Chemical process module partitioning and fault tracing method based on topology-attribute fusion

Abstract

The application discloses a kind of based on topological-attribute fusion chemical process module division and fault tracing method.The application extracts stream attribute and topological connection information by analyzing chemical process simulation data, constructs directed graph model, quantifies the attribute difference of node in and out stream, constructs multidimensional node feature vector;Fusion topological structure similarity and node physical property similarity, construct comprehensive similarity matrix;Calculate node importance score;Reconstruct directed graph and extract connected branch, realize automatic module division.Further map real-time running data to each functional module, identify abnormal module, and determine candidate fault source according to stream flow direction, node input-output deviation change and importance score.The application deeply couples chemical physical mechanism and network topological characteristics, effectively overcomes the defects such as strong subjectivity and low efficiency of artificial division, realizes automatic module division of complex chemical process, fault diagnosis range narrowing and fault source positioning.

Description

Technical Field

[0001] This invention relates to the field of computer-aided chemical engineering design and process systems engineering, specifically to a method for dividing chemical process modules and tracing fault sources based on topology-attribute fusion. Background Technology

[0002] Modern chemical processes typically involve a vast number of chemical equipment and unit operations, which are intertwined through complex material and energy flows, forming a highly coupled and complex network system. In the development and operation of chemical processes, steady-state simulation is fundamental to obtaining the overall material and energy balance. However, as the scale of the process increases, the nonlinearity and coupling of the global system increase dramatically, posing a significant "high-dimensional curse" to subsequent dynamic simulation modeling, the design of distributed process control systems, and the optimization and fault diagnosis of the entire process.

[0003] To reduce the difficulty of analyzing and solving complex systems, the field of chemical systems engineering typically employs the concepts of "system decomposition" or "modularization," dividing a vast end-to-end network into several internally interconnected but externally loosely coupled subsystems (functional modules). Modularization reduces the complexity of end-to-end modeling and optimization to some extent, facilitating engineers' understanding of the process structure, determination of control ranges, and organization of process units. It also enables rapid and accurate analysis of abnormal operating conditions and troubleshooting and tracing of operational failures during actual industrial operation.

[0004] However, existing modular approaches are increasingly revealing significant limitations when dealing with modern, highly integrated chemical processes. Traditional chemical process zoning often relies heavily on human experience, with engineers making subjective judgments and manually dividing processes according to physical areas or approximate reaction and separation sections. This approach is not only inefficient, but also prone to severing previously close implicit physical connections due to limitations in experience when dealing with processes involving a large amount of complex thermodynamic reflux and material circulation, leading to the failure of subsequent dynamic control strategies.

[0005] Currently, research on the deep integration of graph theory with the functional modularization of chemical processes remains very scarce. More importantly, traditional complex network community detection typically equates chemical processes to conventional directed transport networks, focusing solely on whether equipment nodes are "connected" or simply on "flow direction." This approach, biased towards pure mathematical topology, ignores the essential characteristics of chemical processes that distinguish them from general networks, namely the complex state evolution and material transformations occurring within equipment. Because it fails to effectively extract and integrate the differences in thermodynamic properties and physical states of the flow streams before and after passing through nodes, subsystems defined solely by network connectivity often fail to reflect the inherent functional synergy between chemical units. Consequently, the final module partitioning lacks practical guiding significance in subsequent dynamic modeling, control zoning, operational analysis, and fault tracing.

[0006] Fault tracing in chemical processes refers to the process of quickly locating the source of an anomaly or malfunction and finding its root cause during chemical production. However, in the operation of chemical plants, anomalies in local equipment or streams often propagate along the material flow direction, causing fluctuations at multiple measuring points simultaneously. Fault tracing often requires directly addressing all equipment and variables in the entire process, resulting in a large scope and low location efficiency, which is extremely detrimental to process safety and the restoration of normal production. Therefore, physically meaningful modular division should not only serve process understanding and model dimensionality reduction but also be able to incorporate actual operating data to provide a foundation for subsequent anomaly module identification, narrowing down the investigation scope, and locating the fault source.

[0007] Therefore, there is an urgent need in this field to develop an intelligent modular method that can deeply couple the underlying physical mechanisms of chemical engineering with the topological characteristics of spatial networks. This method would overcome the limitations of purely mathematical topological partitioning, which lacks engineering significance, and enable the modular decomposition of complex chemical processes with clear physical meaning. This would provide a scientific and rigorous basis for subsequent distributed optimization, dynamic control, and fault diagnosis throughout the entire process. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent modularization method and system for chemical processes based on topology-attribute fusion. This addresses the problems of current module partitioning relying heavily on human experience, leading to high subjectivity and low efficiency, and the lack of practical engineering guidance due to the detachment of pure network topology algorithms from the underlying mechanisms of chemical processes. This invention deeply couples the flow topology of a chemical process with its underlying physicochemical properties, driving the adaptive reconstruction of complex directed graph networks. This enables automated module decomposition and core node identification of complex chemical processes, significantly improving the scientific rigor and efficiency of subsequent full-process distributed optimization, dynamic control model construction, and fault tracing.

