A monitoring system for gasoline hydrogen refueling facilities based on edge computing

Abstract

This invention relates to the fields of edge computing and high-risk environmental process monitoring technology, specifically a monitoring system for gasoline hydrogen refueling recovery facilities based on edge computing. It includes edge node monitoring, function migration request, proximity resource assessment, shadow model negotiation, and simulation and fidelity evaluation units. The system identifies hard failures by monitoring sensor heartbeat data packets. Its core is to respond to failure signals, negotiate and elect the optimal carrying node in the proximity node cluster, activate the shadow model to continuously estimate the key parameters of the failed node, and perform closed-loop evaluation of simulation fidelity. This invention can accurately identify data vacuums, achieving rapid dynamic migration of failed functions and continuous recovery of key data.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and high-risk environmental process monitoring technology, specifically a monitoring system for gasoline hydrogen refueling recovery facilities based on edge computing. Background Technology

[0002] In high-risk process environments such as gasoline hydrogen refueling facilities, monitoring systems heavily rely on critical sensors deployed at the edge to ensure process safety. However, when these critical sensors experience hard failures due to physical damage or communication interruptions, a data vacuum is created. Existing centralized monitoring systems often struggle to respond promptly to such sudden failures, resulting in monitoring blind spots and potentially serious safety hazards. Therefore, ensuring that monitoring functions can still be taken over instantly and with high fidelity in the event of a hard failure at an edge node, in order to avoid data vacuums and monitoring blind spots, is a pressing technical problem that needs to be solved. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a monitoring system for gasoline hydrogen refueling recovery facilities based on edge computing. Specifically, the technical solution of this invention includes:

[0004] The edge node monitoring unit is used to periodically monitor the sensor heartbeat data packets of the edge nodes and process the heartbeat data packets according to the preset data vacuum triggering rules to identify whether the edge node has experienced a hard failure, obtain a failure signal or a normal signal, and continue to perform periodic monitoring when a normal signal is generated.

[0005] The function migration request unit is used to broadcast a dynamic function migration request to the cluster of neighboring nodes of the failed node in response to a failure signal.

[0006] The neighboring resource assessment unit is used to assess the resource availability of each neighboring node in the neighboring node cluster in response to dynamic function migration requests, and obtain the resource availability index of each neighboring node.

[0007] The shadow model negotiation unit is used to negotiate and elect the optimal bearer node in the neighboring node cluster based on the resource availability index.

[0008] The shadow model inference unit is used by the bearer node to load and activate a pre-trained shadow model from the local cache or edge data center, collect multimodal secondary sensor data of the neighboring node cluster, and use the shadow model to continuously estimate the key parameters of the failed node to obtain the estimated value.

[0009] The simulation fidelity evaluation unit is used to continuously evaluate the estimated value in a closed loop, calculate the simulation deviation index, and execute the corresponding fidelity control operation based on the comparison result between the simulation deviation index and the preset fidelity threshold.

[0010] Preferably, the process by which the edge node monitoring unit determines and processes heartbeat data packets includes:

[0011] Monitor the timestamps, data quality flags, and sensor measurements in the heartbeat data packets;

[0012] Data vacuum triggering rules include:

[0013] Determine whether the time difference between the current system time and the timestamp exceeds the preset maximum tolerable delay threshold;

[0014] Determine whether the data quality indicator is in an invalid state;

[0015] Determine whether the sensor measurement value remains unchanged within a preset number of monitoring cycles;

[0016] When any condition is met, the edge node is determined to have experienced a hard failure, and a failure signal is generated.

[0017] Preferably, the process of assessing resource availability by neighboring resource assessment units includes:

[0018] Real-time measurement of CPU load, memory usage, and network bandwidth utilization of neighboring nodes;

[0019] The resource availability index is obtained by weighting the real-time CPU load rate, real-time memory usage rate, real-time network bandwidth utilization rate, and preset resource type weights.

[0020] Preferably, preset resource type weights are used to represent the resource profile of the shadow model;

[0021] Resource profiles are pre-configured based on whether the shadow model is CPU-intensive, I / O-intensive, or memory-intensive.

[0022] Preferably, the process of using the shadow model to perform continuity estimation in the shadow model inference unit includes:

[0023] Collect heterogeneous data vectors from each neighboring node in the neighboring node cluster;

[0024] Using a pre-trained lightweight transfer function, heterogeneous data vectors are mapped to preliminary estimates;

[0025] The preliminary estimate is fused using a multimodal weighted fusion method, combining the data reliability weights of neighboring nodes and the physical correlation weights between neighboring nodes and failed nodes.

[0026] The final estimated value is obtained by combining the system's bias correction.

[0027] Preferably, the data reliability weight is dynamically adjusted;

[0028] The data quality flags originate from the heartbeat data packets received by the bearer node from neighboring nodes.

[0029] And through a preset state-weight mapping rule, the discrete states of data quality indicators are converted into data reliability weights.

[0030] Preferably, the physical relevance weights are pre-defined;

[0031] The data is derived from the process flow diagram and physical model of the gasoline hydrogen refueling recovery facility, and is calibrated based on the physical distance between adjacent nodes and failure nodes and the correlation of process flow.

