Coriolis mass flowmeter measurement calibration method and system based on deep learning

Abstract

The invention discloses a Coriolis mass flow meter measurement calibration method and system based on deep learning, and the method comprises the following steps: carrying out the adaptive filtering according to a collected vibration signal of a vibration coil during the operation of a flow meter; extracting frequency and phase information in frequency domain characteristics of the filtered signals by using a Hilbert transform method, removing abnormal values of the extracted frequency and phase information to ensure the validity of the data, and fusing the pressure data collected under the same working condition to perform data normalization. The method comprises the following steps: constructing a gated loop network model for data training, firstly, determining the optimal use weight of each characteristic parameter by using a Pearson's correlation coefficient method, and adjusting and optimizing model parameters to obtain an algorithm model with the best robustness and generalization; the method can adapt to the characteristic that the vibration signal of the vibration coil is randomly increased under the two-phase flow working condition, the measurement precision of the flowmeter under the two-phase flow can be improved to a certain extent, and the problem that the measurement precision of the Coriolis mass flowmeter under the gas-liquid two-phase flow is lowered is solved.

Description

Technical Field

[0001] The present invention relates to the field of instrumentation and automation, and particularly to a measurement calibration method and system for a Coriolis mass flowmeter based on deep learning. Technical Background

[0002] Flow measurement is widely used in people's production activities and trade settlements, such as the measurement of chemical industry and petrochemical raw materials. Different from the initial measurement of the volume flow of the flowing medium, modern flowmeters will convert the volume flow into mass flow through various characteristics of the medium, such as temperature, pressure, and density, based on theory. The Coriolis flowmeter (hereinafter referred to as the Coriolis flowmeter for short) was developed by MicroMotion, Inc. in the United States in the late 1970s. The Coriolis flowmeter is insensitive to the pressure, viscosity, and flow velocity distribution of the fluid medium and can directly measure the fluid mass with high precision. With the continuous stacking of industrial technologies, the variety and quantity of flowmeters on the market have been increasing, and the market share of traditional types of flowmeters has been showing a downward trend, while the Coriolis flowmeter has been growing at an annual growth rate of 16%. In terms of usage, the global sales of Coriolis flowmeters in 1991 were 235 million yuan, accounting for about 8% of the total sales of flowmeters at that time. The global sales of Coriolis flowmeters in the past five years were approximately 2.1 billion, accounting for 19% of the total sales of all flowmeters. Thus, it can be seen that the Coriolis flowmeter has gradually become the main product in flow meters with the characteristics of high precision and intelligence. At the same time, the application market for flowmeters in China is huge. Since the country began to introduce Coriolis flowmeters in the 1980s, about 2,000 units have been put into use in various industries by 1992.