[0009] To achieve the above objectives, the present invention proposes the following technical solution:

[0010] This invention first provides a method for dividing chemical process modules and tracing faults based on topology-attribute fusion, which includes the following steps:

[0011] 1) Based on the steady-state operating data of the chemical process to be divided, extract the attribute data of each stream and the topological connection relationship between the equipment; take the equipment in the chemical process as graph nodes, the streams connecting the equipment as directed edges, construct a directed graph model representing the chemical process, and assign the attribute data of the streams to the corresponding directed edges.

[0012] 2) For each node in the directed graph model, extract the attribute differences between its inflow and outflow flows, quantify the attribute differences between the inflow and outflow flows, and construct the feature vector of each graph node.

[0013] 3) Extract the topological connections driven by traffic in the directed graph model, and calculate the structural similarity and attribute similarity between graph nodes by combining the distance between graph nodes, and then fuse them to obtain a comprehensive similarity matrix;

[0014] 4) Based on the comprehensive similarity matrix, and combined with the topological features of the graph structure and the device functional features corresponding to the graph nodes, the final importance score of each graph node is obtained;

[0015] 5) Point each graph node in the directed graph model to the target node with the higher final importance score and the highest comprehensive similarity, thereby reconstructing and generating a simplified graph structure. Extract each connected component in the simplified graph structure as a functional module to obtain the module division result of the chemical process.

[0016] 6) Based on the module division results, monitor the real-time operation data of the chemical process. When an abnormality occurs, first identify the abnormal functional module, then generate a sequence of candidate fault sources and conduct fault investigation in sequence to identify the faulty equipment.

[0017] According to a preferred embodiment of the present invention, step 6) specifically includes: calculating the mean vector and covariance matrix of the state vector of each stream based on historical steady-state operating data; acquiring real-time operating data of each stream in the chemical process, and calculating the stream anomaly score based on the mean vector and covariance matrix; obtaining the module anomaly score based on the stream anomaly score within each functional module, and identifying functional modules whose module anomaly scores exceed a set threshold as abnormal functional modules; for each graph node within an abnormal functional module, constructing a candidate fault source score based on its input-side average anomaly score, output-side average anomaly score, and final importance score, and sorting the candidate fault source scores to obtain a candidate fault source sequence; and then sequentially troubleshooting each device according to the candidate fault source sequence to determine the faulty device.

[0018] The present invention also provides a chemical process module division and fault tracing system for implementing the aforementioned method, the system comprising:

[0019] Data parsing and modeling module: used to parse chemical steady-state process simulation data, extract topological structure and stream attribute data, and construct directed graph models;

[0020] Physical feature calculation module: used to extract the differences in inbound and outbound flow attributes of process nodes and generate multi-dimensional node feature vectors;

[0021] Feature fusion and importance assessment module: used to calculate and fuse the structural similarity and attribute similarity between graph nodes to obtain a comprehensive similarity matrix, and calculate the final importance score of the nodes based on topological redundancy and device functional features;

[0022] Module partitioning and graph reconstruction module: used to reconstruct simplified graph structures based on node importance scores and comprehensive similarity, and extract connected components to output module partitioning results;

[0023] Fault diagnosis and tracing module: Calculates the abnormality score of each stream based on the steady-state operating data and real-time operation data of each stream, and further obtains the module abnormality score, thereby identifying abnormal functional modules. Based on the average abnormality score on the input side, the average abnormality score on the output side, and the final importance score of each graph node in the abnormal functional module, constructs the candidate fault source score, and outputs the candidate fault source sequence according to the candidate fault source score.

[0024] Through the above technical solution, the present invention has the following significant advantages:

[0025] 1. This invention realizes a data-driven automated module partitioning of chemical processes. By directly parsing steady-state process simulation data, it relies on the method to adaptively drive the reconstruction and connection branch extraction of complex directed graph networks, overcoming the shortcomings of traditional chemical process partitioning, which is highly dependent on human experience, subjective, and inefficient.

[0026] 2. This invention achieves a deep integration of topological structure and the underlying physical mechanism of chemical engineering. It innovatively introduces a physicochemical operator based on the principle of three transmissions and one reaction, which accurately quantifies the thermodynamic and compositional differences before and after the equipment node into multi-dimensional node characteristics. This effectively solves the problem that pure mathematical topology algorithms are divorced from chemical engineering practice and lack engineering guidance significance.

[0027] 3. This invention achieves precise location and extraction of core chemical units. By introducing topological redundancy and equipment function penalty mechanisms, it adaptively corrects the importance score of nodes, effectively filtering out interference from non-core equipment and ensuring that each module is built around the core equipment, which is highly consistent with the engineering logic of core operation units in chemical production.

