Chemical process early warning method and electronic device

Abstract

The application discloses a chemical process early warning method and an electronic device. The chemical process early warning method comprises the following steps: acquiring sensor time series data sequence and operation log of a chemical process, and performing time alignment on the sensor time series data sequence based on flow integration; parsing log text of each operation event in the operation log into an operation intention vector; determining an event dynamic influence field strength changing over time according to the operation intention vector; calculating a residual feature of the event dynamic influence field strength and sensor data at a sampling alignment moment, and determining whether to give an early warning. The application solves the false alarm problem caused by the inability to quantify dynamic influence in the prior art, and significantly reduces the false alarm rate. Meanwhile, by using the pre-physical alignment operation, it is ensured that the time when the log occurs and the time when the sensor responds are accurately corresponding in a physical manner, and data misplacement caused by lag is avoided. Finally, the event dynamic influence field strength of real-time operation dynamically suppresses the false alarm caused by normal operation, and intelligent monitoring is realized.

Description

Technical Field

[0001] This invention relates to the field of chemical process-related technologies, and in particular to a chemical process early warning method, electronic equipment, storage medium, and computer program product. Background Technology

[0002] In large-scale continuous process industries such as coal chemical and petrochemical industries, the safety and stability of the production process are of paramount importance. With the development of industrial internet technology, modern factories have accumulated massive amounts of multimodal data, mainly including two categories: One type is structured time-series data (OT data): data from sensors such as temperature, pressure, and flow rate in a DCS (Distributed Control System). Its characteristics are high frequency, strong continuity, and direct reflection of physical state.

[0003] Another type is unstructured text data (IT / management data): operation records, maintenance work orders, inspection logs, etc. from MES (Manufacturing Execution System) or shift handover logs. Its characteristics are discrete and sparse, containing rich semantics of human intervention (such as "switching the standby pump" and "manually adjusting the valve opening").

[0004] Existing early warning systems primarily rely on DCS time-series data for threshold monitoring or statistical anomaly detection. However, in actual production, a large number of data fluctuations are caused by normal process operations.

[0005] For example, when an operator performs a "filter switching" operation, the pressure and flow rate in the relevant pipelines may experience sudden and significant fluctuations. However, a purely data-driven model cannot detect this human intervention and often misidentifies it as a equipment malfunction (such as a leak or blockage), triggering an alarm. These frequent false alarms can lead to a "crying wolf" mentality among operators, causing them to overlook the real risks.

[0006] To address the aforementioned issues, some existing technologies (such as knowledge graph-based fault diagnosis methods) attempt to incorporate operation logs. These methods typically utilize NLP techniques to extract log entities, construct a static "fault-operation" relationship graph, or assist in diagnosis by retrieving the graph after a fault occurs.

[0007] However, this fusion is static and a posteriori. It treats "operational events" as isolated points in time or nodes in a graph, lacking a quantitative description of the dynamic impact process of the events. In fact, the impact of an operation (such as "opening the reflux valve wider") on the system is not instantaneous, but has a duration window and dynamic decay characteristics (such as the impact decaying exponentially over time). Existing technologies cannot convert text logs into continuous numerical curves of the same dimension as sensor data, resulting in low fusion accuracy.

[0008] Finally, existing technologies suffer from multimodal misalignment due to the lack of physical-level data alignment. Multimodal fusion presupposes spatiotemporal data consistency. However, chemical processes exhibit significant long-term physical transport lag. Existing fusion methods often directly use "current time" as a benchmark, forcibly matching DCS data with log timestamps. Because DCS data itself has unsteady physical lag (affected by pipeline volume and real-time flow rate), the "time of operation" and the "time of sensor response" do not physically correspond. Therefore, existing technologies, using "misaligned data" without physical alignment, not only fail to learn correct causal relationships but also introduce severe noise, causing the model to completely fail under varying load conditions.

[0009] For example, Chinese patent CN119477277A discloses a fault diagnosis scheme that integrates multimodal data using a knowledge graph. Its core technical steps include: Entity extraction: Extracting entities such as "components" and "fault types" from unstructured text such as equipment maintenance records and operation logs using Natural Language Processing (NLP) technology. Graph construction: Establishing a static fault knowledge graph containing triple relationships of "component-operating parameters-fault". Cross-modal fusion: Associating real-time data from IoT sensors with node attributes in the knowledge graph. Graph reasoning: Mapping graph nodes to low-dimensional vectors using graph embedding algorithms, and performing logical reasoning on the graph using a rule engine to calculate the similarity between the current fault phenomenon and historical records, thereby outputting the cause of the fault.

[0010] However, this approach has the following inherent drawbacks when dealing with chemical processes that have highly dynamic characteristics: First, there is a lack of quantitative description of the "dynamic impact process": knowledge graphs essentially describe static logical relationships (e.g., "valve A" is associated with "pressure B"), but cannot describe dynamic, time-varying processes. The impact of chemical operations (such as "opening a reflux valve") on a system is a continuous process that decays over time (e.g., the impact lasts 20 minutes, with the first 5 minutes being intense and the next 15 minutes gradual). Existing technologies can only record that "an operation has occurred," but cannot construct a numerical function that changes over time to express this gradual impact.

[0011] Secondly, the fusion mechanism struggles to suppress false alarms: existing technologies are primarily used for causal reasoning "after a fault occurs," but are insufficient for real-time false alarm suppression "before a fault occurs." Existing technologies cannot convert operational intent into a numerical vector with direction and amplitude, and cannot directly perform addition and subtraction operations with sensor data at the signal processing level (e.g., subtracting the expected fluctuations caused by the operation from the abnormal signal). Therefore, when normal operations cause significant data fluctuations, the graph inference often lags, and the system still issues false alarms.

[0012] Finally, this scheme ignores the spatiotemporal misalignment of physical data. It directly correlates current sensor data with logs, neglecting the unsteady physical lags prevalent in long-process chemical plants. Without fluid dynamic alignment, the data foundation for map fusion suffers from severe spatiotemporal misalignment, easily leading to incorrect causal learning.

[0013] In summary, existing technologies lack a collaborative early warning method that can align data based on physical mechanisms and transform discrete text operation logs into a continuous dynamic influence field. They also lack a high-precision alignment mechanism and fail to quantify "human operation intentions" into a numerical vector that can be directly computed with time-series data. Therefore, they are unable to suppress false alarms caused by normal operations. Summary of the Invention

[0014] Therefore, it is necessary to provide a chemical process early warning method, electronic equipment, storage medium, and computer program product to address the technical problem that existing technologies are unable to suppress false alarms caused by normal operation.

