Abnormality analysis method and system applied to diborane production control system

Abstract

The application provides an abnormality analysis method and system applied to a diborane production control system, first acquires a multi-source production data set generated in the operation of the diborane production control system, contains time stamp marked data such as equipment state, raw material conveying and reaction parameters, then constructs a production correlation knowledge graph, takes the production node as the corresponding production equipment or parameter, and takes the production edge as the physical connection between the equipment or the logical correlation between the parameters, then carries out graph feature extraction on the production correlation knowledge graph, calls a pre-trained graph neural network model to carry out graph structure analysis, generates an abnormality detection result containing detection confidence of different abnormality types, determines the abnormality event type and the distribution characteristic information in the knowledge graph according to the abnormality detection result, finally generates a production early warning instruction containing an event positioning identifier and sends it to a production control terminal to trigger an abnormality response operation, so that the abnormality in the diborane production can be comprehensively and accurately analyzed, and the production safety and stability are improved.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, and more specifically, to an anomaly analysis method and system applied to an ethylene borane production control system. Background Technology

[0002] In the production of diborane, an important chemical raw material, the production process involves complex physicochemical reactions and the coordinated operation of equipment. Traditional diborane production control systems mainly rely on human experience and simple threshold alarm mechanisms to monitor production status. Human experience is often limited by the operator's professional level and subjective judgment, making it difficult to comprehensively and accurately grasp various potential anomalies in the production process. Furthermore, simple threshold alarm mechanisms can only judge based on fixed thresholds for single parameters and cannot consider the complex interrelationships between multiple parameters in the production process.

[0003] In actual production, the diborane production control system generates a large amount of multi-source production data, including equipment status data, raw material delivery data, and reaction parameter data. These data are closely correlated, but existing methods lack effective means to mine and utilize this correlational information. This leads to the inability to detect potential anomalies in the production process in a timely manner, which may in turn cause production accidents, affect product quality and production efficiency, and even create safety hazards. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of this application, embodiments of this application provide an anomaly analysis method applied to an ethylborane production control system, the method comprising:

[0005] Acquire a multi-source production data set generated during the operation of the borane production control system. The multi-source production data set includes equipment status data, raw material delivery data, and reaction parameter data with timestamps.

[0006] A production association knowledge graph is constructed based on the multi-source production data set. The production association knowledge graph contains multiple production nodes and production edges connecting the production nodes. The production nodes correspond to borane production equipment or production parameters. The production edges correspond to the physical connection relationship between production equipment or the logical association relationship between production parameters.

[0007] The production-related knowledge graph is subjected to graph feature extraction processing to obtain the node feature set of the production node and the edge feature set of the production edge;

[0008] A pre-trained graph neural network model is invoked to perform graph structure analysis on the node feature set and the edge feature set to generate anomaly detection results for the production-related knowledge graph. The anomaly detection results include detection confidence scores corresponding to different anomaly types.

[0009] Based on the anomaly detection results, determine the types of abnormal events existing in the diborane production process and the distribution characteristics of the abnormal events in the production association knowledge graph;

[0010] Based on the abnormal event type and the distribution characteristic information, a production early warning instruction containing an event location identifier is generated, and the production early warning instruction is sent to the borane production control terminal to trigger an abnormal response operation.

[0011] In another aspect, embodiments of this application also provide an anomaly analysis system applied to an ethylborane production control system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0012] Based on the above, this application embodiment acquires a multi-source production data set generated during the operation of the borane production control system, constructs a production-related knowledge graph based on the multi-source production data set, and can intuitively display the relationship between production equipment and production parameters, which helps to deeply explore the potential information behind the data. The production-related knowledge graph is processed by graph feature extraction, and a pre-trained graph neural network model is called for graph structure analysis. This fully utilizes the advantages of graph neural networks in processing complex graph structure data, accurately identifies abnormal situations in the production process, and generates abnormal detection results containing detection confidence levels corresponding to different abnormal types, improving the accuracy and reliability of abnormal detection. Based on the abnormal detection results, the abnormal event type and its distribution characteristics in the production-related knowledge graph are determined, enabling a comprehensive understanding of the specific circumstances of the abnormal event. Based on the abnormal event type and distribution characteristics, a production early warning instruction containing an event location identifier is generated and sent to the borane production control terminal to trigger an abnormal response operation. This achieves timely early warning and rapid response to abnormal events, effectively avoiding production accidents, improving the safety and stability of borane production, and ensuring product quality and production efficiency. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the execution flow of the anomaly analysis method applied to the borane production control system provided in the embodiments of this application.

[0014] Figure 2 This is a schematic diagram of the hardware architecture of an anomaly analysis system applied to the borane production control system provided in an embodiment of this application. Detailed Implementation

[0015] The present application will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an anomaly analysis method for a diborane production control system provided in one embodiment of this application. The anomaly analysis method for a diborane production control system will be described in detail below.

[0016] Step S110: Obtain a multi-source production data set generated during the operation of the borane production control system. The multi-source production data set includes equipment status data, raw material delivery data, and reaction parameter data with timestamps.

[0017] During the production of diborane, the production control system continuously generates a large amount of production-related data. To conduct effective anomaly analysis of the diborane production process, it is first necessary to acquire this multi-source production data set.

[0018] Equipment status data describes the operational status of diborane production equipment. Different production equipment has different status characteristics. For example, the status of a reactor may include temperature control status, pressure control status, and the operating status of the agitator; the status of a transfer pump may include speed and flow stability. This equipment status data is collected in real time by various sensors installed on the equipment. Each equipment status data point is assigned a timestamp, accurate to a specific moment, which helps in subsequent analysis of the equipment status changing over time. For example, at a certain moment, the temperature sensor in the reactor records the temperature data inside the reactor, and this temperature data is tagged with the precise time for subsequent traceability and analysis. Among these, parameters reflecting the core operating status of the equipment, such as the temperature control status of the reactor, the operating status of the agitator, and the speed stability of the transfer pump, are key data for subsequent anomaly analysis. They will play an important role in constructing the real-time operating status attributes of knowledge graph nodes, generating equipment node feature vectors, and locating the equipment nodes corresponding to abnormal events.

[0019] Raw material transport data primarily concerns the transport status of various raw materials during the production process. Raw material transport is crucial for the production of diborane, as factors such as the transport volume and speed directly affect the reaction. Raw material transport data includes information such as flow rate, pressure, and transport time. This data is collected by sensors installed on the raw material transport pipelines, and each data point is timestamped. For example, flow sensors and pressure sensors are installed on the raw material transport pipelines to monitor the flow rate and pressure of the raw materials in real time and transmit the timestamped data to the production control system. The flow rate, pressure, and transport rate data are used to analyze anomalies in material transport between equipment, playing a key role in determining the transport rate attributes of physical connections between production equipment, generating transport rate characteristics, and analyzing the diffusion trend of abnormal events at equipment connections.

[0020] Reaction parameter data are crucial data reflecting the diborane production reaction process. These parameters include reaction temperature, reaction pressure, reaction time, and reactant concentration. The accuracy of reaction parameter data is essential for controlling the reaction process and ensuring product quality. This data is collected by sensors installed inside or around the reaction equipment, and each data point is timestamped. For example, temperature and pressure sensors are installed inside the reactor to monitor the temperature and pressure in real time and transmit timestamped data to the production control system. Data such as reaction temperature, pressure, reactant concentration, and reaction time are important bases for analyzing logical anomalies in the reaction process. They are used to construct real-time measurement attributes of production parameter nodes, generate parameter node feature vectors, and extract logical relationships between parameters based on reaction kinetic models.

[0021] Step S120: Construct a production association knowledge graph based on the multi-source production data set. The production association knowledge graph contains multiple production nodes and production edges connecting the production nodes. The production nodes correspond to borane production equipment or production parameters, and the production edges correspond to the physical connection relationship between production equipment or the logical association relationship between production parameters.

[0022] After obtaining the multi-source production data set, the next step is to construct a production-related knowledge graph based on this data set. A production-related knowledge graph is a graph structure used to represent the relationships between various production factors in the diborane production process. By constructing a production-related knowledge graph, the relationships between various production equipment and parameters in the diborane production process can be understood more intuitively.

[0023] The production equipment nodes here specifically include reactors (e.g., "Reactor 001"), raw material transfer pumps (e.g., "Transfer Pump 002"), heaters, coolers, gas compressors, purification towers, storage tanks, etc. Each node contains equipment type attributes (e.g., "Reactor") and real-time operating status attributes (e.g., "Normal Operation"). The production parameter nodes specifically include reaction temperature, reaction pressure, reaction time, reactant concentration (e.g., diborane concentration, hydrogen concentration), raw material flow rate (e.g., sodium borohydride solution flow rate, hydrochloric acid flow rate), raw material purity, stirrer speed, etc. Each node contains parameter type attributes (e.g., "Reaction Temperature") and real-time operating status attributes. Measured value attributes (e.g., "300K"); Production edges are specifically divided into three categories: First, physical connection edges between equipment, such as the pipeline connection between the reactor and the transfer pump, whose attributes include connection type (e.g., "pipeline connection") and transmission rate (e.g., "50L / min"); Second, logical association edges between parameters, such as the association between reaction temperature and reaction rate, whose attributes include the direction of influence (e.g., "temperature increases - rate increases") and association strength (e.g., "0.9"); Third, equipment-parameter association edges, such as the measurement association between the reactor and reaction temperature, whose attributes include association type (e.g., "measurement association") and acquisition frequency (e.g., "1 time / second").