[0032] Preferably, the process of calculating the deduction deviation index by the deduction fidelity evaluation unit includes:

[0033] Obtain the estimated value calculated by the shadow model inference unit;

[0034] Obtain physical mechanism reference values, which are derived from a process-based mechanism model running on the bearer node;

[0035] The deduction deviation index is obtained by calculating the average absolute percentage error between the estimated value and the physical mechanism reference value.

[0036] Preferably, the fidelity control operations include:

[0037] Determine whether the inference deviation index is greater than the preset fidelity threshold;

[0038] When the simulation deviation index exceeds the preset fidelity threshold, a level 2 alarm or function renegotiation is triggered.

[0039] When the inference deviation index is not greater than the preset fidelity threshold, the shadow model inference unit continues to perform continuous estimation.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This system makes a comprehensive judgment based on the timestamp of the sensor heartbeat data packet, data quality flags, and the stagnation status of the measurement value from multiple dimensions. It can accurately and quickly identify hard failures such as physical damage to the sensor or communication interruption, effectively avoiding missed or false judgments caused by a single standard, and ensuring an immediate response to data gaps.

[0042] 2. This system evaluates the real-time resource availability of each neighboring node cluster and performs weighted calculations and autonomous negotiations based on the resource profiles of the shadow model to be migrated, which can dynamically elect the most suitable carrying node with the remaining resources. This ensures the success rate of function migration and the stability of shadow model operation, and solves the optimal node matching problem in edge resource-constrained environments.

[0043] 3. This system integrates multimodal secondary sensor data from neighboring node clusters through shadow model inference units; it innovatively adopts a dual weighting mechanism of dynamic data reliability weight and pre-calibrated physical correlation weight, which makes the estimation results not only data-driven but also more in line with physical mechanisms, greatly improving the fidelity and anti-interference ability of the estimation values.

[0044] 4. This system introduces a physical mechanism model based on the process flow as a cross-validation benchmark through a deduction fidelity evaluation unit to continuously evaluate the estimated value of the shadow model in a closed loop. When the deduction deviation exceeds the preset threshold, an alarm or function renegotiation can be triggered. This constructs a self-correcting and evaluation closed-loop control mechanism, which effectively prevents the monitoring blind spot caused by the drift of the shadow model itself and ensures the long-term reliability of the system in failure modes. Attached Figure Description

[0045] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0046] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0048] Example 1:

[0049] Please see Figure 1 A monitoring system for a gasoline hydrogen refueling facility based on edge computing, comprising:

[0050] The edge node monitoring unit is used to periodically monitor the sensor heartbeat data packets of the edge nodes and process the heartbeat data packets according to the preset data vacuum triggering rules to identify whether the edge node has experienced a hard failure, obtain a failure signal or a normal signal, and continue to perform periodic monitoring when a normal signal is generated.

[0051] The function migration request unit is used to broadcast a dynamic function migration request to the cluster of neighboring nodes of the failed node in response to a failure signal.

[0052] The neighboring resource assessment unit is used to assess the resource availability of each neighboring node in the neighboring node cluster in response to dynamic function migration requests, and obtain the resource availability index of each neighboring node.

[0053] The shadow model negotiation unit is used to negotiate and elect the optimal bearer node in the neighboring node cluster based on the resource availability index.

[0054] The shadow model inference unit is used by the bearer node to load and activate a pre-trained shadow model from the local cache or edge data center, collect multimodal secondary sensor data of the neighboring node cluster, and use the shadow model to continuously estimate the key parameters of the failed node to obtain the estimated value.

[0055] The simulation fidelity evaluation unit is used to continuously evaluate the estimated value in a closed loop, calculate the simulation deviation index, and execute the corresponding fidelity control operation based on the comparison result between the simulation deviation index and the preset fidelity threshold.

[0056] This embodiment provides a monitoring system for gasoline hydrogen refueling recovery facilities based on edge computing. Gasoline hydrogen refueling recovery facilities typically contain a large number of critical sensors for process monitoring in high-risk environments. The technical motivation of this invention is that when these critical edge nodes, i.e., sensors, experience hard failures, such as physical damage or communication interruptions leading to data vacuums, traditional centralized monitoring systems struggle to respond in a timely manner, potentially resulting in monitoring blind spots and safety hazards. This invention constructs a dynamic, self-healing shadow monitoring mechanism at the edge to ensure that the monitoring function can be taken over immediately and with high fidelity when the main sensor fails.

[0057] The monitoring system specifically includes an edge node monitoring unit, the purpose of which is to monitor the health status of each edge node within the system in real time and periodically. In this embodiment, the unit is deployed on each edge node or on a nearby monitoring master node; it continuously monitors the target edge node. Heartbeat data packets emitted by key sensors Heartbeat data packets This refers to a small, periodically broadcast data packet that announces its liveness and transmits critical status information to neighboring nodes; its source is sensor nodes. This unit itself is based on a set of preset data vacuum triggering rules. right Perform judgment and processing; data vacuum triggering rules This refers to a set of logical conditions used to determine whether a node has experienced a hard failure. Its function is to accurately identify when a sensor has completely lost its data reporting capability. The source of this is a multi-dimensional threshold based on process safety time calibration. Through the determination of this rule, the unit finally outputs a signal to identify whether the edge node has experienced a hard failure and generated a failure signal, or is in a normal signal state.