[0015] This invention provides a method for early warning of chemical processes, comprising: Acquire sensor time-series data sequences and operation logs for chemical processes, and perform time alignment of the sensor time-series data sequences based on flow integrals to obtain sensor data corresponding to the sampling alignment time; The log text of each operation event in the operation log is parsed into an operation intent vector; Based on the operational intent vector, determine the dynamic influence field strength of the event over time; Calculate the dynamic impact field strength of the event at the sampling alignment time and the residual characteristics of the sensor data. Based on the residual characteristics and sensor data at the same sampling alignment time, determine whether to issue an early warning.

[0016] Further, the step of parsing the log text of each operation event in the operation log into an operation intent vector includes: The log text of each operation event in the operation log is parsed into an entity quadruple, which includes: object, action, direction, and magnitude. Each entity quadruple is encoded as an operation intent vector, and the operation intent vector is associated with the operation time of the corresponding log text. The operation intent vector includes: the object identifier code corresponding to the object, the action type code corresponding to the action type, the direction code corresponding to the direction, and the amplitude quantization value corresponding to the amplitude.

[0017] Furthermore, determining the time-varying event dynamic influence field strength based on the operational intent vector includes: For each operation intent vector, find the response characteristic parameter corresponding to the object encoding, and determine the corresponding response kernel function according to the action type encoding; Based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value, the single event field strength function corresponding to the operation intention vector is obtained; For each sampling alignment time, the superposition value of the function values ​​of one or more individual event field strength functions at that sampling alignment time is calculated as the event dynamic influence field strength at that sampling alignment time.

[0018] Furthermore, determining the corresponding response kernel function based on the action type encoding includes: When the action type is encoded as an adjustment operation, a first-order inertial hysteresis model function is used as the response kernel function; When the action type is encoded as a switch-type operation, a step model function is used as the response kernel function.

[0019] Furthermore, the response characteristic parameters include: static gain coefficient, process time constant, and / or natural decay coefficient. The step of obtaining the single-event field strength function corresponding to the operation intent vector based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value includes: Substitute the process time constant and / or the natural decay coefficient into the response kernel function; Based on the static gain coefficient, the response kernel function, the direction encoding, and the amplitude quantization value, the single-event field strength function corresponding to the operation intention vector is obtained as follows: ,in, For the i-th operation intention vector at time... The field strength of a single event, For a moment, The operation moment associated with the i-th operation intention vector. The direction encoding for the i-th operation intent vector. This is the magnitude quantization value of the i-th operation intent vector. Let be the static gain coefficient of the i-th operational intent vector. The response kernel function for the i-th operation intention vector at time i The response value.

[0020] Furthermore, the calculation of the event dynamics affecting the field strength and the residual characteristics of the sensor data at the sampling alignment time, and the determination of whether to issue an early warning based on the residual characteristics and sensor data at the same sampling alignment time, includes: For each sampling alignment time, perform the following operation: Calculate the residual characteristics of the event dynamics influence field strength and sensor data at the sampling alignment time; The normalized anomaly weights at the sampling alignment time are calculated based on the residual characteristics. An alarm operation is performed when the sensor data at the sampling alignment time meets the alarm conditions and the abnormal weight at the sampling alignment time is greater than or equal to the weight difference.

[0021] Furthermore: The calculation of the residual characteristics of the event dynamic influence field strength and sensor data at the sampling alignment time includes: constructing the residual characteristic function of the event dynamic influence field strength and sensor data at the sampling alignment time as follows: ,in, For at any time The residual characteristics, For time alignment Sensor data, Scaling factor For a moment The event dynamics affect the field strength. Substituting the sampling time into the residual feature function, the residual features at the sampling alignment time are obtained. The step of calculating the normalized anomaly weight at the sampling alignment time based on the residual characteristics includes: The normalized anomaly weight function is constructed as follows: ,in, For a moment Abnormal weights, For normalized exponential functions, The first feature weight, The weights are the second feature weights. For a moment The correlation coefficient between the event dynamics affecting the field strength and the time-aligned sensor data. As a bias term, the sampling alignment time is substituted into the anomaly weight function to obtain the anomaly weight at the sampling alignment time.

[0022] This invention provides an electronic device, comprising: At least one processor; and, A memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that are executed by at least one of the processors to enable the at least one processor to perform the chemical process early warning method as described above.

[0023] The present invention provides a storage medium that stores computer instructions, which, when executed by a computer, are used to perform all steps of the chemical process early warning method described above.

[0024] This invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the chemical process early warning method as described above.

[0025] This invention transforms human operations from qualitative textual descriptions into quantitative numerical curves by constructing a dynamic impact field for events. This allows the system to accurately calculate the expected fluctuations caused by the operation, thereby mathematically separating normal interference from abnormal signals. This solves the false alarm problem caused by the inability to quantify dynamic impact in existing technologies, significantly reducing the false alarm rate. Simultaneously, by utilizing pre-emptive physical alignment, it ensures a precise physical correspondence between the log occurrence time and the sensor response time, avoiding data misalignment due to lag. Finally, this invention calculates the residual characteristics between the dynamic impact field of the event and the sensor data. Based on these residual characteristics and sensor data, it determines whether to issue an alert, thus dynamically suppressing false alarms caused by normal operations through the real-time dynamic impact field strength of the event, achieving intelligent monitoring. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a chemical process early warning method according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating a chemical process early warning method according to another embodiment of the present invention. Figure 3 This relates to the early warning principle of existing technologies when dealing with human-induced interference in chemical processes. Figure 4 This invention illustrates the early warning principle for handling human-induced interference in chemical processes. Figure 5 The preferred embodiment of the present invention is a logical diagram of the architecture of a collaborative early warning system for chemical processes based on physical alignment and dynamic field fusion; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to the present invention. Detailed Implementation

[0027] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. Identical components are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the drawings, while the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component. These terms are used only for the convenience of describing this application and simplifying the description, and are not intended to indicate or imply that the device or element referred to has a specific orientation, or is constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0028] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as open-ended and encompassing, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "exemplary," or "some examples," etc., are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this application. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples.

[0029] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0030] In describing some embodiments, the term "connection" and its derivative expressions may be used. For example, the term "connection" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact with each other. The embodiments claimed herein are not necessarily limited to the content of this document.