[0024] Step S121: Perform data alignment processing on the multi-source production data set to obtain an aligned production data sequence.

[0025] Since the production data in a multi-source production dataset comes from different data sources, and the collection frequency and timing of this data may differ, alignment processing is necessary. The purpose of data alignment is to ensure that production data from different data sources are consistent in time, facilitating subsequent analysis and processing.

[0026] First, it is necessary to unify the timestamps of production data in the multi-source production dataset. Since production data from different data sources may have different timestamp formats, these timestamps need to be converted to a unified format. For example, the timestamps of production data from different data sources should be converted to a standard time format, such as year-month-day hour:minute:second.

[0027] Then, the data is sorted according to the unified timestamps. Sorting the data in chronological order makes the data present a continuous sequence over time.

[0028] Next, the production data needs to be interpolated or sampled to ensure that the data from different data sources are aligned in time. If different data sources have different collection frequencies, data at certain points in time may be missing. These missing data can be supplemented using interpolation methods. For example, linear interpolation can be used to estimate the missing data based on data from adjacent time points. If some data sources have excessively high collection frequencies while others have lower collection frequencies, sampling methods can be used to downsample the high-frequency data to align it in time with the low-frequency data.

[0029] Through the above data alignment steps, an aligned production data sequence is finally obtained. The data in this aligned production data sequence is consistent in time and can be easily analyzed and processed later.

[0030] Step S122: Extract the identification information of the borane production equipment from the aligned production data sequence, and map the data set corresponding to each borane production equipment identification information to a production equipment node. The production equipment node includes equipment type attributes and real-time operating status attributes.

[0031] After obtaining the aligned production data sequence, it is necessary to extract the identification information of the diborane production equipment from it. The identification information is used to uniquely identify each production equipment, such as its serial number and name. Through this identification information, each production equipment can be accurately identified, and its corresponding data set can be mapped to a production equipment node.

[0032] For each diborane production equipment identifier, its corresponding dataset can be integrated and processed to form a production equipment node. The production equipment node contains two important attributes: equipment type and real-time operating status.

[0033] The equipment type attribute describes the type of production equipment, such as reactors, transfer pumps, and heaters. Different types of production equipment have different functions and roles in the diborane production process, therefore, the equipment type attribute is crucial for subsequent analysis and processing.

[0034] Real-time operating status attributes reflect the current operating status of the production equipment. These attributes are determined based on equipment status data, such as whether the equipment is operating normally or whether there are any fault warnings. By understanding these attributes, the operating status of the production equipment can be monitored promptly, allowing for appropriate corrective actions to be taken.

[0035] For example, for a reactor, whose equipment identification information is "Reactor 001", relevant data for the reactor, including equipment status data, is extracted from the aligned production data sequence. This data is then integrated into a production equipment node, whose equipment type attribute is "Reactor". The real-time operating status attribute is determined based on the equipment status data, such as "normal operation" or "fault warning".

[0036] Step S123: Extract the borane production parameter identification information from the aligned production data sequence, and map the data set corresponding to each borane production parameter identification information to a production parameter node. The production parameter node includes parameter type attribute and real-time measurement value attribute.

[0037] In addition to production equipment nodes, it is also necessary to extract the identification information of diborane production parameters from the aligned production data sequence and map the corresponding data set to production parameter nodes. The diborane production parameter identification information is used to uniquely identify each production parameter, such as the parameter name and number. Through this identification information, each production parameter can be accurately identified, and its corresponding data set can be mapped to a production parameter node.

[0038] For each diborane production parameter identifier, its corresponding dataset can be integrated and processed to form a production parameter node. The production parameter node contains two important attributes: parameter type attribute and real-time measurement value attribute.

[0039] The parameter type attribute describes the type of production parameter, such as reaction temperature, reaction pressure, and reactant concentration. Different types of production parameters have different roles and significance in the production of diborane; therefore, the parameter type attribute is crucial for subsequent analysis and processing.

[0040] Real-time measurement attributes reflect the measured values ​​of production parameters at the current moment. These attributes can be determined based on reaction parameter data, such as specific values ​​for reaction temperature and pressure. By using real-time measurement attributes, changes in production parameters can be monitored promptly, allowing for adjustments to the production process.

[0041] For example, for the production parameter of reaction temperature, its parameter identification information is "reaction temperature 001". Relevant data for this reaction temperature, including reaction parameter data, is extracted from the aligned production data sequence. This data is then integrated into a single production parameter node, whose parameter type attribute is "reaction temperature". The real-time measurement value attribute is determined based on the reaction parameter data, such as the specific temperature value.

[0042] Step S124: Extract the physical connection relationship between the diborane production equipment based on the diborane production process flow diagram, and establish production edges between equipment nodes that have direct material transmission or energy exchange. The production edges include connection type attributes and transmission rate attributes.

[0043] In this embodiment, it is necessary to supplement the association edges between production equipment and production parameters. For example, the reaction vessel and reaction temperature and reaction pressure are "measurement associations", and the edge attributes include parameter acquisition frequency; the delivery pump and raw material flow rate are "control associations", and the edge attributes include flow rate adjustment accuracy, etc. Such edges are used to establish direct associations between equipment and parameters to ensure the integrity of the knowledge graph network.

[0044] After determining the production equipment nodes and production parameter nodes, it is necessary to extract the physical connections between the production equipment based on the diborane production process flow diagram. The diborane production process flow diagram details the physical connections and material transfer paths between various production equipment during the production process.

[0045] By analyzing the process flow diagram of diborane production, it is possible to determine which production equipment has direct material transport or energy exchange relationships. For equipment nodes with direct material transport or energy exchange, production edges need to be established between them. Production edges represent the physical connection relationships between production equipment and contain two important attributes: connection type attribute and transport rate attribute.

[0046] The connection type attribute describes the connection method between production equipment, such as pipe connections and cable connections. Different connection types have different characteristics and functions in the production process, therefore, the connection type attribute is very important for subsequent analysis and processing.

[0047] The transfer rate attribute reflects the rate of material transfer or energy exchange between production equipment. It can be determined based on raw material delivery data and equipment operating data, such as raw material flow rate and energy transfer power. By understanding the transfer rate attribute, we can monitor the material transfer and energy exchange between production equipment, thus optimizing the production process.

[0048] For example, in the process flow diagram of diborane production, the reactor and the transfer pump are connected by pipelines, and there is a raw material transfer relationship. In this case, a production edge can be established between the production equipment node corresponding to the reactor and the production equipment node corresponding to the transfer pump. The connection type attribute of this production edge is "pipeline connection", and the transfer rate attribute is determined based on the raw material transfer data, such as the raw material transfer flow rate.

[0049] Step S125: Extract the logical correlation between borane production parameters based on the borane production reaction kinetic model, and establish production edges between parameter nodes that have causal influence or synergistic changes. The production edges include influence direction attributes and correlation strength attributes.

[0050] The correlation strength attribute is calculated using the Pearson correlation coefficient, with the following formula:

[0051] Where xi and yi are the historical measurements of the two parameters. , This is the mean. For example, the correlation coefficient between reaction temperature and reaction rate is calculated to be 0.9, indicating a strong positive correlation.

[0052] In addition to the physical connections between production equipment, it is also necessary to extract the logical relationships between production parameters based on the kinetic model of the diborane production reaction. The kinetic model of the diborane production reaction describes the interaction and variation patterns among various reaction parameters during the production process.

[0053] By analyzing the kinetic model of the diborane production reaction, it is possible to determine which production parameters have causal or synergistic relationships. For parameter nodes with causal or synergistic relationships, production edges need to be established between them. Production edges represent the logical relationships between production parameters and contain two important attributes: the direction of influence and the strength of the relationship.

[0054] The direction of influence attribute describes the causal relationship between production parameters; for example, a change in one parameter may lead to an increase or decrease in another. The direction of influence attribute can be determined based on reaction kinetic models and actual production data, such as by analyzing the trend of parameter changes during the reaction process.

[0055] The correlation strength attribute reflects the degree of correlation between production parameters. The correlation strength attribute can be determined by methods such as calculating the correlation coefficient. For example, by analyzing data from two parameters at multiple time points, the correlation coefficient between them can be calculated, and the magnitude of the correlation coefficient reflects the strength of the correlation between the two parameters.

[0056] For example, in the kinetic model of diborane production, there is a causal relationship between reaction temperature and reaction rate; an increase in reaction temperature leads to an increase in reaction rate. In this case, a production edge can be established between the production parameter node corresponding to reaction temperature and the production parameter node corresponding to reaction rate. The influence direction attribute of this production edge is "an increase in reaction temperature leads to an increase in reaction rate," and the correlation strength attribute is determined by calculating the correlation coefficient.

[0057] Step S126: Construct an initial production association knowledge graph based on the production equipment node, production parameter node and corresponding production edge, perform connectivity verification on the initial production association knowledge graph, delete isolated nodes and corresponding edges that are not connected to the main production process, and generate the final production association knowledge graph.