[0003] The Coriolis flowmeter mainly consists of a measuring tube, an exciter, and a sensor, etc. The measuring tube is the pipeline for fluid flow and is generally designed as U-shaped or straight-shaped. The exciter ensures that the measuring tube vibrates at a certain vibration frequency, and the transducer records the vibration amplitude of the pipeline and transmits it to the hardware device for further analysis. However, when faced with a flowing medium composed of a combination of gas and liquid phases, the Coriolis flowmeter will no longer have the stability and high precision under single-phase flow measurement due to the complex interaction between the wall of the pipe and the fluid caused by the complex interaction at the phase interface between the gas and liquid phases. When the gas content in the fluid medium is greater than a certain value, the increase in damping caused by the friction between the air and the wall of the Coriolis flowmeter will make the internal vibrating tube unable to drive the exciter well, and even cause the flowmeter to stop vibrating. This will seriously affect the wide application of the Coriolis flowmeter in industries such as petroleum and chemical engineering. Therefore, it is necessary to improve the existing mass flow algorithm of the Coriolis flowmeter so that it can still maintain high-precision measurement under the liquid flow state with a high gas content; an appropriate algorithm should be selected to process the extracted data to improve the measurement accuracy of the Coriolis flowmeter under two-phase flow conditions. Hao Zhu [Hao Zhu. From disturbance to measurement: Application of Coriolis meter for two-phase flow with gas bubbles [J]. Flow Measurement and Instrumentation, 2021, 79: 101892.] proposed a formula correction for the measurement of gas-containing liquids based on the bubble flow model and the moving harmonic oscillator theory. The derivation process proved the influence of the pipeline pressure, the vibration frequency of the measured medium, and the vibration frequency of the exciter on the measurement process in the Coriolis flowmeter for two-phase gas-liquid flow and generated a correction formula according to the degree of influence. However, the research emphasized the influence degree of bubble type identification on the measurement process error in the measurement of the mass flow of gas-containing liquids. It is difficult to realize the real-time detection of the gas state in the liquid in the actual production process, resulting in the difficulty of determining the correct parameters for the application of the correction formula. Chunhui Li [Li, Chunhui, et al. Improvement of signal processing in Coriolis mass flowmeters for gas-liquid two-phase flow [J]. Frontiers of Information Technology & Electronic Engineering, 2021, 22(2): 272-286.] analyzed the application of three digital signal processing methods, namely the quadrature demodulation (QD) method, the Hilbert method, and the sliding discrete-time Fourier transform method, in processing the signals of Coriolis sensors, generated simulation signals to analyze the phase difference tracking performance of the three methods, and tested the algorithms on an experimental platform.However, during the experimental process of the research, the two-phase flow rate was fixed, and there was a lack of research on the vibration characteristics of gas-liquid mixtures in sensors with dynamically changing flow rates. Summary of the Invention

[0004] To solve the problem that the measurement accuracy of Coriolis mass flowmeters drops severely during measurements under complex flow conditions such as gas-liquid two-phase flow, the present invention proposes a measurement calibration method and system for Coriolis mass flowmeters based on deep learning. By reasonably training the pipeline pressure signals, sensor vibration signals, and gas mass flowmeter signals collected during the experimental process, the mapping relationship from various types of data to mass flow can be learned. During the measurement of two-phase flow with relatively complex internal interactions of the medium, the Coriolis mass flowmeter can still maintain high measurement accuracy. The present invention uses traditional digital signal algorithms to build a data-driven model after feature extraction of vibration data. The model has a certain physical interpretability and good generalization and robustness.

[0005] The present invention is achieved by at least one of the following technical solutions.

[0006] A measurement calibration method for a Coriolis mass flowmeter based on deep learning, comprising the following steps:

[0007] Using a trained gated recurrent unit model to predict the mass flow rate in the measurement tube; the training of the gated recurrent unit model includes the following steps:

[0008] Step 1: Collect the vibration data, pressure data, gas mass flowmeter data, and liquid mass flowmeter data of the Coriolis mass flowmeter under two-phase flow conditions;

[0009] Step 2: Perform preprocessing operations on the collected various data, including performing adaptive filtering on the vibration data, and segmenting and averaging the pressure data and the data of the liquid mass flowmeter and the gas mass flowmeter to make the data have a unified data length;

[0010] Step 3: Extract time-domain features from the preprocessed vibration signal. The time-domain feature extraction includes the extraction of the main frequency information and the main phase information; after the feature extraction of the vibration signal, data normalization is required;

[0011] Step 4: Train the gated recurrent unit model with the normalized feature data: determine the optimal weight selection through the Pearson correlation coefficient method for the weights of the three types of data, then build a gated recurrent unit model, perform L1 regularization constraints on the model loss function, and use an adaptive moment estimation optimizer based on gradients to control its learning rate to accelerate the convergence speed, and continuously optimize the model parameters so that the finally trained model can effectively predict the mass flow rate.