[0031] Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0032] This document describes exemplary embodiments in conjunction with the accompanying drawings, all of which are idealized schematic diagrams. For clarity, the thickness proportions of various structural layers and regions may be enlarged in the drawings. Those skilled in the art should understand that due to objective factors such as manufacturing processes and measurement tolerances, the shape of the actual product may reasonably differ from that shown in the drawings. Therefore, the understanding of the exemplary embodiments should not be limited to the schematic shapes shown in the drawings. Any reasonable variations caused by the manufacturing process that are not substantially different from the shape shown in these drawings should be considered to fall within the scope of the embodiments disclosed in this specification. The drawings themselves are not intended to limit the precise geometry of the actual product, nor do they constitute a limitation on the scope of patent protection.

[0033] like Figure 1 The diagram shown is a flowchart of a chemical process early warning method according to an embodiment of the present invention, including: Step S101: Obtain the sensor time series data sequence and operation log of the chemical process, and align the sensor time series data sequence based on the flow integral to obtain the sensor data corresponding to the sampling alignment time. Step S102: Parse the log text of each operation event in the operation log into an operation intent vector; Step S103: Determine the event dynamic influence field strength that changes over time based on the operation intention vector; Step S104: Calculate the residual characteristics of the event dynamic influence field strength and sensor data at the sampling alignment time. Based on the residual characteristics and sensor data at the same sampling alignment time, determine whether to issue an early warning.

[0034] Specifically, the present invention can be applied to electronic devices with processing capabilities, such as computers.

[0035] This invention aims to solve the problem of high false alarm rates in existing chemical process fault early warning technologies, which are unable to effectively distinguish between actual equipment faults and normal human error. Specifically, it addresses the following shortcomings of existing technologies (such as fusion methods based on static knowledge graphs): Lack of dynamic quantification: Existing technologies treat operation logs as static time points or isolated graph nodes, which cannot describe the continuous and gradual impact of operation behavior on process variables (such as the long-tail effect after valve adjustment), resulting in a numerical misalignment between "operation" and "response".

[0036] Spatiotemporal misalignment: Existing multimodal fusion does not consider the long lag characteristics of chemical processes and directly associates the current logs with the current sensor data, resulting in a serious physical spatiotemporal misalignment in the training data.

[0037] First, step S101 is executed to obtain the sensor time-series data sequence and operation log of the chemical process. The sensor time-series data sequence is time-aligned based on the flow integral to obtain the sensor data corresponding to the sampling alignment time.

[0038] Specifically, a flow integration mechanism based on a digital material transport network is introduced to perform fluid dynamics backtracking on the original DCS time-series data, eliminating unsteady lags and obtaining a high-fidelity process dataset with strict physical spatiotemporal consistency, which serves as the benchmark for subsequent fusion. The sensor time-series data sequence includes sensor data corresponding to a series of sampling times. After time alignment of the sensor time-series data sequence based on flow integration, the times corresponding to the sensor data are corrected to the sampling alignment times.

[0039] Then, step S102 is executed to parse the log text of each operation event in the operation log into an operation intent vector.

[0040] Specifically, natural language processing technology is used to parse unstructured operation logs, extract the operation object, action direction and magnitude, and generate operation intent vectors (e.g., {object: reflux valve, direction: +open, magnitude: 10%}).

[0041] Then, step S103 is executed to determine the dynamic influence field strength of events that change over time, based on the operation intent vector.

[0042] Specifically, a dynamic impact field for events is constructed. Based on process control mechanisms, a time-decaying response function is defined as a single event field strength function for each type of operational intent vector. This function transforms discrete log text into a continuous, numerical impact field curve, quantifying the expected interference window of human operations on system variables over a future period. Then, the function values ​​of one or more individual event field strength functions at different sampling alignment times are superimposed to obtain the time-varying dynamic impact field strength of the events.

[0043] Finally, step S104 is executed to calculate the residual characteristics of the event dynamic influence field strength and sensor data at the sampling alignment time. Based on the residual characteristics and sensor data at the same sampling alignment time, it is determined whether to issue an early warning.

[0044] Specifically, by calculating the residual characteristics of the field strength and sensor data at the sampling alignment time, the dynamic impact of the event is determined, and the sensor data fluctuations are judged to be related to the operation event, thereby suppressing false alarms caused by normal operation through the residual characteristics.

[0045] This invention transforms human operations from qualitative textual descriptions into quantitative numerical curves by constructing a dynamic impact field for events. This allows the system to accurately calculate the expected fluctuations caused by the operation, thereby mathematically separating normal interference from abnormal signals. This solves the false alarm problem caused by the inability to quantify dynamic impact in existing technologies, significantly reducing the false alarm rate. Simultaneously, by utilizing pre-emptive physical alignment, it ensures a precise physical correspondence between the log occurrence time and the sensor response time, avoiding data misalignment due to lag. Finally, this invention calculates the residual characteristics between the dynamic impact field of the event and the sensor data. Based on these residual characteristics and sensor data, it determines whether to issue an alert, thus dynamically suppressing false alarms caused by normal operations through the real-time dynamic impact field strength of the event, achieving intelligent monitoring.

[0046] like Figure 2 The diagram shown is a flowchart of a chemical process early warning method according to another embodiment of the present invention, including: Step S201: Obtain the sensor time-series data sequence and operation log of the chemical process, and align the sensor time-series data sequence based on the flow integral to obtain the sensor data corresponding to the sampling alignment time.

[0047] Step S202: Parse the log text of each operation event in the operation log into an entity quadruple, wherein the entity quadruple includes: object, action, direction and amplitude; Each entity quadruple is encoded as an operation intent vector, and the operation intent vector is associated with the operation time of the corresponding log text. The operation intent vector includes: the object identifier code corresponding to the object, the action type code corresponding to the action type, the direction code corresponding to the direction, and the amplitude quantization value corresponding to the amplitude.

[0048] Step S203: For each operation intent vector, find the response characteristic parameter corresponding to the object encoding, and determine the corresponding response kernel function according to the action type encoding; Based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value, the single event field strength function corresponding to the operation intention vector is obtained; For each sampling alignment time, the superposition value of the function values ​​of one or more individual event field strength functions at that sampling alignment time is calculated as the event dynamic influence field strength at that sampling alignment time.