[0058] After identifying the production equipment nodes, production parameter nodes, and the production edges between them, these nodes and edges are combined to construct an initial production association knowledge graph. This initial production association knowledge graph is a graph structure containing multiple production nodes and production edges connecting them, which initially reflects the relationships between various production elements in the diborane production process.

[0059] However, the initial production association knowledge graph may contain some isolated nodes and corresponding edges that are not connected to the main production process. These isolated nodes and edges may be due to data acquisition errors, equipment failures, or other reasons, and they are not of practical significance for analyzing anomalies in the diborane production process. Therefore, it is necessary to perform connectivity verification on the initial production association knowledge graph and delete these isolated nodes and corresponding edges.

[0060] Step S1261: Extract the core node set corresponding to the main production process from the initial production association knowledge graph. The core node set includes raw material input nodes, reactor nodes, and product output nodes.

[0061] First, it is necessary to extract the core node set corresponding to the main production process from the initial production association knowledge graph. The main production process is the main flow in the production of diborane, which includes key steps such as raw material input, reaction, and product output. The core node set contains three important nodes: the raw material input node, the reaction vessel node, and the product output node.

[0062] The raw material input node represents the input stage of raw materials in the production process, and it is connected to the raw material conveying equipment and related production parameter nodes. The reactor node represents the reaction stage in the production process; it is the core equipment in the production of diborane and is closely related to other production equipment nodes and reaction parameter nodes. The product output node represents the output stage of the production process, and it is connected to the product conveying equipment and related production parameter nodes.

[0063] By extracting the core node set, the key nodes of the main production process can be identified.

[0064] Step S1262: Use a breadth-first search algorithm to traverse the initial production association knowledge graph starting from the core node set, and mark all reachable nodes and corresponding edges that are connected to the core nodes by paths.

[0065] After determining the core node set, a breadth-first search algorithm is used to traverse the initial production-related knowledge graph, starting from the core node set. Breadth-first search is an algorithm used to traverse graph structures; it starts from the initial node and traverses the nodes in the graph layer by layer until all reachable nodes have been traversed.

[0066] Starting from each core node in the core node set, a breadth-first search algorithm is used to traverse the entire set. During the traversal, all reachable nodes connected to core nodes by paths and their corresponding edges are marked. A reachable node is a node that can be reached from a core node via a generated edge.

[0067] By traversing the data using the breadth-first search algorithm, we can determine which nodes and edges are connected to the main production process and which nodes and edges are isolated.

[0068] Step S1263: Identify the unmarked nodes and corresponding edges in the initial production association knowledge graph. The unmarked nodes and corresponding edges are isolated nodes and edges that are not connected to the main production process.

[0069] After traversing the breadth-first search algorithm, unlabeled nodes and their corresponding edges in the initial production-related knowledge graph can be identified. These unlabeled nodes and their corresponding edges are isolated nodes and edges that are not connected to the main production process, and they have no practical significance for analyzing anomalies in the diborane production process, so they need to be deleted.

[0070] Step S1264: Delete the isolated nodes and their corresponding edges from the initial production association knowledge graph to generate a production association knowledge graph after connectivity verification.

[0071] After identifying isolated nodes and their corresponding edges, these nodes and edges are deleted from the initial production-related knowledge graph. After deleting these isolated nodes and their corresponding edges, a production-related knowledge graph with connectivity verified is obtained. All nodes and edges in this production-related knowledge graph are connected to the main production process, making it more suitable for subsequent anomaly analysis.

[0072] Step S1265: Perform node and edge counts on the production association knowledge graph after connectivity verification to ensure that the remaining node and edge counts meet the basic structural requirements of the borane production process, and generate the final production association knowledge graph.

[0073] After obtaining the production association knowledge graph with connectivity verified, it is necessary to perform node and edge counts. The diborane production process has certain basic structural requirements, such as a certain number of production equipment nodes and production parameter nodes, as well as the connections between them.

[0074] By counting the number of nodes and edges, it can be ensured that the production association knowledge graph after connectivity verification meets the basic structural requirements of the diborane production process. If the number of nodes or edges does not meet the requirements, it may be necessary to further check and adjust the structure of the graph, such as adding missing nodes or edges, or deleting redundant nodes or edges.

[0075] After counting the number of nodes and edges, a final production association knowledge graph that meets the basic structural requirements of the diborane production process is generated. This final production association knowledge graph accurately reflects the relationships between various production factors in the diborane production process.

[0076] Step S130: Perform graph feature extraction processing on the production-related knowledge graph to obtain the node feature set of the production node and the edge feature set of the production edge.

[0077] After obtaining the final production-related knowledge graph, graph feature extraction processing is required to obtain the node feature set of production nodes and the edge feature set of production edges. Graph feature extraction processing can convert the information of nodes and edges in the production-related knowledge graph into feature vectors that can be used for subsequent analysis, thereby improving anomaly analysis.

[0078] Step S131: Perform feature encoding processing on the real-time operating status attributes of the production equipment node to generate a device node feature vector. The feature encoding processing includes hot encoding of status categories and normalization of status change rate.

[0079] The state change rate is calculated as follows: State change rate = (S(t2) - S(t1)) / (t2 - t1), where S(t) is the state value of the equipment at time t ("normal" is denoted as 1, "warning" as 0.5, and "fault" as 0), and t2 - t1 is the time interval. For example, if the reactor is in "normal" (1) at 10:00 and changes to "warning" (0.5) at 10:05, then the state change rate is (0.5 - 1) / 5 min = -0.1 / min.

[0080] First, the real-time operating status attributes of the production equipment nodes undergo feature encoding processing. Real-time operating status attributes describe the current operating status of the production equipment, such as whether the equipment is operating normally or whether there are fault warnings. To convert this information into feature vectors that can be used for analysis, feature encoding processing is required.

[0081] Feature encoding processing involves two steps: hot encoding of state categories and normalization of state change rates.

[0082] Status category hot coding is the process of converting the operating status categories of production equipment into binary vectors. For example, the operating status of production equipment may include several categories such as normal operation, fault warning, and shutdown. For each status category, it can be encoded into a binary vector, where only one bit is 1 and the rest are 0. Through status category hot coding, the operating status information of production equipment can be converted into digital features.

[0083] State change rate normalization is the process of normalizing the rate of change of the operating state of production equipment. The operating state of production equipment may change over time, and the state change rate reflects the speed of this change. To eliminate differences in state change rates between different devices, it is necessary to normalize the state change rate. Normalization converts the state change rate into a value between 0 and 1, making the state change rates of different devices comparable. Normalization typically uses linear normalization to map the state change rate to the interval between 0 and 1. For example, first, the maximum and minimum values ​​of the state change rate of all production equipment are calculated, and then for each device's state change rate, it is converted into a value between 0 and 1. Through these two steps of state category hot encoding and state change rate normalization, a device node feature vector is finally generated. This device node feature vector contains the category information and state change rate information of the production equipment's operating state, and can comprehensively reflect the real-time operating state of the production equipment.

[0084] Step S132: Perform feature encoding processing on the real-time measured value attributes of the production parameter node to generate a parameter node feature vector. The feature encoding processing includes parameter value range binning and parameter change trend symbolization.

[0085] Next, feature encoding is performed on the real-time measured values ​​of the production parameter nodes. These real-time measured values ​​describe the current values ​​of the production parameters, such as reaction temperature and reaction pressure. To convert this information into feature vectors suitable for analysis, feature encoding is required, which includes two steps: parameter value range binning and parameter trend symbolization.

[0086] Parameter binning is the process of dividing the range of values ​​for a production parameter into several intervals. Different production parameters have different ranges. By binning their ranges, continuous values ​​can be converted into discrete categories. For example, the range of a reaction temperature parameter may be within a large range. This range can be divided into several smaller intervals, each corresponding to a category. For each real-time measured value of a production parameter, it is determined which interval it falls into, and then coded into the category corresponding to that interval. In this way, the specific numerical information of the production parameter can be converted into more easily processed category information.

[0087] Symbolizing parameter change trends is the process of representing the changing trends of production parameters using symbols. The values ​​of production parameters may change over time, and the trend reflects the direction of this change, such as increasing, decreasing, or remaining unchanged. To represent this trend simply and intuitively, it is symbolized. For example, a positive sign represents an increase in parameter value, a negative sign represents a decrease in parameter value, and zero represents a unchanged parameter value. By symbolizing parameter change trends, the information about the changing trends of production parameters can be converted into numerical characteristics.

[0088] After feature encoding processes involving parameter value range binning and parameter change trend symbolization, a parameter node feature vector is generated. This parameter node feature vector contains information about the value range and change trend of the production parameter, effectively reflecting the real-time status of the production parameter.

[0089] Step S133: Merge the device node feature vector and the parameter node feature vector to obtain the node feature set of the production node.

[0090] After generating the device node feature vectors and parameter node feature vectors, they need to be merged to obtain the node feature set of the production node. In order for the merged feature vectors to more comprehensively and accurately reflect the information of the production node, some preprocessing operations need to be performed on the device node feature vectors and parameter node feature vectors.

[0091] Step S1331: Perform dimensional expansion processing on the device node feature vector to generate an expanded device node feature vector. The dimensional expansion processing includes adding a device type encoding dimension.

