Mass flow calibration method and system

Abstract

The invention discloses a mass flow calibration method and system, and the method comprises the steps: installing and connecting a Coriolis mass flow meter to a calibration rack pipeline, and axially installing an intelligent sensor comprising an ultrasonic Doppler array at the upstream side; starting a liquid supply pump to pump calibration fluid, and controlling an ultrasonic Doppler array to perform section scanning on the fluid to obtain dynamic flow velocity profile time sequence data; inputting the dynamic flow velocity profile time series data into a pre-trained fluid relaxation state prediction machine learning model, and outputting a steady state arrival time point by the fluid relaxation state prediction machine learning model; and triggering a calibration sampling instruction at the steady-state arrival time point, obtaining a sampling mass flow value, and comparing the sampling mass flow value with a reference standard flow value to complete calibration. The problem that systematic errors exist in calibration sampling due to flow pattern unsteady-state change of high-viscosity fluid in the shear thinning recovery period is solved.

Description

Technical Field

[0001] This invention relates to the field of metering sensing and data processing, and more specifically, to a mass flow rate calibration method and system. Background Technology

[0002] The Coriolis mass flow meter is a high-precision metering instrument that can directly measure the mass flow rate of fluids. It is widely used in industrial production process control and bulk trade settlement. Its core measurement principle utilizes the Coriolis force generated when fluid flows through a vibrating measuring tube. This force forces the measuring tube to produce a slight torsion, and the true mass of the fluid can be calculated by detecting the phase difference of the vibration at both ends of the tube. To ensure absolute measurement accuracy, the instrument must be installed on a dedicated calibration bench during factory shipment or periodic verification, allowing fluid to flow continuously through the instrument, and its readings must be compared and calibrated against a high-precision standard. However, as the industrial demand for metering accuracy of high-viscosity fluids (such as crude oil, polymers, and special calibration oils) continues to increase, existing actual flow calibration techniques are gradually revealing multiple deep-seated physical defects.

[0003] The flow state of high-viscosity fluids in calibration pipelines is extremely complex. To overcome pipeline resistance, calibration systems typically require a high-power supply pump to force the fluid into the pipeline. After being subjected to intense physical shearing by the pump impeller, the microscopic macromolecular chains of the high-viscosity fluid are temporarily broken or aligned, causing a momentary drop in viscosity—a phenomenon known as shear thinning. When this fluid leaves the pump and enters the straight pipeline of the calibration bench, flowing through the flow meter being calibrated, the fluid's microstructure is in a non-steady-state transition period, undergoing re-entanglement and striving to recover its original viscosity. During this relaxation recovery period, the fluid viscosity recovery rate differs between the center of the pipe cross-section and near the pipe wall, resulting in extremely complex and unpredictable spatiotemporal dynamic distortions in the velocity distribution profiles of different liquid layers within the pipe. Because the Coriolis force is extremely sensitive to the mass distribution and velocity of the fluid within the pipe, if calibration sampling is performed under such an unstable, dynamically distorted, non-steady-state flow field, the calibrated instrument coefficients will carry severe transient biases. When users install such a biased flow meter back into stable operating conditions, significant systematic measurement errors will inevitably occur. Existing calibration methods usually rely on experience to set a fixed waiting time or blindly judge based on a single average flow rate, which cannot accurately capture the critical point at which the fluid's microstructure truly reaches physical steady state.

[0004] Furthermore, the mechanical clamping process of installing the flowmeter onto the calibration bench can itself introduce hidden errors. During actual installation, technicians typically secure the flowmeter to the test pipe using flanges and bolts. However, the tightening torque of the bolts, the force distribution among the bolts, and the flange contact state are often difficult to ensure are completely consistent across different installations. For Coriolis mass flowmeters, whose measuring tube operates under vibration, they are highly sensitive to changes in external boundary constraints. Therefore, such clamping differences are not merely simple installation deviations but alter the stress state, boundary stiffness, and structural damping of the measuring tube. This can be compared to a tuning fork; when the degree of fixation at the base of the tuning fork varies, its vibration frequency, amplitude attenuation characteristics, and vibration symmetry all change. Similarly, different degrees of clamping of the flowmeter will also alter the vibration characteristics of the measuring tube. Specifically, under no-flow conditions, the vibration response on both sides of the measuring tube should theoretically remain symmetrical, and the Coriolis phase difference caused by fluid mass flow should theoretically be zero. However, under conditions such as uneven force on both sides of the flange, the dynamic response at both ends of the measuring tube may no longer maintain ideal symmetry, resulting in a slight residual phase difference, i.e., a pseudo-phase difference, still appearing in the flowmeter under no-flow conditions. Traditional full-pipe static water zeroing methods usually only zero out the current output result and cannot identify whether the zero-point offset originates from the instrument's intrinsic state under true static conditions or from structural disturbances introduced by mechanical clamping.

[0005] In highly demanding continuous variable-condition calibration scenarios (such as continuously changing fluid delivery temperature or pressure), the two types of interference mentioned above can become deeply physically coupled. Small changes in temperature or pressure not only alter the microscopic viscosity recovery curve of the high-viscosity fluid itself, causing fluctuations in fluid damping, but also induce thermal expansion and contraction of external metal flanges and test pipelines, causing fluctuations in mechanical structural damping. Conventional single-dimensional monitoring systems cannot separate these two interference signals from their distinct physical sources, easily misinterpreting signal anomalies caused by dynamic drift of structural stress as characteristics indicating that the fluid has not yet returned to a steady state. This coupling confusion frequently leads existing calibration systems to misjudgments, making it impossible to accurately pinpoint the optimal sampling time in complex variable-condition environments, severely limiting the calibration accuracy and automated calibration efficiency of mass flow meters in demanding scenarios. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a mass flow rate calibration method and system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A mass flow rate calibration method includes the following steps: A Coriolis mass flow meter is installed and connected to the calibration bench pipeline, and a smart sensor containing an ultrasonic Doppler array is installed on the upstream side of the Coriolis mass flow meter along the axial direction of the calibration bench pipeline. Start the liquid supply pump to continuously pump the calibration fluid into the calibration bench pipeline, and control the ultrasonic Doppler array in the smart sensor to perform cross-sectional scanning of the calibration fluid to obtain dynamic flow velocity profile time series data. The dynamic velocity profile time series data is input into a pre-trained fluid relaxation state prediction machine learning model, and the fluid relaxation state prediction machine learning model outputs the steady state arrival time point. At the steady-state arrival time point, a calibration sampling command is triggered to obtain the sampled mass flow rate value from the Coriolis mass flow meter, and the sampled mass flow rate value is compared with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter.

