Flow instrument intelligent calibration system based on multi-sensor fusion

Abstract

The invention relates to the technical field of flow measurement calibration, in particular to a flow instrument intelligent calibration system based on multi-sensor fusion, which comprises a data acquisition unit, a deep coupling compensation unit and a closed loop verification unit, a data acquisition unit obtains a differential pressure value, an environment temperature, a pipeline pressure, a vibration frequency spectrum and sensor accumulated working time, a depth coupling compensation unit constructs an aging prediction model based on a Weibull distribution life model, and a temperature-pressure coupling equation and a vibration compensation mechanism are combined to obtain an aging real-time value and a time sequence deviation. Multi-parameter coupling characteristics are extracted through a neural network, an environment disturbance compensation coefficient matrix is constructed, a joint compensation amount is generated through dynamic weight distribution, a closed-loop verification unit optimizes model parameters, the problems that multi-source disturbance coupling analysis is insufficient and calibration precision is low in the prior art are solved, accurate calibration of a flow instrument under complex working conditions is achieved, and the calibration precision is improved. The metering stability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flow metering calibration, in particular to a flow instrument intelligent calibration system based on multi-sensor fusion. BACKGROUND

[0002] Flow metering calibration is an important technology. In industrial processes such as petroleum chemical industry and water supply and drainage, accurate measurement of flow instruments is the core link to ensure production efficiency and control cost loss. Intelligent calibration technology corrects instrument measurement deviation in real time, which is a key means to ensure the reliability of flow data.

[0003] With the improvement of industrial automation level, it is of great significance to build a calibration system that integrates multi-parameter sensing and dynamic compensation to adapt to complex working conditions and prolong the calibration period of instruments. Traditional single-parameter calibration mode has been difficult to cope with the measurement deviation caused by environmental disturbance and equipment aging. However, the traditional flow instrument calibration technology has the core problems of insufficient multi-source disturbance coupling analysis and low calibration accuracy. The existing scheme only relies on single parameter of differential pressure, and does not establish a coupling compensation model of environmental temperature, pipeline pressure and vibration. When the environmental temperature fluctuates obviously, the interference of temperature on differential pressure measurement is not effectively corrected, which leads to a significant increase in measurement error. Fixed period calibration cannot match the aging characteristics of instruments with working time. After the sensor accumulates a long working time, the drift error caused by aging gradually increases, and it is not associated with vibration spectrum analysis. Instantaneous deviation caused by high-frequency vibration further enlarges the total error. In addition, the traditional system lacks a dynamic weight distribution mechanism. In the case of sudden temperature and pressure changes, the compensation strategy cannot be adjusted adaptively, and the calibration response is lagged, which leads to a decrease in the flow control accuracy of industrial processes, and ultimately causes a decrease in product qualification rate. In order to solve this technical problem, we provide a flow instrument intelligent calibration system based on multi-sensor fusion. SUMMARY

[0004] The purpose of the present application is to provide a flow instrument intelligent calibration system based on multi-sensor fusion to solve the problems raised in the background.

[0005] Due to the insufficient multi-source disturbance coupling analysis of traditional technology and the reliance on single parameter of differential pressure, environmental temperature and other interference cause large measurement error. Therefore, the present case acquires multi-parameters through a data acquisition unit, constructs an environmental disturbance compensation coefficient matrix through a deep coupling compensation unit, which can effectively correct environmental interference and improve measurement accuracy.

[0006] Since traditional calibration cannot match the aging characteristics of instruments, fixed period calibration leads to large drift error. Therefore, the present case constructs an aging prediction model through a Weibull distribution life model to generate a joint compensation amount, which can dynamically adapt to the aging characteristics, reduce the drift error and improve the calibration accuracy.

[0007] To achieve the above object, a flow meter intelligent calibration system based on multi-sensor fusion is provided, comprising: The data acquisition unit obtains the differential pressure value, environmental temperature value, pipeline pressure value, vibration frequency spectrum value and sensor cumulative working time of the flow meter through the sensor; The deep coupling compensation unit takes the sensor cumulative working time as input, constructs an aging prediction model based on the Weibull distribution life model and outputs an aging prediction value, inputs the environmental temperature value, pipeline pressure value and vibration frequency spectrum value into a temperature-pressure coupling equation to calculate a theoretical differential pressure value, obtains an aging real-time value according to the differential pressure value and the theoretical differential pressure value, performs time sequence deviation calculation on the aging real-time value and the aging prediction value to obtain a time sequence deviation, extracts the pressure-temperature coupling features obtained from the environmental temperature value and the pipeline pressure value and the frequency energy distribution features of the vibration frequency spectrum value through the convolution kernel in the embedded neural network and the Fourier transform layer respectively, inputs the pressure-temperature coupling features and the frequency energy distribution features into a long short-term memory network, outputs fusion features, and maps the fusion features into multi-dimensional compensation coefficients through a fully connected layer, constructs an environmental disturbance compensation coefficient matrix, and finally performs weighted fusion on the time sequence deviation and the environmental disturbance compensation coefficient matrix to generate a joint compensation amount.

[0008] As a further improvement of the technical solution, a segmented failure period modeling and dynamic parameter optimization mechanism is adopted to construct an aging prediction model based on the Weibull distribution life model and output an aging prediction value: A segmented life prediction model is established based on accelerated degradation test data, and three stages of initial failure period, stable period and wear-out failure period are divided, and the accelerated degradation test data is obtained from sensor historical data; In the initial failure period, an exponential decay function is used to fit the drift characteristics in the running-in stage, in the stable period, a logarithmic function is used to simulate the material fatigue progressive process, and in the wear-out failure period, a power function is used to represent the material structure damage accumulation effect; Finally, according to the online error data collected by the closed-loop verification unit, the gradient descent algorithm is used to dynamically update the slope coefficient and time sequence attenuation factor of each stage function, and the online error data is the absolute deviation of the automatic calibration flow value and the standard value.