[0092] The feature vectors of device nodes undergo dimensionality expansion. Device type is a crucial attribute of production equipment; different types of equipment have different functions and characteristics during production. To reflect device type information in the feature vectors, a device type encoding dimension needs to be added. Device type encoding is the process of digitally representing the device type. For example, one-hot encoding can be used to encode each device type into a binary vector, where only one bit is 1 and the rest are 0. By adding the device type encoding dimension, the dimensionality of the device node feature vectors is expanded, generating extended device node feature vectors. This extended device node feature vector not only contains real-time operating status information of the production equipment but also device type information.

[0093] Step S1332: Perform dimensional expansion processing on the parameter node feature vector to generate an expanded parameter node feature vector. The dimensional expansion processing includes adding a parameter type encoding dimension.

[0094] Similarly, the feature vectors of parameter nodes undergo dimensionality expansion. Parameter type is a crucial attribute of production parameters; different types of parameters have different meanings and functions in the production process. To reflect parameter type information in the feature vectors, a parameter type encoding dimension needs to be added. Parameter type encoding also adopts a similar method to device type encoding, encoding each parameter type as a binary vector. By adding the parameter type encoding dimension, the dimensionality of the parameter node feature vectors is expanded, generating expanded parameter node feature vectors. This expanded parameter node feature vector not only contains real-time measurement values ​​and trend information of production parameters but also parameter type information, enabling it to more accurately reflect the characteristics of production parameters.

[0095] Step S1333: The extended device node feature vector and the extended parameter node feature vector are concatenated to generate a merged node feature vector.

[0096] After obtaining the extended device node feature vector and the extended parameter node feature vector, they are concatenated. Concatenation is the process of joining the two vectors together in a predetermined order. Through concatenation, the feature information of the device node and the parameter node is integrated into a single vector, generating a merged node feature vector. This merged node feature vector contains multifaceted information about the production equipment and production parameters, providing a more comprehensive description of the production node's characteristics.

[0097] Step S1334: Arrange the merged node feature vectors in the order of the timestamps of the production nodes to generate the node feature set of the production nodes.

[0098] Finally, the merged node feature vectors are arranged in the order of the production node timestamps. Since the production process occurs over time, the features of the production nodes also change over time. Arranging the merged node feature vectors in timestamp order allows the node feature set to reflect the changes in production node features over time. This arranged node feature set can be better used for subsequent anomaly analysis, as anomalies may be related to changes in production node features over time. Through the above method, the final node feature set for the production nodes is generated.

[0099] Step S134: Perform feature extraction processing on the transmission rate attribute of the production edge to generate transmission rate features. The feature extraction processing includes calculating the historical average rate and statistically analyzing the rate fluctuation range.

[0100] For the transmission rate attribute of the production edge, feature extraction processing is required to generate transmission rate features. The transmission rate attribute reflects the speed of material transfer or energy exchange between production equipment, and feature extraction can help to better analyze anomalies in the production process. Feature extraction processing includes two steps: calculating the historical average rate and statistically analyzing the rate fluctuation range.

[0101] Historical rate average calculation is the process of calculating the average transmission rate of the production edge over a period of time. The transmission rate of the production edge may fluctuate over time; by calculating the historical rate average, a relatively stable reference value can be obtained. First, all data on the production edge transmission rate within a historical time period are collected. Then, these data are summed and divided by the number of data points to obtain the historical rate average. This historical rate average reflects the overall level of the production edge transmission rate, providing a benchmark for judging whether the current transmission rate is abnormal.

[0102] Rate fluctuation range statistics are the process of statistically analyzing the degree of fluctuation in the transmission rate of the production edge over a period of time. The transmission rate of the production edge may fluctuate within a certain range, and this fluctuation range reflects the stability of the transmission rate. To statistically analyze the rate fluctuation range, we first find the maximum and minimum transmission rates within a historical time period, and then calculate the difference between them. This difference is the rate fluctuation range. The rate fluctuation range reflects the magnitude of changes in the transmission rate of the production edge and is of great significance for determining whether abnormal fluctuations in the transmission rate have occurred.

[0103] Through two steps—calculating the historical average rate and statistically analyzing the rate fluctuation range—a transmission rate characteristic is ultimately generated. This characteristic includes information on the average level and fluctuation range of the production-side transmission rate, providing a comprehensive reflection of the production-side transmission rate situation.

[0104] Step S135: Perform feature extraction processing on the association strength attribute of the production edge to generate association strength features. The feature extraction processing includes historical data correlation coefficient calculation and influence lag time statistics.

[0105] The historical data correlation coefficient here differs from the correlation strength attribute in step S125. The correlation strength in S125 is a "theoretical correlation coefficient" based on reaction kinetic models, such as the temperature-rate relationship derived from formulas. In contrast, the historical data correlation coefficient in this step is a "measured correlation coefficient" calculated based on actual production data, reflecting dynamic correlation deviations in actual production. Combining the two provides a comprehensive description of the correlation strength. Specifically, the impact lag time statistics are analyzed using the sliding window method. This involves: sorting the historical data of nodes A and B by time series; setting a window size (e.g., 10 minutes); calculating the difference |tB-tA| between the change time tA of A and the corresponding change time tB of B; and using the value with the highest frequency of this difference as the impact lag time. For example, if the frequency of B (rate) increasing after 2 minutes is highest after A (temperature), then the lag time is 2 minutes.

[0106] Feature extraction is performed on the correlation strength attributes of production edges to generate correlation strength features. Correlation strength attributes reflect the degree of correlation between production parameters or production equipment; feature extraction helps analyze the inherent connections and anomalies in the production process. Feature extraction processing includes two steps: calculating the correlation coefficient of historical data and statistically analyzing the impact lag time.

[0107] Historical data correlation coefficient calculation is the process of calculating the correlation coefficient between the relevant parameters of two nodes connected by a production edge in historical data. The correlation coefficient is an indicator that measures the strength of the linear relationship between two variables. By calculating the correlation coefficient, we can understand the degree of association between the two nodes connected by the production edge. First, we collect all the relevant parameter data of the two nodes connected by the production edge within a historical time period, and then calculate the correlation coefficient between these two sets of data. The correlation coefficient ranges from -1 to 1; the closer the absolute value is to 1, the stronger the association between the two nodes; the closer the absolute value is to 0, the weaker the association between the two nodes.

[0108] Impact lag time statistics are the process of calculating the time delay between the impacts of two nodes connected by a production edge. In the production process, a change in one node may not immediately affect another node, but rather there is a certain time delay. By calculating impact lag time, we can understand the impact transmission mechanism between the two nodes connected by the production edge. To calculate impact lag time, we need to analyze the time difference between a change in one node and a corresponding change in another node in historical data, identifying the most frequent time difference as the impact lag time. Impact lag time reflects the dynamic relationship between the two nodes connected by the production edge.

[0109] By processing historical data correlation coefficients and statistically analyzing impact lag times, a correlation strength feature is ultimately generated. This feature includes information on the tightness of the correlation between the two nodes connected to the production edge and the impact delay time, comprehensively reflecting the correlation strength of the production edge.

[0110] Step S136: Merge the transmission rate feature and the association strength feature to obtain the edge feature set of the production edge.

[0111] During the merging process, the connection type and influence direction features of the production edges need to be included. The connection type (e.g., pipe connection -1, cable connection -2) is converted into features through one-hot encoding, and the influence direction (positive -1, negative -1) is converted into features through symbol encoding. The final edge feature set includes "transmission rate + connection type + correlation strength + influence direction + historical correlation coefficient + lag time".

[0112] After generating the transmission rate feature and association strength feature, they are merged to obtain the edge feature set of the production edge. The merging process combines the transmission rate feature and association strength feature. Since these two features reflect the characteristics of the production edge from different perspectives, merging them provides a more comprehensive description of the production edge's information. During merging, attention must be paid to the feature dimension matching issue to ensure that the merged edge feature set accurately reflects the various characteristics of the production edge. Through the above merging method, the final edge feature set of the production edge is generated.

[0113] Step S140: Call the pre-trained graph neural network model to perform graph structure analysis processing on the node feature set and the edge feature set, and generate the anomaly detection results of the production-related knowledge graph. The anomaly detection results include the detection confidence scores corresponding to different anomaly types.

[0114] After obtaining the node feature set of the production nodes and the edge feature set of the production edges, a pre-trained graph neural network model is invoked to perform graph structure analysis on these feature sets to generate anomaly detection results for the production-related knowledge graph. The graph neural network model is an artificial intelligence model specifically designed for processing graph-structured data; it can automatically learn the complex relationships between nodes and edges in the graph, thereby performing anomaly detection.

[0115] Step S141: Input the node feature set and the edge feature set into the input layer of the graph neural network model, whereby the input layer maps the node feature vector and the edge feature vector into node embedding vector and edge embedding vector, respectively.

[0116] First, the node feature sets and edge feature sets are input into the input layer of the graph neural network model. The input layer is the first layer of the graph neural network model, and its main function is to map the input node feature vectors and edge feature vectors into node embedding vectors and edge embedding vectors, respectively. Node embedding vectors and edge embedding vectors are low-dimensional vector representations that better preserve the feature information of nodes and edges and facilitate processing in subsequent layers. The input layer transforms the node feature vectors and edge feature vectors through a series of linear transformations and non-linear activation functions, allowing them to enter the subsequent layers of the graph neural network model in a more suitable form for processing.