[0028] 4. This invention effectively reduces the analysis and solution dimensions of complex end-to-end systems, decouples highly coupled complex chemical networks into multiple physical subsystems, alleviates the "curse of high dimensionality" in complex modeling, and provides a scientific and rigorous basis for subsequent end-to-end distributed optimization and dynamic control.

[0029] 5. This invention improves the targeting of fault tracing and diagnosis in chemical processes. By combining the module division results with chemical process operation data, abnormal functional modules can be identified first. Then, within the abnormal functional modules, candidate fault sources can be determined based on material flow direction, changes in node input and output deviations, and the final importance score of the nodes. This narrows down the scope of fault investigation and provides a basis for fault diagnosis and tracing in chemical processes. Attached Figure Description

[0030] Figure 1 This is a flowchart of the intelligent modularization method for chemical processes based on topology-attribute fusion, as described in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the entire process of preparing biofuel from vegetable oil in an example of the present invention;

[0032] Figure 3 This refers to the traditional manually defined modules and partitioning results based on expert experience.

[0033] Figure 4 This is the modular division result of the present invention;

[0034] Figure 5 This is a comparison of the modularity of the vegetable oil-to-biofuel process in the examples of this invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. The specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0036] like Figure 1 As shown, in a preferred embodiment, a method for partitioning chemical process modules and tracing faults based on topology-attribute fusion is provided, comprising the following steps:

[0037] 1) Based on the steady-state operating data of the chemical process to be divided, extract the attribute data of each stream and the topological connection relationship between the equipment; take the equipment in the chemical process as graph nodes, the streams connecting the equipment as directed edges, construct a directed graph model representing the chemical process, and assign the attribute data of the streams to the corresponding directed edges.

[0038] Specifically, this embodiment reads and parses the steady-state operating data of the chemical process to be divided into modules (wherein the steady-state operating data can be steady-state data from actual industrial production or simulation data, such as bkp format files simulated by Aspen Plus). The chemical equipment in the process is extracted as a set of graph nodes. The material flow streams connecting the equipment are treated as a set of directed edges. To satisfy the closed computation requirement of graph theory, virtual nodes are constructed and filled in for the initial feed and final discharge streams of the process (virtual nodes are not assigned feature vectors). The set of directed edges is defined as follows: Each directed edge Defined as a triple , As a unique identifier for the flow, and These are the source device nodes and the target device nodes, respectively. For the initial feed and final discharge streams in a chemical process, missing source or target devices are filled by constructing virtual nodes, and these virtual nodes do not have feature vectors. Thermodynamic and physical properties of the streams are extracted simultaneously to construct a feature matrix. Each row vector of this matrix Corresponding Flow The attribute information (k represents the stream sequence number), specifically determined by the temperature of that stream. ,pressure mass flow rate and the mass fraction vector of each component The edges are constructed by splicing together the components and then assigned as feature weights to the corresponding directed edges. Generate a directed graph model of a chemical process with physical properties. .

[0039] 2) For each node in the directed graph model, extract the attribute differences between its inflow and outflow streams, quantify the attribute differences between the inflow and outflow streams, and construct the feature vector of each graph node.

[0040] Specifically, for each real device node in the directed graph model, this embodiment calls the built-in physicochemical operators to quantify the attribute differences of the inflow and outflow streams, and constructs a multi-dimensional feature vector characterizing the three-transmission-one-reaction mechanism. .

[0041] The specific calculations are as follows:

[0042] 2-1) Calculate the absolute change in mass flow rate of each component in the directed edges before and after the node to obtain the normalized reaction intensity index. :

[0043]

[0044] in, Indicates the component number in a chemical process; This indicates the first of all input streams entering this node. Mass flow rate of each component; This represents the first of all input streams flowing out of this node. Mass flow rate of each component;

[0045] 2-2) Calculate the flow-weighted average pressure of the directed edges before and after the node. and weighted average temperature It refers to the set of directed edges for input or output. The calculation formula is:

[0046]

[0047] in, Used to distinguish between the input side and the output side; Indicates the flow of shares mass flow rate; Indicates the flow of shares The pressure; Indicates the flow of shares Temperature;

[0048] Based on this, the logarithmic ratio of the output to the input is calculated to obtain the pressure gain. and temperature gain :

[0049]

[0050] in, These represent the flow-weighted average pressure and flow-weighted average temperature of the input stream at the node, respectively. These represent the flow-weighted average pressure and flow-weighted average temperature of the streams on the node's output side, respectively.

[0051] 2-3) For a given component vector Its central logarithmic ratio transformation Defined as the sum of the components and the geometric mean The natural logarithm of the ratio, i.e.:

[0052]

[0053] in ;

[0054] Based on the central logarithmic ratio transformation, the component vectors of each component in the directed edge are calculated and output. relative total output weighted average component vector The flow-weighted Aitchison distance was used to obtain the separation strength index. :

[0055]

[0056] in, This represents the set of directed output edges flowing out of this node, that is, the set of all output streams from this node; This indicates the mass flow rate of the stream output by this node;

[0057] 2-4) Mass flow rate of total feed to the node Perform a logarithmic transformation to obtain the flow rate index. .