[0008] Preferably, the fluid relaxation state prediction machine learning model includes a spatiotemporal feature extraction network and a state regression output layer. The spatiotemporal feature extraction network includes a three-dimensional convolutional neural network and a long short-term memory network. The step of inputting the dynamic velocity profile time-series data into the pre-trained fluid relaxation state prediction machine learning model and having the fluid relaxation state prediction machine learning model output the steady-state arrival time point includes: The dynamic velocity profile time series data is input into the three-dimensional convolutional neural network to perform convolution and pooling operations, and outputs a spatial feature sequence. The spatial feature sequence is input into the long short-term memory network to extract temporal dependencies and output a spatiotemporal fusion feature vector. The state regression output layer is controlled to perform a fully connected mapping operation on the spatiotemporal fusion feature vector and output the steady-state arrival time point.

[0009] Preferably, the training process of the pre-trained fluid relaxation state prediction machine learning model includes: Acquire multiple sets of historical dynamic flow velocity profile time-series data under different liquid supply pump output pressure conditions; By continuously monitoring the velocity fluctuation variance of the historical dynamic velocity profile time series data, the moment when the velocity fluctuation variance is less than the preset variance threshold for ten consecutive sampling periods is recorded as the actual steady-state time point. The historical dynamic velocity profile time series data is used as the sample input data, and the actual steady-state time points are used as the sample label data to construct a training dataset. The sample input data is input into the initial machine learning model to obtain the prediction time point, and the mean squared error value between the prediction time point and the sample label data is calculated. The mean squared error value is used as a loss function in the backpropagation algorithm to update the weight parameters of the initial machine learning model until the mean squared error value reaches a preset lower error threshold, thereby obtaining the pre-trained fluid relaxation state prediction machine learning model.

[0010] Preferably, the step of comparing the sampled mass flow rate value with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter includes: At the steady-state arrival time point, the standard mass flow rate value generated by the static weighing scale is read as the reference standard flow rate value; Divide the benchmark flow rate value by the sampled mass flow rate value, and obtain the instrument calibration coefficient from the division result. The instrument calibration coefficient is written into the control storage unit of the Coriolis mass flow meter.

[0011] Preferably, before starting the liquid supply pump, a static stress decoupling step is performed, the static stress decoupling step including: Keep the calibration fluid inside the Coriolis mass flow meter stationary; An external mechanical exciter is controlled to apply a single mechanical pulse to the calibration bench piping. The electromagnetic sensor inside the Coriolis mass flow meter is controlled to collect the free vibration amplitude sequence of the measuring tube, and the logarithmic decay rate of the envelope of the free vibration amplitude sequence is extracted as the free vibration decay rate. Multiply the free vibration attenuation rate by a preset mechanical damping mapping matrix to output the structural damping increment; The structural damping increment is mapped and converted into a mechanical stress pseudo-phase difference using a preset structural dynamics lookup table. The mechanical stress pseudo-phase difference is subtracted from the initial zero-point parameters of the Coriolis mass flow meter to output an updated zero-point reference.

[0012] Preferably, the step of mapping the structural damping increment to a mechanical stress pseudo-phase difference using a preset structural dynamics lookup table includes: The free vibration amplitude sequence is subjected to Fourier transform to obtain a frequency domain signal sequence, and the frequency corresponding to the maximum amplitude peak is extracted from the frequency domain signal sequence as the natural resonant frequency of the measuring tube. In the structural dynamics lookup table, find the phase shift angle value that corresponds to both the natural resonant frequency and the structural damping increment; The found phase shift angle value is set as the mechanical stress pseudo-phase difference.

[0013] Preferably, the step of obtaining the sampled mass flow rate value from the Coriolis mass flow meter includes: At the steady-state arrival time, the updated zero-point reference is input into the phase difference calculation module of the Coriolis mass flow meter; The phase difference calculation module is controlled to perform an internal subtraction correction operation using the updated zero-point reference, and outputs the sampled mass flow rate value.

[0014] Preferably, the fluid relaxation state prediction machine learning model includes a spatiotemporal feature extraction network, a cross-attention mechanism layer, and a state regression output layer; When the delivery temperature or delivery pressure of the calibration fluid changes continuously, the dynamic velocity profile time series data is input into the spatiotemporal feature extraction network to output a spatiotemporal fusion feature vector, and the free vibration attenuation rate is input as a structural reference feature vector into the cross-attention mechanism layer. The cross-attention mechanism layer is controlled to perform attention weight allocation operations between the structural reference feature vector and the spatiotemporal fusion feature vector to isolate the structural damping drift component caused by the continuous change of the conveying temperature or the conveying pressure. The state regression output layer is controlled to perform a fully connected mapping operation after removing the structural damping drift component, and outputs the steady-state arrival time point.

[0015] Preferably, the step of controlling the cross-attention mechanism layer to perform attention weight allocation operations between the structural reference feature vector and the spatiotemporal fusion feature vector includes: Multiply the structural reference feature vector by the first weight matrix to output the query matrix; The spatiotemporal fusion feature vector is multiplied by the second weight matrix and the third weight matrix respectively to output the key matrix and the value matrix; Calculate the dot product between the query matrix and the key matrix, and output the attention score matrix; Multiply the attention score matrix by the value matrix to output the adjusted spatiotemporal fusion feature vector; The state regression output layer performs the fully connected mapping operation using the adjusted spatiotemporal fusion feature vector.