[0009] As a further improvement of the technical solution, when obtaining the aging real-time value according to the differential pressure value and the theoretical differential pressure value, a vibration no-load calibration and vibration frequency band energy compensation double mechanism is executed: Under the zero flow condition of the flow meter, the environmental temperature is adjusted in a stepwise heating mode, the pipeline pressure is adjusted synchronously in a stepwise pressurization mode, and the differential pressure offsets of multiple temperature-pressure combinations are recorded; A compensation polynomial equation containing a temperature first-order term, a pressure first-order term and a temperature-pressure cross term is generated by using a multiple regression analysis method, and for vibration interference, the sliding average value of the vibration energy in the preset frequency band is extracted, multiplied by the vibration sensitivity calibration coefficient, and then superimposed to the theoretical differential pressure value, and the final aging real-time value is the algebraic difference between the original differential pressure value and the theoretical differential pressure value after vibration compensation, and the vibration sensitivity calibration coefficient is obtained by vibration no-load calibration.

[0010] As a further improvement of the technical solution, the time sequence deviation calculation introduces a double-layer fusion architecture of first-order instantaneous deviation and second-order trend deviation: The first-order instantaneous deviation is obtained by subtracting the aging prediction value from the aging real-time value at the current time, and the second-order trend deviation is obtained by calculating the moving average value of the hourly change rate according to the sequence of the first-order instantaneous deviation in the past 24 hours; According to the life cycle stage of the sensor, the deviation weight is configured, the second-order trend deviation is given in the initial failure period, and the first-order instantaneous deviation is given in the stable period, and the weighted first-order instantaneous deviation and second-order trend deviation are linearly superimposed to generate the time sequence deviation.

[0011] As a further improvement of the technical solution, when the convolution kernel extracts the pressure-temperature coupling features obtained from the environment temperature value and the pipeline pressure value, a double-channel cross-correlation operation architecture is adopted: Channel one is configured with a one-dimensional convolution layer with a preset sliding window size to extract the per-minute gradient change features of the environment temperature value, and channel two is configured with a dilated convolution layer with a preset inflation rate to capture the periodic fluctuation mode of the pipeline pressure value; The per-minute gradient change features of the environment temperature value and the periodic fluctuation mode of the pipeline pressure value are cross-correlated element by element to generate a coefficient matrix reflecting the pressure-temperature coupling time lag effect, and the matrix element value represents the response delay length of the pipeline pressure value caused by unit environment temperature value, and the coefficient matrix of the pressure-temperature coupling time lag effect is taken as the pressure-temperature coupling feature.

[0012] As a further improvement of the technical solution, when the Fourier transform layer extracts the frequency energy distribution features, a one-third octave and energy entropy fusion analysis is performed: The vibration frequency spectrum value is divided into one-third octave, the energy proportion values of three center frequency bands are extracted, and the information entropy value of the full-band vibration energy is calculated to represent the randomness intensity of the energy distribution, and the energy proportion values of the three center frequency bands and the information entropy value are concatenated to form a four-dimensional vibration feature vector, and the four-dimensional vibration feature vector is taken as the frequency energy distribution feature.

[0013] As a further improvement of the technical solution, when the pressure-temperature coupling features and the frequency energy distribution features are input into the long short-term memory network to output the fusion features, a gated attention and high-frequency forgetting dual mechanism is adopted: The input gate of the long short-term memory network receives the stress-temperature coupling features and the frequency energy distribution features, the forgetting gate of the long short-term memory network monitors the change rate of the ambient temperature value, the high-frequency oscillation feature discarding protocol is started when the temperature rise of the ambient temperature value per minute exceeds the preset temperature rise rate, the output gate of the long short-term memory network generates a two-dimensional time coding matrix, the row vector of the two-dimensional time coding matrix corresponds to the sliding time window number, the column element of the two-dimensional time coding matrix stores the coupling strength coefficient of the environmental disturbance on the flow signal, and the two-dimensional time coding matrix is the final fusion feature.

[0014] As a further improvement of the technical solution, when the full connection layer constructs the environmental disturbance compensation coefficient matrix, a steady-state and transient dual-path compensation strategy is implemented: The steady-state path output zero drift compensation coefficient and the range slope compensation coefficient are used for basic working condition calibration, and the transient path output pressure shock suppression gain and temperature recovery rate compensation coefficient are used for sudden environmental disturbance. The environmental compensation coefficient matrix is stored in a two-dimensional table according to the line index compensation type and the column index sensor ID, and the table element value is a normalized compensation weight.

[0015] As a further improvement of the technical solution, the generation of the joint compensation quantity adopts a dynamic weight distribution mechanism controlled by a deviation threshold value: When the absolute value of the timing deviation exceeds 0.5% of the full range, the timing deviation accounts for 70% of the compensation quantity weight, and the Euclidean norm of the compensation matrix accounts for 30%; When the absolute value of the timing deviation is lower than the threshold value, the weighted sum of the coefficients of the compensation matrix accounts for 85%, and the timing deviation is only used to adjust the matrix elements. The final joint compensation quantity is the linear superposition value of the two types of compensation components.

[0016] As a further improvement of the technical solution, the closed-loop verification unit comprises a double-channel back propagation and forced triggering rule, which is as follows: The neural network optimization channel takes the root mean square of the calibration error as the loss function, and updates the convolution kernel weight parameter by using the stochastic gradient descent method. The life model optimization channel uses the timing deviation cumulative distribution to adjust the Weibull function segmented parameter group by using the particle swarm algorithm. When the calibration error exceeds the accuracy level requirement for three times in succession or the cumulative running time reaches 100 hours, the double-channel optimization program is forced to start.

