Dynamic compensation and error correction system of high-precision flow instrument

Abstract

The invention belongs to the technical field of flow instruments, and provides a dynamic compensation and error correction system of a high-precision flow instrument, which comprises a multi-dimensional data acquisition and preprocessing module, a module capable of being additionally provided with a sensor, a module for establishing a data stream with a timestamp and processing data, a module for fluid network modeling and priori knowledge construction, and a data processing module. A distributed collaborative optimization and parameter calibration module capable of constructing a directed graph model and generating initial compensation parameters; a network association reasoning and dynamic weight distribution module capable of adding a global penalty term to optimize parameters on the basis of a traditional error function; a working condition adaptation calibration and cross-domain parameter migration module capable of constructing an association graph, detecting abnormity and distributing weights; real-object-free calibration and parameter migration can be realized; according to the system, through multi-module cooperation, the problem that traditional single-table compensation is local and is not global is solved, and high precision and real-time performance of industrial-grade cooperative metering are supported.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of flow meters, and in particular to a dynamic compensation and error correction system for high-precision flow meters. BACKGROUND

[0002] In large industrial production scenarios, fluid delivery and metering systems often work cooperatively with multiple flow meters to achieve accurate monitoring and management of flow in complex pipe networks.

[0003] However, in existing technologies, the compensation algorithm parameters of each flow meter are usually optimized independently, and this independent optimization mode cannot effectively take into account the global flow conservation constraint of the entire fluid network.

[0004] When there are pressure fluctuations, flow state changes, or multiphase flow in the system, the local optimal solution of single-meter compensation parameters is prone to cause total mass balance deviation of the entire network, resulting in significant errors in metering results and actual flow demand, which may seriously affect material proportioning, energy consumption accounting, and process control accuracy in industrial production.

[0005] To overcome the technical problem of inaccurate total mass balance in the traditional independent optimization mode, the present application proposes a dynamic compensation and error correction system for high-precision flow meters, which realizes accurate metering and dynamic correction of global flow through a multi-module collaborative optimization mechanism. SUMMARY

[0006] To compensate for the deficiencies of existing technologies and solve at least one technical problem raised in the background art.

[0007] The technical solution adopted by the present application to solve its technical problems is:

[0008] In a first aspect, the present application provides a dynamic compensation and error correction system for high-precision flow meters, comprising:

[0009] A multi-dimensional data acquisition and preprocessing module: a miniature multiphase flow sensor is installed at the upstream and downstream 5D pipe diameters of each flow meter, a multi-dimensional data stream with time stamp is established through a precise clock protocol, an isolated forest algorithm is used to detect data jumps and sensor failures in real time, and a reliable multi-dimensional data sequence is generated through a redundant sensor voting mechanism;

[0010] A fluid network modeling and prior knowledge construction module: a directed graph model is constructed on an edge server, a real-time balance equation is established, a fluid mechanics simulation model is used to calculate the flow distribution under different working conditions, and a mapping table is generated as prior knowledge of initial compensation parameters;

[0011] Distributed collaborative optimization and parameter calibration module: on the basis of traditional single-table error function, global balance penalty term is added, gradient information of flowmeter compensation algorithm model is uploaded through homomorphic encryption technology, global optimization model is generated, filter parameters are adjusted based on real-time data, polynomial fitting coefficients in flowmeter compensation algorithm model are updated through gradient descent method, instrument coefficients are recalibrated by using federal learning method;

[0012] Network association reasoning and dynamic weight distribution module: a fluid network association graph is constructed, abnormal association detection is performed on the fluid network association graph, interference of associated pipelines is inferred through a graph neural network, compensation parameters of the flowmeter with the highest pressure fluctuation are automatically adjusted, a dynamic weight matrix is generated, and higher weight is given to the high-reliability instrument during total mass balance calculation;

[0013] Working condition adaptive calibration and cross-domain parameter migration module: a virtual standard flow signal is generated through a fluid mechanics simulation model to realize online calibration without physical objects, and the optimal compensation parameters under similar working conditions are quickly migrated by using transfer learning technology.

[0014] As a further improved scheme of the present application, the specific process of establishing a multi-dimensional data stream with a time stamp through a precise clock protocol is:

[0015] The sampling frequencies of the microwave phase sensor, the laser particle size instrument and the flowmeter are unified through the IEEE1588 precise clock protocol, the unified sampling frequency is greater than or equal to 100Hz, the precise time stamp is controlled within 1us, and the spatiotemporal consistency of the data collected by the microwave phase sensor, the laser particle size instrument and the flowmeter is ensured.

[0016] A multi-dimensional data stream with a time stamp is constructed, and the multi-dimensional data stream with a time stamp is specifically: Wherein, is the flow measurement value of the i th flowmeter at time t, is the pressure value of the i th pressure sensor at the corresponding measurement point at time t, is the temperature value of the i th temperature sensor at the corresponding measurement point at time t, is the flow component concentration of the i th microwave phase sensor at time t.