[0049] Step S204: For each sampling alignment time, perform the following operation: Calculate the residual characteristics of the event dynamics influence field strength and sensor data at the sampling alignment time; The normalized anomaly weights at the sampling alignment time are calculated based on the residual characteristics. An alarm operation is performed when the sensor data at the sampling alignment time meets the alarm conditions and the abnormal weight at the sampling alignment time is greater than or equal to the weight difference.

[0050] Specifically, step S201 is executed first to obtain the sensor time-series data sequence and operation log of the chemical process. The sensor time-series data sequence is time-aligned based on the flow integral to obtain the sensor data corresponding to the sampling alignment time.

[0051] Specifically, this step constructs a high-fidelity spatiotemporal aligned dataset based on physical mechanisms, aiming to eliminate the spatiotemporal misalignment of "data-log" caused by unsteady lags in long chemical processes, and to provide an accurate physical benchmark for subsequent fusion.

[0052] First, data acquisition is performed by obtaining high-frequency time-series sensor data sequences from the DCS (Distributed Control System). (such as temperature, pressure, flow rate). It is a matrix, where each time step represents a time step. Sensor data vectors, which include data from multiple sensors; unstructured operation logs obtained from MES or electronic duty log systems. , Includes multiple log lines, each line representing a specific time. Operation events.

[0053] Then, physical lag correction is performed, and a topological network model of the process flow is constructed using a flow integral alignment method based on a digital material transport network. Real-time flow data is then utilized. For pipe capacity Perform inverse integration Calculate the dynamic lag time that changes in real time with the operating conditions. Dynamic lag time, also known as unsteady residence time, refers to the dynamic changes in material flow velocity within a pipeline network caused by the frequent changes in load conditions (such as production increase, production decrease, or disturbances) in chemical production. Therefore, the residence time of material in a certain equipment or pipeline is not a fixed constant (steady), but a variable that fluctuates in real time with time and load (unsteady).

[0054] Specifically, a digital material transport network is constructed: the chemical process flow diagram (PFD) is abstracted into a computer-resolvable logical network. Among these: Equipment with physical volume, such as gasifiers and synthesis towers, is defined as node units and assigned an effective volume attribute (V). Define the pipeline connecting the equipment as a unidirectional transport link and assign it a real-time volumetric flow rate attribute. .

[0055] A digital material handling network can include one or more nodes. When a digital material handling network includes one node, the infeed node and the discharge node can be the same node, meaning that material enters a device, reacts in the device, and then exits from the device. Alternatively, there can be multiple nodes, cascaded from the infeed node to the discharge node. Material enters from the infeed node, passes through one or more intermediate nodes, and then exits from the discharge node.

[0056] Dynamic residence time calculation based on fluid mechanics: Establish a material tracking model based on mass conservation. For the i-th node, obtain the real-time flow of its inlet link. .from Starting from a certain moment, moving forward along the timeline ( The flow rate is accumulated and integrated until the accumulated fluid volume equals the effective volume of the node. That is, it satisfies the following integral equation: ; The solution This refers to the actual physical moment when the material enters the i-th node (equipment), i.e., the entry moment of that node.

[0057] This allows us to calculate unsteady lags, i.e., to obtain the dynamic lag time under the current operating condition. . This is the dynamic lag time, which varies with real-time flow. It fluctuates and changes in real time, rather than being a fixed constant.

[0058] Preferably, the flow rate is a volumetric flow rate.

[0059] Finally, baseline alignment is performed to align the raw sensor data. According to the calculation Perform a backtracking translation, The sampling time in the middle is shifted to the past. The aligned sampling times are obtained, and a physically aligned high-fidelity sensor dataset is generated. As an aligned sensor data sequence, the time corresponding to each sensor data point in the aligned sensor data sequence is the sampling alignment time. At this time, The alignment sampling time physically corresponds strictly to the operation log. The recorded moment.

[0060] Then, step S202 is executed to parse the log text of each operation event in the operation log into an entity quadruple, which includes: object, action, direction and magnitude. Each entity quadruple is encoded as an operation intent vector, and the operation intent vector is associated with the operation time of the corresponding log text. The operation intent vector includes: the object identifier code corresponding to the object, the action type code corresponding to the action type, the direction code corresponding to the direction, and the amplitude quantization value corresponding to the amplitude.

[0061] Specifically, this step performs vectorized parsing of the operational intent of unstructured logs, aiming to transform discrete natural language text into mathematical vectors that can be computed by computers.

[0062] First, entity extraction is performed using a pre-trained Industrial Natural Language Processing (NLP) model (such as a large-scale chemical engineering model based on BERT) to extract key entity quadruples <object(Obj), action(Act), direction(Dir), magnitude(Mag)> from the text of each operation event in the operation log. Here, object represents the target of the operation event, action represents the specific operation performed on the object, direction represents the direction of the action (e.g., increase or decrease, positive or negative), and magnitude quantifies the intensity or degree of the action.

[0063] Example: The log entry "08:00 Manually opened the reflux valve FV-101 by approximately 10%" can be extracted as follows:<FV-101, Switch, +,10%> .

[0064] Then, semantic quantization and encoding are performed: Object identification code ( Objects are represented by an identifier (ID). The object identifier (ID) is embedded and encoded to obtain an embedding vector as the object identifier encoding, or the object ID is hashed and the resulting hash value is used as the object identifier encoding. Action type coding ( ): Classify and encode action entities (Label Encoding). For example: Regulate = 1, Switch = 2. Different action type encodings will correspond to different kernel functions (e.g., action type encoding of 1 corresponds to a first-order inertial response, action type encoding of 2 corresponds to a step response).

[0065] Direction encoding ( The direction is represented by positive and negative 1. Positive or increasing is +1, and negative or decreasing is -1.

[0066] Amplitude quantization value ( ): Establish mapping rules for descriptions without clearly defined numerical values: "Fine adjustment" is mapped to within 5% of the standard operating range; "significant adjustment" is mapped to within 20% of the standard operating range; "full open / full close" is mapped to 100% of the standard operating range.

[0067] Finally, an operation intent vector is generated by converting the extracted quadruples into an operation intent vector. .in An embedding vector or hash value for the object ID; The action type is coded (which determines the shape of the influence curve). It is used for direction coding, with a value of +1 or -1 (which determines the positive or negative influence on the curve). This is the amplitude quantization value, which is the normalized operating amplitude (determining the peak height of the influence curve).