[0017] Compared with the prior art, the present application has the following advantages: In a flow meter intelligent calibration system based on multi-sensor fusion, a deep coupling compensation unit constructs an aging prediction model based on a Weibull distribution life model, combines a temperature-pressure coupling equation and a vibration compensation mechanism, accurately obtains aging real-time values and timing deviations, extracts multi-parameter coupling features through an embedded neural network, constructs an environmental disturbance compensation coefficient matrix, realizes collaborative compensation of aging characteristics and environmental disturbances, generates a joint compensation amount using a dynamic weight distribution mechanism, and a steady-state and transient dual-path compensation strategy respectively adapts to basic working conditions and sudden disturbances, a closed-loop verification unit continuously optimizes model parameters to ensure compensation accuracy, effectively corrects measurement deviations caused by environmental temperature, pipeline pressure, vibration and equipment aging, improves the measurement accuracy and stability of the flow meter under complex working conditions, prolongs the calibration period, reduces cost loss and energy waste in industrial processes, and meets the demand for high-precision flow measurement. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The figure is a whole block diagram of the present application.

[0019] The meanings of various numbers in the figure are as follows: 1, data acquisition unit; 2, deep coupling compensation unit; 3, closed-loop verification unit. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0021] The present application provides a flow meter intelligent calibration system based on multi-sensor fusion, please refer to Figure 1 As shown in the figure, it comprises: The data acquisition unit 1 obtains the differential pressure value, environmental temperature value, pipeline pressure value, vibration spectrum value and sensor cumulative working time of the flow meter through the sensor; The deep coupling compensation unit 2 takes the sensor cumulative working time as input, constructs an aging prediction model based on the Weibull distribution life model and outputs an aging prediction value, inputs the environmental temperature value, the pipeline pressure value and the vibration frequency spectrum value into a temperature-pressure coupling equation to calculate a theoretical differential pressure value, obtains an aging real-time value according to the differential pressure value and the theoretical differential pressure value, performs time sequence deviation calculation on the aging real-time value and the aging prediction value to obtain a time sequence deviation, extracts the pressure-temperature coupling features obtained from the environmental temperature value and the pipeline pressure value and the frequency energy distribution features of the vibration frequency spectrum value through the convolution kernel and the Fourier transform layer in the embedded neural network respectively, inputs the pressure-temperature coupling features and the frequency energy distribution features into a long short-term memory network, outputs fusion features, maps the fusion features to multi-dimensional compensation coefficients through a fully connected layer, constructs an environmental disturbance compensation coefficient matrix, and finally performs weighted fusion on the time sequence deviation and the environmental disturbance compensation coefficient matrix to generate a joint compensation amount.

[0022] Further, in order to accurately predict the aging characteristics of the flow meter with working time and provide a reliable basis for subsequent calibration compensation, the deep coupling compensation unit 2 adopts a segmented failure period modeling and dynamic parameter optimization mechanism to construct an aging prediction model based on the Weibull distribution life model and output an aging prediction value, specifically as follows: A segmented life prediction model is established based on accelerated degradation test data. Because the aging process of the flow meter presents different characteristics in different stages, segmented modeling is needed to accurately reflect its life characteristics. The accelerated degradation test data are derived from the sensor historical database, which stores the degradation data of the same type of meter under different stress conditions. The cumulative working time of each meter and the corresponding differential pressure measurement deviation are extracted to form the original data set. Then, the K-means clustering algorithm is used to divide the data into stages, and the deviation rate is used as the clustering index. The life of the meter is divided into three stages: the initial failure period (the first 3000 hours, the deviation rate is greater than 0.05% / hour), the stable period (3000-15000 hours, the deviation rate is 0.01%-0.05% / hour), and the wear failure period (more than 15000 hours, the deviation rate is greater than 0.05% / hour). Ensure that the characteristics of each stage are significantly distinguished. Different functions are used to fit the aging characteristics in different stages. In the initial failure period, the instrument is in the running-in stage, and the component gap gradually stabilizes. The deviation decreases rapidly with time and then tends to be flat. Therefore, an exponential decay function is used to fit the drift characteristics in the initial failure period by least squares method, so that the mean square error of the fitted curve and the measured data is less than 0.02%. In the stable period, the material fatigue causes the deviation to increase slowly and linearly. A logarithmic function is used to simulate the gradual process of material fatigue, which can accurately simulate the gradual change of the deviation with time. In the wear failure period, the rapid increase of the deviation is caused by the aggravation of the component wear. A power function is used to represent the cumulative effect of structural damage, and the cumulative effect of structural damage is represented by the function. Through iterative optimization, the prediction error of the function is reduced when the cumulative working time exceeds 15000 hours. Finally, the gradient descent algorithm is used to dynamically update the slope coefficient and time sequence attenuation factor of each stage function according to the online error data collected by the closed-loop verification unit 3. Because there are differences between the actual working conditions and the accelerated test, the model needs to be corrected in real time to ensure the prediction accuracy. The online error data is the absolute deviation of the flow indication value and the standard value after each automatic calibration. It is collected once an hour and stored in the database. The gradient descent algorithm is used to update the parameters. The prediction deviation (the difference between the model output value and the online error data) is used as the loss function. The partial derivatives of each parameter are calculated. The slope coefficient and the time sequence attenuation factor are adjusted gradually according to the learning rate η=0.01. After each iteration update, the loss function value is calculated. When the change rate of the loss function value is less than 0.01% for 5 consecutive iterations, the update is stopped.