[0016] The present invention also discloses a mass flow rate calibration system for implementing the above method, comprising: The mounting module is used to mount and connect the Coriolis mass flow meter to the calibration bench pipeline, and to mount a smart sensor containing an ultrasonic Doppler array on the upstream side of the Coriolis mass flow meter along the axial direction of the calibration bench pipeline. The data acquisition module is used to start the liquid supply pump to continuously pump the calibration fluid into the calibration bench pipeline, and control the ultrasonic Doppler array in the smart sensor to perform cross-sectional scanning of the calibration fluid to obtain dynamic flow velocity profile time series data. The model processing module is used to input the dynamic velocity profile time series data into a pre-trained fluid relaxation state prediction machine learning model, and the fluid relaxation state prediction machine learning model outputs the steady state arrival time point. The execution module is used to trigger a calibration sampling command at the steady-state arrival time point, obtain the sampled mass flow rate value from the Coriolis mass flow meter, and compare the sampled mass flow rate value with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter.

[0017] The advantage of this invention over existing technologies lies in its use of a smart sensor containing an ultrasonic Doppler array installed upstream of a Coriolis mass flow meter. This sensor scans the calibration fluid cross-section to obtain high-dimensional dynamic velocity profile time-series data. This data is then input into a pre-trained machine learning model predicting fluid relaxation states. The model outputs the steady-state arrival time point to precisely trigger calibration sampling. This core technology cleverly utilizes the high-resolution imaging capability of the ultrasonic Doppler array to perceive the nonlinear evolution of the flow field inside the pipe. Compared to the blindness of traditional calibration methods that rely solely on a single average flow velocity or a fixed waiting time, this invention leverages the deep spatiotemporal feature extraction and prediction advantages of machine learning models for complex fluid relaxation evolution. This allows for the early and accurate identification of the critical point where, after strong shear thinning, the microscopic macromolecular chains of a high-viscosity fluid completely recover their entanglement, and the velocity profile truly reaches absolute physical steady state. This completely solves the problem of inaccurate calibration sampling caused by the dynamic distortion of the internal flow pattern during the non-steady-state transition of high-viscosity fluids, fundamentally eliminates the systematic calibration deviation introduced by incorrect sampling timing, and significantly improves the absolute accuracy of high-precision instrument actual flow calibration.

[0018] Building upon this foundation, to completely eliminate the hidden interference of mechanical clamping stress on the calibration zero-point reference, this invention applies a single mechanical pulse to the test pipeline via an external mechanical exciter while the fluid is stationary, and controls the electromagnetic sensor inside the instrument to collect the free vibration amplitude sequence of the measuring tube. The system further extracts the logarithmic decay rate of the envelope of this amplitude sequence as the free vibration decay rate, multiplies it by a preset mechanical damping mapping matrix, and then precisely maps it to the mechanical stress pseudo-phase difference via a structural dynamics lookup table. Finally, this error is directly subtracted from the instrument's initial zero-point parameters to output an updated zero-point reference. This technical feature profoundly reveals and utilizes the physical law in structural dynamics that structural damping is directly constrained by assembly preload. By quantitatively analyzing the decay rate of external broadband vibration signals, it successfully achieves the pure physical removal of uncontrollable mechanical stress errors introduced by the rigid clamping of the flange, perfectly solving the industry problem of false zero-point offset caused by external pipeline stress, and ensuring the absolute purity of the calibration starting point.

[0019] For calibration scenarios with extremely high requirements under continuous variable operating conditions, this invention further introduces a cross-attention mechanism layer to overcome complex physical coupling interference. When the system detects continuous changes in the delivery temperature or pressure of the calibration fluid, it not only causes fluid damping fluctuations due to microscopic viscosity changes in the high-viscosity fluid itself, but also inevitably leads to dynamic drift of mechanical structure damping caused by thermal expansion and contraction of the external metal flange. This invention uses the extracted free vibration attenuation rate as a structural reference feature vector, and inputs it into the model together with the spatiotemporal fusion feature vector extracted from the dynamic flow velocity profile, and performs attention weight allocation calculations between the two. This deep fusion of multimodal physical features enables the model to accurately isolate and eliminate structural damping drift components caused by slight temperature or pressure changes. This effectively solves the problem of mutual confusion between the stress distortion signal of the pipe structure and the unstable signal of fluid shear thinning under variable operating conditions, which can easily lead to the system falling into a sampling rejection dead loop. It ensures the absolute purity and reliability of steady-state triggering commands under complex and variable environments, and significantly improves the anti-interference capability and operating efficiency of the automated calibration system. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the method of the present invention; Figure 3 This is a schematic diagram of obtaining the calibration coefficients of the present invention. Figure 4 This is a schematic diagram illustrating the application of mechanical vibration in this invention; Figure 5 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] This invention addresses high-precision mass flow rate calibration scenarios, particularly suitable for calibration fluids with high viscosity, shear-thinning characteristics, or significant flow field relaxation processes after startup. Traditional methods typically employ a fixed waiting time or observe only a single point's average flow velocity and whether the outlet reading stabilizes before sampling. While this approach is still usable for low-viscosity, short-relaxation fluids, in high-viscosity fluids, although continuous flow has begun, the velocity distribution within the cross-section, the boundary layer recovery state, and the molecular chain entanglement recovery state may still be in a transitional phase. If sampling is performed too early, the Coriolis mass flow meter reading may appear stable, but the corresponding actual flow state has not yet truly stabilized, thus introducing unsteady-state errors into the calibration results.

[0023] like Figure 1 and Figure 2 As shown, the core approach of this invention is to install a smart sensor containing an ultrasonic Doppler array upstream of the Coriolis mass flow meter to directly observe the dynamic velocity profile time-series data of the fluid cross-section. Then, a fluid relaxation state prediction machine learning model provides the steady-state arrival time point, which is used to trigger calibration sampling. In this way, calibration sampling no longer relies on empirical waiting time but is based on the identification of the actual flow field evolution process.