[0012] Further, the vibration data refers to the real-time vibration amplitude of the internal vibration coil of the Coriolis mass flowmeter in the gas-liquid two-phase flow scenario; the pressure data refers to the pressure value in the corresponding pipeline of the flowmeter; the gas mass flowmeter data shows the mass flow value of the gas in the current two-phase flow; the liquid mass flowmeter data shows the mass flow value of the liquid in the current two-phase flow; the sum of the gas mass flow value and the liquid mass flow value is the target value that the model should predict.

[0013] Further, the process of adaptive filtering of the vibration data is as follows:

[0014]

[0015] Where x(n) represents the amplitude of the signal to be filtered, y(n) represents the amplitude of the filtered signal, and w(n) is the weight coefficient vector calculated by the adaptive filter; during the filtering process, the output y(n) is the weighted sum of the product of the input signal x(n) and the filter coefficient w(n), and w k (n) is the kth coefficient of the filter at sample n, x(n - k) represents the input samples of the input signal from x(n) to x(n - k + 1), and L represents the number of coefficients in the weight coefficient vector;

[0016] The error value e(n) during the filtering process is:

[0017] e(n) = d(n) - y(n) = d(n) - w T (n)x(n);

[0018] d(n) represents the filtering reference signal, and the coefficients of the adaptive filter are corrected iteratively in real time according to the error value e(n);

[0019] Further, the time-domain feature extraction uses the Hilbert transform to extract the main frequency and phase difference of the signal within one calculation period. The Hilbert transform delays the phase of all frequency components of the signal by 90°, and the Hilbert transform is defined as:

[0020]

[0021] In the formula, x(t) is the filtered signal, t represents the time variable, which is used to describe the time point of the transformation; τ is the integration variable, and x(τ) represents the value of the signal x(t) at time τ;

[0022] The solution process of the main phase at time t is:

[0023]

[0024] θ(t) represents the main phase value of the signal at time t; arctan(.) represents the arctangent function;

[0025] The Coriolis mass flowmeter has two vibration signals, corresponding to two sets of phase difference data. To obtain the phase difference between the two signals, it is only necessary to find the difference between the main phases of the two sets of signals. The solution process of the main frequency f(t) at time t is as follows:

[0026]

[0027] Furthermore, the data normalization is to perform standard deviation normalization on the three types of characteristic data X, namely frequency data, phase data, and pressure data, so as to convert the data into a normal distribution data with a mean of 0 and a variance of 1. The calculation formula is as follows:

[0028]

[0029] In the formula, X' is the normalized data, X is the unnormalized data, μ is the mean of the data, and σ is the standard deviation of the data.

[0030] Furthermore, the Pearson correlation coefficient formula is used to measure the linear correlation between various types of characteristics and the target value, that is, the measured mass flow rate. The calculation formula is as follows:

[0031]

[0032] where x i is the i-th data point of the training data, that is, frequency, phase, and pressure, and y i is the i-th data point of the target prediction data, that is, the mass flow rate. The total number of training data and target prediction data is both m. is the mean of the training data, is the target prediction data. After calculation, the Pearson correlation coefficient r is obtained. Calculate the correlation coefficients r 1 、r 2 、r 3 of frequency, phase, and pressure with the target variable, that is, the mass flow rate. After normalization, the weights of each characteristic data are obtained

[0033] Furthermore, the gated recurrent unit (GRU) model realizes the prediction of the mass flow rate in the measuring pipe by training the characteristic data collected from experiments to adjust the model parameters;

[0034] The data types input into the gated recurrent network are three types, including the phase value, frequency value obtained after Hilbert transform, and the pressure value of the measuring pipe. After being calculated by the gated recurrent network, the predicted value of the model is output. The internal parameters of the model are further adjusted by the difference between the predicted mass flow rate and the true mass flow rate; The Adam activation function is selected to adjust the parameters of the model to minimize the loss function.

