The weight coefficient α of the current time t belonging to different working conditions is predicted by the long short-term memory network (LSTM). k (t). Each operating mode corresponds to a Gaussian function, whose center point is μ k (corresponding to low, medium and high load center points respectively), the variance is According to the difference between the current production load Load(t) and the center point of each working condition, the corresponding weight adjustment value is calculated by Gaussian function, and then combined with the benchmark weight of the equipment associated edge Get the dynamic weight W of the association edge between device i and device j at time t ij For example, under high-load conditions, the weights of the associated edges between certain key equipment will be increased accordingly to enhance the early warning capability for the propagation of these equipment failures, adapt to production needs under different working conditions, and be applicable to the dynamic weight adjustment of all refining and chemical units under different working conditions.

[0208] 6. Topology optimization processing

[0209] Community division based on modularity maximization:

[0210]

[0211] The community optimization goals are:

[0212]

[0213] Where: Q: modularity, used to measure the quality of the topological graph community division. m: the total number of edges in the topological graph. ij : Adjacency matrix element, when there is an edge connecting the devices corresponding to subscripts i and j, A ij =1, otherwise A ij =0. k i and k j : The degrees of the devices corresponding to the subscripts i and j are respectively, that is, the number of edges connected to the device. δ(c i ,c j ): If the devices i and j corresponding to the subscripts i and j belong to the same community (i.e., c i =c j ), then δ(c i ,c j )=1, otherwise δ(c i ,c j )=0. C k :Subscript k represents the number of the community, C kRepresents the kth community, which is a collection of devices. Capacity i :Subscript i represents the i-th device, Capacity i is the production capacity or capacity-related indicator of device i. Γ = 0.3: The upper limit coefficient of community capacity. n: The total number of devices, that is, the number of nodes in the topology.

[0214] In this application, considering that there are many devices and complex relationships in petroleum refining enterprises, dividing the devices into different communities can more effectively manage and analyze faults. Modularity Q is an indicator to measure the quality of community division in the topological graph. Its calculation is based on the adjacency matrix element A. ij , the degree of equipment i and k j , and the indicator function δ(c i ,c j By maximizing the modularity Q, we can find an optimal community division method, which makes the connections between devices in the community dense and the connections between communities relatively sparse. At the same time, considering the limitation of community capacity in actual production, through the constraint condition Ensure that the total capacity or production capacity of each community does not exceed a certain ratio (here Γ = 0.3). For example, after dividing the equipment in a refinery into different communities such as the reaction zone, separation zone, and heating zone, the edge weight within the community is increased, while the cross-community edge weight is reduced. This allows fault propagation analysis to be focused on a local area, improving the accuracy and efficiency of fault analysis. This is applicable to the topological structure community division optimization of all refineries.

[0215] Weighted device association topology

[0216]

[0217] V={v1,v2,...,v n}: The subscript n represents the number of devices, V is the set of device nodes, v i (i ranges from 1 to n) represents the i-th device node.

[0218] E is the set of associated edges, which are composed of node pairs and come from the Cartesian product of the device node set V.

[0219] W:E→[0,1]: W is a dynamic weight function for each associated edge (belonging to set E)

[0220] Optionally, an executable alarm rule base is generated based on the weighted device association topology, including:

[0221] A multi-dimensional labeling system is constructed for the weighted device association topology to generate a device label dataset;

[0222] Perform high-frequency alarm data screening on the equipment label dataset to generate high-frequency item set data;

[0223] Perform tree structure index construction on high-frequency item set data to generate an association rule tree model;

[0224] Perform conditional path backtracking on the association rule tree model to generate a causal rule set;

[0225] The causal rule set is mapped to a business baseline to generate an executable alarm rule base.

[0226] Optionally, a multi-dimensional labeling system is constructed for the weighted device association topology to generate a device label dataset, including:

[0227] Perform label dimension definition processing on the nodes and edges of the weighted device association topology to generate a device label dataset;

[0228] The device label dataset and rule mining basic data are processed with label data filling to generate a device label dataset.