[0017] As a further improved scheme of the present application, the specific process of generating a reliable multi-dimensional data sequence through a redundant sensor voting mechanism is:

[0018] For redundant sensors of the same physical quantity, a two-out-of-three voting rule is adopted, and the two-out-of-three voting rule is specifically:

[0019] In the use scene of the flow meter, three microwave phase sensors or three laser particle size meters are selected for the three redundant sensors at the same position, and a two-out-of-three voting rule is used to screen out reliable multi-dimensional data sequences.

[0020] If the measurement deviation of two or more microwave phase sensors is less than a specified threshold, the average fluid component concentration is taken as valid data, i.e., a reliable multi-dimensional data sequence.

[0021] As a further improved scheme of the present application, the specific process of constructing the directed graph model is:

[0022] Based on the CAD drawing of the fluid system, a directed graph model is constructed on the edge server , wherein V is a node set, corresponding to the intersection of pipes, such as the points of three-way pipes, container interfaces, fluid convergence or flow separation positions, E is an edge set, corresponding to the pipe section where the flow meter is installed, and each edge is labeled with pipe geometric parameters, including pipe diameter D, length L, and inner wall roughness .

[0023] As a further improved scheme of the present application, the specific process of adding a global balance penalty term is:

[0024] Taking a single flow meter i as an optimization unit, a loss function with global constraints is constructed based on multi-dimensional data flow and flow conservation constraints: , wherein is the mean square error of the measurement value of the i-th flow meter and the true value, which is obtained by a fluid mechanics simulation model CFD simulation, is a set of topological nodes to which the i-th flow meter belongs, defined by the directed graph model, and the association between the flow meter and the node is clear, is the flow imbalance deviation of node V at time t, is a weight coefficient, which is dynamically adjusted by reinforcement learning, to preferentially ensure the accuracy of a single meter while constraining the global balance deviation.

[0025] As a further improved scheme of the present application, the specific process of constructing the fluid network correlation graph is:

[0026] A fluid network correlation graph is constructed, and each flow meter is defined as a node of the fluid network correlation graph, and the node features integrate three-dimensional data, including real-time measurement dimensional data, error history dimensional data, and working condition adaptation dimensional data; wherein the real-time measurement dimensional data includes signal-to-noise ratio and data jump frequency, the error history dimensional data includes the cumulative loss function of the edge computing module, and the working condition adaptation dimensional data includes the matching degree of the instrument error-working condition parameter mapping table.

[0027] The pipeline is an edge of a fluid network association graph, and the edge weight is based on the benchmark true value in the fluid network modeling and prior knowledge construction module, the benchmark true value is normalized, the fluid dynamics parameters between adjacent nodes are calculated as the edge weight, and the greater the edge weight, the stronger the dynamic interference generated by the fluid flow coupling of two flowmeters, and the more intense the fluid dynamics interaction.

[0028] As a further improved scheme of the present application, the specific process of abnormal association detection of the fluid network association graph is:

[0029] When the calculation deviation of a certain flowmeter belonging to the topology node exceeds the calculation deviation threshold, an association reasoning signal is triggered; the calculation deviation threshold is set by the industry technical personnel according to experience; the error propagation path is analyzed by using a graph convolution network (GCN), the neighbor node with the highest pressure fluctuation is located, and the error propagation path is learned; and the neighbor node with the highest pressure fluctuation is started according to the interference frequency characteristics.

[0030] If it is a high-frequency interference, the interference frequency is greater than 1Hz, a short-term optimization is triggered, and the filtering parameter is adjusted; if it is a low-frequency interference, the interference frequency is less than 1Hz, a medium-term optimization is triggered, and the polynomial fitting coefficient is updated.

[0031] As a further improved scheme of the present application, the specific process of realizing online calibration without real objects is:

[0032] When the system is in a non-full load stable working condition, input the current working condition parameters, and generate a virtual standard flow signal through a fluid mechanics simulation model , the virtual standard flow signal is used as an input reference value of a flowmeter compensation algorithm model to calculate the output after compensation , and the relative error is verified: If the relative error is greater than 1%, the relative error is greater than 1% for 5 sampling periods, a short-term optimization is triggered; if the relative error is still greater than 1% after the short-term optimization and lasts for 10 minutes, a medium-term optimization is started; if the relative error is greater than 2% after the medium-term optimization, the sensor hardware is marked as a fault, and an alarm is triggered.

[0033] As a further improved scheme of the present application, the specific process of quickly migrating the optimal compensation parameter under similar working conditions by using a transfer learning technology is:

[0034] ​When the pipeline pressure fluctuation transfer function form changes dramatically, it is determined that the flow state is abnormal, the Euclidean distance is calculated based on the instrument error-working condition parameter mapping table, the characteristic distance of the current working condition and the historical working condition is selected, the historical scene with the minimum characteristic distance and the similarity greater than 0.7 is selected as the migration source, the optimal compensation parameter set is extracted, the compensation parameters of the similar working condition are quickly loaded into the current algorithm, and after migration, the time cost of retraining is avoided through short-term optimization fine-tuning.

