Safety monitoring and fault tracing method for petrochemical instrument data abnormity

Abstract

The invention discloses a safety monitoring and fault tracing method for petrochemical instrument data abnormity. The method comprises the following steps: S1, constructing a petrochemical instrument data change dynamic knowledge graph; s2, performing instrument state change monitoring according to the intrinsic change of the instrument data; and S3, inference is carried out by fusing the petrochemical instrument data variation dynamic knowledge graph and the instrument state variation monitoring result, and safety monitoring and fault tracing of petrochemical instrument data abnormality are realized. By adopting the technical scheme of the invention, the weak link of the existing petrochemical instrument safety monitoring technology is broken through, and the problems that the instrument abnormity and the equipment fault are difficult to separate, the tiny fault is difficult to capture in time and the false alarm and the missed alarm are frequent in the monitoring process are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of petrochemical production safety, and in particular relates to a method for safety monitoring and fault tracing of abnormal petrochemical instrument data. Background Art

[0002] Petrochemical plants are typical large and complex plants. Instruments are an important part of petrochemical plants. Different types of equipment and instruments in different links of the petrochemical production process have different safety risks. In the petrochemical production process, instruments are not only an important basis for understanding the equipment status, operating conditions, material quality and emissions, but also an important basis for monitoring the operation failures and accident hazards of the entire plant, and discovering abnormal conditions and catastrophic risks. In particular, interlocking instruments and control instruments bear the important responsibility of regulating the operating status and preventing major safety hazards. If the instrument fails or fails itself, it may induce serious consequences or even catastrophic accidents. Clarifying whether the instrument is normal or not is an important link to effectively prevent misjudgment of faults and avoid misoperation.

[0003] Due to the large number of instruments in petrochemical process industry equipment, the many hidden dangers that affect instrument safety, the many signs and manifestations of instrument failure, and the long process flow, long material transportation pipelines, and the involvement of flammable and explosive dangerous goods in the entire process, there are mostly complex interlocking relationships, feedback relationships, or causal relationships between instruments. The impacts of failures are intertwined and change dynamically over time, making it difficult for conventional fault tree analysis, causal tree analysis, and simple hierarchical network topology diagrams to fully reflect the complex multiple coupling and mutual cause and effect situations, increasing the difficulty of safety risk analysis of ethylene equipment and process.

[0004] In the production process of existing large-scale petrochemical plants, no matter whether the monitoring platform adopts the DCS mode or the FCS mode, due to the above-mentioned complexity and technical difficulty, the most widely used over-limit alarm method based on fixed threshold or over-limit graded alarm based on multi-level threshold in the field of instrument data change monitoring is not only difficult to accurately set the threshold, but also the monitoring method itself inevitably has limitations such as misjudgment, false alarm or missed judgment, missed alarm, etc., especially in the early stage of the change or when the amplitude of the change is less than 3 times the root variance of the measurement error, early fault signs are difficult to be discovered and weak faults are difficult to detect. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for safely monitoring petrochemical instrument data anomalies and tracing fault sources, so as to solve the problems that instrument anomalies are difficult to separate from equipment failures, minor failures are difficult to capture in time, and false alarms and missed alarms are frequent during the monitoring process.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A method for safety monitoring and fault tracing of abnormal petrochemical instrument data, comprising:

[0008] Step S1, constructing a dynamic knowledge graph of petrochemical instrument data changes;

[0009] Step S2, monitoring the abnormal change of the instrument state according to the intrinsic change of the instrument data;

[0010] Step S3: Inferring the petrochemical instrument data anomaly dynamic knowledge graph and the instrument state anomaly monitoring results to achieve safe monitoring of petrochemical instrument data anomalies and fault tracing;

[0011] Step S1 includes:

[0012] Based on petrochemical big data, an expandable data set of structured and unstructured data is obtained;

[0013] Based on the expandable data set, form the material, working condition, mode, and production process entity set E related to the instrument; the relationship set R that reflects the process relationship, structural relationship, and association relationship; and the attribute set A that reflects the quantification degree of the reaction process working condition and the normality of the object;

[0014] Using the three sets E, R, and A, we construct triples {(ě,§,ǔ):ě,ǔ∈E,§∈R} and {(ě,ζ,ǔ):ě,ǔ∈E;ζ∈A}, and perform entity disambiguation, coreference resolution, and knowledge merging on the triples to obtain a dynamic knowledge graph of petrochemical instrument data anomalies.