[0023] Further, in order to accurately obtain the aging real-time value of the flow meter, the interference of environmental factors such as vibration, temperature and pressure needs to be eliminated. Therefore, when the aging real-time value is obtained according to the differential pressure value and the theoretical differential pressure value, the vibration empty load calibration and vibration frequency band energy compensation double mechanism are executed, which are as follows: Vibration no-load calibration is performed on the flow meter under zero flow conditions. Since there is no medium flow interference at zero flow, the differential pressure offset caused by pure environmental factors can be accurately obtained. During the specific operation, the ambient temperature is regulated in a step-by-step heating mode. Starting from the lower limit of the instrument's operating temperature, the temperature is increased by 5°C for 10 minutes each time until the upper limit is reached. The pipeline pressure is adjusted by synchronous step-by-step pressurization, and the differential pressure offset of multiple temperature and pressure combinations is recorded. Starting from 0.1 MPa, the temperature is increased by 0.1 MPa for 5 minutes each time until 1.0 MPa, forming 90 temperature and pressure combinations containing 13 temperature points and 10 pressure points. For each combination, the deviation between the output value of the differential pressure sensor and the theoretical zero value, that is, the differential pressure offset, is recorded, and the vibration spectrum value under the corresponding working condition is collected and stored in the calibration database. This process ensures that the environmental interference in the entire working range of the instrument is covered, providing basic data for subsequent compensation. The multivariate regression analysis method is used to generate a compensation polynomial equation containing a temperature linear term, a pressure linear term and a temperature-pressure cross term. Since the influence of temperature and pressure on the differential pressure is synergistic, the coupling effect needs to be reflected through the cross term. The temperature-pressure combination is used as the independent variable, and the temperature T, pressure P and differential pressure offset are used as the dependent variables. The equation is constructed: ΔP=k0+k1T+k2P+k3TP, where k0 is a constant term, k1 and k2 are the temperature and pressure linear term coefficients respectively, and k3 is the temperature-pressure cross term coefficient. 90 groups of data are fitted by the least squares method, and the values ​​of each coefficient are calculated to make the average error between the equation prediction value and the measured offset less than 0.02 kPa. This equation can effectively correct the differential pressure deviation caused by the separate and combined effects of temperature and pressure, improve the accuracy of the theoretical differential pressure value, and implement frequency band energy compensation for vibration interference, because pipeline vibration will produce additional The external differential pressure signal needs to be quantified and its influence eliminated. The sliding average of the vibration energy within the preset frequency band is extracted to smooth the instantaneous vibration fluctuations. The sliding average is multiplied by the vibration sensitivity calibration coefficient to obtain the vibration compensation amount, which is superimposed on the theoretical differential pressure value after temperature-pressure compensation to correct the deviation caused by vibration and reduce the error caused by vibration interference. The final aging real-time value is the algebraic difference between the original differential pressure value and the theoretical differential pressure value after vibration compensation. Because the theoretical differential pressure value after dual-mechanism compensation has eliminated environmental interference, the difference between the two is the actual deviation caused by the aging of the instrument itself. This value can accurately reflect the drift of the instrument due to aging, and provide a reliable real-time reference vibration for subsequent timing deviation calculations. The vibration sensitivity calibration coefficient is obtained by vibration no-load calibration, providing accurate real-time comparison data for the aging prediction model, ensuring the reliability of subsequent joint compensation.

[0024] Furthermore, to accurately quantify the deviation between the real-time aging value and the predicted value, taking into account both instantaneous fluctuations and long-term trends, the time series deviation calculation introduces a two-layer fusion architecture of first-order instantaneous deviation and second-order trend deviation, as follows: First, the first-order instantaneous deviation is calculated, because it can directly reflect the real-time deviation degree of the aging state at the current time, and the first-order instantaneous deviation is obtained by subtracting the aging prediction value from the aging real-time value at the current time. Then, the second-order trend deviation is calculated. Since the first-order instantaneous deviation is easily affected by short-term noise, it needs to be smoothed by long-term trend analysis. Therefore, the moving average of the hourly change rate is calculated according to the first-order instantaneous deviation sequence of the past 24 hours. The second-order trend deviation is obtained by calculating the moving average of the hourly change rate according to the first-order instantaneous deviation sequence of the past 24 hours. The specific steps are as follows: The deviation data of the past 24 hours is divided by hour to obtain 24 hourly deviation values. The average value of the deviation in each hour is calculated, the difference between the deviation values of adjacent hours is calculated, and 23 hourly change rates are obtained, such as the difference between the deviation values of the 2nd hour and the 1st hour. A moving average method with a window size of 3 hours is used to smooth the change rate, that is, the trend value of each point is the average value of the change rate of itself and the change rate of the previous and next 1 hour. Finally, 21 trend values are obtained, and the last trend value is taken as the second-order trend deviation. This value can effectively filter out instantaneous noise and reflect the long-term evolution trend of the deviation. Then, the deviation weight is configured according to the sensor life cycle stage. In the initial failure period, the second-order trend deviation is given a weight, because the dominant factors of deviation are different in different stages: in the initial failure period (first 3000 hours), the instrument aging state is unstable, and the trend change is more significant. Therefore, the second-order trend deviation is given a weight of 60%, and the first-order instantaneous deviation is given a weight of 40%. In the stable period (3000-15000 hours), the aging state is flat, and the instantaneous deviation can better reflect the true state, so the first-order instantaneous deviation is given a weight of 60%, and the second-order trend deviation is given a weight of 40%. In the wear-out failure period (more than 15000 hours), the trend deterioration is intensified, and the weight configuration is consistent with that in the initial failure period. The weight values are stored in the stage weight table. The system automatically matches the weight of the corresponding stage according to the cumulative working time of the sensor, and the response time of the switching is less than 100 milliseconds. The first-order instantaneous deviation and the second-order trend deviation after weighting are linearly superimposed to generate the time series deviation. This value not only retains the accuracy of the real-time state, but also integrates the stability of the long-term trend, providing a reliable deviation benchmark for the subsequent weighted fusion of the environmental disturbance compensation matrix.