[0035] In the second aspect, the application provides a dynamic compensation and error correction method for a high-precision flow meter, comprising:

[0036] S1: A micro multi-phase flow sensor is installed on the upstream and downstream 5D pipe diameters of each flow meter, a multi-dimensional data stream with a time stamp is established through a precise clock protocol, an isolated forest algorithm is used to detect data jumps and sensor failures in real time, and a reliable multi-dimensional data sequence is generated through a redundant sensor voting mechanism;

[0037] S2: A directed graph model is constructed on an edge server, a real-time metering equation is established, a fluid mechanics simulation model is used to calculate the flow distribution under different working conditions, a mapping table is generated, and the mapping table is used as prior knowledge of the initial compensation parameters;

[0038] S3: On the basis of the traditional single-table error function, a global metering penalty term is added, gradient information of the flow meter compensation algorithm model is uploaded through homomorphic encryption technology, a global optimization model is generated, filter parameters are adjusted based on real-time data, polynomial fitting coefficients in the flow meter compensation algorithm model are updated through the gradient descent method, and instrument coefficients are recalibrated using the federal learning method;

[0039] S4: A fluid network association graph is constructed, abnormal association detection is performed on the fluid network association graph, interference of associated pipelines is inferred through a graph neural network, compensation parameters of the flow meter with the highest pressure fluctuation are automatically adjusted, a dynamic weight matrix is generated, and high-reliability instruments are given higher weights during total metering;

[0040] S5: A virtual standard flow signal is generated through a fluid mechanics simulation model, online calibration without physical objects is realized, and optimal compensation parameters under similar working conditions are quickly migrated using transfer learning technology.

[0041] The application has the following beneficial effects:

[0042] 1. A micro multi-phase flow sensor is installed on the upstream and downstream 5D pipe diameters of the flow meter, auxiliary parameters such as fluid components and bubble rates can be obtained, the short board of traditional single-dimensional data is made up, more comprehensive flow field feature data is provided for dynamic compensation, the IEEE1588 protocol is used to unify the sampling frequency and control the time stamp accuracy, the data space consistency is ensured, the isolated forest algorithm and the redundant sensor voting mechanism are combined, data jumps and sensor failures can be effectively detected, and a reliable multi-dimensional data sequence is generated.

[0043] 2. Construct a directed graph model and establish a real-time weighing equation to provide global consistency conditions for multi-table collaborative metrology as a constraint of mass conservation; generate reference true values and error mapping tables based on fluid mechanics simulation to provide prior knowledge for compensation; increase a global weighing penalty term in a traditional single-table error function, combine homomorphic encryption and federated learning, realize multi-time scale optimization, and solve the problems of local incompleteness and parameter update lag in single-table compensation.

[0044] 3. Construct a fluid network association graph, use a graph neural network to quantify implicit associations, locate error propagation paths, and generate a dynamic weight matrix to make high-reliability instruments dominant in total quantity weighing, and solve the problem of low-precision tables dragging the global. Generate a virtual standard signal through fluid mechanics simulation to realize online calibration without physical objects, combine transfer learning to quickly transfer optimal compensation parameters under similar working conditions, avoid retraining, and improve system adaptability and calibration efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0045] The application will be further described below with reference to the accompanying drawings.

[0046] Figure 1 is a system block diagram of a dynamic compensation and error correction system for a high-precision flow meter according to the application;

[0047] Figure 2 is a step flow chart of a dynamic compensation and error correction method for a high-precision flow meter according to the application. DETAILED DESCRIPTION

[0048] In order to make the technical means, creative features, purposes and effects achieved by the application easy to understand, the application will be further described below with reference to the specific embodiments.

[0049] Example 1

[0050] As shown in Figure 1 , the dynamic compensation and error correction system for a high-precision flow meter according to the embodiments of the application comprises:

[0051] Multi-dimensional data acquisition and preprocessing module: install a miniature multiphase flow sensor at the 5D pipe diameter upstream and downstream of each flow meter, establish a multi-dimensional data stream with a time stamp through a precise clock protocol, use an isolation forest algorithm to detect data jumps and sensor failures in real time, and generate a reliable multi-dimensional data sequence through a redundant sensor voting mechanism;

[0052] The specific process of installing a miniature multiphase flow sensor at the 5D pipe diameter upstream and downstream of each flow meter is as follows:

[0053] A micro multiphase flow sensor is installed in the upstream and downstream areas of each flow meter, which are 5 times the pipe diameter (5D). The micro multiphase flow sensor consists of a microwave phase sensor and a laser particle size analyzer.

[0054] The microwave phase sensor analyzes fluid composition by detecting changes in dielectric constant, and the laser particle size analyzer collects auxiliary parameters, including fluid composition, bubble rate, and particle concentration. The addition of a micro-multiphase flow sensor complements traditional technologies that only acquire single-dimensional data on flow rate, pressure, and temperature, providing more comprehensive flow field characteristic data for subsequent dynamic compensation.