[0035] Furthermore, the gated recurrent network includes an update gate and a reset gate. The calculation formula of the update gate is:

[0036]

[0037] where x t is the input data, i.e., three types of feature data, h t-1 is the hidden state of the model at the previous moment, W z and b z are the weights and biases inside the update gate, is the value of the activation function in the model, which is calculated from the data features and parameter settings of the model input. z t reflects the degree to which the current state needs to be retained. The closer z t is to 1, the more the aforementioned calculation results are retained;

[0038] The calculation formula of the reset gate is:

[0039]

[0040] where W r and b r are the weights and biases inside the reset gate. The closer r t is to 0, the more the previous data is discarded;

[0041] After calculating the update gate and the reset gate, a candidate state is obtained

[0042]

[0043] where W h and b h are the weights and biases of the candidate state. After calculation, the formula can generate a temporary state by combining the current input and historical data;

[0044] Combining multiple calculation results in the model, the gated recurrent network calculates the final hidden state. The calculation process is as follows:

[0045]

[0046] To make the model training proceed in the correct prediction direction, the following regulations are made for the model loss function:

[0047]

[0048] where E(k) is the loss function, k represents the identifier of the current parameter set calculation process, P is the total number of samples participating in the loss function calculation, y ρ is the mass flow rate value predicted by the model, y ρis the target prediction data, and λ is the regularization parameter used to control the strength of the regularization term, and w j is the j-th parameter inside the model. The loss function combines the mean squared error loss function and the regularization function to ensure the correct training of the model and avoid overfitting.

[0049] A system for implementing the above-mentioned Coriolis mass flowmeter measurement calibration method based on deep learning includes a data acquisition device and a gated recurrent network model module;

[0050] The data acquisition device includes multiple flowmeters, a pressure transmitter for monitoring the pressure signal in the flow pipeline, and a single-chip microcomputer; the multiple flowmeters include a flowmeter for monitoring the gas pipeline, a flowmeter for monitoring the liquid pipeline, etc., and the single-chip microcomputer synchronously acquires the output signals of each flowmeter and the pressure gauge;

[0051] The gated recurrent network model module trains the feature data collected from the experiment through a GRU neural network to adjust the model parameters to achieve the prediction of the mass flow rate in the measuring tube.

[0052] A computer device of the present invention includes: a memory, a processor, and a computer program stored on the memory. When the computer program is executed on the processor, the above-mentioned Coriolis mass flowmeter measurement calibration method based on deep learning is implemented.

[0053] Compared with the existing technology, the beneficial effects of the present invention are:

[0054] The present invention collects various types of data under the two-phase flow condition of the Coriolis mass flowmeter, and combines the traditional digital signal processing algorithm and the data-driven machine learning model to complete the mapping from data to mass flow rate. Compared with the requirements of large signal data volume and full-cycle conditions for using the traditional digital signal algorithm, the trained model of the present invention can also ensure the prediction accuracy and response speed for small batches of data. At the same time, the present invention selects appropriate hardware modules, which can realize the data acquisition required by the model and the function of carrying the model, and has great application value. Description of the Drawings

[0055] Figure 1 is the flowchart of the Coriolis mass flowmeter measurement calibration method based on deep learning according to the embodiment of the present invention;

[0056] Figure 2 is the hardware platform framework diagram of the calibration method described in the embodiment;

[0057] Figure 3 is the application effect diagram of the algorithm model in the embodiment. Detailed Embodiments

[0058] To enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0059] An embodiment of the present invention provides a Coriolis mass flowmeter measurement and calibration system based on deep learning, including a data acquisition device and a gated recurrent network model module.

[0060] The data acquisition device includes a variety of flowmeters, a pressure transmitter for monitoring the pressure signal in the flow pipeline, and a single-chip microcomputer. The variety of flowmeters include a flowmeter for monitoring the gas pipeline, a flowmeter for monitoring the liquid pipeline, etc. The single-chip microcomputer synchronously collects the output signals of each flowmeter and the pressure gauge. In this embodiment, the gas-liquid two-phase experimental platform needs to monitor the gas mass flow and the liquid mass flow in real time. The 5MM-5LPM-D model flowmeter in ALICAT is used to monitor the gas pipeline, the TMASS15T model flowmeter of Emerson is used to monitor the liquid pipeline, and the OHR-M2 pressure transmitter of Hongrun is used to monitor the pressure signal in the flow pipeline. After the gas and liquid pipeline media converge, they jointly flow through the Coriolis mass flowmeter of Emerson's TMASS15T, and the AD7606 is used to synchronously collect the output signals of each flowmeter and the pressure gauge.