[0025] Further, in order to accurately extract the coupling characteristics between the environmental temperature and the pipeline pressure, and reflect the influence of the interaction between the two on the flow measurement, when the convolution kernel extracts the pressure-temperature coupling characteristics obtained by the environmental temperature value and the pipeline pressure value, a double-channel cross-correlation operation architecture is adopted, which is as follows: First, a dual-channel convolution layer is configured to process the ambient temperature value and the pipeline pressure value respectively. Because the gradient change of temperature and the periodic fluctuation of pressure are the main manifestations of the coupling effect between the two, targeted feature extraction is required. Channel 1 configures a one-dimensional convolution layer with a preset sliding window size to extract the gradient change characteristics of the ambient temperature value per minute. The sliding window size is set to 5 minutes, that is, each convolution operation covers 5 consecutive minutes of temperature data. The convolution kernel size is 1×5, and the step size is 1 minute. By calculating the first-order difference of the temperature value in the window and taking the absolute value, the temperature change rate per minute is obtained. The first-order difference is the temperature of the next minute minus the temperature of the previous minute, and then the convolution operation is aggregated into a gradient change feature vector. This vector can clearly reflect the trend of temperature rise and fall. Channel 2 is configured with a dilated convolution layer with a preset expansion rate to capture the periodic fluctuation mode of the pipeline pressure value. Considering that the pressure fluctuation may have a long period, the expansion rate is set to 3, that is, the interval between the convolution kernel elements is 3 sampling points, and the convolution kernel size is 1×7. By expanding the receptive field to cover a longer time range of pressure data, for the pressure values ​​collected every hour, the dilated convolution layer calculates the matching degree between the pressure value in the window and the preset period template, and outputs the periodic intensity feature to avoid the loss of periodic information caused by the sampling interval. Then, the features output by the two channels are cross-correlated element by element. Because there is a time delay in the coupling of temperature change and pressure fluctuation, such as the expansion of the pipeline due to heat after the temperature rises, the pressure rise has a lag. This time lag effect needs to be quantified. The specific steps are as follows: An element-by-element cross-correlation operation is performed on the minute-by-minute gradient variation characteristics of the ambient temperature value and the periodic fluctuation mode of the pipeline pressure value. Specifically, the temperature eigenvector is fixed, and the pressure eigenvector is slid from left to right, one element at a time. The sum of the products of the elements at the corresponding positions is calculated. The position where the maximum cross-correlation result is located is the response delay time of the pressure to temperature. This generates a coefficient matrix reflecting the pressure-temperature coupling time lag effect. The matrix element values ​​represent the response delay time of the pipeline pressure value caused by a unit ambient temperature value. The coefficient matrix of the pressure-temperature coupling time lag effect is used as the pressure-temperature coupling feature. Finally, this coefficient matrix is ​​used as the pressure-temperature coupling feature. The matrix dimension is consistent with the length of the two channel eigenvectors. It fully records the time lag relationship between temperature changes and pressure fluctuations at different times. For example, the element values ​​in a certain area of ​​the matrix are concentrated at 5 minutes, indicating that the pressure response to temperature during this period is generally delayed by 5 minutes. This feature representation method can accurately quantify the dynamic coupling relationship between the two, providing a basis for time lag quantification for subsequent fusion analysis and providing key coupling feature input for the construction of the environmental disturbance compensation coefficient matrix.

[0026] Furthermore, in order to fully capture the energy distribution characteristics of the vibration signal and accurately quantify the interference of vibration on flow measurement, the Fourier transform layer performs a one-third octave band and energy entropy fusion analysis when extracting the frequency domain energy distribution characteristics, as follows: The vibration spectrum value is segmented by one-third octave, because the distribution of vibration energy in different frequency bands directly affects the instrument measurement deviation, and the octave segmentation can focus on the key frequency band according to the human ear perception characteristics. The vibration spectrum value is obtained by the data acquisition unit 1 through the vibration sensor, and is converted into a frequency domain signal through Fourier transform. The one-third octave segmentation is performed according to the international standard frequency sequence. The center frequency f0 of each frequency band and the upper and lower limit frequencies f1 and f2 satisfy f2 / f0=f0 / f1=2^(1 / 3), that is, the width of each frequency band is 2^(1 / 3)-2^(-1 / 3) times of the center frequency. Three key center frequency bands are selected from the segmented frequency bands, which are 20-50 Hz (dominated by pipeline fluid pulsation), 100-300 Hz (dominated by pump body vibration), and 500-1000 Hz (dominated by internal component resonance of the instrument). These three frequency bands cover more than 90% of the vibration interference energy through experimental verification. The energy proportion of the three center frequency bands is extracted. In order to quantify the contribution of each frequency band energy in the total energy, the vibration energy in each center frequency band is calculated. The vibration energy is obtained by integrating the power spectrum of the frequency domain signal in the frequency band, and then divided by the total vibration energy of the full frequency band (1-10000 Hz) to obtain the energy proportion. The information entropy value of the full frequency band vibration energy is calculated to represent the randomness intensity of the energy distribution. Since the randomness of the vibration energy distribution will affect the measurement stability, the information entropy can effectively represent this randomness. The information entropy calculation steps are as follows: the full frequency band is segmented into 30 frequency bands (covering 1-10000 Hz) by one-third octave, the energy proportion pᵢ of each frequency band is calculated, and the information entropy is calculated according to the formula H=-Σpᵢ・ln(pᵢ). The greater the entropy value, the more dispersed the energy distribution (the stronger the randomness), and vice versa. Finally, the energy proportion values of the three center frequency bands and the information entropy value are concatenated to form a four-dimensional vibration feature vector, i.e. [20-50 Hz proportion, 100-300 Hz proportion, 500-1000 Hz proportion, information entropy value]. This vector contains the energy contribution of specific frequency bands and covers the randomness of the overall distribution, fully representing the frequency domain characteristics of vibration. The four-dimensional vibration feature vector is used as the frequency energy distribution feature, which provides a key input of vibration dimension for the subsequent fusion analysis of long short-term memory network.