[0229] Optionally, high-frequency alarm data is filtered on the device tag dataset to generate high-frequency item set data, including:

[0230] Perform statistical processing on the device tag data set and the alarm records of the specified time period of five working days to generate an alarm frequency statistics table;

[0231] Perform frequency weight filtering on the alarm frequency statistics table to generate a high-value alarm data set;

[0232] The high-value alarm dataset is frequency-reordered to generate high-frequency item set data.

[0233] Optionally, a tree structure index construction process is performed on the high-frequency item set data to generate an association rule tree model, including:

[0234] Perform item header table construction processing on high-frequency item set data to generate the item header table structure;

[0235] Perform FP-Growth tree construction on high-frequency item set data and item header table structure to generate a basic rule tree;

[0236] The basic rule tree and the weighted device association topology structure are subjected to topology dependency fusion processing to generate an association rule tree model.

[0237] Optionally, conditional path backtracking is performed on the association rule tree model to generate a causal rule set, including:

[0238] Perform prefix path backtracking on the association rule tree model and the specified target alarm item to generate the original conditional path set;

[0239] Perform timing constraint annotation processing on the original conditional path set and the weighted device association topology structure to generate a conditional path set with timing;

[0240] The rule pruning optimization process is performed on the conditional path set with time sequence to generate a causal rule set.

[0241] Optionally, a business baseline mapping process is performed on the causal rule set to generate an executable alarm rule base, including:

[0242] Perform business threshold mapping on the causal rule set and the oil refining production business baseline configuration to generate an intermediate set of quantitative rules;

[0243] Performing rule expression generation processing on the intermediate set of quantified rules to generate structured rule candidates;

[0244] Rule weight assignment is performed on the structured rule candidates and the weighted device association topology to generate an executable alarm rule base.

[0245] The technical benefits of the above solution are detailed as follows:

[0246] First, a multidimensional labeling system is constructed for the weighted device association topology to generate a device label dataset. By defining the label dimensions for nodes and edges and populating the label data, comprehensive information about the devices in the topology is captured. For example, in addition to the basic attributes of the device itself, this includes its position in the process flow, its relationships with other devices, and its weight. This multidimensional labeling system provides a rich data foundation for subsequent rule mining, enabling the generated alarm rules to reflect the connections and impacts between devices from multiple perspectives, ensuring comprehensiveness.

[0247] Next, the causal rule set is mapped to a business baseline to generate an executable alarm rule library. This causal rule is then integrated with the oil refining production business baseline configuration to perform business threshold mapping, rule expression generation, and rule weight assignment. This process fully considers the business needs and standards of actual production. The generated alarm rules are not only based on the device association topology and data mining results, but are also closely integrated with production operations, ensuring that the rules can be effectively applied in actual production and cover various scenarios and requirements during the production process.

[0248] Furthermore, high-frequency alarm data is filtered from the device tag dataset to generate high-frequency item sets. By traversing and counting alarm data, filtering by frequency weight, and reordering the frequency, high-value alarm data can be accurately screened and generated into high-frequency item sets. This step focuses on frequently occurring alarm information in actual production, eliminating low-frequency, low-value interference data. This allows subsequent rule mining to focus more on key issues, improving the relevance and accuracy of the rules.

[0249] Furthermore, the association rule tree model is processed through conditional path backtracking to generate a causal rule set. Through prefix path backtracking, timing constraint annotation, and rule pruning optimization, the conditional path leading to the target alarm item can be accurately traced back, and the rules are pruned and optimized based on timing relationships. This process ensures that the generated causal rules accurately reflect the causal relationships between devices, avoids the generation of false associations and redundant rules, and improves the accuracy and reliability of the rules.

[0250] Furthermore, a tree-structured index is constructed on the high-frequency item set data to generate an association rule tree model. By constructing an item header table, building an FP-Growth tree, and integrating topological dependencies, the high-frequency item set data is organized into a tree-structured index. This tree-structured index can quickly locate and access relevant data, greatly accelerating the rule mining process and improving algorithm efficiency. Compared with traditional traversal search methods, the tree-structured index significantly reduces computing time and resource consumption, making the generation of the alarm rule base more efficient.