[0061] The gated recurrent network model module trains the feature data collected from the experiment through a GRU neural network to adjust the model parameters to achieve the prediction of the mass flow in the measuring tube.

[0062] Figure 1 The Coriolis mass flowmeter measurement and calibration method based on deep learning shown includes the following steps:

[0063] Figure 1 The Coriolis mass flowmeter measurement and calibration method based on deep learning shown includes the following steps:

[0064] Step 1: Collect the vibration data, pressure data, gas mass flowmeter data, and liquid mass flowmeter data of the Coriolis mass flowmeter under two-phase flow conditions from the gas-liquid two-phase experiment.

[0065] Use a single-chip microcomputer of model AD7606 to collect the voltage signals of the vibration signal ports of the Coriolis mass flowmeter and the voltage port signals of the pressure gauge at the same time, and convert them into the vibration amplitude value of the flowmeter and the measured pipeline pressure value through the use manuals of each hardware. As an example, during the experiment, the hardware can collect the vibration signal of the flowmeter at a frequency of 3000HZ, and the digital signal processing method can extract the periodic signal into 10 main frequency data and 10 main phase data.

[0066] Step 2: Perform preprocessing operations on the collected multiple data, including performing adaptive filtering on the vibration data, and segmenting and averaging the pressure data and the data of the liquid mass flowmeter and the gas mass flowmeter to make the data have a unified data length. The specific steps are as follows:

[0067] Perform adaptive filtering on the vibration data: Adaptive filtering is a filtering method with strong adaptability and filtering performance, and it has been widely used in signal digital processing. The specific filtering process is as follows:

[0068]

[0069] Among them, x(n) represents the amplitude of the signal to be filtered, y(n) represents the amplitude of the filtered signal, and w(n) is the weight coefficient vector calculated by the adaptive filter. During the filtering process, the output y(n) is the weighted sum of the product of the input signal x(n) and the filter coefficient w(n), and w k (n) is the kth coefficient of the filter at sample n, x(n-k) represents k input samples of the input signal from x(n) to x(n-k+1), and L represents the number of coefficients in the weight coefficient vector.

[0070] The error value e(n) during the filtering process is:

[0071] e(n) = d(n) - y(n) = d(n) - w T (n)x(n);

[0072] d(n) represents the filtering reference signal, and the coefficients of the adaptive filter are corrected iteratively in real time according to the error value e(n).

[0073] Step 3: Extract the time-domain characteristics of the vibration signal: The time-domain characteristic extraction is to use the Hilbert transform to extract the main frequency and phase difference of the signal within a calculation period. The Hilbert transform delays the phase of all frequency components of the signal by 90°, and the specific calculation process is that x(t) is the filtered signal, and its Hilbert transform is defined as the convolution of x(t) and , that is:

[0074]

[0075] x(t) is the filtered signal, where t represents the time variable, which is used to describe the time point of the transformation; τ is the integral variable, and x(τ) represents the value of the signal x(t) at time τ.

[0076] The process of solving the main phase is:

[0077]

[0078] θ(t) represents the main phase value of the signal at time t; arctan(.) represents the inverse tangent function.

[0079] The Coriolis mass flowmeter has two vibration signals, corresponding to two sets of phase difference data. The phase difference between the two signals is calculated by calculating the difference between the main phases of the two sets of signals. The process of solving the main frequency f(t) at time t is:

[0080]

[0081] Step 4, data normalization: After feature extraction of the vibration signal, data normalization is required to make the data suitable for the machine learning training framework. First, the validity of the current group of data is judged according to the value of the main frequency, and invalid data is eliminated. Then the training data is normalized. The data normalization is to standardize the standard deviation of the data X of the three types of feature data (frequency data, phase data, and pressure data) in the experiment to convert the data into a normal distribution data with a mean of 0 and a variance of 1 suitable for the machine learning algorithm. The calculation formula is as follows:

[0082]

[0083] In the formula, X' is the standardized data, X is the unstandardized data, μ is the mean of the data, and σ is the standard deviation of the data.