[0015] Iterative optimization of the knowledge graph is achieved based on the continuous reception of instrument data and process data sent by the petrochemical DCS system;

[0016] In step S2, the characteristic change data extracted from the structured data and the unstructured data are taken as the object, and the sliding window fault tolerance rate wave algorithm and the filter residual generation algorithm are used to realize the data-driven instrument state abnormal change monitoring;

[0017] In step S3, according to the petrochemical instrument data anomaly dynamic knowledge graph and the instrument state anomaly monitoring results, multi-mode fusion reasoning based on graph convolutional network GCN is adopted to realize the safety monitoring and fault tracing of petrochemical instrument data anomalies.

[0018] Preferably, the expandable data set for obtaining structured and unstructured data includes:

[0019] Build a data flow structure classifier;

[0020] Through the data stream structure classifier, the data streams of structured data and unstructured data are extracted from petrochemical big data;

[0021] The data stream is preprocessed by cleaning, deduplication, and formatting to form corresponding expandable data sets of structured and unstructured data.

[0022] Preferably, step S3 comprises:

[0023] Step S31, according to the petrochemical instrument data abnormality dynamic knowledge graph and the instrument state abnormality monitoring results, through the graph convolution network GCN, build a GCN reasoning engine based on the dynamic knowledge graph;

[0024] Step S32: Use the GCN reasoning engine based on the dynamic knowledge graph to construct the premise set and consequence set of the fault risk to achieve fault tracing and location.

[0025] Preferably, step S31 includes:

[0026] Based on the dynamic knowledge graph, a set of triples of abnormal knowledge graphs describing the fault sources of petrochemical equipment and instruments, the abnormal patterns of instrument data, and the relationship between fault sources and data abnormalities is constructed to form a graph dataset.

[0027] Based on the graph dataset, a multi-layer perceptron is constructed and trained to implement GCN modeling of instrument data anomalies and instrument failures, and a priori logical rules are introduced in the modeling process;

[0028] A relational graph convolutional network R-GCN is constructed as an editor to implement specific convolutional learning of the relationship between the fault entity and the neighboring entities in the knowledge graph, obtain the corresponding feature vector representation, and build a GCN reasoning engine based on the dynamic knowledge graph.

[0029] Preferably, step S32 includes:

[0030] Adopt the GCN reasoning engine based on dynamic knowledge graph to construct the premise set and consequence set of failure risk;

[0031] The historical data and historical failure cases of instruments and equipment are used for model training to obtain the fuzzy correlation between the premise set and the consequence set.

[0032] According to the premise set and consequence set, a supervised machine learning method is used to establish a fault detection and diagnosis model based on the combination of analytical redundancy and premise set triggering, and determine the fault attribute membership of each element of the consequence set;

[0033] Establish a fuzzy reasoning engine based on fuzzy correlation and fault membership;

[0034] The fuzzy inference engine is used to perform reasoning layer fusion to achieve fault source tracing and location.