[0024] In one embodiment, the entire calibration system includes a calibration bench piping, a supply pump, a storage tank, a Coriolis mass flow meter, a smart sensor, a static weighing scale, a controller, and a data processing unit. The calibration bench piping is preferably a rigid metal pipe, with an inner diameter selected from 10mm to 100mm depending on the diameter of the instrument being calibrated. The Coriolis mass flow meter is installed in the middle of the calibration bench piping and serves as the object being calibrated. The smart sensor is installed upstream of the Coriolis mass flow meter and arranged axially along the calibration bench piping, ensuring its scanning cross-section maintains a fixed geometric relationship with the pipe axis. To reflect the actual inflow state before it enters the Coriolis mass flow meter while avoiding installation interference due to excessive proximity to the instrument inlet, the axial distance between the smart sensor and the Coriolis mass flow meter inlet can be set to 3 to 10 times the inner diameter of the calibration bench piping, and further, to 5 to 8 times.

[0025] The smart sensor in this embodiment includes a sensor housing, a signal conditioning circuit, a control board, and an ultrasonic Doppler array. The ultrasonic Doppler array can consist of 8 to 64 transducer units, each distributed circumferentially along the pipe, and collects fluid scattering signals from different depths and orientations within the pipe through a preset transmission and reception sequence. Since the Doppler signal is sensitive to velocity changes in the scattering particles within the fluid, the controller can recover the local axial velocity at each measurement location based on the echo frequency shift, and then, combined with the array's geometric position and beam pointing relationship, reconstruct the velocity profile at each sampling moment. What is obtained here is not a single average velocity, but a velocity distribution covering multiple spatial locations across the entire cross-section. To facilitate subsequent machine learning model processing, each frame of the velocity profile can be discretized into a velocity matrix composed of multiple grid points. The sampling frequency is preferably 50Hz to 1000Hz. For high-viscosity fluids, a sampling frequency of 100Hz to 500Hz typically balances temporal resolution and computational burden. The controller continuously stores these velocity matrices in chronological order, thus obtaining the dynamic velocity profile time-series data.

[0026] At the start of calibration, first connect the Coriolis mass flow meter to the calibration bench piping. A coaxial connection is preferred during installation, and the parallelism error of the flange connection surface should be controlled to be no greater than 0.05 mm to avoid initial assembly deviations causing additional errors in subsequent zero-point and structural damping identification. After installation, if the static stress decoupling scheme is not enabled, the supply pump can be started directly to continuously pump the calibration fluid into the calibration bench piping. The supply pump can be a variable frequency gear pump, screw pump, or diaphragm pump, and the controller outputs the corresponding drive signal according to the set flow rate. After the supply pump starts, the ultrasonic Doppler array in the intelligent sensor begins to scan the cross-section of the calibration fluid and outputs dynamic velocity profile time-series data in real time.

[0027] To enable the model to visualize both the spatial distribution of the flow field and its recovery trend over time, this embodiment employs a fluid relaxation state prediction machine learning model to process the time-series data of the dynamic velocity profile. This model includes a spatiotemporal feature extraction network and a state regression output layer. In a further embodiment, the spatiotemporal feature extraction network comprises a three-dimensional convolutional neural network and a long short-term memory network. The three-dimensional convolutional neural network is responsible for extracting local spatial structure and short-term change patterns from data blocks composed of multiple consecutive time frames, while the long short-term memory network is responsible for modeling the temporal continuity of these spatial features. This design is based on the fact that the fluid reaching steady state is not determined solely by the velocity distribution of a single frame, but rather by the gradual stabilization of the velocity profile over a period of time; therefore, it is necessary to utilize both spatial and temporal information simultaneously.

[0028] In practical implementation, velocity profiles from multiple consecutive sampling times can be combined into an input sample block. This sample block is input into a 3D convolutional neural network, undergoing multiple convolution and pooling operations to extract local flow field morphology features and compress redundant information. The pooling result forms a spatial feature sequence, which is then input into a long short-term memory network to extract the time dependencies in the flow field recovery process. In this way, the model can learn the changing patterns related to the fluid relaxation process, such as the recovery of the central flow kernel, the stabilization of the velocity gradient near the wall, and the weakening of eccentric flow at the cross-section. After the long short-term memory network outputs a spatiotemporal fusion feature vector, the state regression output layer performs a fully connected mapping operation on this feature vector, outputting the steady-state arrival time point. This steady-state arrival time point can be set as the time value calculated from the start of the self-supplying pump, or it can be set as the remaining time value that still needs to be waited for relative to the current sampling window. In this embodiment, the time value calculated from the start of the self-supplying pump is preferred, as it facilitates direct comparison between the controller and the current clock to trigger sampling.

[0029] To ensure the usability of the above model, it needs to be pre-trained. During training, multiple sets of historical dynamic velocity profile time-series data under different supply pump output pressure conditions are acquired. The reason for covering different supply pump output pressure conditions is that the starting impact intensity, the initial flow ramp-up rate, and the flow field recovery path are all related to the supply pump output pressure. If the training samples only cover a single starting condition, the model may only learn the flow field evolution under specific operating conditions during actual calibration, resulting in insufficient generalization ability. In a further embodiment, the supply pump output pressure can cover the range of 0.05 MPa to 1.5 MPa, and samples are collected at no less than 10 discrete pressure levels. At each pressure level, no less than 30 sets of historical dynamic velocity profile time-series data are repeatedly collected to enhance sample diversity.

[0030] Sample labels can be constructed as follows. The controller continuously monitors the time-series data of each set of historical dynamic velocity profiles. Within each sampling period, it first calculates the deviation of all grid velocity values ​​at the current cross-section relative to the average velocity of the cross-section in that period, and then calculates the velocity fluctuation variance based on these deviations. To avoid inconsistencies in thresholds due to excessive dimensional differences between different flow points, in one embodiment, the velocity values ​​of each grid are first divided by the average velocity of the cross-section at that moment to obtain the normalized velocity, and then the velocity fluctuation variance is calculated based on the normalized velocity. At this time, the preset variance threshold can be set to 0.0005 to 0.01, and further set to 0.001 to 0.005. When the velocity fluctuation variance is less than the variance threshold for 10 consecutive sampling periods, the first period of these 10 consecutive sampling periods can be recorded as the actual steady-state time point. The reason for using 10 consecutive sampling periods is to prevent the flow field from occasionally exhibiting a stable illusion in a local short period of time. If the sampling frequency is 200Hz, then 10 consecutive sampling periods correspond to 0.05s. This duration is usually sufficient to filter out most random fluctuations, while not delaying the actual steady-state time point too much.