[0058] 3) Extract the topological connection relationships based on flow in the directed graph model, and calculate the structural similarity and attribute similarity between graph nodes by combining the distance between graph nodes, and then fuse them to obtain a comprehensive similarity matrix.

[0059] 3-1) Structural similarity: The flow-driven topological overlap matrix (Flow-TOM) method is used to calculate the similarity. The flow quality flow between graph nodes is logarithmically normalized to construct the flow weight matrix. Traffic weight matrix The middle corresponds to the graph node With graph nodes elements And calculate the common neighbor strength matrix accordingly. ; Calculate the weighted degree of the nodes Based on the common neighbor strength matrix Traffic weight matrix Calculating the weighted degree of graph nodes With graph nodes The topological overlap similarity between them is used as the structural similarity matrix. matrix elements :

[0060]

[0061] Structural similarity matrix Each element is restricted to Within the interval, the diagonal elements are set to 1.

[0062] 3-2) Attribute Similarity: Obtained by transforming the distance between the node feature vectors extracted in step 2) using a Gaussian kernel function, resulting in the attribute similarity matrix. The matrix elements are:

[0063]

[0064] in The squared Euclidean distance between the feature vectors of two nodes after alignment.

[0065] 3-3) Network Fusion: Comprehensive Similarity Matrix From the structural similarity matrix Similarity matrix with attributes The linearly weighted fusion is obtained, and its element representation is as follows:

[0066]

[0067] The weighting factor is set.

[0068] 4) Based on the comprehensive similarity matrix, and combined with the topological features of the graph structure and the functional features of the devices corresponding to the graph nodes, the final importance score of each graph node is obtained. .

[0069] Final Importance Score The basic score is determined by the magnitude of the feature vectors of the graph nodes. To filter out topology artifacts from high-connectivity non-core devices in the pipeline network, a penalty mechanism (topology metric penalty coefficient) is introduced, based on a weighted sum of similarity with neighbors. Penalty coefficient for auxiliary functions of equipment ), final importance score The formula is as follows:

[0070]

[0071] 4-1) Topology metric penalty coefficient For auxiliary nodes in the process that mainly serve to divert, merge, or connect, but do not undertake the main reaction or separation functions, if their connectivity... If the value is high, a negative exponential function is used for reduction; where... Indicates entering the node The number of upstream streams and the nodes The sum of the number of downstream streams flowing out. The redundancy penalty coefficient is... It is a minimal constant:

[0072]

[0073] 4-2) Equipment auxiliary function penalty coefficient Based on the reaction intensity in the eigenvector With separation strength If both indicators are less than a set threshold, a reduction coefficient is assigned; otherwise... .

[0074] 5) Point each graph node in the directed graph model to the target node with the higher final importance score and the highest overall similarity, thereby reconstructing and generating a simplified graph structure. Extract each connected component in the simplified graph structure as a functional module to obtain the module division result of the chemical process.

[0075] This embodiment drives adaptive decoupling of the entire network based on a comprehensive similarity matrix and the final importance score: First, virtual nodes are filtered out; for each retained real node... In the set of potential target nodes with records that have similarity to it In the process, the final importance score is higher than And the best target node with the highest overall similarity and establish by point to Directed connections are used to generate a simplified graph structure:

[0076]

[0077] If node If there is no potential target node with a final importance score higher than itself, no further pointing relationship will be established for it, and it will be regarded as a candidate core node; other nodes will be assigned to the established pointing relationships until they reach the candidate core node.

[0078] The simplified directed graph is then converted into an undirected graph, and connected components are extracted using a connected component algorithm. Each connected component is output as an independent functional module. Furthermore, within each module, the core nodes with the highest importance scores are extracted. This is used to characterize the core chemical operations of this module:

[0079] .

[0080] 6) Based on the module division results, monitor the real-time operation data of the chemical process. When an abnormality occurs, first identify the abnormal functional module, then generate a sequence of candidate fault sources and conduct fault investigation in sequence to identify the faulty equipment.

[0081] Step 6) specifically includes: calculating the mean vector and covariance matrix of the state vector of each stream based on historical steady-state operating data; obtaining real-time operating data of each stream in the chemical process, and calculating the stream anomaly score based on the mean vector and covariance matrix; obtaining the module anomaly score based on the stream anomaly score within each functional module, and identifying functional modules whose module anomaly scores exceed a set threshold as abnormal functional modules; for each graph node within an abnormal functional module, constructing a candidate fault source score based on its input-side average anomaly score, output-side average anomaly score, and final importance score, and sorting the candidate fault source scores to obtain a candidate fault source sequence; and then sequentially troubleshooting each device according to the candidate fault source sequence to determine the faulty device.

[0082] In one specific embodiment of the present invention, step 6 is carried out according to the following process. It should be noted that the following specific implementation steps are merely illustrative.