[0035] The present invention starts with petrochemical big data, and realizes safety monitoring and cause tracing for instrument data mutations through the construction of instrument fault dynamic knowledge graph, instrument data sliding fault-tolerant filtering, and the fusion and reasoning of instrument fault dynamic knowledge graph and instrument data mutation monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0037] Figure 1 This is a flow chart of a method for safety monitoring and fault tracing of abnormal petrochemical instrument data according to an embodiment of the present invention;

[0038] Figure 2 Schematic diagram of the process of building a knowledge graph for petrochemical instruments;

[0039] Figure 3 This is a schematic diagram of the instrument abnormality monitoring process. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Embodiment 1:

[0043] like Figure 1 As shown, an embodiment of the present invention provides a method for safety monitoring and fault tracing of abnormal petrochemical instrument data, including:

[0044] Step S1, constructing a dynamic knowledge graph of petrochemical instrument data changes;

[0045] Step S2, monitoring the abnormal change of the instrument state according to the intrinsic change of the instrument data;

[0046] Step S3: Inferring the petrochemical instrument data anomaly dynamic knowledge graph and the instrument status anomaly monitoring results to achieve safe monitoring of petrochemical instrument data anomalies and fault tracing.

[0047] As an implementation method of an embodiment of the present invention, in step S1, a petrochemical instrument data mutation dynamic knowledge graph is constructed from the bottom up, specifically: based on petrochemical big data, a petrochemical data stream structure classifier is established; structured data and semi-structured data are extracted from the petrochemical big data through the data stream structure classifier; the data stream is cleaned, deduplicated, and formatted to form corresponding expandable data sets of structured and unstructured data respectively; knowledge is extracted from the data set to form an entity set E of materials, working conditions, modes, production processes, etc. related to petrochemical instruments, a relationship set R that reflects process relationships, structural relationships, and association relationships, and an attribute set A of the quantification degree of reaction process conditions and the normality of objects; using E, R, and A Three sets are constructed to construct triples {(ě,§,ǔ):ě,ǔ∈E,§∈R} and {(ě,ζ,ǔ):ě,ǔ∈E;ζ∈A}, and entity disambiguation, coreference resolution and knowledge merging are performed on the above triple sets to achieve knowledge fusion; the concepts, terms and their relationships in the petrochemical instrument fault knowledge graph are defined and constructed to form the fault knowledge base architecture of petrochemical instruments; the existing knowledge of instrument faults is expanded and new knowledge is obtained through information extraction, knowledge fusion and knowledge reasoning; the obtained knowledge graph is stored in a graph database; while continuously receiving instrument data and process data sent by the petrochemical DCS system, the incremental data information is combined to achieve iterative optimization of the knowledge graph and complete the dynamic knowledge graph of the petrochemical instrument in-service process, such as Figure 2 shown.

[0048] As an implementation method of the embodiment of the present invention, Figure 3 As shown, in step S2, for the petrochemical instrument sampling data sequence {Y(t i ), i = 1, 2, ..., N}, the embodiment of the present invention adopts a safe and reliable sliding window fault-tolerant filtering and a residual generation and data anomaly monitoring algorithm based on fault-tolerant filtering:

[0049] Design a sliding window three-segment middle-segment fault-tolerant filtering method: Specifically, construct a sliding window with a window width of s=3k+2, where k is an adjustable parameter (the default value is 5). For the half-open interval time window [t i ,t i+s ) within the instrument sampling data fragment X i = {Y (t j ),t j ∈D i}Sort in ascending order, and record the sorted data fragment as calculate Mean filter Y of the middle k points i med and filter residual

[0050]

[0051] The sliding filter has strong fault tolerance for outliers and local spots. The residual sequence formed by the residual obtained by the sliding window fault-tolerant filter is The threshold parameter C is constructed using a recursive algorithm i

[0052]

[0053] In the formula

[0054]

[0055] Based on this, the abnormal changes of instrument data can be monitored and diagnosed:

[0056] when The data is considered normal when The data is suspected to be abnormal. When the data is reported abnormally high, When the data is high, it is abnormal; when The data is suspected to be abnormal. The data is underreported abnormally. The high alarm is abnormal.

[0057] The monitoring method of the embodiment of the present invention can quickly realize early symptom detection, small abnormal change identification and fault isolation, and effectively overcome the problems of false alarm and missed alarm existing in the fixed threshold alarm method commonly used in the existing DCS or FCS system instrument data monitoring, as well as the limitations such as difficulty in realizing abnormal change monitoring within the threshold.