[0031] After label construction, historical dynamic velocity profile time-series data are used as input data, and actual steady-state time points are used as label data to construct a training dataset. The input data is then fed into the initial machine learning model to obtain predicted time points, and the mean squared error (MSE) between the predicted time points and the label data is calculated. During training, a backpropagation algorithm is used, with the MSE value as the loss function to update the weight parameters of the initial machine learning model. The optimizer can be the Adam optimizer, with an initial learning rate of 0.0001 to 0.001, a batch size of 16 to 64, and 50 to 300 training epochs. A preset lower error threshold can be set according to the calibration accuracy requirements. For high-precision real-flow calibration, if time error is used as the evaluation metric, the preset lower error threshold can be set to 5ms to 50ms, and further to 10ms to 20ms. When the MSE value decreases and stabilizes near this lower error threshold, the pre-trained fluid relaxation state prediction machine learning model is obtained. After training is complete, the model parameters are stored in the model processing module and can be directly called during the actual calibration process.

[0032] In actual operation, the controller continuously inputs the dynamic flow velocity profile time-series data collected by the smart sensors into the fluid relaxation state prediction machine learning model. When the model outputs the steady-state arrival time point, the execution module triggers the calibration sampling command at that time point. The reason for using triggered sampling here, rather than retrospectively extracting from long-term records, is that the calibration bench typically needs to be synchronized with the static weighing scale, valve switching timing, and the internal sampling clock of the Coriolis mass flow meter. Triggering directly at the steady-state arrival time point ensures that the reference standard flow rate value and the sampled mass flow rate value originate from the same physical time period.

[0033] In one embodiment, for obtaining the reference standard flow rate value, a collection container is disposed downstream of the calibration bench and placed on a static weighing scale. The static weighing scale continuously outputs a cumulative mass signal, and the controller calculates the standard mass flow rate value based on the change of this cumulative mass signal within a preset time window. This preset time window is preferably consistent with the internal output period of the Coriolis mass flow meter, or an integer multiple thereof, such as 20ms, 50ms, or 100ms. Thus, at the steady-state arrival time, the controller can read the standard mass flow rate value generated by the static weighing scale as the reference standard flow rate value.

[0034] like Figure 3 As shown, the sampled mass flow rate value is read from the Coriolis mass flow meter, and then the reference standard flow rate value is divided by the sampled mass flow rate value to obtain the instrument calibration coefficient. This calibration coefficient is then written into the control storage unit of the Coriolis mass flow meter. Subsequently, during normal measurements, the internal phase difference conversion result of the instrument can be corrected according to this calibration coefficient. To improve statistical robustness, multiple samplings can be triggered at the same flow point to obtain multiple instrument calibration coefficients, and then the average or median value is written into the control storage unit.

[0035] like Figure 4 As shown, in some high-precision scenarios, simply addressing the unsteady state of the flow field is insufficient, because Coriolis mass flow meters are also affected by external assembly stresses. Especially when the calibration bench uses flange connections, has long rigid pipes, or has high clamping preload, the measuring tube will bear additional loads. These loads may not necessarily change the actual mass flow rate, but they will alter the vibration damping and phase response of the measuring tube, thus creating a false phase difference at zero flow. To address this issue, this invention performs a static stress decoupling step before starting the supply pump.

[0036] During static stress decoupling, the calibration fluid inside the Coriolis mass flow meter is first kept in a static state. This static state means the supply pump is not operating, and there is no macroscopic flow within the measuring tube by closing relevant valves or using other isolation methods. At this time, the Coriolis mass flow meter does not exhibit a true flow-induced phase difference, retaining only the vibration response of the structure itself. Then, an external mechanical exciter is used to apply a single mechanical pulse to the calibration bench piping. The external mechanical exciter can be an electromagnetic hammer, a piezoelectric impactor, or a small pulse vibrator. To elicit a clear free vibration response without damaging the installation structure, the peak force of the single mechanical pulse can be set from 5N to 200N, and the pulse width can be set from 1ms to 20ms. After the pulse is applied, the electromagnetic sensor inside the Coriolis mass flow meter begins to acquire the free vibration amplitude sequence of the measuring tube.

[0037] The extraction of the free vibration amplitude sequence can be achieved using the existing drive and detection circuit of the Coriolis mass flow meter. Since the measuring tube generates damped vibrations after the pulse, the electromagnetic sensor outputs a series of amplitude data that gradually decays over time. The controller calculates the envelope of this free vibration amplitude sequence and then calculates the logarithmic decay rate of the envelope based on the decay relationship between adjacent peaks, thus obtaining the free vibration decay rate. The reason for using the logarithmic decay rate of the envelope is that this parameter is more sensitive to damping changes caused by mechanical preload, but not sensitive to instantaneous impact phases and single noise spikes; therefore, it is more suitable for characterizing the influence of static assembly stress on structural dynamics.

[0038] After obtaining the free vibration attenuation rate, it is multiplied by a preset mechanical damping mapping matrix to output the structural damping increment. In one embodiment, the mechanical damping mapping matrix can be a calibration matrix obtained through offline calibration experiments. This calibration matrix is ​​derived from measured data of the same model of instrument under different known assembly stress conditions, and is used to reflect the correspondence between the change in free vibration attenuation rate and the change in structural damping. The reason for using a mapping matrix instead of directly using a simple proportional coefficient is to leave room for subsequent expansion. In a simpler implementation, the mapping matrix can be degenerated into a single calibration coefficient; in a more complex implementation, the free vibration attenuation rate can be combined with other auxiliary structural features to form an eigenvector, and then the structural damping increment can be obtained through a high-dimensional mapping matrix.