[0084] Step 4: Train the feature data through a machine learning algorithm model, including the following steps:

[0085] S1. Pearson correlation coefficient calculates the weight of each data type in the algorithm model: By using the Pearson correlation coefficient formula to measure the linear correlation between each type of feature and the target value, that is, the measured mass flow. The calculation formula is as follows:

[0086]

[0087] where x i is the i-th data point of the training data, namely frequency, phase and pressure, y i is the target prediction data, i.e., the i-th data point of mass flow, and the total number of training data and target prediction data is m. is the mean of the training data, is the target prediction data, and the Pearson correlation coefficient r is obtained after calculation. Calculate the correlation coefficients r 1 、r 2 、r 3 of frequency, phase, and pressure with the target variable, i.e., mass flow rate, and then perform normalization to obtain the weights of each feature data

[0088] S2. Select a suitable network model. In this embodiment, the gated recurrent unit (GRU) model is adopted, which performs excellently in processing sequence data by introducing the update gate and reset gate mechanisms. There are three types of input data, including the phase value, frequency value obtained after Hilbert transform, and the pressure value of the measurement pipeline. The difference between the mass flow rate predicted by the model and the true mass flow rate is used to further adjust the internal parameters of the model.

[0089] S3. Train and update the network parameters: The gated recurrent network consists of two parts, the update gate and the reset gate. The calculation formula of the update gate is:

[0090]

[0091] where x t refers to the input data of the model, i.e., the three types of feature data, and h t-1 refers to the hidden state of the previous moment in the model. W z and b z are the weights and biases inside the update gate, is the value of the activation function in the model, which is calculated from the data characteristics and parameter settings of the model input. z t reflects the degree to which the current state needs to be retained. The closer z t is to 1, the more the aforementioned calculation results are retained.

[0092] The calculation formula of the reset gate is:

[0093]

[0094] In the formula, W r and b r are the weights and biases inside the reset gate, and r t determines which previous parameters need to be discarded. The closer r t is to 0, the more previous data is discarded.

[0095] After the calculation of the update gate and the reset gate, the candidate state is obtained, and its calculation formula is as follows:

[0096]

[0097] W hWith b h is the weight and bias in the candidate state. After calculation by the formula, it can combine the current input and historical data to generate a temporary state.

[0098] Combining multiple calculation results in the model, the gated recurrent network layer calculates the final hidden state. The calculation process is as follows:

[0099]

[0100] z t is calculated for the update gate, h t-1 refers to the hidden state of the previous moment in the model. is the candidate state, h t refers to the final hidden state of the current moment in the model.

[0101] S4. Constraint network training direction: Perform L1 regularization constraint on the model loss function to increase its robustness, ensure generalization and robustness, and use the adaptive moment estimation optimizer based on gradients to control its learning rate to accelerate the convergence speed. Continuously optimize the model parameter settings so that the finally trained model can effectively predict the mass flow rate.

[0102] To make the model training proceed in the correct prediction direction, the following regulations are made for the model loss function:

[0103]

[0104] where E(k) is the loss function, k represents the identifier of the current parameter set calculation process, P is the total number of samples participating in the loss function calculation, y ρ is the mass flow rate value predicted by the model, y ρ is the target prediction data, λ is the regularization parameter used to control the strength of the regularization term, w j is the jth parameter inside the model. The loss function combines the mean squared error loss function and the regularization function to ensure the correct training of the model and avoid overfitting.