[0058] As an implementation method of the present invention, in step S3, the source of multiple causes of abnormal changes in petrochemical instrument data based on the deep fusion of graph neural network and dynamic knowledge graph is traced, and the advantages of graph neural network models such as graph convolution network and graph attention network in processing graph structure data are utilized to realize the source of the causes of instrument data abnormal changes. Specifically, the causes of instrument in-service data abnormal changes include:

[0059] Step S31, design of inference engine based on graph convolutional network GCN. Based on dynamic knowledge graph, construct a set of triples of mutation knowledge graphs describing the fault sources of petrochemical equipment and instruments, the mutation patterns of instrument data, and the relationship between fault sources and data mutations to form a graph data set; based on the graph data set, construct and train multi-layer perceptrons to realize GCN modeling of instrument data mutations and instrument faults, and introduce prior logic rules in the modeling process; based on the historical data of petrochemical instruments and the dynamic knowledge graph of faults, give a summary module of the relationship between entities and a reasoning module for state inference based on this; construct a relational graph convolutional network R-GCN as an editor to realize the convolutional learning of the specific relationship between the fault entity and the neighbor entity in the knowledge graph, obtain the corresponding feature vector representation, and then construct a GCN inference engine based on the dynamic knowledge graph.

[0060] Step S32, multi-mode fusion reasoning engine design. The embodiment of the present invention adopts a multi-mode fusion reasoning method to achieve the tracing and positioning of instrument and equipment faults. Specifically, a GCN reasoning engine based on a dynamic knowledge graph is used to construct a premise set and a consequence set of fault risks; historical data and historical fault cases of instruments and equipment are used for model training to obtain the fuzzy correlation of the connection relationship between the premise set and the consequence set; based on the premise set and the consequence set, a supervised machine learning method is used to establish a fault detection and diagnosis model based on the combination of analytical redundancy and premise set triggering, thereby determining the fault attribute membership of each element of the consequence set; the technical kernel of the fault reasoning engine based on Boolean operations is absorbed to establish a fuzzy reasoning engine based on fuzzy correlation and fault membership, and the reasoning layer is integrated with the machine learning model reasoning engine and the Boolean operation reasoning engine to achieve fault tracing and positioning. The diagnostic reasoning method of reasoning layer fusion can effectively avoid the limitations of conventional forward reasoning and reverse reasoning that are difficult to use in situations such as network associations between causes and effects, and effectively overcome the influence of multiple coupling mutual cause and effect relationships between detection devices and between detection devices and the objects and controlled processes, thereby achieving the tracing and positioning of instrument and equipment faults.

[0061] Embodiment 2:

[0062] The embodiment of the present invention provides a method for safety monitoring and fault tracing of abnormal petrochemical instrument data, which is divided into three stages:

[0063] Phase 1: Building a dynamic knowledge graph for petrochemical instrument data mutation, including:

[0064] Step 11: Use the Clementine data mining platform to mine the variation patterns of different types of instrument data, and sort out the causes of instrument data variation and the root causes of related equipment failures.

[0065] Step 12: Using a bottom-up approach, extract fault sources (entities), data anomaly patterns (entities), and the relationship between fault sources and data anomalies (attributes), as well as the relationship between entities and between entities and attributes from petrochemical big data, and add them to the object knowledge base to form a data layer.

[0066] Step 13: Form a model layer through data layer concept abstraction, and build a dynamic knowledge graph system of abnormalities / failures with instrument data anomalies and equipment failures as core nodes.

[0067] Step 14: Comply with Gruber's completeness and consistency principles, combine data-driven incremental ontology construction technology, and start from operating data mutations and failure cases, use Protégé5.5 ontology editing tool to realize ontology model construction.