[0058] The system also includes a function migration request unit, designed to immediately initiate a function takeover process when a data vacuum is confirmed; this unit is designed to activate in response to a failure signal generated by the edge node monitoring unit; once a failure signal is received, such as a rule... Determined as The unit immediately sent a message to the failed node. A predefined cluster of neighboring nodes Broadcast a dynamic feature migration request Neighboring node cluster This refers to the point of failure in physical space or process flow. A closely related set of edge nodes that can provide data redundancy; dynamic function migration requests. This refers to a node containing failed nodes. The message, which identifies the identifier and the KPI (Key Performance Indicator) type to be derived, serves to notify... All nodes in the cluster, nodes It has become invalid and a node is needed to take over its function;

[0059] Meanwhile, the system stipulates that when the edge node monitoring unit generates a normal signal, i.e. If not, the function migration will not be triggered, and the edge node monitoring unit will continue its regular periodic monitoring tasks.

[0060] The system also includes a proximity resource assessment unit, whose purpose is to quickly assess which node in the cluster is capable of taking over a failed function upon receiving a function migration request; this unit is deployed in Each neighboring node in the cluster Above; it responds to received dynamic function migration requests. Triggered; the unit immediately evaluates the node. Its own resource availability is technically motivated by quantifying the resource constraints of edge nodes, ensuring that function migration, i.e., loading the shadow model, does not affect neighboring nodes. The normal monitoring tasks were severely impacted; ultimately, the unit calculated and output a quantified resource availability index. ;

[0061] The system also includes a shadow model negotiation unit, the purpose of which is to determine the capabilities of each node. To efficiently determine the sole bearer; this unit in Running within the cluster, it collects data from all neighboring nodes. Resource availability index calculated from neighboring resource assessment units This unit uses a self-negotiation mechanism, such as broadcasting its own... The maximum value is selected and used to determine the migration target, ultimately in the neighboring node cluster. The optimal bearer node is negotiated and elected. Optimal bearer node It refers to China has the highest The winning node of the value, i.e. This indicates that the node It has the most remaining resources and is best suited to host the shadow model;

[0062] The system also includes a shadow model derivation unit, the purpose of which is to simulate the shadow host, i.e., the carrying node. Reconstructing the failed node KPI estimation; this unit determined in the shadow model negotiation unit. Afterwards, Activated above; shadow model This refers to a custom model used for continuity estimation, derived from a pre-trained model package; the unit consists of carrier nodes. Load and activate this pre-trained shadow model from its local cache or edge data center. After activation, the unit begins collecting data from neighboring node clusters. include Its own multimodal secondary sensor data; multimodal secondary sensor data refers to data with heterogeneity and asynchronicity, such as data from neighboring nodes. temperature Acoustic signals Vibration frequency or pressure This unit utilizes a shadow model. Calculations are performed on this fused data to determine the key parameters of the failed nodes. For example, pressure is continuously estimated, ultimately yielding an estimated value. ;

[0063] The system also includes a projection fidelity evaluation unit, the purpose of which is to ensure the accuracy of the estimates generated by the shadow model projection unit. It is inherently reliable, preventing misjudgment of systemic disasters due to its own drift; this unit also carries the node. On the run, it is for Continuous closed-loop evaluation is performed; to this end, the unit calculates a projection deviation index. ; Deduction Deviation Index It refers to a real-time, quantitative metric used for evaluation. The degree of deviation from the actual physical state; this unit is based on the deduction deviation index. With a preset fidelity threshold The comparison results are used to perform corresponding fidelity control operations;

[0064] This invention provides a complete closed-loop monitoring system based on edge computing. Through an edge node monitoring unit and a function migration request unit, it achieves immediate identification and rapid response to sensor hard failures. Through a proximity resource evaluation unit and a shadow model negotiation unit, it dynamically and autonomously elects the optimal bearer node among resource-constrained edge nodes, solving the resource matching problem of where to migrate functions. Through a shadow model deduction unit, it implements a lightweight, multimodal fusion shadow model, ensuring continuous estimation of failure KPIs. Through a deduction fidelity evaluation unit, it introduces a closed-loop self-evaluation and control mechanism for the shadow model.

[0065] In summary, this system can effectively solve the problems of data vacuum and monitoring blind spots caused by hard failure of key sensors in high-risk environments such as gasoline hydrogen refueling facilities, and significantly improve the robustness, self-healing ability and estimation fidelity of the entire monitoring system.

[0066] Example 2:

[0067] The process by which the edge node monitoring unit judges and processes heartbeat data packets includes:

[0068] Monitor the timestamps, data quality flags, and sensor measurements in the heartbeat data packets;

[0069] Data vacuum triggering rules include:

[0070] Determine whether the time difference between the current system time and the timestamp exceeds the preset maximum tolerable delay threshold;

[0071] Determine whether the data quality indicator is in an invalid state;

[0072] Determine whether the sensor measurement value remains unchanged within a preset number of monitoring cycles;

[0073] When any condition is met, the edge node is determined to have experienced a hard failure, and a failure signal is generated.

[0074] This embodiment is a preferred and more specific implementation of the edge node monitoring unit in Example 1; the core function of the edge node monitoring unit is to trigger the data vacuum rule. To determine the node Has a hard failure occurred?