[0068] Then, step S203 is executed: for each operation intent vector, the response characteristic parameter corresponding to the object encoding is found, and the corresponding response kernel function is determined according to the action type encoding; Based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value, the single event field strength function corresponding to the operation intention vector is obtained; For each sampling alignment time, the superposition value of the function values ​​of one or more individual event field strength functions at that sampling alignment time is calculated as the event dynamic influence field strength at that sampling alignment time.

[0069] Specifically, this step is based on the construction of the event dynamic influence field of multi-morphological response kernel, aiming to use the operation intention vector generated in step S202 to construct a dynamic influence field that is highly consistent with the physical mechanism.

[0070] First, physical parameters are retrieved based on the object identifier in the operation intent vector. Retrieve the response characteristic parameters corresponding to the object identifier from the preset device mechanism parameter library: { , , },in, This is the static gain coefficient; The process time constant (which determines the rate of rise); This is the natural decay coefficient.

[0071] Then select the response kernel function and use action type coding. Select the corresponding response kernel function.

[0072] In one embodiment, determining the corresponding response kernel function based on the action type encoding includes: When the action type is encoded as an adjustment operation, a first-order inertial hysteresis model function is used as the response kernel function; When the action type is encoded as a switch-type operation, a step model function is used as the response kernel function.

[0073] Specifically, when =1, meaning when the action type is coded as a regulation operation, and the action type is a regulation operation (such as valve fine-tuning), a "first-order inertial hysteresis model" is used to describe the gradual process: ,in, For a first-order inertial hysteresis model function, The operation time is e, where e is the natural constant and e is the process time constant. and natural decay coefficient These are the response characteristic parameters retrieved based on the object identifier in the operation intent vector.

[0074] when =2, meaning the action type is encoded as a switch operation, the action type is a switch operation (such as pump start / stop), and a step model function is used, preferably an "S-shaped step model" (Sigmoid fitting), to describe the sudden change process: ,in, This is the step model function.

[0075] Then, single-event field generation is performed, combined with the static gain coefficient in the response characteristic parameters. Direction coding and amplitude quantization value Construct a single-event field strength function for a single operation event.

[0076] In one embodiment, the response characteristic parameters include: a static gain coefficient, a process time constant, and / or a natural decay coefficient. The step of obtaining the single-event field strength function corresponding to the operation intent vector based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value includes: Substitute the process time constant and / or the natural decay coefficient into the response kernel function; Based on the static gain coefficient, the response kernel function, the direction encoding, and the amplitude quantization value, the single-event field strength function corresponding to the operation intention vector is obtained as follows: ,in, For the i-th operation intention vector at time... The field strength of a single event, For a moment, The operation moment associated with the i-th operation intention vector. The direction encoding for the i-th operation intent vector. This is the magnitude quantization value of the i-th operation intent vector. Let be the static gain coefficient of the i-th operational intent vector. The response kernel function for the i-th operation intention vector at time i The response value.

[0077] Specifically, the single event field strength function of the i-th operation intention vector is: ,in, For the i-th operation intention vector at time... The field strength of a single event, For a moment, The operation moment associated with the i-th operation intention vector. The direction encoding for the i-th operation intent vector. This is the magnitude quantization value of the i-th operation intent vector. Let be the static gain coefficient of the i-th operational intent vector. This is the response kernel function for the i-th operation intent vector, i.e., the response value of the response kernel function determined by the aforementioned action type encoding.

[0078] The field strength function for a single event is a piecewise function; before the operation time, the output is 0, and after the operation time, the output is 0. The field strength function of a single event is based on the static gain coefficient and the response kernel function at time t. The response value (i.e., the function value of the response kernel function), the direction code, and the amplitude quantization value are calculated and output at time t. The field strength of a single event.

[0079] Finally, for each sampling alignment time, the superposition value of the function values ​​of one or more individual event field strength functions at that sampling alignment time is calculated as the event dynamic influence field strength at that sampling alignment time.

[0080] Specifically, the composite field superposition is based on the principle of linear superposition, which superimposes the influence of all N valid operational events within the current time window. This involves adding the field strength function of a single valid event within the current time window at the sampling alignment time to the function value of that single event at the sampling alignment time (i.e., the field strength of a single event), thus obtaining the dynamic influence field strength of the event at the sampling alignment time. Here, the valid operational events are those upstream of the node where the sensor is located.

[0081] Aligned sensor data sequence This includes sensor data corresponding to multiple sampling alignment moments. For each sampling alignment moment, the event dynamic impact field strength is calculated. A preset duration before or after each sampling alignment moment is used as the time window for that sampling alignment moment. When calculating the event dynamic impact field strength for a given sampling alignment moment, the effective single event field strength function within the current time window of that sampling alignment moment is determined. A valid operation event is defined as: the minimum value of the single event field strength function determined by the operation intent vector corresponding to that operation event within the current time window of that sampling alignment moment is greater than a preset minimum threshold. Simultaneously, a single event field strength function whose minimum value within the current time window of the sampling alignment moment is greater than the preset minimum threshold is considered a valid single event field strength function.

[0082] Since the field strength function of a single event is a piecewise function, it is 0 when it is less than the operation time and gradually decreases after it is greater than the operation time. Therefore, a preset minimum threshold is set to exclude operation intent vectors whose minimum value of the field strength function of a single event is less than or equal to the preset minimum threshold, so as to avoid over-computation.

[0083] Specifically, the function for constructing the dynamic influence field strength of the event is as follows: ,in, For a moment The dynamic influence of events on field strength, For a moment The number of valid single event field strength functions within the defined time window. When calculating the event dynamic influence field strength at each sampling alignment moment, substitute the corresponding moment to determine the valid single event field strength function within the time window defined by the sampling alignment moment. Calculate the function value of each valid single event field strength function at the sampling alignment moment. Add the number of function values ​​of the valid single event field strength functions to obtain the event dynamic influence field strength at that sampling alignment moment.

[0084] Finally, step S204 is executed, and for each sampling alignment time, the following operation is performed: Calculate the residual characteristics of the event dynamics influence field strength and sensor data at the sampling alignment time; The normalized anomaly weights at the sampling alignment time are calculated based on the residual characteristics. An alarm operation is performed when the sensor data at the sampling alignment time meets the alarm conditions and the abnormal weight at the sampling alignment time is greater than or equal to the weight difference.