[0117] Step S142: The node embedding vector and the edge embedding vector are processed by the graph convolutional layer of the graph neural network model to generate an aggregated node vector containing local structural information. The neighborhood information aggregation process includes a weighted summation of the node's own features and the features of its neighboring nodes.

[0118] Next, the graph convolutional layer of the graph neural network model performs neighborhood information aggregation processing on the node embedding vectors and edge embedding vectors. The graph convolutional layer is one of the core layers of the graph neural network model, capable of capturing local structural information between nodes in the graph. Neighborhood information aggregation is a process of weighted summation of the features of a node itself and the features of its neighboring nodes.

[0119] Step S1421: Define a neighborhood range for each production node, wherein the neighborhood range includes first-order neighbor nodes directly connected to the production node and their corresponding edges.

[0120] First, define the neighborhood range for each producer node. The neighborhood range refers to the set of nodes and edges directly connected to that producer node; here, we mainly consider first-order neighbors, i.e., nodes directly connected to the producer node. Each producer node's neighborhood range includes its directly connected first-order neighbors and the edges connecting them. By defining the neighborhood range, the local environment of each node in the graph can be clearly defined.

[0121] Step S1422: Perform a weighted summation on the node embedding vector of each production node and the node embedding vectors of all its neighboring nodes to obtain a weighted summation result. The weights of the weighted summation are calculated from the edge embedding vectors of the corresponding edges.

[0122] Then, a weighted summation is performed on the node embedding vector of each producer node and the node embedding vectors of all its neighboring nodes. The weights for the weighted summation are calculated from the edge embedding vectors of the corresponding edges. The edge embedding vectors contain the feature information of the edges, and the weights used for the weighted summation can be obtained from the edge embedding vectors. For each producer node, its own node embedding vector is multiplied by a weight, and the node embedding vectors of each of its neighboring nodes are also multiplied by their respective weights. These weighted vectors are then summed to obtain the weighted summation result. This weighted summation result combines the features of the producer node itself and the features of its neighboring nodes, and can better reflect the local structural information of the nodes in the graph.

[0123] Step S1423: The weighted summation result is concatenated with the original node embedding vector of the production node to generate an intermediate node vector containing the production node's own features and neighborhood features.

[0124] Next, the weighted summation result is concatenated with the original node embedding vector of the production node. Concatenation is the process of joining two vectors in a predetermined order. Through concatenation, the production node's own features and neighborhood features are integrated into a single vector, generating an intermediate node vector. This intermediate node vector contains both the production node's original features and the feature information of its neighboring nodes, providing a more comprehensive reflection of the production node's local features in the graph.

[0125] Step S1424: Perform nonlinear activation processing on the intermediate node vector to generate activated intermediate node vectors. The nonlinear activation processing uses the ReLU activation function.

[0126] To enhance the nonlinear expressive power of the graph neural network model, nonlinear activation is applied to the intermediate node vectors. The ReLU activation function is used here. ReLU is a commonly used nonlinear activation function that changes negative values ​​in the intermediate node vectors to 0, while leaving positive values ​​unchanged. This nonlinear activation introduces nonlinear factors, enabling the graph neural network model to learn more complex feature relationships and improving its expressive power and performance. The ReLU activation function generates the activated intermediate node vectors.

[0127] Step S1425: Perform layer normalization on the activated intermediate node vector to generate an aggregate node vector containing local structural information.

[0128] Finally, layer normalization is applied to the activated intermediate node vectors. Layer normalization is a technique used to normalize the output of layers in a neural network, which makes the feature vectors of each node have a similar distribution, helping to improve the training stability and convergence speed of the model. Through layer normalization, the activated intermediate node vectors are normalized, making the values ​​of their various dimensions comparable. After layer normalization, an aggregated node vector containing local structural information is finally generated. This aggregated node vector integrates the features of the producing node itself and the features of its neighboring nodes, and after nonlinear activation and layer normalization, it can effectively reflect the local structural information of the producing node in the graph.

[0129] Step S143: Perform global correlation analysis on the aggregated node vector through the attention mechanism layer of the graph neural network model to generate attention node vectors containing global dependencies. The global correlation analysis includes calculating the attention weights between any two nodes and performing feature fusion based on the weights.

[0130] The attention mechanism layer of the graph neural network model performs global association analysis on the aggregated node vectors. This layer captures global dependencies between nodes in the graph, enabling better analysis of anomalies in the generated knowledge graph. The global association analysis process involves two steps: calculating the attention weights between any two nodes and fusing features based on these weights.

[0131] Calculating the attention weight between any two nodes is the process of determining the degree of association between them. The attention weight reflects the degree of attention one node pays to another; a larger weight indicates a stronger association between the two nodes. The attention mechanism layer calculates the attention weight between any two aggregated node vectors. This calculation process is based on the feature information of the aggregated node vectors; different node vectors will produce different attention weights.

[0132] Weight-based feature fusion is the process of combining node features according to attention weights. For each aggregated node vector, the aggregated node vectors of other nodes are multiplied by their corresponding attention weights, and then these weighted vectors are summed to obtain a new vector. This new vector integrates the feature information of all nodes and is adjusted according to the attention weights, thus better reflecting the global dependencies between nodes. Through this method, an attention node vector containing global dependencies is finally generated. This attention node vector can comprehensively reflect the global relationships between nodes in the knowledge graph.

[0133] Step S144: The attention node vector is processed by the classification layer of the graph neural network model to calculate the anomaly probability, and an anomaly detection result containing the detection confidence of different anomaly types is generated. The anomaly probability calculation process includes inputting the attention node vector into a fully connected network and outputting the probability distribution through the softmax function.

[0134] The classification layer of a graph neural network model calculates anomaly probabilities on the attention node vectors to generate anomaly detection results that include detection confidence for different anomaly types. The classification layer is the last layer in the graph neural network model, and its main function is to determine whether anomalies exist in the generated knowledge graph and what types of anomalies they are based on the input attention node vectors.

[0135] The anomaly probability calculation process involves two steps: inputting the attention node vector into a fully connected network and outputting a probability distribution through a softmax function. A fully connected network is a common neural network architecture that takes the attention node vector as input and processes it through a series of linear transformations and non-linear activation functions to obtain an intermediate output vector. This intermediate output vector contains preliminary information about different anomaly types.

[0136] Next, the intermediate output vector is input into the softmax function. The softmax function is a commonly used probability transformation function that converts each element of the intermediate output vector into a probability value between 0 and 1, and the sum of all probability values ​​is 1. These probability values ​​are the detection confidence scores for different anomaly types, with each probability value representing the likelihood of the corresponding anomaly type appearing in the knowledge graph. Through this process, anomaly detection results containing detection confidence scores for different anomaly types are finally generated.

[0137] Step S150: Determine the types of abnormal events existing in the production process of diborane and the distribution characteristics of the abnormal events in the production association knowledge graph based on the abnormal detection results.

[0138] After obtaining the anomaly detection results from the production association knowledge graph, it is necessary to determine the types of abnormal events existing in the diborane production process and the distribution characteristics of these abnormal events in the production association knowledge graph based on these results. This will help to further understand the scope and extent of the impact of abnormal situations on the production process.

[0139] Step S151: Analyze the detection confidence in the anomaly detection results and extract the anomaly types with detection confidence exceeding a preset threshold as target anomaly types.

[0140] First, the detection confidence scores in the anomaly detection results are analyzed. Detection confidence scores represent the probability of different anomaly types occurring in the production-related knowledge graph. To determine the actual anomaly types, a preset threshold needs to be set. The preset threshold is a pre-defined probability value; when the detection confidence score of a certain anomaly type exceeds this preset threshold, that anomaly type is considered to have actually occurred. By traversing all detection confidence scores in the anomaly detection results, anomaly types with detection confidence scores exceeding the preset threshold are extracted and designated as target anomaly types. Target anomaly types are the actual anomalies occurring in the diborane production process, and subsequent analysis and processing will revolve around these anomaly types.

[0141] Step S152: Locate the production node or production edge corresponding to the target anomaly type, extract the timestamp information of the production node or production edge, and determine the start and end times of the anomaly event.

[0142] After identifying the target anomaly type, it is necessary to locate the corresponding production nodes or edges. Each production node and edge in the production association knowledge graph is associated with a specific production device or production parameter, and abnormal events often manifest on certain production nodes or edges. By analyzing the association between the target anomaly type and the production nodes and edges, the corresponding production nodes or edges can be located.

[0143] Next, extract the timestamp information from these production nodes or edges. The timestamp information records the collection time of each data point. By analyzing the timestamp information, the start and end times of the abnormal event can be determined. The start time is when the abnormal event begins to occur, and the end time is when the abnormal event ends. By determining these two times, the duration of the abnormal event can be understood.

[0144] Step S153: Perform change trend analysis on the historical feature vectors of the production node or production edge to determine the diffusion direction of abnormal events in the production-related knowledge graph. The change trend analysis includes calculating the difference value of the feature vectors in the time dimension.