[0027] Further, in order to effectively fuse the pressure-temperature coupling feature and the frequency energy distribution feature, and accurately capture the dynamic influence of environmental disturbance on the flow signal, the pressure-temperature coupling feature and the frequency energy distribution feature are input into the long short-term memory network, and a fusion feature is output. At this time, the gate attention and high-frequency forgetting double mechanisms are adopted. The input gate of the long short-term memory network receives the pressure-temperature coupling features and the frequency energy distribution features, because the two types of features respectively reflect the time delay effect of the temperature-pressure coupling and the distribution characteristics of the vibration energy, and need to be integrated through the input gate. The input gate uses a learnable weight matrix to perform linear transformation on the two types of features. The pressure-temperature coupling features are expanded into a one-dimensional vector (length 10000) according to the time step, and then are spliced with the frequency energy distribution features (length 4). Then, the input gate value is calculated by a sigmoid activation function. The closer the value is to 1, the more significant the feature is to the current time step. The forget gate of the long short-term memory network monitors the rate of change of the environmental temperature value. Since rapid changes in temperature can easily cause high-frequency vibrations and other disturbances, it is necessary to filter out invalid high-frequency features to avoid model overfitting. The forget gate receives the environmental temperature value from the data acquisition unit 1 to calculate the temperature rise rate per minute. When the temperature rise of the environmental temperature value per minute exceeds the preset temperature rise rate, the high-frequency oscillation feature discard protocol is started. In specific operation, when the forget gate value (calculated by the sigmoid function) is reduced to below 0.3, the energy proportion of the 500-1000 Hz frequency band in the frequency energy distribution features is attenuated, that is, multiplied by a coefficient of 0.5. At the same time, the high-frequency fluctuation components with a time delay of less than 1 minute in the pressure-temperature coupling features are suppressed, so that the model focuses on low-frequency stable features. Finally, the output gate of the long short-term memory network generates a two-dimensional time coding matrix. The row vectors of the two-dimensional time coding matrix correspond to the sliding time window sequence number, and the column elements of the two-dimensional time coding matrix store the coupling strength coefficients of the environmental disturbance to the flow signal. In order to quantitatively express the coupling strength of the environmental disturbance at different times, the output gate combines the cell state (stored historical feature information) and the current input feature to generate a candidate output through a tanh activation function. After being filtered by the output gate value (calculated by the sigmoid function), the two-dimensional time coding matrix is formed. The row vectors of the matrix correspond to the sliding time window sequence number (window size 10 minutes, step 5 minutes, a total of 24 rows covering 2 hours), and the column elements are divided into temperature, pressure and vibration columns according to the disturbance type, which respectively store the coupling strength coefficients (range -1 to 1, positive value indicating signal enhancement, negative value indicating signal weakening) of the corresponding environmental factors to the flow signal. The two-dimensional time coding matrix is the final fusion feature, which records the disturbance coupling law at different times, and provides a basis for the subsequent full connection layer to generate compensation coefficients. Compared with ordinary long short-term memory networks, it can better adapt to dynamic coupling relationships under complex working conditions, and provides high-quality feature input for the construction of the environmental disturbance compensation coefficient matrix.

[0028] Further, in order to comprehensively cope with the influence of the basic working condition and the sudden disturbance on the flow measurement, when the full connection layer constructs the environmental disturbance compensation coefficient matrix, the steady-state and transient dual-path compensation strategy is implemented. The steady-state path output zero drift compensation coefficient and the range slope compensation coefficient are set for the basic working condition calibration. Because in the basic working condition, that is, the working condition in which the environmental parameters are stable and have no sharp fluctuations, the metering deviation of the instrument is mainly caused by the zero point shift and the range linearity change caused by long-term use, and needs to be compensated, the zero point drift compensation coefficient is used to correct the output deviation of the instrument at zero flow, and the calculation is based on the average differential pressure deviation in the past 72 hours under the zero flow condition. The fusion features are mapped to the coefficient value by the first group of neurons of the fully connected layer, the range is-0.5 to 0.5 kPa, the positive value represents positive compensation, and the negative value represents negative compensation. The range slope compensation coefficient is used to correct the linear relationship deviation between the flow and the differential pressure, and is determined according to the ratio of the measured differential pressure to the theoretical value at 80% of the full scale, which is output by the second group of neurons of the fully connected layer, and the range is 0.9 to 1.1. Less than 1 indicates that the actual output is low and needs to be amplified, and greater than 1 needs to be attenuated. Secondly, the transient path output pressure shock suppression gain and temperature recovery rate compensation coefficient are set to cope with sudden environmental disturbances. Since sudden environmental disturbances will cause transient metering deviation, they need to be quickly compensated. The pressure shock suppression gain is for the case where the pressure change rate exceeds 0.1 MPa / s, which is output by the third group of neurons of the fully connected layer according to the high-frequency energy proportion in the vibration spectrum, and the range is 0 to 2. The greater the gain, the stronger the suppression of the instantaneous pressure shock. For example, when the pressure shock caused by pump starting is detected, the gain automatically rises to 1.5, and the fluctuation amplitude of the differential pressure signal is reduced through a smoothing algorithm. The temperature recovery rate compensation coefficient is used for the scene where the temperature change rate exceeds 2℃ / min, which is generated by the fourth group of neurons in combination with the time delay coefficient in the pressure-temperature coupling feature, and the range is 0.5 to 1.5. The coefficient increases with the temperature recovery speed. The environmental compensation coefficient matrix is stored as a two-dimensional table according to the row index compensation type and the column index sensor ID. The table element value is the normalized compensation weight. Finally, the above four types of coefficients are stored as an environmental compensation coefficient matrix in the form of a two-dimensional table. The row index is the zero drift, the range slope, the pressure shock suppression, and the temperature recovery rate, respectively. The column index is the sensor ID, which supports up to 32-way sensor parallel compensation. The table element value is mapped to the 0 to 1 interval through L2 normalization processing, ensuring that the weights of different types of coefficients are comparable. The matrix storage uses a real-time database, which is updated every 100 milliseconds. The fully connected layer dynamically adjusts the neuron weights through the back propagation algorithm, so that the matching degree of the coefficient value and the actual disturbance is improved, providing a structured compensation basis for the generation of subsequent joint compensation quantities.