[0085] Specifically, this step involves dynamic threshold decision-making based on collaborative attention. It abandons the traditional fixed threshold alarm approach and instead dynamically suppresses false alarms by calculating the collaborative consistency between "DCS measured data" and the "event impact field." The dynamic threshold does not directly modify the physical hard threshold set in the DCS control system. Instead, it refers to dynamically adjusting the "system's overall decision sensitivity" by calculating anomaly weights. When a known operation exists, the field strength is high, and the anomaly weight is lowered. In this case, even if the measured value exceeds the physical threshold, the system will still "suppress the alarm." In effect, this is equivalent to dynamically raising the system's alarm threshold. First, obtain the dual-channel features: Channel A (Physical Measurement): Aligned Sensor Data Sequence .

[0086] Channel B (Expected Interference): The generated events dynamically affect the field strength. .

[0087] Then, a collaborative attention mechanism is constructed, along with an anomaly weight function to calculate the "anomaly confidence weight" at each sampling alignment time. This anomaly weight function reflects whether the fluctuations at the current sampling alignment time cannot be explained by the operating field.

[0088] In one embodiment: The calculation of the residual characteristics of the event dynamic influence field strength and sensor data at the sampling alignment time includes: constructing the residual characteristic function of the event dynamic influence field strength and sensor data at the sampling alignment time as follows: ,in, For at any time The residual characteristics, For time alignment Sensor data, Scaling factor For a moment The event dynamics affect the field strength. Substituting the sampling time into the residual feature function, the residual features at the sampling alignment time are obtained. The step of calculating the normalized anomaly weight at the sampling alignment time based on the residual characteristics includes: The normalized anomaly weight function is constructed as follows: ,in, For a moment Abnormal weights, For normalized exponential functions, The first feature weight, The weights are the second feature weights. For a moment The correlation coefficient between the event dynamics affecting the field strength and the time-aligned sensor data. As a bias term, the sampling alignment time is substituted into the anomaly weight function to obtain the anomaly weight at the sampling alignment time.

[0089] Specifically, the residual characteristic is defined as the deviation between the measured value and the expected field: ,in, For at any time The residual characteristics, To align the time in the sensor data sequence Sensor data, β is a scaling factor used to unify the dimensions and amplitude gradient. Since the physical quantities measured by the DCS sensor (such as flow rate) and the generated event influence field (normalized or dimensionless values) are not on the same order of magnitude, the value of β can be adaptively and dynamically calculated based on the normal fluctuation variance of the DCS variable under historical operating conditions (e.g., set as the reciprocal of the variance). The purpose is to normalize the dynamic influence field strength of the event. Dynamic mapping to actual measurement The same magnitude of amplitude, thus the residual characteristics obtained by subtracting the two are... It is physically comparable. For a moment The event dynamics influence field is obtained by substituting the sampling time into the residual feature function to obtain the residual features at the sampling alignment time.

[0090] Substitute each sampling alignment time into the time in the above formula. The residual characteristics at the sampling alignment time are obtained.

[0091] Then, the normalized anomaly weight function is constructed using the Sigmoid function as follows: ,in: For a moment Abnormal weights; It is a normalized exponential function; The first feature weight can be preset based on experience, because... Including time Data from multiple sensors, therefore, For a moment The vector, therefore, can also be... The learnable weight matrix serves as a feature of the residual. Used to characterize the degree to which the model pays attention to "data amplitude differences"; The second feature weight can be preset based on experience, or it can be... Learnable weight matrix as a relevance feature, Used to characterize the degree of attention the model pays to "waveform trend similarity"; For a moment The cross-correlation coefficient between the event's dynamic influence field strength and the time-aligned sensor data is given, where X represents the aligned sensor data and F represents the event's dynamic influence field strength. Specifically, the cross-correlation coefficient is calculated using time... Using a preset time period as a baseline, calculate the cross-correlation coefficient between the dynamic influence field strength of events within the local sliding window and the time-aligned sensor data (if the waveforms of the two are similar and the correlation is high, then...). (Small strain).

[0092] This is a bias term.

[0093] During model training, and By iteratively optimizing the model using historical samples of normal operation and abnormal failures, the model can automatically learn and balance the influence of "absolute bias" and "trend consistency" on the final false alarm decision. Specifically, if the waveforms of the two parameters are highly similar, even with some residuals, the model will be able to adjust the weights accordingly. It will still get smaller. A machine learning model is built using anomaly weight functions, and the neural network parameters (i.e., weight matrices) in the model are adjusted. , and bias Supervised learning is performed using historical operational data. Historically confirmed real equipment failure segments (labeled 1) and data fluctuation segments caused by normal human operation (labeled 0) are selected. The aligned DCS temporal residual features and generated event field features are simultaneously input into the model. A binary classification loss function (e.g., calculating the error based on the predicted value and the true label) is used, and the weight parameters are iteratively updated through backpropagation. This ensures that the model's output αt approaches 0 when facing "normal operational fluctuations," and its output α approaches 1 when facing "real failures." Finally, when the sensor data at the sampling alignment time meets the alarm conditions and the abnormal weight at the sampling alignment time is greater than or equal to the weight difference, an alarm operation is performed.

[0094] Specifically, the dynamic early warning decision-making logic is implemented as follows: Set the baseline alarm threshold as The system according to The following judgment shall be executed:

[0095] Specifically, when the measured fluctuation at a certain sampling alignment time is large ( However, the residuals are very small and highly correlated, leading to If the fluctuation is very small, it indicates that the fluctuation is caused by the operation, thus suppressing the alarm. When the measured fluctuation is large and cannot be offset by the influencing field, it leads to... A large value indicates an unexpected fault, thus triggering an alarm. Among these, For the preset threshold, Preferably, it is a very small percentage multiplied by a base.

[0096] This embodiment parses operation logs into entity quadruples and determines a single time field strength function corresponding to each operation event, thereby constructing an event dynamic impact field strength that conforms to the characteristics of the operation event. This accurately transforms human operation from a qualitative text description into a quantitative numerical curve. This allows the system to accurately calculate the expected fluctuations caused by the operation, thus mathematically separating normal interference from abnormal signals and solving the false alarm problem caused by the inability to quantify dynamic impact in existing graph-based methods. Simultaneously, by utilizing a pre-emptive physical alignment operation, it ensures that the time of log occurrence and the time of sensor response are physically precisely correlated, avoiding data misalignment due to lag. Finally, unlike traditional fixed threshold alarms, this invention utilizes a collaborative attention mechanism to dynamically adjust alarm sensitivity based on the field strength of real-time operations, achieving intelligent monitoring that is lenient when there is human operation and strict when there is no human operation.