[0145] To determine the propagation direction of abnormal events in the production-related knowledge graph, it is necessary to analyze the changing trends of the historical feature vectors of production nodes or edges. These historical feature vectors record the characteristic information of production nodes or edges at different points in time. By analyzing the changing trends of these feature vectors, the propagation path of abnormal events can be understood.

[0146] Trend analysis involves calculating the differences in feature vectors over time. For each production node or edge's historical feature vector, the difference between adjacent time points is calculated. These differences reflect the degree of change in the feature vector over time. By analyzing the magnitude and direction of these differences, the direction of anomaly propagation can be determined. If the difference in the feature vector of a production node or edge increases significantly after a certain time point, it indicates that the anomaly may have started from that node or edge and spread to other nodes or edges. By performing trend analysis on the historical feature vectors of all relevant production nodes and edges, the direction of anomaly propagation within the production-related knowledge graph is ultimately determined.

[0147] Step S154: Construct an abnormal event propagation model based on the start time point, end time point, and propagation direction. The abnormal event propagation model includes a propagation rate parameter in the time dimension and a propagation range parameter in the spatial dimension.

[0148] After determining the start time, end time, and propagation direction of the abnormal event, an abnormal event propagation model is constructed based on this information. The abnormal event propagation model describes the propagation patterns of abnormal events within the production-related knowledge graph, and it includes a propagation rate parameter in the time dimension and a propagation range parameter in the spatial dimension.

[0149] The propagation rate parameter in the time dimension reflects the speed at which an anomaly propagates over time. It is obtained by calculating the duration of the anomaly from its start time to its end time, as well as the distance it travels during that time. The propagation rate parameter can help predict how an anomaly will propagate in the future.

[0150] The spatial propagation range parameter reflects the spatial propagation range of anomaly events within the production-related knowledge graph. This parameter can be determined by analyzing the set of production nodes and edges involved in the propagation direction of the anomaly. The propagation range parameter helps to understand the extent of the impact of anomaly events on the production process.

[0151] By combining information on the start time, end time, and propagation direction, an anomaly event propagation model is constructed, incorporating both time-dimension propagation rate parameters and spatial-dimension propagation range parameters. This model can more comprehensively describe the propagation patterns of anomaly events within the production-related knowledge graph.

[0152] Step S155: Generate the distribution feature information of the abnormal event in the production-related knowledge graph through the abnormal event propagation model. The distribution feature information includes the set of nodes affected by the abnormal event, the set of edges, and the corresponding time coverage interval.

[0153] An anomaly propagation model is used to generate distribution characteristic information of anomaly events in the production-related knowledge graph. The distribution characteristic information is a detailed description of the scope and time of the impact of anomaly events in the production-related knowledge graph, which includes the set of nodes affected by the anomaly, the set of edges affected by the anomaly, and the corresponding time coverage interval.

[0154] The set of nodes affected by an anomaly refers to the set of production nodes affected during the propagation of the anomaly event. These affected production nodes can be identified using the spatial dimension propagation range parameter of the anomaly propagation model. The set of edges affected by an anomaly refers to the set of production edges connecting the affected production nodes, and these edges can also be identified using the anomaly propagation model.

[0155] The corresponding time coverage interval refers to the time range during which the abnormal event affects each affected production node and production edge. By using the time dimension propagation rate parameter of the abnormal event propagation model and the information of the start and end time points, the time coverage interval of each affected production node and production edge can be determined.

[0156] By employing an anomaly propagation model that comprehensively considers the propagation patterns of anomalies and the structure of the production-related knowledge graph, a distribution feature information is generated, including the set of nodes and edges affected by anomalies, as well as the corresponding time coverage intervals. This distribution feature information clearly demonstrates the distribution of anomalies within the production-related knowledge graph.

[0157] Step S160: Generate a production early warning instruction containing an event location identifier based on the abnormal event type and the distribution feature information, and send the production early warning instruction to the borane production control terminal to trigger an abnormal response operation.

[0158] After determining the types of abnormal events and their distribution characteristics in the production-related knowledge graph, a production early warning instruction containing an event location identifier needs to be generated based on this information. This instruction is then sent to the diborane production control terminal to trigger anomaly response operations. The production early warning instruction is used to notify the production control terminal of an abnormal event and guide it to take corresponding measures. The event location identifier accurately indicates the location of the abnormal event, facilitating processing by the production control terminal.

[0159] Step S161: Query the preset anomaly type response rule base and extract the response priority identifier and emergency handling strategy code associated with the target anomaly type.

[0160] To formulate reasonable production early warning instructions, it is necessary to query a pre-built exception type response rule base. This rule base is a pre-established database that stores response rules corresponding to different exception types. Each exception type is associated with a response priority identifier and an emergency handling strategy code. The response priority identifier indicates the processing priority of that exception type; different exception types may have different priorities, with higher-priority exceptions requiring priority handling. The emergency handling strategy code is a code used to identify the emergency handling strategy for that exception type; each code corresponds to a specific handling method. By querying the exception type response rule base, the response priority identifier and emergency handling strategy code associated with the target exception type are extracted. This information serves as a crucial component of the production early warning instructions, guiding the production control terminal to take appropriate measures.

[0161] Step S162: Extract the set of nodes and the set of edges affected by the anomaly from the distribution feature information, and obtain the device identifier or parameter identifier corresponding to the set of nodes, and the connection identifier or association identifier corresponding to the set of edges.

[0162] The set of nodes and edges representing the impact of anomalies is extracted from the distribution feature information. These sets record the scope of the impact of the anomaly event within the production-related knowledge graph. For the set of nodes, each production node is associated with a specific production device or parameter; the corresponding device identifier or parameter identifier can be obtained from the node set. The device identifier uniquely identifies the production device, and the parameter identifier uniquely identifies the production parameter. Similarly, for the set of edges, each production edge is associated with a specific device connection or parameter association; the corresponding connection identifier or association identifier can be obtained from the edge set. The connection identifier uniquely identifies the connection relationship between devices, and the association identifier uniquely identifies the association relationship between parameters. By extracting these identifiers, the location of the anomaly event can be accurately pinpointed.

[0163] Step S163: Convert the device identifier, parameter identifier, connection identifier and association identifier into positioning coordinate information that the production control system can recognize. The positioning coordinate information includes the physical location coordinates of the device and the logical location coordinates of the parameters.

[0164] To enable the production control system to accurately identify the location of abnormal events, equipment identifiers, parameter identifiers, connection identifiers, and associated identifiers need to be converted into positioning coordinate information that the production control system can recognize. Positioning coordinate information includes the physical location coordinates of the equipment and the logical location coordinates of the parameters. The physical location coordinates of the equipment are the physical location information of the production equipment in the actual production environment; the corresponding physical location coordinates can be found through the equipment identifier. The logical location coordinates of the parameters are the logical location information of the production parameters within the production control system; the corresponding logical location coordinates can be found through the parameter identifier. Similarly, connection identifiers and associated identifiers can also be converted into corresponding position coordinate information. By converting this identifier information into positioning coordinate information, the production control system can accurately locate the location of abnormal events.

[0165] Step S164: Based on the propagation rate parameters and propagation range parameters of the abnormal event propagation model, predict the extended node set and extended edge set of the abnormal event in the future time period, and generate extended positioning coordinate information containing the predicted impact range.

[0166] By utilizing the propagation rate and propagation range parameters of an anomaly propagation model, we can predict the set of expanded nodes and edges of an anomaly event within a future time period. The propagation rate parameter reflects the speed at which the anomaly event propagates over time, while the propagation range parameter reflects the spatial extent of its propagation. By combining these two parameters, we can predict the set of production nodes and production edges that the anomaly event may affect within a future time period. These sets of nodes and edges are then used as the expanded node set and expanded edge set, respectively.

[0167] Then, extended positioning coordinate information containing the predicted impact range is generated for the extended node set and extended edge set. Similar to the previous positioning coordinate information, this extended positioning coordinate information includes the physical location coordinates of the equipment and the logical location coordinates of the parameters; however, this coordinate information is generated based on the predicted impact range of the anomaly. This extended positioning coordinate information helps the production control terminal understand the possible expansion range of the abnormal event in advance, so that appropriate preventative measures can be taken.

[0168] Step S165: Aggregate the response priority identifier, emergency handling strategy code, location coordinate information and extended location coordinate information to generate a production early warning instruction containing an event location identifier chain with timestamp alignment. Each identifier node in the event location identifier chain corresponds one-to-one with a node or edge in the production association knowledge graph.

[0169] The response priority identifier, emergency handling strategy code, location coordinate information, and extended location coordinate information are aggregated to generate a production early warning instruction. To ensure the production early warning instruction accurately indicates the location and handling method of the abnormal event, an event location identifier chain with timestamp alignment needs to be generated. The event location identifier chain is a chain composed of multiple identifier nodes, each corresponding one-to-one with a node or edge in the production association knowledge graph. Timestamp alignment ensures that each identifier node in the event location identifier chain is consistent with its corresponding production node or edge in time. The response priority identifier, emergency handling strategy code, location coordinate information, and extended location coordinate information are integrated into the event location identifier chain to generate a production early warning instruction containing this information. The production early warning instruction includes information such as the handling priority of the abnormal event, the emergency handling strategy, the location of the abnormal event, and the predicted extension range.