[0029] Further, in order to make the joint compensation quantity dynamically adapt to different working conditions according to the deviation degree, and take into account the comprehensive influence of instrument aging and environmental disturbance, the generation of the joint compensation quantity adopts a dynamic weight distribution mechanism controlled by a deviation threshold: Firstly, the comparison threshold of the absolute value of the time deviation and 0.5% of the full scale is determined, because the time deviation reflects the actual deviation degree of the aging of the instrument, and 0.5% of the full scale is a critical value determined based on a large number of experiments, when the value is exceeded, the aging dominates the measurement deviation, and when the value is below, the influence of environmental disturbance is more significant, the full scale value is determined by the model parameters of the flow instrument, when the absolute value of the time deviation exceeds 0.5% of the full scale, it indicates that the deviation caused by the aging of the instrument has become the main error source, therefore, the time deviation is given a weight of 70% in the compensation amount, and the Euclidean norm of the environmental disturbance compensation coefficient matrix accounts for 30%, the specific calculation steps are as follows: the sign of the time deviation is reserved, a positive value indicates a positive deviation, and a negative value indicates a negative deviation, multiplied by 0.7 to obtain the aging compensation component, the Euclidean norm of the environmental disturbance compensation coefficient matrix is calculated, the Euclidean norm is the square root of the sum of the squares of all elements of the matrix, which reflects the overall strength of the environmental disturbance, multiplied by 0.3 to obtain the environmental compensation component, and the two are linearly superimposed to generate a joint compensation amount, the formula is: compensation amount = 0.7 x time deviation + 0.3 x matrix norm; When the absolute value of the time deviation is below the threshold, it indicates that the influence of environmental disturbance on measurement is greater, therefore, the weighted sum of each coefficient of the environmental disturbance compensation coefficient matrix accounts for 85%, and the time deviation is only used to adjust the matrix elements, the specific operation includes: the four types of compensation coefficients (zero drift, range slope, pressure shock suppression and temperature recovery rate) in the matrix are given fixed weights according to their importance, the weighted sum is calculated: matrix weighted sum = 0.4 x zero drift coefficient + 0.3 x range slope coefficient + 0.2 x pressure shock suppression gain + 0.1 x temperature recovery rate coefficient, the matrix weighted sum is corrected by the time deviation, if the time deviation is positive, the matrix weighted sum is multiplied by (1+0.15 x time deviation / threshold), if it is negative, it is multiplied by (1-0.15 x |time deviation| / threshold), the correction amplitude is controlled within ±15% to avoid excessive adjustment, the final joint compensation amount is the linear superposition value of the two types of compensation components, the final joint compensation amount = 0.85 x corrected matrix weighted sum + 0.15 x time deviation, the output range of the joint compensation amount remains consistent with the full scale of the instrument, after the calculation is completed, it is sent to the correction module of the flow instrument in real time, used to adjust the measurement value of the instrument, updated every 100 milliseconds to ensure the real-time of the compensation, effectively ensuring the accuracy of the flow measurement under complex working conditions.

[0030] Further, in order to continuously optimize the model parameters to adapt to the changes of working conditions and ensure the long-term stability of the calibration accuracy, the closed-loop verification unit 3 contains a double-channel back propagation and a forced triggering rule, which is as follows: Firstly, the neural network optimization channel is set, since the convolution kernel weight parameter directly affects the feature extraction accuracy, it needs to be dynamically adjusted through calibration error feedback, the root mean square of error after each calibration is taken as the loss function, the root mean square error calculation method is that 100 groups of flow values collected in the calibration process are taken, the flow values include measured values and standard values, the sum of squares of errors of each group is calculated, the average value is taken and the square root is taken, the random gradient descent method is used to update the parameters: 20% of the error samples are randomly selected for each iteration, the partial derivative of the loss function to the convolution kernel weight is calculated, the weight is adjusted by the learning rate 0.001, the new weight value=old weight value-learning rate x partial derivative, and the iteration is stopped when the root mean square error is less than 1 / 3 of the instrument accuracy level, secondly, the life model optimization channel is set, the life model optimization channel uses the time sequence deviation cumulative distribution, because the segmented parameter group of the Weibull distribution life model will change with the aging characteristics of the instrument, it needs to be dynamically calibrated based on the time sequence deviation, the time sequence deviation cumulative distribution of the past 72 hours is used, that is, the frequency of occurrence is counted according to the deviation value interval, the segmented parameter group of the Weibull function is adjusted through the particle swarm algorithm, 50 particles are initialized, each particle represents a group of parameters, the Bhattacharyya distance between the cumulative distribution and the theoretical distribution is taken as the fitness function to update the position, and the particle parameter with the highest fitness after 30 iterations is taken as the optimal solution, the segmented parameter group of the Weibull function is adjusted through the particle swarm algorithm, when the calibration error exceeds the accuracy level requirement for three times in a row or the cumulative running time reaches 100 hours, the double-channel optimization program is forcibly started, after triggering, the neural network optimization channel is iterated for 100 times first, and then the life model optimization channel is iterated for 50 times, the parameter is updated cooperatively, the precision decay problem caused by the fixed parameters of the traditional calibration system is effectively solved, and a reliable guarantee is provided for the long-term stable operation of the flow instrument.

[0031] The data acquisition unit 1 of the present application obtains the differential pressure value, the environment temperature, the pipeline pressure, the vibration spectrum and the cumulative working time of the sensor, the aging prediction model is constructed based on the Weibull distribution life model, the aging real-time value and the time sequence deviation are obtained by combining the temperature-pressure coupling equation and the vibration compensation mechanism, the multi-parameter coupling features are extracted through the neural network, the environment disturbance compensation coefficient matrix is constructed, the joint compensation quantity is generated through dynamic weight distribution, and the model parameters are optimized by the closed-loop verification unit 3, so that the problems of insufficient multi-source disturbance coupling analysis and low calibration accuracy in the prior art are solved, the accurate calibration of the flow instrument under complex working conditions is realized, and the measurement stability is improved.