[0083] Specifically, let the set of functional modules obtained in step 5) be:

[0084]

[0085] in, Indicates the first Each functional module is assigned to a specific flow stream in the chemical process.

[0086] For flowing stock Define its time at time The running state vector is:

[0087]

[0088] in, , , and They represent the flow strands respectively. Temperature, pressure, flow rate and the first The content of each component.

[0089] Furthermore, based on the flow Sample set under normal operating conditions Calculate the mean vector and covariance matrix of the flow state vector:

[0090]

[0091]

[0092] in, For flow Normal operating condition mean vector For flow Normal operating condition covariance matrix;

[0093] Based on the mean vector and covariance matrix, the flow stream is defined. At any moment The abnormality score is:

[0094]

[0095] in, Here is the regularization constant. It is an identity matrix.

[0096] Set up functional modules The corresponding set of streams is Then the module anomaly score for this functional module is:

[0097]

[0098] in, For functional modules The number of midstream stocks. When the module receives an abnormal rating. Exceeding the set threshold At that time, the corresponding functional modules will be... This module has been identified as an abnormal function module.

[0099] For any graph node in the abnormal function module ,set up To enter the node The upstream stream collection, For the node For the set of downstream streams flowing out, the average anomaly score on the input side and the average anomaly score on the output side of the node are respectively:

[0100]

[0101]

[0102] Therefore, nodes are defined. The abnormal amplification is as follows:

[0103]

[0104] in, Used to characterize abnormal states at nodes The degree of enhancement along the material flow direction from the input side to the output side.

[0105] Combined with the final importance score of the graph nodes obtained in step 4) The importance of normalization is defined as follows:

[0106]

[0107] Further construct candidate fault source scores:

[0108]

[0109] Scoring based on candidate fault sources The graph nodes within the abnormal function module are sorted from largest to smallest to obtain a sequence of candidate fault sources; then, the corresponding devices are sequentially checked for faults based on the candidate fault source sequence to determine the faulty device.

[0110] To implement the aforementioned method, in another embodiment of the present invention, a chemical process intelligent modular system is provided. This system is used to implement the topology-attribute fusion-based intelligent modular method for chemical processes described in the aforementioned embodiment. The system of this embodiment consists of multiple modules, including:

[0111] Data parsing and modeling module: used to parse chemical steady-state process simulation data, extract topological structure and stream attribute data, and construct directed graph models;

[0112] Physical feature calculation module: used to extract the differences in inbound and outbound flow attributes of process nodes and generate multi-dimensional node feature vectors;

[0113] Feature fusion and importance assessment module: used to calculate and fuse the structural similarity and attribute similarity between graph nodes to obtain a comprehensive similarity matrix, and calculate the final importance score of the nodes based on topological redundancy and device functional features;

[0114] Module partitioning and graph reconstruction module: used to reconstruct simplified graph structures based on node importance scores and comprehensive similarity, and extract connected components to output module partitioning results;

[0115] Fault diagnosis and tracing module: Calculates the abnormality score of each stream based on the steady-state operating data and real-time operation data of each stream, and further obtains the module abnormality score, thereby identifying abnormal functional modules. Based on the average abnormality score on the input side, the average abnormality score on the output side, and the final importance score of each graph node in the abnormal functional module, constructs the candidate fault source score, and outputs the candidate fault source sequence according to the candidate fault source score.

[0116] The effectiveness of this invention will be verified using a specific case study of a complete chemical process simulation.

[0117] Example: Modular division of the entire process of "vegetable oil to biofuel" based on the method of the present invention.

[0118] This embodiment aims to verify the applicability of the method and system described in this invention in handling complex chemical processes with highly nonlinear thermodynamic behavior, as well as the effectiveness of module division. The test system uses the classic case "Biodiesel production from vegetable oil" built into the process simulation software Aspen Plus for full-process simulation. The process is as follows: Figure 2 As shown. Figure 3 The diagram illustrates the traditional manually defined modules and their division results based on expert experience. Under the traditional module division method, the process mainly includes reaction modules, methanol purification, water washing process, ester purification, product glycerol purification, catalyst separation and other sections. The module division is mainly based on physical regions or general reaction and separation sections.

[0119] The aforementioned method of the present invention is now used to... Figure 2 The process shown is modularized and includes fault tracing analysis. The specific implementation process is as follows:

[0120] (1) Import and preliminary analysis of steady-state simulation data:

[0121] Start the data parsing and modeling module to read Figure 2 The process is simulated in Aspen Plus using a .bkp format file. The system automatically extracts all unit operations (such as reactors, distillation columns, extraction columns, mixers, etc.) and connecting pipelines in the process. Virtual nodes are added to the feed streams (vegetable oil, methanol, etc.) and discharge streams (biofuel, glycerol, etc.) of the process, and a stream attribute matrix is ​​generated based on the extracted stream temperature, pressure, flow rate, and component data to complete the directed graph model. The construction.