[0097] like Figure 3 and Figure 4 The diagram illustrates a comparison of the early warning principles of existing technologies and embodiments of the present invention in handling human-induced interference in chemical processes. Figure 3 In the existing technology illustration, when the operator performs valve adjustment or other operations at Top1 (e.g., 2.0 seconds), the DCS sensor signal, due to physical transmission lag, only begins to rise significantly and exceed the alarm threshold at Talarm1, generating a false alarm (time zone 31). Because the existing technology only records the operation as an isolated static time point, it cannot establish a numerical correlation between the operation and the subsequent delayed response, causing the system to incorrectly determine the fluctuation as a device malfunction and trigger a false alarm (shown in red in the diagram). However... Figure 4 In this embodiment, after the system identifies the operation intention at the Top 2 moment, it immediately generates an event dynamic influence field 41 that decays over time based on the mechanism model. Figure 4 The blue dashed line (envelope curve) represents the influence field that completely covers the subsequent fluctuation trajectory of the DCS signal in the time domain. When the measured signal exceeds the original threshold but remains within the envelope of this influence field, the system determines that the fluctuation is the expected response caused by the operation, and thus automatically executes alarm suppression. Figure 4(As shown in the blue area), it effectively distinguishes between real faults and operational interference.

[0098] Figure 5 The diagram below illustrates the architecture of a collaborative early warning system for chemical processes based on physical alignment and dynamic field fusion, representing the preferred embodiment of this invention. This architecture includes a dual-parallel data processing flow: the upper data stream 510 processes raw DCS time-series data 511, introducing a physical alignment module 512 to eliminate unsteady lags using a flow integral algorithm, generating a high-fidelity process dataset with physical spatiotemporal consistency as a spatiotemporal reference; the lower data stream 520 processes unstructured operation logs 521, transforming discrete text into operation intent vectors with direction and amplitude through semantic parsing, and further constructing a continuous event dynamic influence field 522. Both data streams ultimately converge in a collaborative attention network module 530. The model dynamically adjusts the attention weight for DCS data fluctuations based on the intensity of the real-time generated event influence field, achieving tolerant monitoring when there is human operation and strict adaptive monitoring when there is no human operation, ultimately outputting accurate early warning decisions after false alarm suppression.

[0099] One example of this invention is a synergistic early warning system for a long process of coal-to-methanol gasification-synthesis.

[0100] Operating Scenario: A coal-to-methanol plant is undergoing a load increase operation (coal slurry feed rate increased from 80% to 100%). At 08:50:00, the temperature / flow rate DCS sensor at the downstream "synthesis tower" inlet shows abnormal fluctuations, exceeding the warning threshold.

[0101] Objective: To determine whether the fluctuation is a genuine equipment malfunction or an expected physical response caused by normal operation of the upstream gasifier, and thus decide whether to trigger an alarm.

[0102] S1. Constructing a high-fidelity spatiotemporal alignment dataset based on physical mechanisms 1.1 Physical baseline establishment: First, the system performs spatiotemporal alignment of chemical process data by constructing a system that includes a gasifier (node ​​N1, volume 200m³). 3 ) and the transformation and purification unit (node ​​N2, equivalent volume 3000m³) 3 A digital material transport network to acquire syngas flow rate in real time. .

[0103] 1.2 Dynamic Alignment Calculation: When downstream fluctuations are detected at 08:50:00, the system uses the above alignment method to perform flow integral calculation. .

[0104] Because it is currently operating at 100% full load (flow rate 1200 m³ / h) 3The physical lag time was calculated in real time to be 2.67 minutes, thus accurately locating the upstream operation time as 08:47:20.

[0105] S2. Vectorized parsing of operation intent in unstructured logs 2.1 Log matching: The system retrieved an operation log at 08:47:20: "Fine adjustment of the central oxygen valve of the gasification furnace by about 5%".

[0106] 2.2 Intent Extraction: The intent was parsed into a quadruple using an NLP model: <Central oxygen valve, Regulate, +1 (open larger), 5%>.

[0107] 2.3 Vector Generation: Generating Standardized Operation Intent Vectors This clarifies the target of the operation, the type of action, and the expected direction of the impact.

[0108] S3. Constructing the Dynamic Impact Field of Events 3.1 Kernel Function Selection: Upon identifying the action type as "adjustment type," the system invokes the first-order inertial hysteresis response model. .

[0109] 3.2 Field Strength Calculation: Combining the physical gain coefficient of the oxygen valve, the system generates a numerical influence curve starting from 08:47:20, with the peak pointing downstream to around 08:50:00. This curve quantifies the envelope of expected flow and temperature fluctuations due to oxygen valve regulation.

[0110] S4. Dynamic threshold decision based on collaborative attention 4.1 Feature Fusion: The system performs collaborative analysis on the DCS out-of-limit fluctuations measured at 08:50:00 and the event influence field generated in S3.

[0111] 4.2 Suppression Decision: Due to the high degree of agreement between the measured fluctuation shape and amplitude and the predicted influence field (minimal residuals and high waveform correlation), the anomaly weighting function... The function value at that moment (i.e., the outlier weight at that moment) drops to an extremely low level.

[0112] 4.3 Final Judgment: The system determines that the fluctuation is an expected response caused by normal process adjustment, and outputs the result as Status=Suppress (suppression alarm), and prompts on the monitoring interface that "the fluctuation is affected by the oxygen valve operation at 08:47:20, and no action is required".

[0113] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0114] like Figure 6 The diagram shown is a hardware structure schematic of an electronic device according to the present invention, comprising: At least one processor 601; and, A memory 602 is communicatively connected to at least one of the processors 601; wherein, The memory 602 stores instructions that are executed by at least one of the processors to enable the at least one of the processors to perform the chemical process early warning method as described above.

[0115] Figure 6 Take the 601 processor as an example.

[0116] The electronic device may also include an input device 603 and a display device 604.

[0117] The processor 601, memory 602, input device 603 and display device 604 can be connected by a bus or other means. The figure shows an example of connection by a bus.

[0118] The memory 602, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the chemical process early warning method in the embodiments of this application, for example, Figure 1 , Figure 2 The method flow is shown. The processor 601 executes various functional applications and data processing by running non-volatile software programs, instructions, and modules stored in the memory 602, thereby realizing the chemical process early warning method in the above embodiments.