[0055] Calibrate the spatiotemporal synchronization of the added microwave phase sensor and laser particle size analyzer;

[0056] The specific process of establishing a multi-dimensional data stream with a timestamp through the precision clock protocol is as follows:

[0057] The sampling frequencies of the microwave phase sensor, laser particle size analyzer, and flowmeter are unified using the IEEE1588 precision clock protocol, ensuring a uniform sampling frequency greater than or equal to 100 Hz to accurately capture high-frequency changes in the dynamic flow field. The timestamp accuracy is controlled within 1 μs, ensuring the spatiotemporal consistency of the data collected by the microwave phase sensor, laser particle size analyzer, and flowmeter.

[0058] On this basis, a multi-dimensional data stream with timestamps is constructed. The multi-dimensional data stream with timestamps is specifically as follows: ,in, is the flow measurement value of the i-th flow meter at time t, is the pressure value of the corresponding measurement point of the i-th pressure sensor at time t, is the temperature value of the corresponding measurement point of the i-th temperature sensor at time t, is the fluid component concentration of the i-th microwave phase sensor at time t;

[0059] The established multi-dimensional data stream with time stamp can be used to detect data jumps and sensor failures;

[0060] The specific process of using the isolation forest algorithm to detect data jumps and sensor failures in real time is as follows:

[0061] Using Isolation Forest to analyze multidimensional data streams Conduct online monitoring,

[0062] By constructing a random binary tree, data points that deviate from the normal distribution, such as instantaneous flow mutations and pressure jumps, are isolated;

[0063] Random binary trees can identify data jumps, i.e. multi-dimensional data that exceeds the normal fluctuation range and sensor faults, such as constant value output and over-range;

[0064] The specific process of generating a reliable multidimensional data sequence through the redundant sensor voting mechanism is as follows:

[0065] For redundant sensors with the same physical quantity, a two-out-of-three voting rule is used. The specific two-out-of-three voting rule is:

[0066] In the flow meter usage scenario, for three redundant sensors at the same location, three microwave phase sensors or three laser particle size analyzers are selected, and a two-out-of-three voting rule is used to screen out reliable multidimensional data sequences;

[0067] If the deviation of the measured values ​​of two or more microwave phase sensors is less than the specified threshold, the average fluid component concentration is taken as the valid data, which is a reliable multidimensional data sequence;

[0068] Micro-multiphase flow sensors containing microwave phase sensors and laser particle size analyzers are installed in the 5D pipe diameter area upstream and downstream of each flow meter. In addition to traditional flow, pressure, and temperature data, auxiliary parameters such as fluid composition, bubble rate, and particle concentration are also obtained to supplement the shortcomings of single-dimensional data. This provides comprehensive flow field characteristics for subsequent dynamic compensation, making the compensation basis richer and more accurate.

[0069] Fluid network modeling and prior knowledge building module: Build a directed graph model on the edge server, establish real-time balance equations, use fluid mechanics simulation models to calculate flow distribution under different working conditions, and generate a mapping table as prior knowledge for initial compensation parameters;

[0070] The specific process of constructing the directed graph model is as follows:

[0071] Based on the CAD drawings of the fluid system, a directed graph model is built on the edge server. , where V is a node set, corresponding to the intersection of pipelines, such as tees, container interfaces, and other points where fluids converge or diverge. E is an edge set, corresponding to the pipeline section where the flow meter is installed. Each edge is marked with pipeline geometric parameters, including: pipe diameter D, length L, inner wall roughness ;

[0072] The specific process of establishing the real-time balance equation is as follows:

[0073] According to the core equation of mass conservation, a real-time balance equation is established for any node V in the directed graph model: ,in, is the volume of the pipeline section where the node V is located at time t, is the total flow into node V at time t, The total flow rate of the node V at time t, the rate of change reflects the process of filling and emptying the pipeline;

[0074] When the system flow is stable, the volume change rate tends to 0, at this time the flow conservation constraint is forced to be satisfied, that is: ;

[0075] Use the fluid mechanics simulation model to calculate the flow distribution under different working conditions, and generate a mapping table;

[0076] Use the fluid mechanics simulation model (CFD) to simulate typical working conditions, including normal flow, extreme load variation, pressure mutation, multiphase flow, etc. By solving the Navier-Stokes (N-S) equation, output the simulation reference value distribution under each working condition as the benchmark true value;

[0077] Based on the benchmark true value, an error mapping table is constructed. Compare the flow meter measurement value with the benchmark true value output by the fluid mechanics simulation model CFD simulation under the same working condition, and calculate the flow meter error , wherein, The error of the i-th flow meter measurement value, that is, the deviation between the measurement result and the true value, reflects the accuracy of the flow meter measurement, The measurement value of the i-th flow meter, which is the flow data actually detected and output by the flow meter, is obtained through the sensing, conversion and other links of the flow meter, The true flow value corresponding to the i-th flow meter, which is the flow data objectively existing and reflecting the actual fluid flow situation, can be obtained through high-precision standard devices, and is used as a reference benchmark for measuring the accuracy of the flow meter; Generate an instrument error-working condition parameter mapping table, in which the working condition parameters include temperature T, pressure P, fluid component C, and flow state;

[0078] Establish the correlation between error and working condition parameters, and use the instrument error-working condition parameter mapping table as the prior basis for initial compensation parameters;