[0039] After obtaining the structural damping increment, the controller maps the structural damping increment into a mechanical stress pseudo-phase difference using a preset structural dynamics lookup table. This lookup table can be established using two types of data. One type of data comes from finite element structural dynamics simulations, used to obtain the phase response variation trend of the measuring tube under different damping and preload conditions. The other type of data comes from physical calibration experiments, used to correct the differences between simulation results and actual instruments. To improve mapping accuracy, this embodiment also uses the natural resonant frequency corresponding to the free vibration amplitude sequence as an additional search condition. Specifically, the free vibration amplitude sequence is first subjected to a Fourier transform to obtain a frequency domain signal sequence; then, the frequency corresponding to the peak value of the maximum amplitude is extracted from the frequency domain signal sequence and used as the natural resonant frequency of the measuring tube. Subsequently, the phase shift angle value corresponding to both the natural resonant frequency and the structural damping increment is found in the structural dynamics lookup table, and this phase shift angle value is set as the mechanical stress pseudo-phase difference.

[0040] In a further embodiment, the structural dynamics lookup table can be in two-dimensional form, with one dimension corresponding to the natural resonant frequency and the other dimension corresponding to the structural damping increment. The unit values ​​in the table are phase shift angle values. The discrete step size of the natural resonant frequency can be set to 0.1Hz to 2Hz, and the discrete step size of the structural damping increment can be set to 0.0001 to 0.01. During actual table lookup, if the current operating condition value falls between two discrete nodes, interpolation can be used to calculate the phase shift angle value to avoid introducing significant step errors due to the limited resolution of the lookup table.

[0041] In one embodiment, the structural dynamics lookup table is pre-established using a historical calibration method. Specifically, multiple prototypes with the same structural form, installation method, and material parameters as the Coriolis mass flowmeters used in actual calibration are selected and statically installed under different known mechanical assembly stress conditions, while keeping the fluid inside the flowmeter in a flow-free state. For the Coriolis mass flowmeter, in the flow-free state, the vibration response on both sides of the measuring tube should theoretically remain symmetrical, and the theoretical value of the Coriolis phase difference caused by fluid mass flow should be zero. However, under actual installation conditions, mechanical assembly stress, external pipeline traction force, changes in structural damping, and slight changes in the relative position of the sensor may cause the vibration symmetry of the measuring tube to be disrupted, thus detecting a residual phase difference even in the flow-free state. This residual phase difference does not correspond to the true mass flow rate but is a pseudo-phase difference caused by mechanical stress coupling. Under each known mechanical assembly stress condition, a single mechanical pulse is applied using an external mechanical exciter, and the corresponding free vibration amplitude sequence is collected to calculate the free vibration attenuation rate and natural resonant frequency. Simultaneously, the flow-free residual phase difference output by the flowmeter in this state is read and recorded as a phase shift angle value. Subsequently, using the inherent resonant frequency and structural damping increment as input dimensions and the corresponding phase shift angle value as output, a structural dynamics lookup table is established. This table reflects the correspondence between the flow-free pseudo-phase difference and structural parameters caused by changes in structural dynamic characteristics under different mechanical stress states for a specific instrument structure. This provides a practical calibration basis for the mechanical stress pseudo-phase difference obtained from subsequent table lookups.

[0042] After obtaining the pseudo-phase difference of mechanical stress, this pseudo-phase difference is subtracted from the initial zero-point parameters of the Coriolis mass flow meter to output the updated zero-point reference. Physically, this removes the spurious zero-point offset caused by mechanical assembly stress from the initial zero-point parameters, making the zero-point reference used in subsequent flow conversions closer to the intrinsic zero point under true no-flow conditions.

[0043] In embodiments employing the static stress decoupling scheme, upon reaching the steady-state arrival time, the original sampled mass flow rate value is not directly read. Instead, the updated zero-point reference is first input into the phase difference calculation module of the Coriolis mass flow meter. The phase difference calculation module performs an internal subtraction correction operation on the measured instantaneous phase difference signal, that is, it first subtracts the updated zero-point reference and then calculates the sampled mass flow rate value based on the instrument's internal calibration constant. The purpose of this design is to eliminate the influence of static assembly stress before it enters the mass flow rate calculation, rather than passively absorbing it during the final calibration coefficient stage. This is because although the latter may temporarily correct the error at a certain flow rate point, the error may reappear once the installation stress changes or the structural response characteristics change due to the flow rate point switching.

[0044] For calibration scenarios involving continuous changes in transport temperature or pressure, using only the aforementioned single-modal fluid relaxation state prediction machine learning model may misidentify dynamic drift of structural damping as an unstable fluid, or misidentify the actual relaxation process caused by changes in fluid viscosity as structural changes. Therefore, in a further embodiment, this invention introduces a fluid relaxation state prediction machine learning model that includes a spatiotemporal feature extraction network, a cross-attention mechanism layer, and a state regression output layer.

[0045] Here, we first explain the method for determining continuous temperature changes. The controller can receive real-time outputs from the conveying temperature sensor and the conveying pressure sensor. When the conveying temperature changes in the same direction for 5 to 20 consecutive sampling periods, and the cumulative change reaches 1°C to 5°C, it can be determined that the conveying temperature has changed continuously. When the conveying pressure changes in the same direction for 5 to 20 consecutive sampling periods, and the cumulative change reaches 0.02MPa to 0.2MPa, it can be determined that the conveying pressure has changed continuously. As long as either of these conditions is met, the system enters the continuous variable operating condition recognition mode.

[0046] In this mode, the time-series data of the dynamic velocity profile is still input into the spatiotemporal feature extraction network, outputting a spatiotemporal fusion feature vector. Simultaneously, the free vibration attenuation rate is input as a structural reference feature vector into the cross-attention mechanism layer. To ensure the representativeness of this structural reference feature vector under continuous varying operating conditions, in one implementation, the system can re-execute a small-amplitude structural excitation at preset time intervals, acquiring a new free vibration amplitude sequence and updating the free vibration attenuation rate; in another implementation, the equivalent free vibration attenuation rate can be extracted using the attenuation segment of the instrument's original drive signal. Thus, the input to the cross-attention mechanism layer is not a one-time parameter acquired only before startup, but a reference feature reflecting the current structural state changes. The free vibration attenuation rate is used as the structural reference feature vector because it directly reflects the changing trend of the measuring tube's structural damping. Continuous changes in transport temperature or pressure often alter fluid viscosity and shear response, and also change structural damping through flange thermal expansion and contraction, shell thermal stress, or pressure loads. Without separately introducing structural features into the model, the model only sees the mixed overall response, making it difficult to distinguish which part comes from fluid relaxation and which part from structural drift.