[0251] Finally, the causal rule set is mapped to a business baseline to generate an executable alarm rule library. Through business threshold mapping, rule expression generation, and rule weight assignment, the causal rules are transformed into structured, quantifiable, and executable rules. The generated rule expressions are clear and unambiguous, facilitating system integration and implementation. The rule weight assignment takes into account the weight information associated with devices, enabling practical application of rules based on the importance and relevance of devices. This executable alarm rule library can be directly applied to alarm monitoring and processing in intelligent manufacturing systems, improving system response speed and processing efficiency.

[0252] As can be seen from the above, the entire rule generation process fully considers the actual conditions and needs of oil refining production. From data screening, rule mining, to rule mapping, it is closely integrated with production practices. Furthermore, through processes such as rule pruning optimization and topology dependency fusion, the generated rules can adapt to changes and uncertainties in the production process. For example, when the production process or equipment status changes, the rule base can be dynamically adjusted and optimized to maintain high accuracy and effectiveness, ensuring that the alarm rules always adapt to actual production needs.

[0253] In particular, in a specific application scenario, the specific implementation details of the above solution are as follows:

[0254] 1. Construction and processing of multi-dimensional labeling system

[0255] This application device label vector tensor product construction:

[0256]

[0257] in: The multi-dimensional label vector of device i (dimension d = 768), integrating the full-dimensional features of the device, The mapping function of the kth feature space (such as device type, topological location, etc.), D i : Description vector of equipment i (including physical parameters, process roles, etc.), x i : real-time sensor data vector of device i (such as temperature, pressure and other 5-dimensional parameters), μ k ,σ k : Statistical characteristic distribution parameters of the kth type of equipment (based on historical data fitting).

[0258] It can be seen that this application integrates the six-dimensional feature space of the equipment (equipment type, topological location, real-time data, historical statistics, process role, and failure mode) into a unified vector through tensor product operation. Gaussian kernel function Implementing data-driven adaptive weight allocation: When the device real-time data x i Deviation from the mean value of similar equipment μ k More than 2 times the standard deviation σ k When the feature dimension is increased exponentially, the weight of the corresponding feature dimension increases exponentially. For example, if the temperature sensor data of a reactor in a petroleum refining process exceeds the normal range, the weight of the "fault association" dimension will be automatically increased, providing more sensitive feature input for subsequent rule mining, and improving feature discrimination by 35% compared to traditional labeling methods.

[0259] 2. High-frequency alarm data screening and processing

[0260] This application applies to alarm event scores based on spatiotemporal density:

[0261]

[0262] Where: S(a j ): Alarm event a j The comprehensive score of f(a j ): alarm frequency (count the number of occurrences within 5 working days), w k : The service weight of the kth type alarm (e.g. reactor failure w = 1.5, pump abnormality w = 1.0), Rel(a j ,C k ): Alarm aj With business category C k Semantic relevance (calculated by BERT model), λ: spatiotemporal clustering coefficient (λ = 0.3 for oil refining scenario to suppress short-term noise), t j -t i : Alarm time interval, used to calculate time density.

[0263] It can be seen that the above formula numerator of this application highlights the key equipment alarms (such as the reactor-related alarms) by multiplying the frequency and the semantic weight. k Higher and the score improves), the denominator is calculated through the space-time density function Suppressing repeated noise alarms within a short period of time. In an oil refining scenario, if a heating furnace issues three consecutive temperature fluctuation alarms within 30 minutes, the spatiotemporal density term in the denominator will reduce the scores of subsequent alarms by 40%, preventing the generation of redundant rules. Furthermore, the BERT semantic matching mechanism ensures a 25% increase in the scores of semantically related alarms, such as "reactor pressure anomaly" and "feed rate fluctuation," improving the process relevance of the rules.

[0264] 3. Tree structure index construction processing

[0265] FP-Growth tree node update for this application topology enhancement:

[0266]

[0267] Where: ΔW node : Weight update amount of FP-Growth tree node, W tree : Support weight based on data statistics in the tree structure, W topo : Equipment topology importance (calculated by PageRank algorithm, reflecting the criticality of the equipment in the process), The set of path edges from the root node to the current node, w e : The process dependency weight of edge e (such as pipeline flow impact coefficient, material transmission priority), α, β: adjustment parameters (α = 0.6, β = 2.0 in the oil refining scenario).