[0122] (2) Generation of process node feature vectors:

[0123] The physical characteristic calculation module is activated, traversing the real nodes of the directed graph (filtering out virtual nodes), and calling physicochemical operators to quantify the differences in the properties of each node's inflow and outflow streams in multiple dimensions such as reaction, energy, separation, and flow rate. This generates a multi-dimensional feature vector for each node representing the "three-transfer-one-reaction" mechanism. .

[0124] (3) Similarity calculation and node importance assessment:

[0125] By fusing flow-driven structural similarity with Euclidean distance based on node attributes, a comprehensive similarity matrix is ​​generated. When calculating the base score based on vector magnitude, the system automatically identifies auxiliary nodes with high connectivity (such as methanol recirculation mixers) that lack substantial reaction / separation characteristics due to the large amount of methanol recirculation in this process, thereby triggering a topology and auxiliary function penalty mechanism. This mechanism effectively reduces the interference of auxiliary nodes on structural connectivity and obtains the final importance score for each node. .

[0126] (4) Graph network reconstruction and modular decomposition:

[0127] Virtual nodes in the directed graph model are filtered out. Based on the calculated comprehensive similarity matrix and the final importance score of each node, for each retained real device node, a target node with a higher final importance score and the highest comprehensive similarity is found in its set of potential target nodes, and a directed connection is established from the current node to the target node. If a node does not have a potential target node with a higher final importance score, no further connection is established for it, and it is designated as a candidate core node; the remaining nodes are progressively assigned to the corresponding candidate core nodes along the established connection relationships.

[0128] Subsequently, the system transforms the reconstructed simplified graph into an undirected graph and automatically extracts the connected components, with each connected component representing a chemical engineering functional module. For each functional module, the system further selects the device node with the highest final importance score within the module as the core node of that module, used to characterize the core chemical engineering unit operations within that module.

[0129] Figure 4 The module partitioning results of this example are shown. The thick solid lines with color represent the modular partitioning results of the method of this invention (different colors represent different modules), and the red pentagrams represent the core nodes of each module automatically identified by this method. The system of this invention automatically outputs four functional modules with clear physical meanings:

[0130] Oil phase feeding and pretreatment module: The core unit is the feed pump, which covers the equipment for mixing and preheating fresh and circulating raw oil materials;

[0131] Methanol Reaction and Recovery Coupling Module: Centered on a methanol recovery tower, this module encompasses operations such as methanol feed mixing, transesterification reaction, post-reaction stream transportation, and methanol recovery. This module reflects the strong coupling relationship between methanol recycling, reaction conversion, and recovery separation.

[0132] Glycerol Phase Separation and Purification Module: Centered on a filter, this module encompasses neutralization, solid impurity separation, glycerol phase cooling, and glycerol product separation. It primarily addresses the separation and purification of reaction byproducts and impurities.

[0133] Biodiesel washing and refining module: With the ester distillation column as the core unit, this module covers operations such as washing, heat exchange, removal of light components, and refining of ester products. This module primarily undertakes the washing, purification, and final refining functions of biodiesel products.

[0134] To further verify the decoupling effect of the partitioning results of this invention, the classic evaluation metric of complex networks, modularity (Q-value), is introduced for comparative analysis. A larger Q-value indicates tighter internal connections within modules, weaker coupling between modules, and a more significant decoupling effect across the entire system. Figure 5 This paper compares the modularity results calculated using the traditional manual module partitioning method with those calculated using the intelligent partitioning method of this invention. It can be seen that the modularity of the traditional manual partitioning method is Q = 0.301 in a graph theory sense, while the modularity of the present invention is significantly improved to Q = 0.454. This result demonstrates that the method of this invention can more reasonably identify highly coupled functional regions and reduce the interference of highly connected auxiliary devices on the module partitioning results.

[0135] Therefore, the method of this invention can achieve automatic modular partitioning of complex chemical processes without the need for manual pre-specification of partition boundaries. The resulting modules can be used not only for subsequent dynamic simulation, distributed control, and full-process optimization, but also as the functional partitioning basis for abnormal module identification and fault tracing.

[0136] (5) Application process of fault tracing based on module partitioning:

[0137] After completing the module division, this invention further applies the module division results to the abnormal operation analysis and fault tracing of the biofuel preparation process. This process first narrows down the scope of investigation based on the abnormality of the modules, and then identifies candidate faulty equipment within the abnormal modules.

[0138] The system acquires data such as temperature, pressure, flow rate, and component content during process operation, and assigns this data to the corresponding module based on the correspondence between flow streams, equipment nodes, and functional modules. For each flow stream, using normal operating condition data as a reference, a flow stream state vector containing temperature, pressure, flow rate, and component content is generated, and an anomaly score for that flow stream relative to the normal state is calculated. Then, the anomaly scores of all flow streams within the same functional module are summarized to obtain the anomaly score for that module.