[0170] Step S166: Parse the event location identifier chain in the production early warning instruction and extract the device identifier, parameter identifier, connection identifier or association identifier corresponding to each identifier node.

[0171] After generating a production early warning command, it needs to be parsed so that the production control terminal can understand and execute it. Parsing the event location identifier chain in the production early warning command extracts the device identifier, parameter identifier, connection identifier, or association identifier corresponding to each identifier node. Each identifier node in the event location identifier chain contains information about a specific production node or production edge. By parsing these identifier nodes, the corresponding device identifier, parameter identifier, connection identifier, or association identifier can be obtained. This identifier information is crucial for the production control terminal to process; through this identifier information, the production control terminal can accurately locate the position of the abnormal event and take appropriate measures.

[0172] Step S167: Query the equipment control table according to the equipment identifier to obtain the corresponding equipment adjustment instruction template. The equipment adjustment instruction template includes equipment start / stop control parameters and operating parameter adjustment range.

[0173] The equipment control table is queried based on the extracted equipment identifier. The equipment control table is a pre-built database that stores equipment adjustment instruction templates for each production device. Each equipment adjustment instruction template includes equipment start-stop control parameters and operating parameter adjustment ranges. The start-stop control parameters control the start and stop of the equipment, while the operating parameter adjustment ranges define the adjustable range of the equipment's operating parameters. By querying the equipment control table, the equipment adjustment instruction templates corresponding to the equipment identifiers are obtained. These templates serve as the basis for the production control terminal to adjust the equipment.

[0174] Step S168: Query the parameter correction table according to the parameter identifier to obtain the corresponding parameter correction instruction template, wherein the parameter correction instruction template includes the parameter target value and the correction rate limit.

[0175] The parameter correction table is queried based on the extracted parameter identifier. The parameter correction table is a pre-built database that stores parameter correction instruction templates for each production parameter. Each parameter correction instruction template includes the parameter target value and the correction rate limit. The parameter target value is the value the parameter should achieve under normal circumstances, and the correction rate limit specifies the maximum speed at which the parameter is corrected. By querying the parameter correction table, the parameter correction instruction templates corresponding to the parameter identifiers are obtained. These parameter correction instruction templates will serve as the basis for the production control terminal to correct the parameters.

[0176] Step S169: Query the association adjustment table according to the connection identifier or association identifier to obtain the corresponding association adjustment instruction template. The association adjustment instruction template includes the collaborative adjustment strategy of the associated device and the synchronous correction rules of the associated parameters.

[0177] The associated adjustment table is queried based on the extracted connection identifier or association identifier. The associated adjustment table is a pre-built database storing the associated adjustment instruction templates corresponding to each connection identifier or association identifier. Each associated adjustment instruction template contains the coordinated adjustment strategy for associated devices and the synchronization correction rules for associated parameters. The coordinated adjustment strategy for associated devices specifies how multiple associated devices should coordinate to ensure the stability of the production process. The synchronization correction rules for associated parameters specify how multiple associated parameters should be synchronized to ensure the correlation between parameters. By querying the associated adjustment table, the associated adjustment instruction templates corresponding to the connection identifier or association identifier are obtained. These associated adjustment instruction templates will serve as the basis for the production control terminal to adjust associated devices and associated parameters.

[0178] Step S1610: Combine the equipment adjustment instruction template, parameter correction instruction template and associated adjustment instruction template according to the time sequence of the event location identifier chain to generate a comprehensive adjustment instruction containing multi-stage operations.

[0179] The equipment adjustment instruction template, parameter correction instruction template, and associated adjustment instruction template are combined in chronological order according to the event location identifier chain. The event location identifier chain records the development process of the abnormal event over time. Combining these instruction templates in chronological order can generate a comprehensive adjustment instruction containing multi-stage operations. Based on the development of the abnormal event, the comprehensive adjustment instruction guides the production control terminal to adjust the equipment, parameters, and related relationships in stages, ensuring that the abnormal event is handled in a timely and effective manner.

[0180] Step S1611: Send the comprehensive adjustment command to the production control terminal so that the production control terminal executes equipment adjustment operation, parameter correction operation and related adjustment operation in sequence according to the time order of the comprehensive adjustment command, until the detection confidence of the abnormal event is lower than the preset threshold.

[0181] Finally, the integrated adjustment command is sent to the diborane production control terminal. Upon receiving the integrated adjustment command, the production control terminal executes equipment adjustment operations, parameter correction operations, and related adjustment operations sequentially according to the command's time order. The equipment adjustment operation controls the start-up and shutdown of production equipment and adjusts operating parameters according to the equipment adjustment command template. The parameter correction operation corrects production parameters according to the parameter correction command template. The related adjustment operation coordinates the adjustment of related equipment and parameters according to the related adjustment command template. Through these operations, the impact of abnormal events is gradually eliminated. During the execution of these operations, the detection confidence level of abnormal events is continuously monitored. When the detection confidence level of an abnormal event falls below a preset threshold, it indicates that the abnormal event has been effectively controlled, and the execution of the integrated adjustment command can be stopped. This process effectively detects and handles abnormal events during diborane production, ensuring the stability and safety of the production process.

[0182] Figure 2 The present application provides a schematic diagram of the hardware structure of an anomaly analysis system 100 for implementing the above-described anomaly analysis method for an ethylene borane production control system, as shown in the embodiment of this application. Figure 2 As shown, the anomaly analysis system 100 applied to the diborane production control system may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0183] In one possible design, the anomaly analysis system 100 applied to the borane production control system can be a single server or a group of servers. The server group can be centralized or distributed (e.g., the anomaly analysis system 100 applied to the borane production control system can be a distributed system). In some embodiments, the anomaly analysis system 100 applied to the borane production control system can be local or remote. For example, the anomaly analysis system 100 applied to the borane production control system can access information and / or data stored in machine-readable storage medium 120 via a network. As another example, the anomaly analysis system 100 applied to the borane production control system can be directly connected to machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the anomaly analysis system 100 applied to the borane production control system can be implemented on an anomaly analysis system for the borane production control system. By way of example only, the anomaly analysis system applied to the borane production control system can include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-layered cloud, etc., or any aggregation thereof.

[0184] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by an anomaly analysis system 100 applied to the diborane production control system to perform or use in order to accomplish the exemplary methods described herein.

[0185] In a specific implementation, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, enabling the processor 110 to execute the anomaly analysis method applied to the borane production control system as described in the above method embodiment. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected via a bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0186] The specific implementation process of processor 110 can be found in the various method embodiments executed by the anomaly analysis system 100 applied to the borane production control system. The implementation principle and technical effect are similar, and will not be repeated here.

[0187] Furthermore, this application embodiment also provides a readable storage medium containing computer-executable instructions. When the processor executes the computer-executable instructions, the above-mentioned anomaly analysis method applied to the borane production control system is implemented.

[0188] It should be noted that, in order to simplify the description disclosed in this application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of this application sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description disclosed in this application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of this application sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. An abnormality analysis method applied to a diborane production control system, characterized by, The method comprises: obtaining a multi-source production data set generated in the operation process of a diborane production control system, the multi-source production data set comprising device state data, raw material delivery data and reaction parameter data marked with a time stamp; constructing a production correlation knowledge graph based on the multi-source production data set, the production correlation knowledge graph comprising a plurality of production nodes and production edges connecting the production nodes, wherein the production nodes correspond to diborane production devices or production parameters, and the production edges correspond to physical connection relationships between production devices or logical correlation relationships between production parameters; performing graph feature extraction processing on the production correlation knowledge graph to obtain a node feature set of the production nodes and an edge feature set of the production edges; calling a pre-trained graph neural network model to perform graph structure analysis processing on the node feature set and the edge feature set to generate an abnormality detection result of the production correlation knowledge graph, the abnormality detection result comprising detection confidence of different abnormality types; determining an abnormal event type existing in the diborane production process and distribution feature information of the abnormal event in the production correlation knowledge graph according to the abnormality detection result; generating a production warning instruction comprising event positioning identification based on the abnormal event type and the distribution feature information, and sending the production warning instruction to a diborane production control terminal to trigger an abnormal response operation; The determination of the abnormal event type existing in the diborane production process and the distribution feature information of the abnormal event in the production correlation knowledge graph according to the abnormality detection result comprises: analyzing the detection confidence in the abnormality detection result and extracting an abnormal type with a detection confidence exceeding a preset threshold as a target abnormal type; locating a production node or a production edge corresponding to the target abnormal type, extracting time stamp information of the production node or the production edge, and determining a start time point and an end time point of the abnormal event; performing change trend analysis processing on a historical feature vector of the production node or the production edge to determine a diffusion direction of the abnormal event in the production correlation knowledge graph, the change trend analysis processing comprising calculating a difference value of the feature vector in the time dimension; constructing an abnormal event propagation model based on the start time point, the end time point and the diffusion direction, the abnormal event propagation model comprising a propagation rate parameter in the time dimension and a propagation range parameter in the space dimension; generating the distribution feature information of the abnormal event in the production correlation knowledge graph through the abnormal event propagation model, the distribution feature information comprising a node set, an edge set and a corresponding time coverage interval affected by the abnormality.