[0268] It can be seen that this application uses the S-type function Implement nonlinear enhancement of topology importance to rule weight: When the topology importance of the device W topo When it is higher than 0.5, the S-type function output is >0.7, making the data-driven support weight W tree For example, in an ethylene cracking unit, the W topo =0.8, the weight of its related alarm nodes will be enhanced to the original W tree 1.8 times that of ordinary pumps, while W topo= 0.3 only increases to 1.2 times. At the same time, the mean term of the path edge weight Ensure that the weight of long path rules is reasonably attenuated (the weight of three-hop paths is attenuated to 60% of the original weight), avoid interference from long chain rules of non-critical equipment, and improve the process matching degree of the rule tree by 50%.

[0269] 4. Conditional Path Backtracking Processing

[0270] This application extracts temporal rules based on Granger causality test:

[0271]

[0272] Granger(X→Y): Granger causal strength from event X to Y. The larger the value, the stronger the causality.

[0273] MSE: Mean square error, which measures the prediction accuracy (the smaller the MSE, the more accurate the prediction), Y t+1 |Y 1:t :Use only the historical data of Y to predict the future value, Y t+1 |Y 1:t ,X 1:t : Use historical data of X and Y to predict future values, τ XY : The actual time lag between X and Y (e.g., the interval from device B vibrating to device C experiencing temperature anomaly), μ τ ,σ τ : The mean and standard deviation of the historical lag distribution (such as μ τ =8 minutes, σ τ = 1.5 minutes), φ: cumulative normal distribution function, used for the probabilistic processing of time constraints. It can be seen that the left side of the above rules in this application The information gain is used to quantify the improvement in the predictive ability of X on Y: when the historical data of X is added, the more the mean square error of the prediction of Y decreases, the stronger the causal relationship between X and Y. The actual lag time τ XY Converted to probability value, when τ XY Falls at the historical mean μ τ When the probability value is within 1 standard deviation of the causal relationship, it is >0.8, confirming a valid causal lag. In petroleum refining, this model accurately identifies the causal relationship (Granger = 1.8) from "abnormal pump vibration (X) to pipeline pressure fluctuation (Y)" and filters out the spurious correlation (Granger = 0.3) from "lighting failure to reactor temperature anomaly," improving the accuracy of causal rules by 60%.

[0274] 5. Business Baseline Mapping Processing

[0275] This application is based on the rule execution probability of fuzzy Petri net:

[0276]

[0277] Where: P(R j ):Rule R j The execution probability (range [0,1]) determines the rule triggering priority, μ i (x i ): Membership function of the i-th condition (e.g., membership μ = 0.9 for temperature > 800°C), w j : topological weight of the rule (reflecting the importance of device association), w0: baseline weight (take 0.5 as the reference point for weight deviation), σ w : Standard deviation of topological weight (reflecting the weight fluctuation range), ΔQ / Q base : The relative deviation between the current load and the design load (such as ΔQ / Q base = +20% means 20% overload), γ: load sensitivity coefficient (in the petroleum refining scenario, γ = 0.3).

[0278] It can be seen that this application formula realizes the intelligent scheduling of rules through a triple mechanism:

[0279] Fuzzy membership product: The degree of satisfaction of multiple conditions is converted into a joint probability. For example, the joint probability of "temperature > 800°C (μ = 0.9) and pressure > 10 MPa (μ = 0.8)" is 0.72, which avoids false triggering of rules that do not fully meet the conditions; weighted normal distribution: Make the rules with topological weight close to the benchmark value w0=0.5 obtain higher probability, suppress the interference of extreme weight rules (such as weight w j = 0.2 rule probability is reduced by 30%); Dynamic load adjustment: Automatically adjust the rule priority based on the current load. When the refining unit is overloaded by 20%, This model increased the probability of executing equipment protection rules by 6%, triggering early warnings. After being applied at a refinery, the model improved the efficiency of rule adaptation under different operating conditions by 45%.

[0280] 6. Formal Expression of Executable Rule Base

[0281] The tensor decomposition representation of this application's rule base is:

[0282]

[0283] in: Rule base three-dimensional tensor (M is the number of rules, N is the number of conditions, P is the number of attributes), Regular pattern matrix (r is the latent factor dimension, take r = 32), Conditional feature matrix, which stores the semantic representation of conditions, Parameter matrix, which stores parameters such as rule weights and time windows, × k : k-th mode tensor product to achieve multi-dimensional feature fusion.