[0068] Step 15: Use the Neo4j graph database and Cypher query language tool to model the information of graph data nodes, relationships between nodes, node-included attributes, and relationship-included attributes, and form a graph database of potential fault hazards, fault symptoms, and their associated relationships for the ethylene production process, as well as a fast and efficient query mechanism for the graph database.

[0069] Step 16. According to the Neo4j database storage method, the basic elements in the ontology model and the Neo4j graph model are mapped through node mapping, relationship mapping and attribute mapping to establish a mapping mechanism from ontology to Neo4j. The knowledge representation form of the ontology model is mapped to a knowledge representation form that can be stored in the Neo4j database. The ontology model is combined with the Neo4j graph database to realize the construction of the fault dynamic knowledge graph.

[0070] Step 17. Develop a safety knowledge base for equipment by constructing a dynamic knowledge graph of instrument and equipment failures in petrochemical plants, and use a set of graph elements such as "nodes" and "edges" to connect different types of information together, to obtain a GCN network that can fully reflect the multiple coupling, mutual cause and effect relationships of failures in the petrochemical production process, which is used to overcome technical difficulties such as the difficulty in identifying and tracing safety hazards caused by multiple accident causes and complex impact relationships.

[0071] Phase 2: Taking the feature change data extracted from structured data and unstructured data as the object, using the sliding window fault tolerance wave algorithm and filter residual generation algorithm, realize data-driven instrument state change monitoring and diagnosis. Specifically include:

[0072] Step 21. Considering that there may be outliers and patches in the field instrument data, which may affect the actual monitoring, a sliding window fault-tolerant filtering algorithm with strong fault tolerance for outliers and patches is adopted. Through the sliding fault-tolerant filtering of the instrument data, the separation of the intrinsic changes of the instrument data and the disturbance components (including outliers and patches) is realized, the identification and cleaning of outliers and patches in the multi-instrument sampling data stream is completed, the intrinsic change curve of the instrument data is extracted, and the fault-tolerant filtering residual sequence of the instrument data is generated.

[0073] Step 22: Use industrial big data mining and machine learning algorithms, combined with process characteristic models and association relationship models, to determine the dynamic change envelope of instrument data; based on the envelope, extract the time-varying thresholds and change curves based on historical data for different types of instruments under different production processes and working conditions, and perform intelligent reconstruction of the time-varying thresholds; through transfer learning, it is used for online monitoring of abnormal changes in data of field instruments in the production process of petrochemical plants.

[0074] Step 23. Combined with the dynamic knowledge graph, analyze the changes in instrument data such as temperature, pressure, flow and liquid level in the petrochemical production process, and their relationship with material input, process control, material output and external disturbance, and construct a recursive neural network (DA-RNN) based on a dual-stage attention mechanism. In the first stage, the attention mechanism adaptively extracts relevant features, and in the second stage, the attention mechanism selects the encoder hidden state related to it. By integrating the above two types of attention in the LSTM-based recursive neural network RNN ​​and performing joint training through back propagation, adaptive learning modeling and model training of features are realized.

[0075] Step 24: Adopt the DA-RNN model, and combine it with adaptive selection of the most relevant input and appropriate capture of the long-range dependency of the time series, input the actual operating data before the current moment and the exogenous drive and disturbance sequence, and form a multi-dimensional instrument data sequence based on the model and data drive; form a comparison residual with the intrinsic change sequence of the instrument data, and through the transfer learning of the residual sequence, intelligently reconstruct the time-varying threshold, judge the situation that the residual exceeds the threshold, and realize the online detection of process abnormalities.

[0076] Step 25, using the original measurement data of the instrument and the sliding window fault-tolerant filtering result to form a residual sequence, the residual sequence contains the measurement data disturbance and outlier and spot information; the fault-tolerant 3σ criterion is used to identify the outliers and spots for the residual sequence with outliers and spots; the residual graphic analysis and statistical hypothesis testing method are used to perform statistical diagnosis on the outlier residual sequence to achieve separation of residual non-stationary components and inference of disturbance amplitude.