[0047] The implementation process of the cross-attention mechanism layer can be described as follows: First, the structural reference feature vector is mapped to query features. Then, the spatiotemporal fusion feature vector is mapped to key features and value features, respectively. Attention weights are then calculated based on the matching degree between the query features and key features. These attention weights are then used to reweight the value features, resulting in an adjusted spatiotemporal fusion feature vector. Since the cross-attention mechanism layer updates its parameters during training with the goal of minimizing the regression error at the steady-state time point, its output is no longer a simple rearrangement of the original spatiotemporal features, but rather an adjusted result that has weakened the influence of the strongly coupled component of structural damping drift. In other words, this layer effectively isolates and suppresses the structural damping drift component. The state regression output layer then uses the adjusted spatiotemporal fusion feature vector to perform a fully connected mapping operation, outputting the steady-state arrival time point.

[0048] To ensure the trainability of the aforementioned cross-attention model, in a further embodiment, a continuously variable operating condition sample set can be constructed in addition to the basic training dataset. This sample set covers dynamic velocity profile time-series data under different rates of change in transport temperature and transport pressure, and simultaneously records the free vibration decay rate. The rate of change in transport temperature can cover 0.5℃ / min to 10℃ / min, and the rate of change in transport pressure can cover 0.01MPa / min to 0.5MPa / min. The sample labels still use the actual steady-state time points obtained when the aforementioned velocity fluctuation variance continuously meets the condition. During training, the regression error of the steady-state time points is still the main loss, and regularization constraints can be added to avoid excessive divergence of the cross-attention parameters. The model trained in this way can more stably provide the steady-state arrival time point in scenarios where the transport temperature or transport pressure is continuously changing, avoiding the system from refusing to sample for a long time because it mistakenly regards structural drift as fluid unsteadiness.

[0049] like Figure 5As shown, the mass flow calibration system corresponding to this invention can be implemented using a hardware and software collaborative approach. The installation module mainly consists of a mounting bracket, positioning clamps, flange connection components, and a smart sensor mounting base. It is used to install the Coriolis mass flow meter to the calibration bench pipeline and to install a smart sensor containing an ultrasonic Doppler array along the axial direction on the upstream side. The data acquisition module includes a liquid supply pump drive unit, a smart sensor control unit, a signal acquisition board, and a clock synchronization unit. It is used to start the liquid supply pump, control the ultrasonic Doppler array to perform cross-sectional scanning, and acquire dynamic velocity profile time-series data. The model processing module includes an industrial computer, a graphics processing unit or an edge computing chip, and a memory for storing parameters of the fluid relaxation state prediction machine learning model. This module is responsible for inputting the dynamic velocity profile time-series data into the pre-trained fluid relaxation state prediction machine learning model and outputting the steady-state arrival time point. The execution module includes timed triggering logic, a Coriolis mass flow meter communication interface, a static weighing scale interface, and coefficient writing control logic. It is used to trigger a calibration sampling command at the steady-state arrival time point, obtain the sampled mass flow rate value from the Coriolis mass flow meter, and compare the sampled mass flow rate value with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter.

[0050] In a further embodiment, the execution module is also connected to a static stress decoupling unit. The static stress decoupling unit includes an external mechanical exciter control circuit, a free vibration amplitude acquisition circuit, a structural dynamics lookup table storage area, and zero-point reference update logic. Thus, before the liquid supply pump is officially started, the system first completes the calculation of updating the zero-point reference; at the steady-state arrival time, the updated zero-point reference is input into the phase difference calculation module to obtain the sampled mass flow rate value, which has already deducted the influence of the pseudo-phase difference caused by mechanical stress. For continuous variable operating condition calibration scenarios, the model processing module internally enables a model branch containing a cross-attention mechanism layer and simultaneously receives information on free vibration attenuation rate, delivery temperature, and delivery pressure to improve the anti-interference capability of steady-state determination.

[0051] In summary, this invention does not simply extend the waiting time, nor does it merely improve the accuracy of the instrument itself. Instead, it integrates upstream real flow field observation, machine learning steady-state time prediction, structural stress zero-point decoupling, and multimodal characteristics under continuous variable operating conditions into a complete closed loop. In this way, calibration sampling occurs both when the fluid truly reaches a steady state and on the basis that structural zero-point errors have been eliminated, thus making the calibration results of the Coriolis mass flow meter closer to the actual mass flow rate.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A mass flow rate calibration method, characterized in that, Includes the following steps: A Coriolis mass flow meter is installed and connected to the calibration bench pipeline, and a smart sensor containing an ultrasonic Doppler array is installed on the upstream side of the Coriolis mass flow meter along the axial direction of the calibration bench pipeline. Start the liquid supply pump to continuously pump the calibration fluid into the calibration bench pipeline, and control the ultrasonic Doppler array in the smart sensor to perform cross-sectional scanning of the calibration fluid to obtain dynamic flow velocity profile time series data. The dynamic velocity profile time series data is input into a pre-trained fluid relaxation state prediction machine learning model, and the fluid relaxation state prediction machine learning model outputs the steady state arrival time point. At the steady-state arrival time point, a calibration sampling command is triggered to obtain the sampled mass flow rate value from the Coriolis mass flow meter, and the sampled mass flow rate value is compared with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter.

2. The mass flow rate calibration method as described in claim 1, characterized in that, The fluid relaxation state prediction machine learning model includes a spatiotemporal feature extraction network and a state regression output layer. The spatiotemporal feature extraction network comprises a three-dimensional convolutional neural network and a long short-term memory network. The step of inputting the dynamic velocity profile time-series data into the pre-trained fluid relaxation state prediction machine learning model and having the model output the steady-state arrival time point includes: The dynamic velocity profile time series data is input into the three-dimensional convolutional neural network to perform convolution and pooling operations, and outputs a spatial feature sequence. The spatial feature sequence is input into the long short-term memory network to extract temporal dependencies and output a spatiotemporal fusion feature vector. The state regression output layer is controlled to perform a fully connected mapping operation on the spatiotemporal fusion feature vector and output the steady-state arrival time point.