[0079] The time-stamped multi-dimensional data stream is used as real-time input to drive the calculation of the balance equation of the directed graph model;

[0080] The mass conservation equation is the physical cornerstone of fluid network balancing, and the real-time balancing equation is established as a constraint to provide global consistency conditions for subsequent multi-meter collaborative metering;

[0081] Directed graph model Define the calculation boundary of the equation, and use the instrument error-working condition parameter mapping table to provide offline prior reference for error compensation, which cooperates to support the convergence of the subsequent distributed optimization algorithm;

[0082] Distributed collaborative optimization and parameter calibration module: on the basis of traditional single-table error function, a global balance penalty term is added, the gradient information of the flowmeter compensation algorithm model is uploaded through homomorphic encryption technology, a global optimization model is generated, the filtering parameters are adjusted based on real-time data, the polynomial fitting coefficients in the flowmeter compensation algorithm model are updated through gradient descent method, and the instrument coefficient is recalibrated by using federal learning method;

[0083] The specific process of adding a global balance penalty term on the basis of a traditional single-table error function is as follows:

[0084] Taking a single flowmeter i as an optimization unit, a loss function with global constraints is constructed based on multi-dimensional data flow and flow conservation constraints: , wherein, is the mean square error of the measurement value and the true value of the ith flowmeter, which is obtained by simulating a fluid mechanics simulation model CFD, is a set of topological nodes to which the ith flowmeter belongs, which is defined by a directed graph model, and the association between the flowmeter and the node is clear, is the flow imbalance deviation of node V at time t, is a weight coefficient, which is dynamically adjusted by reinforcement learning, and the single-table precision is preferentially guaranteed while the global balance deviation is constrained;

[0085] The specific process of uploading the gradient information of the flowmeter compensation algorithm model through homomorphic encryption technology is as follows:

[0086] Each edge computing module uploads the gradient information of the flowmeter compensation algorithm model parameter to the edge server through homomorphic encryption technology;

[0087] Each flowmeter edge ECU calculates the gradient of the flowmeter compensation algorithm model parameter based on the loss function with global constraints , instead of the original measurement data, to avoid multi-dimensional data flow leakage;

[0088] The gradient information of the flowmeter compensation algorithm model is encrypted by using homomorphic encryption technology, and the encrypted gradient information is uploaded to the edge server, to ensure data privacy;

[0089] The specific process of generating a global optimization model is as follows:

[0090] The cloud aggregates the gradient information to generate a global optimization model of the flowmeter compensation algorithm model;

[0091] The cloud management platform aggregates the encrypted gradient information of all edge nodes, decrypts it to generate a global optimization model through federated averaging (FederatedAveraging), and then encrypts the updated parameters and sends them to each edge ECU, to realize global optimization of the flowmeter compensation algorithm model;

[0092] Based on the frequency of change of the flowmeter compensation algorithm model parameters and the response requirement of the flowmeter system, three-level optimization is performed, including short-term optimization, medium-term optimization and long-term optimization.

[0093] The short-term optimization is to adjust the filtering parameters based on real-time multi-dimensional data flow, and the response time is less than 1s.

[0094] The optimization object is the filtering algorithm parameters of the flowmeter compensation algorithm model. Real-time multi-dimensional data flow is input into the flowmeter compensation algorithm model to dynamically adjust the filtering strength, suppress sensor noise and ensure the real-time performance of single meter measurement.

[0095] The medium-term optimization is to update the polynomial coefficients in the flowmeter compensation algorithm model by the gradient descent method, and the optimization period is 10s-10min.

[0096] The optimization object is the polynomial coefficients of the algorithm of the flowmeter compensation algorithm model. Real-time multi-dimensional data flow is input into the flowmeter compensation algorithm model to iteratively update the polynomial coefficients by the gradient descent method.

[0097] The federal learning method is used to recalibrate the instrument coefficients.

[0098] The long-term optimization is to recalibrate the instrument coefficients by using the federal learning method, and the calibration period is more than 1h.

[0099] The long-term optimization object is the flowmeter core parameters. The global flow conservation constraint deviation is calculated. When the global flow conservation constraint deviation exceeds 1% for 30 minutes, the federal learning full iteration is triggered. The slow-varying parameters including instrument coefficients and polynomial fitting coefficients are recalibrated in combination with the instrument error-working condition parameter mapping table to offset the effects of long-term drift such as sensor aging and flow passage fouling.

[0100] Federal learning is used for multi-node collaborative optimization without sharing original data.

[0101] The compensation parameters of the optimized flowmeter compensation algorithm model are synchronized to the network association reasoning and dynamic weight distribution module as one of the input dimensions for flowmeter reliability evaluation.

[0102] Through homomorphic encryption combined with federal learning, the data privacy restriction is broken through to match the dynamic characteristics of the parameters with multiple time scales. The pain points of traditional single meter compensation such as local incompleteness and parameter update lag are solved to support high precision and real-time performance of industrial collaborative metering.