[0075] In this embodiment, the edge node monitoring unit monitors heartbeat data packets. The process of making a judgment includes:

[0076] This unit monitors heartbeat data packets in real time. timestamps in Data quality indicators and sensor measurements ;

[0077] Timestamp Refers to nodes The local time that generated the heartbeat packet serves to indicate the timeliness of the data;

[0078] Data quality indicators This refers to a flag generated and reported by the sensor's built-in self-diagnostic program, which indicates the reliability of the sensor's self-assessment data, such as VALID or INVALID.

[0079] Sensor measurement value This refers to the sensor in Physical measurements at any given time;

[0080] The above , , Together, they constitute the rules for determining data vacuum triggering. Parent characteristics;

[0081] This data vacuum triggering rule The decision logic is designed to be multi-dimensional, including but not limited to any of the following:

[0082] Latency Exceeded: Determine the current system time. With timestamp Does the time difference exceed the preset maximum tolerable delay threshold? That is, to determine Whether it is true or not; among them, This refers to the current system time of the monitoring master node or neighboring nodes; It is a preset maximum tolerable delay threshold, the source of which and the method of determination are: calibrated according to the process safety time of the gasoline hydrogenation process to ensure that failure identification is completed before potential cascading effects occur; this condition is used to identify data vacuums caused by communication interruptions or node failures.

[0083] Quality mark invalidation: Determine data quality mark Whether it is in a failed state; that is, to determine Is this condition met? This condition is used to identify situations where the sensor's self-diagnostic program has explicitly reported a physical fault.

[0084] Data stagnation: Determining sensor measurements Is it within the preset monitoring cycle? The internal state remains stagnant; that is, it is determined that... The state continued One monitoring cycle; It is the length of the monitoring cycle. It is a preset threshold; the method for determining N is: to calibrate according to the process characteristics of the gasoline hydrogenation recovery facility; in a highly dynamic and high-risk chemical process, key parameters should not remain absolutely constant physically for a long time, so N can be calibrated to a small value, which is greater than the normal measurement noise fluctuation period, but much smaller than the minimum failure response time that may lead to safety risks, so as to ensure that it can be identified in time when the sensor is physically stuck; this condition is used to identify latent failures where the sensor may be physically stuck but is still sending valid heartbeat packets;

[0085] When any one of the above three conditions is met, that is, when at least one of the three conditions is met, the edge node monitoring unit determines the edge node. A hard failure occurs, generating a failure signal and triggering subsequent function migration request units;

[0086] By employing the aforementioned multi-dimensional data vacuum triggering rules based on latency, quality, and numerical values, this invention achieves more accurate and robust identification of hard failures in sensors. It can not only identify communication interruptions but also complex failure modes such as sensor false liveness or known faults, effectively avoiding missed or false judgments that may result from a single judgment criterion, thereby greatly improving the accuracy and timeliness of failure identification.

[0087] Example 3:

[0088] The process of assessing resource availability by neighboring resource assessment units includes:

[0089] Real-time measurement of CPU load, memory usage, and network bandwidth utilization of neighboring nodes;

[0090] The resource availability index is obtained by weighting the real-time CPU load rate, real-time memory usage rate, real-time network bandwidth utilization rate, and preset resource type weights.

[0091] This embodiment is a preferred and more specific implementation of the neighboring resource assessment unit in Example 1; the core function of this unit is to assess neighboring nodes. The resource availability is used to obtain the resource availability index. ;

[0092] In this embodiment, the process by which the neighboring resource assessment unit assesses resource availability is as follows:

[0093] This unit measures neighboring nodes in real time. Real-time CPU load rate Real-time memory usage and real-time network bandwidth utilization ; This refers to three ranges. to The real-time performance metrics between nodes are derived from the nodes. Real-time measurements are obtained from the local operating system or monitoring agent;

[0094] This unit is based on real-time CPU load rate. Real-time memory usage Real-time network bandwidth utilization and preset resource type weights The resource availability index is obtained by performing a weighted calculation. ;

[0095] To achieve the above weighted calculation, this embodiment introduces a resource availability index. The technical motivation behind this is to quantify the resource constraints of edge nodes, providing a standardized and comparable decision-making basis for subsequent shadow model negotiation units; the calculation method is as follows:

[0096]

[0097] in, For nodes The availability index is a dimensionless value, calculated using this formula; The higher the value, the more likely the node is to be active. The more remaining resources a model has, the better it is suited to support a shadow model. For nodes Real-time CPU load rate For nodes Real-time memory usage For nodes Real-time network bandwidth utilization; All to Interval values, determined by nodes Real-time measurement and acquisition from the local operating system; The preset resource type weights are also dimensionless, and The role and source of these weights will be further explained in the implementation method.

[0098] This formula is a custom construction of the present invention; it is achieved by modifying nodes. The remaining availability of the three core resources, namely , , A weighted sum is then performed to obtain a comprehensive resource availability index. ; The computational logic is executed in the nearest resource assessment unit, and its calculation results... The output will be sent to the shadow model negotiation unit as an election. The basis;

[0099] Through the above-described formulaic and standardized calculation process, this invention will determine the neighboring nodes. The multi-dimensional and heterogeneous real-time resource status is converged into a single, quantifiable, and comparable resource availability index. This greatly simplifies the decision-making complexity of the shadow model negotiation unit, making it possible to... Electing the optimal host node in the cluster The process becomes efficient and reliable.