[0105] As an embodiment, the trained data model is imported into the Raspberry Pi 4B to realize the deployment and application of the data model on the hardware. Through the communication between the Raspberry Pi 4B and the AD7606, data transmission can be achieved between the two. After the Raspberry Pi 4B receives various data, it realizes the prediction of the mass flow rate according to the loaded data model. Some experimental result graphs are as Figure 3 shown.

[0106] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A Coriolis mass flowmeter measurement and calibration method based on deep learning, characterized in that: The following steps are involved: Prediction of mass flow in a measuring tube using a trained gated recurrent network model; The training of the gated recurrent network model consists of the following steps: Step 1: Collect vibration data, pressure data, gas mass flow meter data, and liquid mass flow meter data of the Coriolis mass flow meter under two-phase flow conditions; Step 2: preprocessing the collected data, including adaptive filtering of vibration data, averaging the pressure data and the data of liquid mass flow meter and gas mass flow meter in sections, so that the data has a uniform data length; Step 3: Extract time domain features from the preprocessed vibration signal. The time domain feature extraction includes main frequency information extraction and main phase information extraction. After feature extraction of the vibration signal, data normalization is required. Step 4: Use the normalized feature data to train the gated recurrent network model: Use the Pearson correlation coefficient method to weight the three types of data to determine the optimal weight selection, then build the gated recurrent network model, perform L1 regularization constraints on the model loss function, and use a gradient-based adaptive moment estimation optimizer to control its learning rate to accelerate convergence, and continuously optimize the model parameters so that the final trained model can effectively predict mass flow.

2. The Coriolis mass flowmeter measurement and calibration method based on deep learning according to claim 1 is characterized in that: The vibration data refers to the real-time vibration amplitude of the internal vibration coil of the Coriolis mass flowmeter in the gas-liquid two-phase flow scenario; the pressure data refers to the pressure value in the corresponding pipeline of the flowmeter; the gas mass flowmeter data displays the mass flow value of the gas in the current two-phase flow; the liquid mass flowmeter data displays the mass flow value of the liquid in the current two-phase flow; the sum of the gas mass flow value and the liquid mass flow value is the target value that the model should predict.

3. The Coriolis mass flowmeter measurement and calibration method based on deep learning according to claim 1, characterized in that: The process of adaptive filtering of vibration data is as follows: Where x(n) represents the amplitude of the filtered signal, y(n) represents the amplitude of the filtered signal, and w(n) is the weight coefficient vector calculated by the adaptive filter. During the filtering process, the output y(n) is the weighted sum of the product of the input signal x(n) and the filter coefficient w(n). k (n) is the kth coefficient of the filter on sample n, x(nk) represents the input samples of the input signal from x(n) to x(n-k+1), and L represents the number of coefficients in the weight coefficient vector; The error value e(n) during the filtering process is: e(n)=d(n)-y(n)=d(n)-w T (n)x(n); d(n) is represented as the filtering reference signal, and the coefficients of the adaptive filtering are iteratively modified in real time according to the error value e(n).

4. The Coriolis mass flowmeter measurement and calibration method based on deep learning according to claim 1, characterized in that: The time domain feature extraction is to use Hilbert transform to extract the main frequency and phase difference of the signal within a calculation cycle. The Hilbert transform delays the phase of all frequency components of the signal by 90°. Defined as: Where x(t) is the filtered signal, t represents the time variable, which is used to describe the time point of the transformation; τ is the integral variable, and x(τ) represents the value of the signal x(t) at time τ; The solution process of the main phase at time t is: θ(t) represents the main phase value of the signal at time t; arctan(.) represents the inverse tangent function; The Coriolis mass flowmeter has two vibration signals, corresponding to two sets of phase difference data. The phase difference of the two signals is calculated by calculating the difference of the main phases of the two sets of signals. The process of solving the main frequency f(t) at time t is as follows:

5. The Coriolis mass flowmeter measurement and calibration method based on deep learning according to claim 1, characterized in that: The data normalization is to normalize the standard deviation of the data X of the three types of characteristic data, frequency data, phase data and pressure data, so as to convert the data into normal distribution data with a mean of 0 and a variance of 1. The calculation formula is as follows: In the formula, X' is the standardized data, X is the unstandardized data, μ is the mean of the data, and σ is the standard deviation of the data.