[0139] When issues arise in the process, such as decreased biodiesel product purity, fluctuations in the composition of the methanol recovery stream, or deviations in the composition of post-reaction materials, the system first compares the anomaly scores of each functional module. If the anomaly scores of the methanol feed stream, reactor outlet stream, and stream near the methanol recovery tower are significantly higher, while the anomaly scores of the oil phase feed and pretreatment module, glycerol phase separation and purification module, and biodiesel water washing and refining module are relatively lower, then the anomaly is preferentially attributed to the methanol reaction and recovery coupling module. In this way, the scope of investigation can be narrowed down from the entire process to this functional module.

[0140] After identifying the abnormal module, the system compares the average abnormality scores of the input and output streams of each equipment node along the material flow direction within that module. If the abnormality on the input side of a certain equipment node is weak, while the abnormality on the output side is significantly enhanced, it indicates that the abnormality may have originated or been amplified at that equipment. If both the input and output sides are clearly abnormal, then that equipment is more likely just an intermediate node in the abnormality propagation path. Subsequently, the system combines the final importance score of the equipment node to calculate the candidate fault source score and provides a sequence of candidate fault sources according to the score.

[0141] Taking the methanol reaction and recovery coupling module as an example, if the methanol and oil phase feed streams at the reactor inlet are close to normal, but the reactor outlet stream and its downstream stream entering the methanol recovery tower show significant compositional deviations, it indicates that an anomaly may have occurred or been amplified at the reactor. In this case, the candidate fault source score for the corresponding node of the reactor is high, and the system will prioritize the reactor for troubleshooting. Operators can first check factors such as reactor temperature, residence time, catalyst addition, methanol to oil phase feed ratio, and mixing and mass transfer status.

[0142] If the reactor outlet stream is generally normal, but there are significant compositional deviations in the inlet and outlet streams, the methanol recovery stream at the top of the tower, or the stream at the bottom of the tower, the anomaly is more likely related to the methanol recovery tower and its reflux and condensation units. In this case, the system will prioritize the methanol recovery tower and related units in the candidate fault source sequence. Operators can then prioritize checking the separation status within the tower, reflux ratio, condensation load, tower pressure control, top product output, and the operating status of the trays or packing.

[0143] Through the above process, the module division results can be used for abnormal module identification and candidate fault source ranking. Fault troubleshooting can be carried out in the order of abnormal module, candidate faulty device, and specific operating parameters, avoiding item-by-item troubleshooting across the entire process from the beginning.

[0144] The above examples are merely one of the preferred embodiments of the present invention, and their descriptions are relatively specific and detailed. They should not be used to limit the scope of protection of the present invention. Any modifications or improvements made without departing from the main design and concept of the present invention, and whose technical problems are still consistent with the present invention, should be included within the scope of protection of the present invention.

Claims

1. A method for module partitioning and fault tracing in chemical processes based on topology-attribute fusion, characterized in that, Includes the following steps: 1) Based on the steady-state operating data of the chemical process to be divided, extract the attribute data of each stream and the topological connection relationship between the equipment; take the equipment in the chemical process as graph nodes, the streams connecting the equipment as directed edges, construct a directed graph model representing the chemical process, and assign the attribute data of the streams to the corresponding directed edges. 2) For each node in the directed graph model, extract the attribute differences between its inflow and outflow flows, quantify the attribute differences between the inflow and outflow flows, and construct the feature vector of each graph node. In step 2), the feature vectors of each graph node are represented as follows: ,in, The reaction intensity index is obtained by calculating and normalizing the absolute change in the mass flow rate of each component in the directed edges before and after the node. For pressure gain, it is the logarithmic ratio of the flow-weighted average pressure of the node's output directed edge to the input directed edge. For temperature gain, it is the logarithmic ratio of the weighted average temperature of the directed edge output by the node to that of the directed edge input. The separation strength index is characterized by the degree of deviation of the component composition of each output directed edge of the node from the average component composition of the total output of the node, and is obtained by weighting according to the mass flow rate of each output directed edge. The flow rate is a quantity index, obtained by logarithmic transformation of the total mass flow rate of the feed at the node; 3) Extract the topological connections driven by traffic in the directed graph model, and calculate the structural similarity and attribute similarity between graph nodes by combining the distance between graph nodes, and then fuse them to obtain a comprehensive similarity matrix; 4) Based on the comprehensive similarity matrix, and combined with the topological features of the graph structure and the device functional features corresponding to the graph nodes, the final importance score of each graph node is obtained; In step 4), the final importance score is calculated. It is obtained by calculation using the following formula: ； in, Let be the magnitude of the feature vector of the graph node; , Represents graph nodes. Similarity matrix The elements in the graph represent the overall similarity between graph nodes; This is a topology metric penalty coefficient, used to penalize topology redundancy of devices that play a role in splitting, merging, or connecting in the process. This is the penalty coefficient for auxiliary functions of the equipment, and its value is based on the reaction intensity index in the eigenvector. With separation strength index set up; 5) Point each graph node in the directed graph model to the target node with the higher final importance score and the highest comprehensive similarity, thereby reconstructing and generating a simplified graph structure. Extract each connected component in the simplified graph structure as a functional module to obtain the module division result of the chemical process. 6) Based on the module division results, monitor the real-time operation data of the chemical process. When an abnormality occurs, first identify the abnormal functional module, then generate a sequence of candidate fault sources and conduct fault investigation in sequence to identify the faulty equipment.