[0100] Example 4:

[0101] The preset resource type weights are used to represent the resource profile of the shadow model;

[0102] Resource profiles are pre-configured based on whether the shadow model is CPU-intensive, I / O-intensive, or memory-intensive.

[0103] This embodiment is a further limitation based on the embodiment in Example 3, and its purpose is to clarify the preset resource type weights. The technical content and setup methods;

[0104] In this embodiment, the preset resource type weight The technical essence is: a shadow model used to represent the model to be transferred. Resource profile;

[0105] The resource profile R(Ψi) refers to the characteristic description of the shadow model Ψi's dependence on different computing resources, namely CPU, memory, I / O, or network, at runtime; this resource profile is based on the shadow model. The characteristics exhibited during offline training and testing are pre-configured as, for example, CPU-intensive, I / O-intensive, or memory-intensive.

[0106] Adjustable parameters The method for determining it is as follows:

[0107] These weights are based on the shadow model to be transferred. The resource profile is used for pre-configuration;

[0108] like If it is a CPU-intensive model, such as one involving complex mathematical operations or shallow neural network inference, then... The weight should be pre-configured to be relatively high;

[0109] like It is a memory-intensive model, for example, if a large historical data window needs to be loaded. The weight should be pre-configured to be relatively high;

[0110] By using shadow models The resource profile code is Weights, and apply them to Example 3 In the calculation formula, the neighboring resource assessment unit is used in the calculation. At that time, one will automatically and selectively consider... Each node in the cluster Resource abundance;

[0111] This implementation method enables resource assessment, i.e. The computation is no longer homogeneous, but task-aware; evaluation Not only considering nodes What resources are available, and more importantly, whether those resources match the task to be migrated. This specific requirement; this profile-aware resource assessment ensures that the optimal bearer node elected by the shadow model negotiation unit is selected. It is truly capable of bearing The nodes effectively avoid placing a CPU-intensive task on the node. Migrate to a node where CPU load is already high. This significantly improves the success rate of function migration and the operational stability of the shadow model after migration.

[0112] Example 5:

[0113] The process of using the shadow model for continuity estimation in the shadow model derivation unit includes:

[0114] Collect heterogeneous data vectors from each neighboring node in the neighboring node cluster;

[0115] Using a pre-trained lightweight transfer function, heterogeneous data vectors are mapped to preliminary estimates;

[0116] The preliminary estimate is fused using a multimodal weighted fusion method, combining the data reliability weights of neighboring nodes and the physical correlation weights between neighboring nodes and failed nodes.

[0117] The final estimated value is obtained by combining the system's bias correction.

[0118] This embodiment is a preferred and more specific implementation of the shadow model deduction unit in Example 1; the core function of this unit is to utilize the shadow model. For failed nodes Key parameters Perform continuity estimation;

[0119] In this embodiment, the shadow model inference unit utilizes the shadow model. The process of performing continuity estimation includes:

[0120] Collect neighboring node clusters Each neighboring node heterogeneous data vectors Heterogeneous data vectors This refers to the neighboring nodes Collected multimodal secondary sensor data, such as It may contain nodes temperature Acoustic signals Vibration frequency or pressure These data are heterogeneous (different types) and asynchronous (different timestamps), hence the shadow model. Internally, it includes mechanisms for performing time window interpolation or event-driven sampling on this data to achieve temporal alignment and form a fused feature vector. ;

[0121] Utilizing pre-trained lightweight transfer functions heterogeneous data vectors Mapping to preliminary estimate Lightweight transfer function This refers to a model, such as a multinomial regression or a shallow neural network, which is obtained through offline training and testing using historical operational data from gasoline hydrogen refueling facilities and then solidified into a model package; for example, if the failure parameters... For pressure, neighboring nodes heterogeneous data vectors Including temperature and vibration Then an unrestricted lightweight transfer function It can be trained offline as a second-order polynomial: , where a, b, c, d are model coefficients determined during offline training; The function is to connect nodes Secondary data Mapping to failure parameters The preliminary estimate, therefore The output dimensions must be consistent with same; Its lightweight characteristics are designed to meet the needs of edge computing, i.e., the support nodes. Resource constraints;

[0122] Combine neighboring nodes Data reliability weight and neighboring nodes With failure node Physical correlation weight between For the preliminary estimate Perform multimodal weighted fusion; and All to The dimensionless weights of the intervals, their specific definitions and sources will be detailed in the implementation method;

[0123] Combined with system bias correction This yields the final estimated value. ; It is a system correction bias, and its dimensions are the same as... The same, its source is During offline training, the bias term, which is a trainable parameter of the model, is determined along with the training parameter.