2. The abnormality analysis method for a diborane production control system according to claim 1, characterized by, The construction of the production correlation knowledge graph based on the multi-source production data set comprises: performing data alignment processing on the multi-source production data set to obtain an aligned production data sequence; extracting diborane production device identification information from the aligned production data sequence, and mapping a data set corresponding to each diborane production device identification information into a production device node, the production device node comprising a device type attribute and a real-time running state attribute; Extracting diborane production parameter identification information from the aligned production data sequence, mapping each diborane production parameter identification information corresponding data set to a production parameter node, the production parameter node containing parameter type attribute and real-time measurement value attribute; Based on the diborane production process flow diagram, the physical connection relationship between the diborane production equipment is extracted, and the production edge is established between the equipment nodes that exist direct material transmission or energy exchange, the production edge contains connection type attribute and transmission rate attribute; Based on the diborane production reaction kinetics model, the logical association relationship between the diborane production parameters is extracted, and the production edge is established between the parameter nodes that exist causal influence or collaborative change, the production edge contains influence direction attribute and association strength attribute; According to the production equipment node, production parameter node and corresponding production edge, an initial production association knowledge graph is constructed, and a connectivity verification process is performed on the initial production association knowledge graph, and isolated nodes and corresponding edges not connected to the main production process are deleted, and a final production association knowledge graph is generated.

3. The abnormality analysis method for a diborane production control system according to claim 2, characterized by, The connectivity verification process is performed on the initial production association knowledge graph, and the isolated nodes and corresponding edges not connected to the main production process are deleted, and a final production association knowledge graph is generated, including: Extracting a core node set corresponding to the main production process from the initial production association knowledge graph, the core node set containing raw material input node, reaction kettle node and product output node; Using breadth-first search algorithm to traverse the initial production association knowledge graph from the core node set, and marking all reachable nodes and corresponding edges connected to the core node through path; Identifying the nodes and corresponding edges in the initial production association knowledge graph that are not marked, the nodes and corresponding edges that are not marked are isolated nodes and edges not connected to the main production process; Deleting the isolated nodes and corresponding edges from the initial production association knowledge graph, and generating a production association knowledge graph after connectivity verification; The number of nodes and edges of the production association knowledge graph after connectivity verification is counted to ensure that the number of remaining nodes and edges meets the basic structure requirements of the diborane production process, and a final production association knowledge graph is generated.

4. The abnormality analysis method for a diborane production control system according to claim 2, characterized by, The graph feature extraction process is performed on the production association knowledge graph to obtain the node feature set of the production node and the edge feature set of the production edge, including: Feature coding processing is performed on the real-time running state attribute of the production equipment node to generate a device node feature vector, the feature coding processing including state category hot coding and state change rate normalization; Feature coding processing is performed on the real-time measurement value attribute of the production parameter node to generate a parameter node feature vector, the feature coding processing including parameter value range binning and parameter change trend symbolization; The device node feature vector and the parameter node feature vector are combined to obtain the node feature set of the production node; Feature extraction processing is performed on the transmission rate attribute of the production edge to generate transmission rate feature, the feature extraction processing including rate historical mean calculation and rate fluctuation range statistics; The correlation strength attribute of the production edge is subjected to feature extraction processing to generate a correlation strength feature, and the feature extraction processing includes historical data correlation coefficient calculation and influence lag time statistics; The transmission rate feature and the correlation strength feature are combined to obtain an edge feature set of the production edge.

5. The abnormality analysis method for a diborane production control system according to claim 4, characterized by, The device node feature vector and the parameter node feature vector are combined to obtain a node feature set of the production node, including: The device node feature vector is subjected to dimension expansion processing to generate an expanded device node feature vector, and the dimension expansion processing includes adding a device type coding dimension; The parameter node feature vector is subjected to dimension expansion processing to generate an expanded parameter node feature vector, and the dimension expansion processing includes adding a parameter type coding dimension; The expanded device node feature vector and the expanded parameter node feature vector are subjected to splicing processing to generate a combined node feature vector; The combined node feature vector is arranged in a timestamp order of the production node to generate a node feature set of the production node.

6. The abnormality analysis method for a diborane production control system according to Claim 1, characterized by, The node feature set and the edge feature set are subjected to graph structure analysis processing by calling a pre-trained graph neural network model to generate an abnormality detection result of the production correlation knowledge graph, including: The node feature set and the edge feature set are input into an input layer of the graph neural network model, and the input layer maps the node feature vector and the edge feature vector into a node embedding vector and an edge embedding vector, respectively; The node embedding vector and the edge embedding vector are subjected to neighborhood information aggregation processing by a graph convolution layer of the graph neural network model to generate an aggregated node vector containing local structure information, and the neighborhood information aggregation processing includes weighted summation of the node's own features and adjacent node features; The aggregated node vector is subjected to global correlation analysis processing by an attention mechanism layer of the graph neural network model to generate an attention node vector containing global dependency relationships, and the global correlation analysis processing includes calculating attention weights between any two nodes and performing feature fusion based on the weights; The attention node vector is subjected to abnormality probability calculation processing by a classification layer of the graph neural network model to generate an abnormality detection result containing detection confidence of different abnormal types, and the abnormality probability calculation processing includes inputting the attention node vector into a fully connected network and outputting a probability distribution by a softmax function.

7. The abnormality analysis method for a diborane production control system according to claim 6, characterized by, The node embedding vector and the edge embedding vector are subjected to neighborhood information aggregation processing by the graph convolution layer of the graph neural network model to generate an aggregated node vector containing local structure information, including: A neighborhood range is defined for each production node, and the neighborhood range includes first-order neighbor nodes and corresponding edges directly connected to the production node; The node embedding vector of each production node and the node embedding vectors of all neighbor nodes in its neighborhood are subjected to weighted summation processing to obtain a weighted summation result, and the weights of the weighted summation processing are calculated from the edge embedding vectors of the corresponding edges; The weighted sum result is spliced with the original node embedding vector of the production node to generate an intermediate node vector containing the self characteristics and neighborhood characteristics of the production node; The intermediate node vector is subjected to nonlinear activation processing to generate an activated intermediate node vector, and the nonlinear activation processing uses a ReLU activation function; The activated intermediate node vector is subjected to layer normalization processing to generate an aggregated node vector containing local structure information.

8. The abnormality analysis method for a diborane production control system according to Claim 1, characterized by, The production pre-warning instruction containing the event positioning identifier is generated based on the abnormal event type and the distribution characteristic information, and the production pre-warning instruction is sent to the diborane production control terminal to trigger an abnormal response operation, which includes: Querying a preset abnormal type response rule library to extract a response priority identifier and an emergency processing strategy code associated with the target abnormal type; From the distribution characteristic information, a node set and an edge set affected by the abnormality are extracted, and device identifiers or parameter identifiers corresponding to the node set and connection identifiers or association identifiers corresponding to the edge set are obtained; The device identifiers, parameter identifiers, connection identifiers, and association identifiers are converted into positioning coordinate information recognizable by the production control system, and the positioning coordinate information includes device physical location coordinates and parameter logical location coordinates; Based on the propagation rate parameter and the propagation range parameter of the abnormal event propagation model, an expanded node set and an expanded edge set of the abnormal event in a future time period are predicted, and expanded positioning coordinate information containing a predicted impact range is generated; The response priority identifier, the emergency processing strategy code, the positioning coordinate information, and the expanded positioning coordinate information are aggregated to generate a production pre-warning instruction containing an event positioning identifier chain aligned with a timestamp, and each identifier node in the event positioning identifier chain corresponds to a node or an edge in the production association knowledge graph one by one; The event positioning identifier chain in the production pre-warning instruction is parsed to extract device identifiers, parameter identifiers, connection identifiers, or association identifiers corresponding to each identifier node; According to the device identifier, a device control table is queried to obtain a corresponding device adjustment instruction template, and the device adjustment instruction template contains device start-stop control parameters and operation parameter adjustment ranges; According to the parameter identifier, a parameter correction table is queried to obtain a corresponding parameter correction instruction template, and the parameter correction instruction template contains a parameter target value and a correction rate limit; According to the connection identifier or the association identifier, an association adjustment table is queried to obtain a corresponding association adjustment instruction template, and the association adjustment instruction template contains a cooperative adjustment strategy of associated devices and a synchronous correction rule of associated parameters; The device adjustment instruction template, the parameter correction instruction template, and the association adjustment instruction template are combined in the time sequence of the event positioning identifier chain to generate a comprehensive adjustment instruction containing multi-stage operations; The comprehensive adjustment instruction is sent to the production control terminal, so that the production control terminal sequentially performs device adjustment operations, parameter correction operations, and association adjustment operations in the time sequence of the comprehensive adjustment instruction until the detection confidence of the abnormal event is lower than a preset threshold.

9. An abnormality analysis system applied to a diborane production control system, characterized by, The abnormality analysis system applied to the diborane production control system comprises a processor and a memory, the memory is connected with the processor, the memory is used for storing programs, instructions or codes, and the processor is used for running the programs, instructions or codes in the memory to realize the abnormality analysis method applied to the diborane production control system in any one of claims 1-8.

Images