[0284] Thus, this application uses Tucker tensor decomposition to compress the high-dimensional rule base into a low-dimensional representation. The core innovation lies in capturing the potential correlation between rules through the latent factor dimension r=32. For example, the two rules "reactor temperature abnormal → reduce feed rate" and "reactor pressure abnormal → reduce feed rate" are Similar row vectors are shared in the matrix, reflecting similar control logic. This decomposition reduces the storage space of the rule base by 80% while maintaining 95% of the information integrity. In the real-time system of petroleum refining, the query delay of a single rule is less than 1ms, and it can be updated incrementally. The matrix enables online learning of the rule base, increasing the efficiency of integrating new rules by 3 times.

[0285] Taking the "reactor temperature abnormality" alarm as an example, the rule generation process is as follows:

[0286] Label construction: multi-dimensional label vectors for reactors In the temperature sensor data x i Deviation from the mean μ k When the standard deviation reaches 2.5 times, the Gaussian kernel function weight is enhanced, and the weight of the "fault correlation" dimension is increased to 1.8 times;

[0287] High frequency screening: The frequency of the alarm under normal production conditions f(a j ) = 12 times / 5 days, the semantic relevance to "feed quantity fluctuation" calculated by BERT is Rel = 0.78, the spatiotemporal density term suppression coefficient is 0.6, and the final score S = 8.3 (higher than the threshold of 5.0), entering the high-frequency item set;

[0288] Tree structure construction: In the FP-Growth tree, the topological weight W of the node topo =0.8, after being enhanced by the S-type function, the node weight ΔW node =0.72, becoming a high-frequency node in the tree structure;

[0289] Causal backtracking: Granger causality test shows that the causal strength of "abnormal feed pump vibration → abnormal reactor temperature" is Granger = 1.6, and the time lag τ = 8.5 minutes falls within the historical mean μ τ =8±1.5 minutes, confirmed as a valid causal rule;

[0290] Business mapping: current load ΔQ / Q base = +15%, rule execution probability P = 0.9 × 0.85 × (1 + 0.3 × 0.15) = 0.82, generating executable rules:

[0291] IF feed pump vibration value>75mm / s AND 5min≤t≤10min

[0292] THEN Reactor

[0293] The above is merely one specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for mining alarm event rules in the intelligent manufacturing industry, applied to intelligent production in petroleum refining, characterized by: include: Acquire multi-source heterogeneous operating data of equipment in intelligent production workshops of petroleum refining; Normalize multi-source heterogeneous operation data to form basic data for rule mining; Obtain the petroleum refining intelligent production process description file and parse it to obtain the process flow sorting and correlation relationship characteristics; Based on the process sequencing and association relationship characteristics, the association topology structure between the intelligent production equipment of petroleum refining is constructed, and the production impact weight is added to the association topology result to generate a weighted equipment association topology structure; Generate an executable alarm rule base based on the weighted device association topology.

2. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 1 is characterized in that: Acquire multi-source heterogeneous operating data of equipment in the intelligent production workshop of petroleum refining, including: Based on the deployed edge data collection nodes, the operating parameter data generated by the equipment operation in the petroleum refining intelligent production workshop and the log data of the control system are collected from different data sources; Based on the deployed centralized data collection platform, the adaptive communication protocol adapter component is called to build a data sampling channel to sample operation parameter data and log data from the edge data collection node and generate multi-source heterogeneous operation data based on this.

3. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 2 is characterized in that: Based on the deployed centralized data collection platform, the adaptive communication protocol adapter component is called to build a data sampling channel to sample operating parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operating data based on this, including: Send a preset general detection data packet to the node edge data collection node to receive the response data packet returned by the node and extract the protocol characterization features; Match the extracted protocol characterization features with the constructed communication protocol multi-dimensional feature library to determine the matching protocol type; Extracting protocol features of the matching protocol type to generate protocol conversion rules. Protocol features include data frame format, data encoding method, command set, and communication parameters. A data sampling channel is constructed based on protocol conversion rules to sample operation parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operation data accordingly.

4. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 2 is characterized in that: Based on the deployed centralized data collection platform, the adaptive communication protocol adapter component is called to build a data sampling channel to sample operating parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operating data based on this, including: Conduct multi-level business semantic mining on industrial communication protocols used in intelligent production in petroleum refining to build a semantic model at the communication protocol layer; Based on the constructed communication protocol layer semantic model, a semantic rule base for different industrial communication protocols is generated to construct a semantic association graph of protocol semantic elements; Generate protocol conversion rules based on the semantic association graph of protocol semantic elements. Protocol features include data frame format, data encoding method, command set and communication parameters. A data sampling channel is constructed based on protocol conversion rules to sample operation parameter data and log data from edge data collection nodes and generate multi-source heterogeneous operation data accordingly.

5. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 1 is characterized in that: Normalize multi-source heterogeneous operation data to form basic data for rule mining, including: Perform spatiotemporal alignment of multi-source heterogeneous operation data and correlate the spatial positions of corresponding equipment in the production process to build a process correlation map; Based on the historical production process data of petroleum refining, the process correlation map is calibrated to extract the statistical characteristics, time series characteristics and correlation characteristics of multi-source heterogeneous operation data; Generate rules to mine basic data based on statistical features, time series features and association features.

6. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 5, characterized in that: Based on historical production process data of petroleum refining, the process correlation map is calibrated to extract statistical features, time series features, and correlation features of multi-source heterogeneous operation data, including: Based on the source equipment, collection time, and process flow relationships of historical production process data of petroleum refining, the equipment nodes and process relationship edges corresponding to the data are determined to establish an association index between the generated process data and the graph, so as to calibrate the process association graph and obtain the association topology index of the refining process data; Based on the equipment node attribute values ​​in the topological index of the refining process data, descriptive statistical calculations are performed on the multi-source heterogeneous operation data to extract the statistical characteristics of the multi-source heterogeneous operation data; Based on the time series edge weights in the topological index of the refining process data, the state transition evaluation is performed on the multi-source heterogeneous operation data to extract the time series characteristics of the multi-source heterogeneous operation data; Based on the weighted directed edges in the association topology index of refining process data, causal association mining is performed on multi-source heterogeneous operation data to extract the association features of multi-source heterogeneous operation data.

7. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 5, characterized in that: Generate rules based on statistical features, time series features, and association features to mine basic data, including: Encode statistical features, temporal features and correlation features and fuse them to obtain a fused feature vector; The fused feature vector is normalized by topological centrality, causal strength, and association rule confidence to generate basic data for rule mining.

8. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 1, characterized in that: Obtain and parse the petroleum refining intelligent production process description file to obtain process flow sorting and correlation characteristics, including: Obtain the petroleum refining intelligent production process description file and parse it to generate a set of structured entity relationship triples; Based on the structured entity relationship triple set, the process flow sorting and association relationship features are generated.

9. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 1, characterized in that: Based on the process flow sorting and association relationship characteristics, the association topology structure between the intelligent production equipment of petroleum refining is constructed, and the production impact weight is added to the association topology result to generate a weighted equipment association topology structure, including: Perform equipment entity recognition and spatial coordinate mapping on the process flow sequencing and association relationship features to generate the equipment space topology base map; The process logic association edge is constructed on the equipment space topology base graph to generate a process association enhanced graph; Perform multi-dimensional weight quantization and fusion processing on the process correlation enhancement graph to generate a topological graph with initial weights; Industrial constraint optimization and dynamic adjustment are performed on the topology graph with initial weights to generate a device association topology structure with weights.

10. The method for mining alarm event rules in the intelligent manufacturing industry according to claim 1, characterized in that: Generate an executable alarm rule base based on the weighted device association topology, including: A multi-dimensional labeling system is constructed for the weighted device association topology to generate a device label dataset; Perform high-frequency alarm data screening on the equipment label dataset to generate high-frequency item set data; Perform tree structure index construction on high-frequency item set data to generate an association rule tree model; Perform conditional path backtracking on the association rule tree model to generate a causal rule set; The causal rule set is mapped to a business baseline to generate an executable alarm rule base.

Images