2. The method for dividing chemical process modules and tracing faults according to claim 1, characterized in that, In step 1), the directed graph model is represented as follows: Among them, the set of directed edges Each directed edge Defined as a triple , As a unique identifier for the flow, and These are the source device node and the target device node, respectively. For the initial feed and final discharge streams in a chemical process, the missing source or target device is filled by constructing virtual nodes, and the virtual nodes do not have feature vectors.

3. The method for dividing chemical process modules and tracing faults according to claim 2, characterized in that, In step 1), the attribute data of the stream includes the thermodynamic and physical properties of the stream; the thermodynamic and physical properties of the stream are extracted, and a feature matrix is ​​constructed. The k-th row vector of the characteristic matrix Corresponding Flow The attribute information is determined by the temperature of the stream. ,pressure mass flow rate and the mass fraction vector of each component Composed of splicing elements; The row vectors in matrix M are mapped and assigned corresponding directed edges, which are then used as the feature weight attributes of those directed edges.

4. The method for dividing chemical process modules and tracing faults according to claim 1, characterized in that, Step 3) involves calculating structural similarity, including: The mass flow rates between graph nodes are logarithmically normalized to construct a flow weight matrix. And calculate the common neighbor strength matrix accordingly. ; Graph nodes are calculated based on the common neighbor strength matrix J, the flow weight matrix W, and the weighted degree of the graph nodes. With graph nodes The topological overlap similarity between them is used as the structural similarity matrix. matrix elements Structural similarity matrix Each element is restricted to Within the interval, the diagonal elements are set to 1.

5. The method for dividing chemical process modules and tracing faults according to claim 4, characterized in that, In step 3), attribute similarity is obtained by transforming feature distance using a Gaussian kernel function, and the attribute similarity between graph nodes constitutes an attribute similarity matrix. ; Comprehensive similarity matrix From the structural similarity matrix Similarity matrix with attributes The result is obtained by linear weighted fusion.

6. The method for dividing chemical process modules and tracing faults according to claim 1, characterized in that, Step 5) specifically includes: Filter out virtual nodes in a directed graph model; for each retained graph node... Find the best target node with the highest final importance score and the greatest overall similarity. : At the node and the best target node Establish directed connections between them to generate a simplified graph structure; The simplified graph structure is converted into an undirected graph, and the connected components are extracted. Each connected component is then divided into a functional module. For each functional module The final importance score of each graph node within the functional module is used as the criterion for determination. The graph node with the highest final importance score is determined as the core node of the module, which is used to characterize the core chemical unit operation of the functional module.

7. The method for dividing chemical process modules and tracing faults according to claim 1, characterized in that, Step 6) specifically includes: Based on historical steady-state operating data, the mean vector and covariance matrix of the state vector of each stream are calculated; real-time operating data of each stream in the chemical process are obtained, and stream anomaly scores are calculated based on the mean vector and covariance matrix; module anomaly scores are obtained based on the stream anomaly scores within each functional module, and functional modules whose module anomaly scores exceed a set threshold are identified as abnormal functional modules; for each graph node within an abnormal functional module, candidate fault source scores are constructed based on its input-side average anomaly score, output-side average anomaly score, and final importance score, and the candidate fault source sequence is obtained by sorting the candidate fault source scores; then, each device is sequentially checked for faults based on the candidate fault source sequence to identify the faulty device.

8. A chemical process module division and fault tracing system for implementing the method according to any one of claims 1-7, characterized in that, The system includes: Data parsing and modeling module: used to parse chemical steady-state process simulation data, extract topological structure and stream attribute data, and construct directed graph models; Physical feature calculation module: used to extract the differences in inbound and outbound flow attributes of process nodes and generate multi-dimensional node feature vectors; Feature fusion and importance assessment module: used to calculate and fuse the structural similarity and attribute similarity between graph nodes to obtain a comprehensive similarity matrix, and calculate the final importance score of the nodes based on topological redundancy and device functional features; Module partitioning and graph reconstruction module: used to reconstruct simplified graph structures based on node importance scores and comprehensive similarity, and extract connected components to output module partitioning results; Fault diagnosis and tracing module: Calculates the abnormality score of each stream based on the steady-state operating data and real-time operation data of each stream, and further obtains the module abnormality score, thereby identifying abnormal functional modules. Based on the average abnormality score on the input side, the average abnormality score on the output side, and the final importance score of each graph node in the abnormal functional module, constructs the candidate fault source score, and outputs the candidate fault source sequence according to the candidate fault source score.

Images