[0124] To achieve the aforementioned multimodal weighted fusion, this embodiment introduces shadow estimation values. The technical motivation lies in the need for a lightweight model capable of integrating multi-source heterogeneous data in high-risk chemical environments, to provide uninterrupted KPI estimation and avoid monitoring blind spots when the main sensor fails; the calculation method is as follows:

[0125]

[0126] in, Failed node The estimated KPI value is the shadow value, and its physical dimensions are the same as... Same, for example It is calculated using this formula; Failed node The cluster of neighboring nodes; For nodes Data reliability weight, to Dimensionless; For nodes With nodes The weight of physical correlation between them to Dimensionless; For the pre-trained lightweight transfer function, Mapping to AND Preliminary estimation of units of the same dimension; To start from neighboring nodes Collected heterogeneous data vectors; To correct the bias of the system, the dimensions and The same applies; it is determined during offline training.

[0127] This formula is the core custom construct of this invention, and it is a multimodal weighted fusion algorithm; it runs in the shadow model inference unit. Together they constitute Calculate the parent feature; by fusing multiple neighboring nodes The weighted estimate is made by and Dual regulation significantly improves the fidelity of the shadow model and avoids the risks of single-point extrapolation;

[0128] Through the above steps and core formulas, this invention provides a lightweight and robust shadow model. It innovatively adopted That is, data reliability and The dual weighting mechanism based on physical correlation ensures accurate estimation. It not only relies on multi-source data, but also treats this data intelligently and differentially, thereby achieving high-fidelity continuous estimation of failed KPIs under conditions of limited edge resources.

[0129] Example 6:

[0130] The data reliability weight is dynamically adjusted;

[0131] The data quality flags originate from the heartbeat data packets received by the bearer node from neighboring nodes.

[0132] And through a preset state-weight mapping rule, the discrete states of data quality indicators are converted into data reliability weights.

[0133] This embodiment is a further limitation based on the embodiment in Example 5, and its purpose is to clarify the data reliability weight. The technical content and determination methods;

[0134] In this implementation, data reliability weight It is designed to adjust dynamically, rather than as a static value;

[0135] Its source is: carrier node Real-time received from neighboring nodes Heartbeat data packets Data quality indicators included ; Refers to neighboring nodes The sensor's self-diagnostic status is defined as in Example 2. same;

[0136] The method for determining it is: through a preset state-weight mapping rule. Data quality indicators Discrete states, such as VALID, INVALID, etc., are transformed into to Mathematical weight within the interval ;

[0137] The state-weight mapping rule ,Right now Examples are as follows:

[0138] like If the data quality is good, then Map it to This indicates that the carrying node Complete trust comes from nodes Data ;

[0139] like If the data quality is poor, then Map it to a preset low weight, for example ;this means think Data Its credibility is low and should be reduced. Contribution in the calculation;

[0140] like If the data is invalid, then Map it to ;this means Data It has already been declared invalid by itself;

[0141] If the carrying node Node not received Heartbeat Pack , i.e., node If it itself is also out of contact or has exceeded the timeout, then its It was also placed as ;

[0142] pass This mechanism, data reliability weight Dynamic adjustment was achieved; in each estimation of the shadow model derivation unit In the iteration, They will all use the latest Refresh Thus, real-time regulation The proportion it accounts for in the final estimate;

[0143] This invention introduces dynamically adjusted This makes the shadow model It has the ability to perceive its data source, i.e., neighboring nodes, in real time. The ability to maintain a healthy state; Like a dynamic valve, it automatically and in real time reduces or shuts off contributions from neighboring nodes whose data quality has deteriorated or which have failed; this greatly improves the estimated value. Its instant accuracy and anti-interference capabilities prevent the shadow model from being... Other nodes in the cluster that are failing are causing contamination.

[0144] Example 7:

[0145] The physical relevance weights are pre-defined;

[0146] The data is derived from the process flow diagram and physical model of the gasoline hydrogen refueling recovery facility, and is calibrated based on the physical distance between adjacent nodes and failure nodes and the correlation of process flow.

[0147] This embodiment is a further limitation based on the embodiment in Example 5, and its purpose is to clarify the weight of physical correlation. The technical content and determination methods;

[0148] and In contrast to its dynamic nature, in this embodiment, the physical correlation weight... Designed to be pre-calibrated static values;

[0149] Its source is: the process flow diagram and physical model of the gasoline hydrogenation recovery facility; It represents a kind of prior knowledge;

[0150] The calibration method is: based on neighboring nodes and failed nodes The physical distance and process flow correlation between them are calibrated;

[0151] The calibration is an offline process, completed in collaboration between process engineers and data scientists;

[0152] For example, in a gasoline hydrogenation reactor, if the failure node... It is a sensor that measures the reactor outlet temperature, while the neighboring node It is a sensor that measures the partial pressure of hydrogen at the reactor inlet; according to the physical model and process flow diagram, these two have a very strong process flow correlation in chemical reaction kinetics; therefore, It should be labeled as a high value, for example ;

[0153] If neighboring nodes It is a sensor that measures the vibration of a pump at a distant location. The physical distance is considerable, and its correlation with the reactor outlet temperature is almost nonexistent in the physical model. It should be labeled as a low value, for example ;

[0154] As A fixed weight in the estimation formula ensures the shadow model... The estimation always focuses on those neighboring nodes that are truly related in terms of physical mechanism, while always suppressing those unrelated nodes;

[0155] By introducing pre-calibrated This invention explicitly injects domain knowledge, namely process flow and physical model, into the shadow model. This makes the shadow model no longer a purely black-box data-driven model that depends on correlations, but a gray-box model that combines data-driven approaches with mechanistic constraints.