3. The mass flow rate calibration method as described in claim 1, characterized in that, The training process of the pre-trained machine learning model for predicting fluid relaxation states includes: Acquire multiple sets of historical dynamic flow velocity profile time-series data under different liquid supply pump output pressure conditions; By continuously monitoring the velocity fluctuation variance of the historical dynamic velocity profile time series data, the moment when the velocity fluctuation variance is less than the preset variance threshold for ten consecutive sampling periods is recorded as the actual steady-state time point. The historical dynamic velocity profile time series data is used as the sample input data, and the actual steady-state time points are used as the sample label data to construct a training dataset. The sample input data is input into the initial machine learning model to obtain the prediction time point, and the mean squared error value between the prediction time point and the sample label data is calculated. The mean squared error value is used as a loss function in the backpropagation algorithm to update the weight parameters of the initial machine learning model until the mean squared error value reaches a preset lower error threshold, thereby obtaining the pre-trained fluid relaxation state prediction machine learning model.

4. The mass flow rate calibration method as described in claim 1, characterized in that, The step of comparing the sampled mass flow rate value with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter includes: At the steady-state arrival time point, the standard mass flow rate value generated by the static weighing scale is read as the reference standard flow rate value; Divide the benchmark flow rate value by the sampled mass flow rate value, and obtain the instrument calibration coefficient from the division result. The instrument calibration coefficient is written into the control storage unit of the Coriolis mass flow meter.

5. The mass flow rate calibration method as described in claim 1, characterized in that, Before starting the liquid supply pump, a static stress decoupling step is performed, which includes: Keep the calibration fluid inside the Coriolis mass flow meter stationary; An external mechanical exciter is controlled to apply a single mechanical pulse to the calibration bench piping. The electromagnetic sensor inside the Coriolis mass flow meter is controlled to collect the free vibration amplitude sequence of the measuring tube, and the logarithmic decay rate of the envelope of the free vibration amplitude sequence is extracted as the free vibration decay rate. Multiply the free vibration attenuation rate by a preset mechanical damping mapping matrix to output the structural damping increment; The structural damping increment is mapped and converted into a mechanical stress pseudo-phase difference using a preset structural dynamics lookup table. The mechanical stress pseudo-phase difference is subtracted from the initial zero-point parameters of the Coriolis mass flow meter to output an updated zero-point reference.

6. The mass flow rate calibration method as described in claim 5, characterized in that, The step of mapping the structural damping increment to a mechanical stress pseudo-phase difference using a preset structural dynamics lookup table includes: The free vibration amplitude sequence is subjected to Fourier transform to obtain a frequency domain signal sequence, and the frequency corresponding to the maximum amplitude peak is extracted from the frequency domain signal sequence as the natural resonant frequency of the measuring tube. In the structural dynamics lookup table, find the phase shift angle value that corresponds to both the natural resonant frequency and the structural damping increment; The found phase shift angle value is set as the mechanical stress pseudo-phase difference.

7. The mass flow rate calibration method as described in claim 5, characterized in that, The step of obtaining the sampled mass flow value from the Coriolis mass flow meter includes: At the steady-state arrival time, the updated zero-point reference is input into the phase difference calculation module of the Coriolis mass flow meter; The phase difference calculation module is controlled to perform an internal subtraction correction operation using the updated zero-point reference, and outputs the sampled mass flow rate value.

8. The mass flow rate calibration method as described in claim 5, characterized in that, The machine learning model for predicting fluid relaxation states includes a spatiotemporal feature extraction network, a cross-attention mechanism layer, and a state regression output layer. When the delivery temperature or delivery pressure of the calibration fluid changes continuously, the dynamic velocity profile time series data is input into the spatiotemporal feature extraction network to output a spatiotemporal fusion feature vector, and the free vibration attenuation rate is input as a structural reference feature vector into the cross-attention mechanism layer. The cross-attention mechanism layer is controlled to perform attention weight allocation operations between the structural reference feature vector and the spatiotemporal fusion feature vector to isolate the structural damping drift component caused by the continuous change of the conveying temperature or the conveying pressure. The state regression output layer is controlled to perform a fully connected mapping operation after removing the structural damping drift component, and outputs the steady-state arrival time point.

9. The mass flow rate calibration method as described in claim 8, characterized in that, The step of controlling the cross-attention mechanism layer to perform attention weight allocation operations between the structural reference feature vector and the spatiotemporal fusion feature vector includes: Multiply the structural reference feature vector by the first weight matrix to output the query matrix; The spatiotemporal fusion feature vector is multiplied by the second weight matrix and the third weight matrix respectively to output the key matrix and the value matrix; Calculate the dot product between the query matrix and the key matrix, and output the attention score matrix; Multiply the attention score matrix by the value matrix to output the adjusted spatiotemporal fusion feature vector; The state regression output layer performs the fully connected mapping operation using the adjusted spatiotemporal fusion feature vector.

10. A mass flow rate calibration system for implementing the method of claim 1, characterized in that, include: The mounting module is used to mount and connect the Coriolis mass flow meter to the calibration bench pipeline, and to mount a smart sensor containing an ultrasonic Doppler array on the upstream side of the Coriolis mass flow meter along the axial direction of the calibration bench pipeline. The data acquisition module is used to start the liquid supply pump to continuously pump the calibration fluid into the calibration bench pipeline, and control the ultrasonic Doppler array in the smart sensor to perform cross-sectional scanning of the calibration fluid to obtain dynamic flow velocity profile time series data. The model processing module is used to input the dynamic velocity profile time series data into a pre-trained fluid relaxation state prediction machine learning model, and the fluid relaxation state prediction machine learning model outputs the steady state arrival time point. The execution module is used to trigger a calibration sampling command at the steady-state arrival time point, obtain the sampled mass flow rate value from the Coriolis mass flow meter, and compare the sampled mass flow rate value with the corresponding reference standard flow rate value to complete the calibration of the Coriolis mass flow meter.

Images