[0119] The memory 602 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the chemical process early warning method. Furthermore, the memory 602 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 602 may optionally include memory remotely located relative to the processor 601, and these remote memories can be connected via a network to the apparatus performing the chemical process early warning method. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0120] The input device 603 can receive user clicks and generate signal inputs related to user settings and function control of the chemical process early warning method. The display device 604 may include a display screen or other display equipment.

[0121] When one or more modules are stored in the memory 602, and are run by one or more processors 601, the chemical process early warning method in any of the above method embodiments is executed.

[0122] This invention transforms human operations from qualitative textual descriptions into quantitative numerical curves by constructing a dynamic impact field for events. This allows the system to accurately calculate the expected fluctuations caused by the operation, thereby mathematically separating normal interference from abnormal signals. This solves the false alarm problem caused by the inability to quantify dynamic impact in existing technologies, significantly reducing the false alarm rate. Simultaneously, by utilizing pre-emptive physical alignment, it ensures a precise physical correspondence between the log occurrence time and the sensor response time, avoiding data misalignment due to lag. Finally, this invention calculates the residual characteristics between the dynamic impact field of the event and the sensor data. Based on these residual characteristics and sensor data, it determines whether to issue an alert, thus dynamically suppressing false alarms caused by normal operations through the real-time dynamic impact field strength of the event, achieving intelligent monitoring.

[0123] One embodiment of the present invention provides a storage medium that stores computer instructions, which, when executed by a computer, are used to perform all steps of the chemical process early warning method described above.

[0124] In the context of this disclosure, a storage medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. The storage medium can be a machine-readable signal medium or a machine-readable storage medium. Optionally, the storage medium can be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), compact disc ROM (CD-ROM), magnetic tape, floppy disk, and optical data storage device.

[0125] One embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the chemical process early warning method as described above.

[0126] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for early warning of chemical processes, characterized in that, include: Acquire sensor time-series data sequences and operation logs for chemical processes, and perform time alignment of the sensor time-series data sequences based on flow integrals to obtain sensor data corresponding to the sampling alignment time; The log text of each operation event in the operation log is parsed into an operation intent vector; Based on the operational intent vector, determine the event dynamics field strength that changes over time; Calculate the dynamic impact field strength of the event at the sampling alignment time and the residual characteristics of the sensor data. Based on the residual characteristics and sensor data at the same sampling alignment time, determine whether to issue an early warning.

2. The chemical process early warning method according to claim 1, characterized in that, The step of parsing the log text of each operation event in the operation log into an operation intent vector includes: The log text of each operation event in the operation log is parsed into an entity quadruple, which includes: object, action, direction, and magnitude. Each entity quadruple is encoded as an operation intent vector, and the operation intent vector is associated with the operation time of the corresponding log text. The operation intent vector includes: the object identifier code corresponding to the object, the action type code corresponding to the action type, the direction code corresponding to the direction, and the amplitude quantization value corresponding to the amplitude.

3. The chemical process early warning method according to claim 2, characterized in that, The step of determining the event dynamic influence field strength changing over time based on the operation intent vector includes: For each operation intent vector, find the response characteristic parameter corresponding to the object encoding, and determine the corresponding response kernel function according to the action type encoding; Based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value, the single event field strength function corresponding to the operation intention vector is obtained; For each sampling alignment time, the superposition value of the function values ​​of one or more individual event field strength functions at that sampling alignment time is calculated as the event dynamic influence field strength at that sampling alignment time.

4. The chemical process early warning method according to claim 3, characterized in that, The step of determining the corresponding response kernel function based on the action type encoding includes: When the action type is encoded as an adjustment operation, a first-order inertial hysteresis model function is used as the response kernel function; When the action type is encoded as a switch-type operation, a step model function is used as the response kernel function.

5. The chemical process early warning method according to claim 3, characterized in that, The response characteristic parameters include: static gain coefficient, process time constant, and / or natural decay coefficient. The step of obtaining the single-event field strength function corresponding to the operation intent vector based on the response characteristic parameters, the response kernel function, the direction encoding, and the amplitude quantization value includes: Substitute the process time constant and / or the natural decay coefficient into the response kernel function; Based on the static gain coefficient, the response kernel function, the direction encoding, and the amplitude quantization value, the single-event field strength function corresponding to the operation intention vector is obtained as follows: ,in, For the i-th operation intention vector at time... The field strength of a single event, For a moment, The operation moment associated with the i-th operation intention vector. The direction encoding for the i-th operation intent vector. This is the magnitude quantization value of the i-th operation intent vector. Let be the static gain coefficient of the i-th operational intent vector. The response kernel function for the i-th operation intention vector at time i The response value.

6. The chemical process early warning method according to claim 1, characterized in that, The calculation of the event dynamics impact field strength and sensor data residual characteristics at the sampling alignment time, and the determination of whether to issue an early warning based on the residual characteristics and sensor data at the same sampling alignment time, includes: For each sampling alignment time, perform the following operation: Calculate the residual characteristics of the event dynamics influence field strength and sensor data at the sampling alignment time; The normalized anomaly weights at the sampling alignment time are calculated based on the residual characteristics. An alarm operation is performed when the sensor data at the sampling alignment time meets the alarm conditions and the abnormal weight at the sampling alignment time is greater than or equal to the weight difference.

7. The chemical process early warning method according to claim 6, characterized in that: The calculation of the residual characteristics of the event dynamic influence field strength and sensor data at the sampling alignment time includes: constructing the residual characteristic function of the event dynamic influence field strength and sensor data at the sampling alignment time as follows: ,in, For at any time The residual characteristics, For time alignment Sensor data, Scaling factor For a moment The event dynamics affect the field strength. Substituting the sampling time into the residual feature function, the residual features at the sampling alignment time are obtained. The step of calculating the normalized anomaly weight at the sampling alignment time based on the residual characteristics includes: The normalized anomaly weight function is constructed as follows: ,in, For a moment Abnormal weights, For normalized exponential functions, The first feature weight, The weights are the second feature weights. For a moment The correlation coefficient between the event dynamics affecting the field strength and the time-aligned sensor data. As a bias term, the sampling alignment time is substituted into the anomaly weight function to obtain the anomaly weight at the sampling alignment time.

8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that are executed by at least one of the processors to enable the at least one of the processors to perform the chemical process early warning method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium stores computer instructions, which, when executed by the computer, are used to perform all steps of the chemical process early warning method as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the chemical process early warning method as described in any one of claims 1 to 7.

Images