[0103] The network association reasoning and dynamic weight distribution module: constructs a fluid network association graph, performs abnormal association detection on the fluid network association graph, infers the interference of the associated pipeline through a graph neural network, automatically adjusts the compensation parameters of the flowmeter with the highest pressure fluctuation, generates a dynamic weight matrix, and gives higher weight to high-reliability instruments during total quantity weighing.

[0104] The specific process of constructing the fluid network correlation graph is:

[0105] Constructing the fluid network correlation graph, defining each flow meter as a node of the fluid network correlation graph , the node features integrating three-dimensional data, the three-dimensional data including: real-time measurement dimension data, error history dimension data, and working condition adaptation dimension data; wherein the real-time measurement dimension data includes: signal-to-noise ratio and data jump frequency, the error history dimension data includes: cumulative loss function of the edge computing module, and the working condition adaptation dimension data includes: instrument error-working condition parameter mapping table matching degree;

[0106] Pipes as edges of the fluid network correlation graph , i and j are node numbers, and the edge weight is based on the benchmark true value in the fluid network modeling and prior knowledge construction module, the benchmark true value is normalized, the fluid dynamics parameters between adjacent nodes are calculated as the edge weight, and the greater the edge weight, the stronger the dynamic interference generated by the two flow meters due to the fluid flow coupling in the pipe, and the more intense the fluid dynamics interaction;

[0107] The specific process of abnormal correlation detection on the fluid network correlation graph is:

[0108] Abnormal correlation detection, using a graph neural network to trace the abnormal correlation, and the abnormal triggering rule is:

[0109] When the calculation deviation of a certain flow meter of the topological node exceeds the calculation deviation threshold, an association reasoning signal is triggered; the calculation deviation threshold is set by the person skilled in the art according to experience;

[0110] A graph convolution network (GCN) is used to analyze the error propagation path, locate the neighbor node with the highest pressure fluctuation, learn the error propagation path, and start the corresponding optimization according to the interference frequency characteristics for the neighbor node with the highest pressure fluctuation;

[0111] If it is high-frequency interference, the interference frequency is greater than 1 Hz, then short-term optimization is triggered to adjust the filtering parameters;

[0112] If it is low-frequency interference, the interference frequency is less than 1 Hz, then medium-term optimization is triggered to update the polynomial fitting coefficients;

[0113] The specific process of giving higher weight to high-reliability instruments during total mass calculation is:

[0114] For each flow meter , reliability is quantified from three dimensions to generate a weight factor; the three dimensions include: real-time signal-to-noise ratio of the flow meter , loss function of the edge computing module of the flow meter , and working condition adaptation dimension data of the flow meter Instrument error - working condition parameter mapping table matching degree;

[0115] It should be noted that the instrument error - working condition parameter mapping table matching degree is to extract the real-time working condition parameters of the current flowmeter, such as the current temperature, pressure, and fluid composition, to the instrument error - working condition parameter mapping table, find the most similar working condition, compare the actual error of the current flowmeter with the pre-stored error of the similar working condition in the instrument error - working condition parameter mapping table, and calculate the instrument error - working condition parameter mapping table matching degree by using the cosine similarity calculation method.

[0116] By normalizing the fusion of the above three dimensions, a weight factor is generated, and a dynamic weight matrix is generated for global total weight calculation, ensuring that high-reliability instruments dominate in total calculation.

[0117] Using graph neural networks to quantify implicit correlations and accurately locate error propagation paths, dynamic weight adaptation for measurement reliability, solving the problem of low-precision tables dragging the whole caused by traditional equal-weight accumulation, supporting high-precision calculation under complex working conditions.

[0118] Working condition adaptation calibration and cross-domain parameter transfer module: generate virtual standard flow signals through fluid mechanics simulation model to realize online calibration without real objects, and use transfer learning technology to quickly transfer the optimal compensation parameters under similar working conditions.

[0119] Under non-full load working conditions, generate virtual standard flow signals through fluid mechanics simulation model to realize online calibration without real objects.

[0120] When the system is in a non-full load stable working condition, input the current working condition parameters, and generate virtual standard flow signals through fluid mechanics simulation model , and the virtual standard flow signals are used as the input reference value of the flowmeter compensation algorithm model to calculate the output after compensation, and verify the relative error: ;

[0121] If the relative error is greater than 1%, and the relative error is greater than 1% for 5 sampling periods, then trigger short-term optimization.

[0122] If the relative error is still greater than 1% after short-term optimization and lasts for 10 minutes, then start medium-term optimization.

[0123] If the relative error is greater than 2% after medium-term optimization, then mark it as a sensor hardware failure and trigger an alarm.

[0124] When the system working condition changes suddenly, use transfer learning technology to quickly transfer the optimal compensation parameters under similar working conditions.

[0125] Transfer learning is used for cross-working condition parameter transfer to avoid repeated training.

[0126] When the graph neural network identifies the mutation of the correlation feature, such as the fluctuation function of the pipeline pressure wave The morphology changes dramatically, the flow state is abnormal, and the system working condition changes suddenly.