[0077] Phase 3: Combine the dynamic knowledge graph of instrument abnormality with the monitoring and diagnosis of instrument status abnormality to build a multifunctional integrated system and realize the source tracing of instrument abnormality causes driven by petrochemical big data. Specifically include:

[0078] Step 31, design of inference engine based on graph convolutional network GCN. Based on the dynamic knowledge graph of instrument abnormality, construct a set of triples of abnormality knowledge graphs describing the fault sources of petrochemical equipment and instruments, the abnormality patterns of instrument data, and the relationship between the fault sources and data abnormalities to form a graph data set; based on the graph data set, construct and train a multi-layer perceptron to implement GCN modeling of instrument abnormalities and equipment failures and introduce prior logic rules in the modeling process; based on the historical data of petrochemical instruments and the dynamic knowledge graph of faults, give a summary module of the relationship between entities and an inference module for state inference based on this; construct a relational graph convolutional network R-GCN as an editor to implement convolutional learning of fault entities and specific link relationships in the knowledge graph, obtain the corresponding feature vector representation, and then construct a GCN inference engine based on the dynamic knowledge graph.

[0079] Step 32, design of multi-mode fusion inference engine. Based on the GCN inference engine of dynamic knowledge graph, construct the premise set and consequence set of fault risk; according to the historical data of instruments and related equipment and the historical fault case training model, obtain the fuzzy correlation between the premise set and the consequence set; based on the premise set and the consequence set, adopt the supervised machine learning method to establish a fault detection and diagnosis model based on the combination of analytical redundancy and premise set triggering, and determine the fault attribute membership of each element of the consequence set; absorb the technical core of the inference engine based on Boolean operation, establish a fuzzy inference engine based on fuzzy correlation and fault membership, realize the inference layer fusion of instrument data-driven GCN inference engine, machine learning inference engine and Boolean operation inference engine, overcome the limitation that conventional forward reasoning and reverse reasoning are difficult to use in situations such as mesh association between cause and effect, and realize the fast reasoning of multi-mode fusion of inference layer.

[0080] Step 33, design of a multi-cause anomaly tracing system integrating instrument data and dynamic knowledge graph. Starting from petrochemical big data, through the construction of dynamic knowledge graph of instrument failure, sliding fault-tolerant filtering of instrument data, and the integration and reasoning of dynamic knowledge graph of instrument failure and instrument data anomaly monitoring, a safety monitoring and cause tracing system for instrument data anomalies is established, forming an intelligent support platform for instrument safety detection and cause tracing, which is used for tracing instrument failures and equipment failures and locating different source failures.

[0081] Step 34, system integration testing and verification. Considering that the entire petrochemical process is inseparable from flammable, explosive, and corrosive dangerous goods, in line with the principle of "safe experiment to ensure the effect", after the system is built, the process post-playback and the actual measured data of the working condition are combined to implement the research process experiment and results verification. First, the instrument data safety monitoring and inducement tracing process are experimentally verified by post-playback of historical data to determine the effectiveness and safety of the monitoring method and the tracing algorithm; secondly, the real-time data in the furnace and the tube are collected for the key components of the cracking furnace during shutdown and burning, and combined with the monitoring data before the shutdown, the abnormal changes of the equipment local hidden danger instruments are experimentally verified; then, the OPC interface software is used to read the DCS data, analyze the dynamic response, steady-state response and chain response of complex objects such as control instruments and chain instruments, and conduct physical in-loop testing and verification to comprehensively evaluate and verify the safety and stability of the system.