6. The Coriolis mass flowmeter measurement and calibration method based on deep learning according to claim 1, characterized in that: The Pearson correlation coefficient formula is used to measure the linear correlation between various characteristics and the target value, that is, the measured mass flow rate. The calculation formula is as follows: where x i is the i-th data point of the training data, namely frequency, phase and pressure, y i is the target prediction data, i.e., the i-th data point of mass flow. The total number of training data and target prediction data is m. is the mean of the training data, The target prediction data is obtained after calculation, and the Pearson correlation coefficient r is obtained. After calculating the correlation coefficients r1, r2, and r3 between the frequency, phase, and pressure and the target variable, namely, the mass flow rate, the weight of each feature data is obtained by normalization.

7. The Coriolis mass flowmeter measurement and calibration method based on deep learning according to claim 1, characterized in that: The gated recurrent network (GRU) model is trained on the feature data collected from the experiment to adjust the model parameters to predict the mass flow in the measuring tube. There are three types of data types for the input of the gated recurrent network, including the phase value, frequency value and pressure value of the measuring pipeline obtained after the Hilbert transform. The predicted value of the model is output after calculation by the gated recurrent network, and the internal parameters of the model are further adjusted by the difference between the predicted mass flow rate and the actual mass flow rate; the Adam activation function is used to adjust the parameters of the model to minimize the loss function.

8. The deep learning-based Coriolis mass flowmeter measurement and calibration method according to claim 7, characterized in that: The gated recurrent network includes an update gate and a reset gate. The calculation formula of the update gate is: where x t is the input data, i.e., three types of feature data, h t-1 is the hidden state of the model at the previous moment, W z With b z is the weight and bias inside the update gate, is the value of the activation function in the model, which is calculated based on the data features and parameter settings of the model input. t Reflects the degree to which the current state needs to be preserved, z t The closer it is to 1, the more the above calculation results are retained; The calculation formula for the reset gate is: Where W r With b r is to reset the weights and biases inside the gate, r t The closer it is to 0, the more previous data is discarded; After the update gate and reset gate calculation, the candidate state is obtained Where W h With b h are the weights and biases of the candidate states. After calculation, the formula can combine the current input and historical data to generate a temporary state. Combining multiple calculation results in the model, the gated recurrent network calculates the final hidden state. The calculation process is as follows: In order to make the model training proceed in the correct prediction direction, the model loss function is specified as follows: Where E(k) is the loss function, k represents the identifier of the current parameter set calculation process, P is the total number of samples involved in the loss function calculation, and y ρ is the mass flow value predicted by the model, y ρ is the target prediction data, λ is the regularization parameter used to control the strength of the regularization term, and w j It is the jth parameter inside the model. The loss function combines the mean square error loss function with the regularization function to ensure the correct training of the model while avoiding overfitting.

9. A system for implementing the deep learning-based Coriolis mass flowmeter measurement and calibration method according to claim 1, characterized in that: It includes a data acquisition device and a gated recurrent network model module; The data acquisition device includes a variety of flow meters, a pressure transmitter for monitoring the pressure signal in the flow pipeline, and a single-chip microcomputer; the various flow meters include a flow meter for monitoring the gas pipeline, a flow meter for monitoring the liquid pipeline, etc., and the single-chip microcomputer synchronously collects the output signals of each flow meter and the pressure gauge; The gated recurrent network model module trains the characteristic data collected from the experiment through a GRU neural network to adjust the model parameters to achieve the prediction of the mass flow in the measuring tube.

10. A computer device, characterized in that: It comprises: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, it implements the deep learning-based Coriolis mass flowmeter measurement and calibration method as described in any one of claims 1 to 8.

Images