[0127] Based on the instrument error-working condition parameter mapping table, the characteristic distance between the current working condition and the historical working condition is calculated by using the Euclidean distance, the historical scene with the minimum characteristic distance and the similarity greater than 0.7 is selected as the migration source, and the optimal compensation parameter set is extracted. The compensation parameters of similar working conditions are quickly loaded into the current algorithm, and after migration, the short-term optimization is fine-tuned to avoid the time cost of retraining.

[0128] Embodiment 2

[0129] As shown in Figure 2 Based on embodiment 1, the application provides a dynamic compensation and error correction method for high-precision flow instruments, comprising:

[0130] S1: A miniature multiphase flow sensor is installed on the upstream and downstream 5D pipe diameter of each flowmeter, a multi-dimensional data stream with a time stamp is established through a precise clock protocol, an isolated forest algorithm is used to detect data jump and sensor failure in real time, and a reliable multi-dimensional data sequence is generated through a redundant sensor voting mechanism;

[0131] S2: A directed graph model is constructed on the edge server, a real-time weighing equation is established, a fluid mechanics simulation model is used to calculate the flow distribution under different working conditions, a mapping table is generated, and used as prior knowledge of the initial compensation parameters;

[0132] S3: On the basis of the traditional single-table error function, a global weighing penalty term is added, the gradient information of the flowmeter compensation algorithm model is uploaded through homomorphic encryption technology, a global optimization model is generated, the filter parameters are adjusted based on real-time data, the polynomial fitting coefficients in the flowmeter compensation algorithm model are updated through the gradient descent method, and the instrument coefficients are recalibrated using the federal learning method;

[0133] S4: A fluid network correlation graph is constructed, abnormal correlation detection is performed on the fluid network correlation graph, the interference of the associated pipeline is inferred through a graph neural network, the compensation parameters of the flowmeter with the highest pressure fluctuation are automatically adjusted, a dynamic weight matrix is generated, and high-reliability instruments are given higher weights in total weighing;

[0134] S5: A virtual standard flow signal is generated through a fluid mechanics simulation model to realize online calibration without physical objects, and the optimal compensation parameters under similar working conditions are quickly migrated using transfer learning technology.

[0135] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A dynamic compensation and error correction system for a high-precision flow meter, characterized by: include: Multi-dimensional data acquisition and pre-processing module: A miniature multiphase flow sensor is installed 5D upstream and downstream of each flow meter. A precision clock protocol is used to establish a multi-dimensional data stream with a timestamp. The isolation forest algorithm is used to detect data jumps and sensor failures in real time. A redundant sensor voting mechanism is used to generate reliable multi-dimensional data sequences. Fluid network modeling and prior knowledge building module: Build a directed graph model on the edge server, establish real-time balance equations, use fluid mechanics simulation models to calculate flow distribution under different working conditions, and generate a mapping table as prior knowledge for initial compensation parameters; Distributed collaborative optimization and parameter calibration module: Based on the traditional single-meter error function, a global balance penalty term is added, and the gradient information of the flow meter compensation algorithm model is uploaded through homomorphic encryption technology to generate a global optimization model. The filter parameters are adjusted based on real-time data, and the polynomial fitting coefficients in the flow meter compensation algorithm model are updated through the gradient descent method. The instrument coefficients are recalibrated using the federated learning method.

2. The dynamic compensation and error correction system for a high-precision flow meter according to claim 1, characterized in that: The specific process of establishing a multi-dimensional data stream with a timestamp through the precision clock protocol is as follows: The sampling frequencies of the microwave phase sensor, laser particle size analyzer, and flowmeter are unified using the IEEE1588 precision clock protocol, ensuring a uniform sampling frequency greater than or equal to 100 Hz to accurately capture high-frequency changes in the dynamic flow field. The timestamp accuracy is controlled within 1 μs, ensuring the spatiotemporal consistency of the data collected by the microwave phase sensor, laser particle size analyzer, and flowmeter. Construct a multidimensional data stream with timestamps. The multidimensional data stream with timestamps is specifically as follows: ,in, is the flow measurement value of the i-th flow meter at time t, is the pressure value of the corresponding measurement point of the i-th pressure sensor at time t, is the temperature value of the corresponding measurement point of the i-th temperature sensor at time t, is the fluid component concentration of the i-th microwave phase sensor at time t.

3. The dynamic compensation and error correction system for a high-precision flow meter according to claim 1, characterized in that: The specific process of generating a reliable multidimensional data sequence through the redundant sensor voting mechanism is as follows: For redundant sensors with the same physical quantity, a two-out-of-three voting rule is used. The specific two-out-of-three voting rule is: In the flow meter usage scenario, for the three redundant sensors at the same location, three microwave phase sensors or three laser particle size analyzers are selected, and a two-out-of-three voting rule is used to screen out reliable multidimensional data sequences; If the measurement value deviations of two or more microwave phase sensors are less than the specified threshold, the average fluid component concentration is taken as the valid data, which is a reliable multidimensional data sequence.