[0156] This combination enables estimation The results are more consistent with physical reality and technological logic, especially in neighboring nodes. Secondary data When it contains noise or transient spurious correlations, The calibration can effectively suppress the interference of irrelevant nodes, greatly improving the robustness and interpretability of the shadow model under complex working conditions.

[0157] Example 8:

[0158] The process of calculating the deduction deviation index by the deduction fidelity assessment unit includes:

[0159] Obtain the estimated value calculated by the shadow model inference unit;

[0160] Obtain physical mechanism reference values, which are derived from a process-based mechanism model running on the bearer node;

[0161] The deduction deviation index is obtained by calculating the average absolute percentage error between the estimated value and the physical mechanism reference value.

[0162] This embodiment is a preferred and more specific implementation of the deduction fidelity evaluation unit in Example 1; the core function of this unit is to calculate the deduction deviation index. ;

[0163] In this embodiment, the deduction fidelity evaluation unit calculates the deduction deviation index. The process includes:

[0164] Obtain the estimated value calculated by the shadow model inference unit. ; The shadow value is calculated in real time according to the formula in Example 5;

[0165] Obtain physical mechanism reference value Physical mechanism reference value The technological motivation lies in providing a real-time, quantitative metric for evaluation. The degree of deviation from the actual physical state serves as the basis for warnings and corrections; The source is a gasoline hydrogenation process, which has slightly higher computational overhead but is useful at key nodes. Mechanistic models that can still function, such as equations based on mass balance or energy balance; for example, if failure parameters... Given the reactor outlet pressure, a non-restrictive mechanistic model is proposed. It can be derived based on simplified fluid dynamics equations or ideal gas equations of state, for example... ,in, and Real-time inlet flow and temperature measurements obtained from other neighboring nodes. , and These are pre-calibrated process parameters or physical constants;

[0166] Used as The cross-validation benchmark, it is with They have the same physical dimensions;

[0167] By calculating the estimated value Compared with physical mechanism reference value Mean absolute percentage error between The deduction deviation index is obtained. ;

[0168] To achieve the above deviation calculation, this embodiment introduces a deduction deviation index. The calculation method is as follows:

[0169]

[0170] in, To derive the deviation index, a dimensionless percentage error value is calculated using this formula; The estimated value for the shadow model originates from the shadow model derivation unit. These are reference values ​​for the physical mechanism, derived from a mechanism model based on the process flow.

[0171] This formula is a standard The calculations are employed in this invention; they are performed within the inference fidelity evaluation unit. and yes The parent feature is computed; this unit continues to compute. To achieve Continuous closed-loop evaluation;

[0172] By introducing an independent physical mechanism reference value As a benchmark, and calculate This invention is a shadow model. A self-correcting and credibility quantitative evaluation mechanism was introduced;

[0173] This is no longer an open-loop system where estimation is done once and for all, but a closed-loop system where the estimation results can be monitored and verified in real time; this effectively prevents shadow models. The estimation itself is affected by model mismatch or input data drift. By deviating from the true physical state over a long period and insidiously, the risk of misjudging systemic disasters was avoided.

[0174] Example 9:

[0175] Fidelity control operations include:

[0176] Determine whether the inference deviation index is greater than the preset fidelity threshold;

[0177] When the simulation deviation index exceeds the preset fidelity threshold, a level 2 alarm or function renegotiation is triggered.

[0178] When the inference deviation index is not greater than the preset fidelity threshold, the shadow model inference unit continues to perform continuous estimation.

[0179] This embodiment is a further limitation based on the embodiments of Example 1 and Example 8, and its purpose is to clarify the situation when the deduction deviation index is... Once calculated, the fidelity control operations performed by the system;

[0180] In this embodiment, the fidelity control operation is a closed-loop control process based on threshold judgment; the process is as follows:

[0181] The system presets a fidelity threshold. ; It is an engineering parameter preset according to process safety requirements, for example... This corresponds to a minimum acceptable fidelity of 80%.

[0182] The simulation fidelity assessment unit determines the simulation deviation index in real time. Is it greater than the preset fidelity threshold? That is, to determine Is it valid?

[0183] When the deduction deviation index Greater than the preset fidelity threshold At that time, it indicates the shadow model The fidelity is insufficient, and the estimation results are no longer reliable; at this point, the system triggers a level two alarm, immediately sending an alarm to maintenance personnel or the upper-level system, indicating that... The function has failed and requires manual intervention or a switch to a safer, more conservative operating mode; alternatively, it may trigger functional renegotiation, in which case the system may selectively and automatically trigger the shadow model negotiation unit for dynamic renegotiation. Functional renegotiation refers to, for example, replacing the current bearer node. or Introducing more neighboring nodes into the calculation In order to improve Fidelity;

[0184] When the deduction deviation index Not greater than the preset fidelity threshold At that time, that is This indicates the shadow model The fidelity is within an acceptable range; at this point, the shadow model inference unit continues to perform continuous estimation, and the system remains in a normal shadow operation state.