[0082] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for safety monitoring and fault tracing of abnormal petrochemical instrument data, characterized in that: include: Step S1, constructing a dynamic knowledge graph of petrochemical instrument data changes; Step S2, monitoring the abnormal change of the instrument state according to the intrinsic change of the instrument data; Step S3: Inferring the petrochemical instrument data anomaly dynamic knowledge graph and the instrument state anomaly monitoring results to achieve safe monitoring of petrochemical instrument data anomalies and fault tracing; Step S1 includes: Based on petrochemical big data, an expandable data set of structured and unstructured data is obtained; Based on the expandable data set, form the material, working condition, mode, and production process entity set E related to the instrument; the relationship set R that reflects the process relationship, structural relationship, and association relationship; and the attribute set A that reflects the quantification degree of the reaction process working condition and the normality of the object; Using the three sets E, R, and A, we construct triples {(ě,§,ǔ):ě,ǔ∈E,§∈R} and {(ě,ζ,ǔ):ě,ǔ∈E;ζ∈A}, and perform entity disambiguation, coreference resolution, and knowledge merging on the triples to obtain a dynamic knowledge graph of petrochemical instrument data anomalies. Iterative optimization of the knowledge graph is achieved based on the continuous receipt of instrument data and process data sent by the petrochemical DCS; In step S2, the characteristic change data extracted from the structured data and the unstructured data are taken as the object, and the sliding window fault tolerance rate wave algorithm and the filter residual generation algorithm are used to realize the data-driven instrument state abnormal change monitoring; In step S3, according to the petrochemical instrument data anomaly dynamic knowledge graph and the instrument state anomaly monitoring results, multi-mode fusion reasoning based on graph convolutional network GCN is adopted to realize the safety monitoring and fault tracing of petrochemical instrument data anomalies.

2. The petrochemical instrument data anomaly safety monitoring and fault tracing method according to claim 1, characterized in that: Expandable datasets that provide structured and unstructured data include: Build a data flow structure classifier; Through the data stream structure classifier, the data streams of structured data and unstructured data are extracted from petrochemical big data; The data stream is preprocessed by cleaning, deduplication, and formatting to form corresponding expandable data sets of structured and unstructured data.

3. The petrochemical instrument data anomaly safety monitoring and fault tracing method according to claim 2, characterized in that: Step S3 includes: Step S31, according to the petrochemical instrument data abnormality dynamic knowledge graph and the instrument state abnormality monitoring results, through the graph convolution network GCN, build a GCN reasoning engine based on the dynamic knowledge graph; Step S32: Use the GCN reasoning engine based on the dynamic knowledge graph to construct the premise set and consequence set of the fault risk to achieve fault tracing and location.

4. The method for safety monitoring and fault tracing of abnormal petrochemical instrument data as claimed in claim 3, characterized in that: Step S31 includes: Based on the dynamic knowledge graph, a set of triples of abnormal knowledge graphs describing the fault sources of petrochemical equipment and instruments, the abnormal patterns of instrument data, and the relationship between fault sources and data abnormalities is constructed to form a graph dataset. Based on the graph dataset, a multi-layer perceptron is constructed and trained to implement GCN modeling of instrument data anomalies and instrument failures, and a priori logical rules are introduced in the modeling process; A relational graph convolutional network R-GCN is constructed as an editor to implement specific convolutional learning of the relationship between the fault entity and the neighboring entities in the knowledge graph, obtain the corresponding feature vector representation, and build a GCN reasoning engine based on the dynamic knowledge graph.

5. The method for safety monitoring and fault tracing of abnormal petrochemical instrument data as claimed in claim 4, characterized in that: Step S32 includes: Adopt the GCN reasoning engine based on dynamic knowledge graph to construct the premise set and consequence set of failure risk; The historical data and historical failure cases of instruments and equipment are used for model training to obtain the fuzzy correlation between the premise set and the consequence set. According to the premise set and consequence set, a supervised machine learning method is used to establish a fault detection and diagnosis model based on the combination of analytical redundancy and premise set triggering, and determine the fault attribute membership of each element of the consequence set; Establish a fuzzy reasoning engine based on fuzzy correlation and fault membership; The fuzzy inference engine is used to integrate the reasoning layers and achieve fault tracing and location.

Images