4. The dynamic compensation and error correction system for a high-precision flow meter according to claim 1, characterized in that: The specific process of constructing the directed graph model is as follows: Based on the CAD drawings of the fluid system, a directed graph model is built on the edge server. , where V is a node set corresponding to the pipe intersection, such as the point where the fluid converges or diverges at the tee or container interface, and E is an edge set corresponding to the pipe section where the flow meter is installed. Each edge is marked with the pipe geometric parameters, including: pipe diameter D, length L, inner wall roughness .

5. The dynamic compensation and error correction system for a high-precision flow meter according to claim 1, characterized in that: The specific process of adding the global balance penalty term is as follows: Taking a single flow meter i as the optimization unit, based on multidimensional data flow and flow conservation constraints, a loss function with global constraints is constructed: ,in, is the mean square error between the measured value and the true value of the i-th flow meter, obtained through CFD simulation of the fluid mechanics simulation model, is the set of topological nodes to which the i-th flow meter belongs, which is defined by a directed graph model to clearly define the association between flow meters and nodes. is the flow imbalance deviation of node V at time t, is the weight coefficient, which is dynamically adjusted through reinforcement learning to prioritize the accuracy of a single table while constraining the global balance deviation.

6. A dynamic compensation and error correction system for a high-precision flow meter, characterized by: Also includes the following modules: Network association reasoning and dynamic weight allocation module: This module constructs a fluid network association graph, performs abnormal association detection on the fluid network association graph, uses graph neural network to infer interference in associated pipelines, automatically adjusts the compensation parameters of flow meters with the highest pressure fluctuations, and generates a dynamic weight matrix to assign higher weights to high-reliability instruments during total volume balancing. Working condition adaptation calibration and cross-domain parameter migration module: Generates a virtual standard flow signal through a fluid mechanics simulation model to achieve online calibration without physical objects, and uses transfer learning technology to quickly migrate the optimal compensation parameters under similar working conditions.

7. The dynamic compensation and error correction system for a high-precision flow meter according to claim 6, characterized in that: The specific process of constructing the fluid network association graph is as follows: A fluid network association graph is constructed, with each flow meter defined as a node in the graph. Node features are integrated with three-dimensional data, including real-time measurement data, error history data, and operating condition adaptation data. Real-time measurement data includes signal-to-noise ratio and data transition frequency, error history data includes the cumulative loss function of the edge computing module, and operating condition adaptation data includes the matching degree of the instrument error-operating condition parameter mapping table. The pipeline serves as the edge of the fluid network association graph. The edge weight is based on the reference truth value in the fluid network modeling and prior knowledge construction module. The reference truth value is normalized, and the fluid dynamics parameters between adjacent nodes are calculated as the edge weight. The larger the edge weight, the stronger the dynamic interference generated by the coupling of fluid flow in the pipeline between the two flow meters, reflecting the intensity of the fluid dynamics interaction.

8. The dynamic compensation and error correction system for a high-precision flow meter according to claim 6, characterized in that: The specific process of performing abnormal association detection on the fluid network association graph is as follows: When the balance deviation of a topological node belonging to a flow meter exceeds the balance deviation threshold, an associated inference signal is triggered. The balance deviation threshold is set by technical personnel in this industry based on experience. The graph convolutional network (GCN) is used to analyze the error propagation path, locate the neighboring node with the highest pressure fluctuation, learn the error propagation path, and initiate corresponding optimization for the neighboring node with the highest pressure fluctuation based on the interference frequency characteristics. If it is high-frequency interference, with a frequency greater than 1Hz, short-term optimization will be triggered to adjust the filter parameters; If it is low-frequency interference, with an interference frequency less than 1 Hz, mid-term optimization is triggered and the polynomial fitting coefficients are updated.

9. The dynamic compensation and error correction system for a high-precision flow meter according to claim 6, characterized in that: The specific process of achieving the non-physical online calibration is as follows: When the system is in a non-full load stable working condition, input the current working condition parameters and generate a virtual standard flow signal through the fluid mechanics simulation model. , the virtual standard flow signal As the input reference value of the flow meter compensation algorithm model, the output after calculation and compensation , verify the relative error: If the relative error is greater than 1% and remains greater than 1% for five sampling periods, short-term optimization is triggered. If the relative error remains greater than 1% after short-term optimization and persists for 10 minutes, mid-term optimization is initiated. If the relative error is greater than 2% after mid-term optimization, it is marked as a sensor hardware failure and an alarm is triggered.

10. The dynamic compensation and error correction system for a high-precision flow meter according to claim 6, characterized in that: The specific process of using transfer learning technology to quickly transfer the optimal compensation parameters under similar working conditions is as follows: When the pipeline pressure fluctuation transfer function changes dramatically, the flow state is determined to be abnormal. Based on the instrument error-operating parameter mapping table, the Euclidean distance is used to calculate the characteristic distance between the current operating condition and the historical operating condition. The historical scene with the smallest characteristic distance and a similarity greater than 0.7 is selected as the migration source, and the optimal compensation parameter set is extracted. The compensation parameters of similar operating conditions are quickly loaded into the current algorithm. After migration, short-term optimization and fine-tuning are performed to avoid the time cost of retraining.

Images