[0185] The above-mentioned fidelity control operations constitute a system based on and Closed-loop feedback; the inference fidelity evaluation unit not only evaluates It also controls the subsequent behavior of the system;

[0186] Through this based This invention implements an adaptive fidelity management mechanism for closed-loop environmental fidelity control operations; the system no longer blindly trusts shadow estimation. Instead, it endows it with the capabilities of failure-alarm and failure-reconstruction;

[0187] It can not only promptly alert to unreliable estimates, but also proactively trigger system reconfiguration, i.e., functional renegotiation, to ensure that the entire monitoring system always operates at an acceptable, preset fidelity level, even in shadow mode. This achieves the ultimate balance between robustness and reliability in the monitoring system.

[0188] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A monitoring system for a gasoline hydrorefining recovery facility based on edge computing, characterized by, The method comprises the following steps: An edge node monitoring unit is configured to periodically monitor sensor heartbeat data packets of the edge node, and perform determination processing on the heartbeat data packets according to a preset data vacuum trigger rule, so as to identify whether the edge node has a hard failure, and obtain a failure signal or a normal signal, and continue the periodic monitoring when the normal signal is generated; A function migration request unit is configured to broadcast a dynamic function migration request to a neighboring node cluster of the failed node in response to the failure signal; A neighboring resource evaluation unit is configured to evaluate resource availability of each neighboring node in the neighboring node cluster in response to the dynamic function migration request, and obtain a resource availability index of each neighboring node; A shadow model negotiation unit is configured to negotiate and elect an optimal bearing node in the neighboring node cluster according to the resource availability index; A shadow model deduction unit is configured to load and activate a pre-trained shadow model from a local cache or an edge data center by the bearing node, collect multi-modal secondary sensor data of the neighboring node cluster, and continuously estimate key parameters of the failed node by using the shadow model, and obtain an estimated value; A deduction fidelity evaluation unit is configured to continuously evaluate the estimated value in a closed loop, calculate a deduction deviation index, and perform a corresponding fidelity control operation according to a comparison result of the deduction deviation index and a preset fidelity threshold; The fidelity control operation comprises: determining whether the deduction deviation index is greater than the preset fidelity threshold; when the deduction deviation index is greater than the preset fidelity threshold, triggering a secondary alarm or function re-negotiation; when the deduction deviation index is not greater than the preset fidelity threshold, the shadow model deduction unit continues to perform continuous estimation.

2. The monitoring system for a gasoline hydrorefining recovery facility based on edge computing according to claim 1, characterized by, The process of the edge node monitoring unit performing determination processing on the heartbeat data packet comprises: monitoring a time stamp, a data quality flag and a sensor measurement value in the heartbeat data packet; The data vacuum trigger rule comprises: determining whether the time difference between the current system time and the time stamp exceeds a preset maximum tolerance delay threshold; determining whether the data quality flag is in a failure state; determining whether the sensor measurement value remains unchanged within a preset number of monitoring periods; when any condition is met, determining that the edge node has a hard failure, and generating a failure signal.

3. The monitoring system for a gasoline hydrorefining recovery facility based on edge computing according to claim 1, characterized by, The process of the neighboring resource evaluation unit evaluating resource availability comprises: real-time measurement of real-time CPU load rate, real-time memory usage rate and real-time network bandwidth occupancy rate of the neighboring node; and according to the real-time CPU load rate, the real-time memory usage rate, the real-time network bandwidth occupancy rate and a preset resource type weight, performing weighted calculation to obtain the resource availability index.

4. The monitoring system for a gasoline hydrorefining recovery facility based on edge computing according to claim 3, characterized by, The preset resource type weight is used to represent a resource portrait of the shadow model; The resource portrait is pre-configured according to whether the shadow model is CPU-intensive, IO-intensive or memory-intensive.

5. The monitoring system for a gasoline hydrorefining facility based on edge computing according to claim 1, characterized by, The process of the shadow model deduction unit using the shadow model to perform continuous estimation comprises: collecting heterogeneous data vectors of each neighboring node in the neighboring node cluster; using a pre-trained lightweight transfer function to map the heterogeneous data vectors to a preliminary estimate; combining the data reliability weight of the neighboring node and the physical correlation weight between the neighboring node and the failed node to perform multi-modal weighted fusion on the preliminary estimate; The final estimated value is obtained by combining the system correction bias.

6. The monitoring system for a gasoline hydrorefining recovery facility based on edge computing according to claim 5, characterized by, The data reliability weight is dynamically adjusted; The data quality flag contained in the heartbeat packet received by the bearing node from the adjacent node; The discrete state of the data quality flag is converted into the data reliability weight through the preset state-weight mapping rule.

7. The monitoring system for a gasoline hydrorefining recovery facility based on edge computing according to claim 5, characterized by, The physical correlation weight is pre-calibrated; The physical correlation weight is derived from the process flow diagram and physical model of the gasoline hydrogenation recovery facility, and is calibrated according to the physical distance and process flow correlation between the adjacent nodes and the failed node.

8. The monitoring system for a gasoline hydrorefining recovery facility based on edge computing according to claim 1, characterized by, The process of calculating the inference deviation index by the inference fidelity evaluation unit includes: Obtain the estimated value calculated by the shadow model inference unit; Obtain the physical mechanism reference value, which is derived from the mechanism model based on the process flow running on the bearing node; The inference deviation index is obtained by calculating the mean absolute percentage error between the estimated value and the physical mechanism reference value.

Images