[0032] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent flow meter calibration system based on multi-sensor fusion, characterized in that: include: The data acquisition unit (1) obtains the differential pressure value, ambient temperature value, pipeline pressure value, vibration spectrum value and accumulated working time of the flow meter through the sensor; The deep coupling compensation unit (2) takes the accumulated working time of the sensor as input, and constructs an aging prediction model based on the Weibull distribution life model and outputs an aging prediction value. The ambient temperature value, pipeline pressure value and vibration spectrum value are input into the temperature-pressure coupling equation to calculate the theoretical differential pressure value. The real-time aging value is obtained according to the differential pressure value and the theoretical differential pressure value. The timing deviation of the real-time aging value and the aging prediction value is calculated to obtain the timing deviation. The convolution kernel and Fourier transform layer in the embedded neural network are used to extract the pressure-temperature coupling characteristics obtained from the ambient temperature value and pipeline pressure value and the frequency domain energy distribution characteristics of the vibration spectrum value respectively. The pressure-temperature coupling characteristics and the frequency domain energy distribution characteristics are input into the long short-term memory network, and the fusion characteristics are output. The fusion characteristics are mapped into multi-dimensional compensation coefficients through the fully connected layer, and the environmental disturbance compensation coefficient matrix is ​​constructed. Finally, the timing deviation and the environmental disturbance compensation coefficient matrix are weighted fused to generate a joint compensation amount.

2. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized in that: The aging prediction model is constructed based on the Weibull distribution life model using segmented failure period modeling and dynamic parameter optimization mechanism, and the aging prediction value is output: A segmented life prediction model is established based on accelerated degradation test data, which is divided into three stages: initial failure period, stable period, and wear-out failure period. The accelerated degradation test data is obtained from historical sensor data. In the initial failure period, an exponential decay function is used to fit the drift characteristics of the running-in stage. In the stable period, a logarithmic function is used to simulate the progressive fatigue process of the material. In the wear-out failure period, a power function is used to characterize the cumulative effect of material structural damage. Finally, the gradient descent algorithm is used to dynamically update the slope coefficient and time series attenuation factor of each stage function based on the online error data collected by the closed-loop verification unit (3). The online error data is the absolute deviation between the automatic calibration flow indication and the standard value.

3. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized in that: When obtaining the real-time aging value based on the differential pressure value and the theoretical differential pressure value, the dual mechanisms of vibration no-load calibration and vibration frequency band energy compensation are implemented: Under the zero flow condition of the flow meter, the ambient temperature is controlled by the step-by-step heating mode, the pipeline pressure is adjusted by the synchronous step-by-step pressurization mode, and the differential pressure offset of multiple temperature and pressure combinations is recorded; The multivariate regression analysis method is used to generate a compensation polynomial equation containing a temperature linear term, a pressure linear term, and a temperature-pressure cross term. For vibration interference, the sliding average of the vibration energy within the preset frequency band is extracted, multiplied by the vibration sensitivity calibration coefficient, and then added to the theoretical differential pressure value. The final aging real-time value is the algebraic difference between the original differential pressure value and the theoretical differential pressure value after vibration compensation. The vibration sensitivity calibration coefficient is obtained by the vibration no-load calibration.

4. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized in that: The time series deviation calculation introduces a two-layer fusion architecture of first-order instantaneous deviation and second-order trend deviation: The first-order instantaneous deviation is obtained by subtracting the aging prediction value from the current real-time aging value. The second-order trend deviation is obtained by calculating the moving average of the hourly change rate based on the first-order instantaneous deviation sequence over the past 24 hours. The bias weight is configured according to the sensor life cycle stage. The initial failure period is assigned a second-order trend bias weight, and the stable period is assigned a first-order instantaneous bias weight. The weighted first-order instantaneous bias and the second-order trend bias are linearly superimposed to generate the time series bias.

5. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized in that: When the convolution kernel extracts the pressure-temperature coupling features obtained by the ambient temperature value and the pipeline pressure value, a dual-channel cross-correlation operation architecture is adopted: Channel 1 is configured with a one-dimensional convolutional layer with a preset sliding window size to extract the minute-by-minute gradient change characteristics of the ambient temperature value, and channel 2 is configured with a dilated convolutional layer with a preset expansion rate to capture the periodic fluctuation mode of the pipeline pressure value. The minute-by-minute gradient change characteristics of the ambient temperature value and the periodic fluctuation mode of the pipeline pressure value are cross-correlated element by element to generate a coefficient matrix reflecting the pressure-temperature coupling time-lag effect. The matrix element values ​​represent the response delay time of the pipeline pressure value caused by a unit ambient temperature value. The coefficient matrix of the pressure-temperature coupling time-lag effect is used as the pressure-temperature coupling feature.

6. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized by: When the Fourier transform layer extracts the frequency domain energy distribution characteristics, it performs a one-third octave band and energy entropy fusion analysis: The vibration spectrum values ​​are divided into one-third octave bands, the energy proportion values ​​of the three central frequency bands are extracted, and the information entropy value of the full-band vibration energy is calculated to characterize the randomness intensity of the energy distribution. The energy proportion values ​​of the three central frequency bands and the information entropy value are connected in series to form a four-dimensional vibration eigenvector, which is used as the frequency domain energy distribution feature.

7. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized by: The pressure-temperature coupling feature and the frequency domain energy distribution feature are input into the long short-term memory network, and when the fusion feature is output, a dual mechanism of gated attention and high-frequency forgetting is adopted: The input gate of the long short-term memory network receives the pressure-temperature coupling characteristics and frequency domain energy distribution characteristics. The forget gate of the long short-term memory network monitors the rate of change of the ambient temperature value. When the temperature rise of the ambient temperature value per minute exceeds the preset temperature rise rate, the high-frequency oscillation feature discarding protocol is initiated. The output gate of the long short-term memory network generates a two-dimensional time coding matrix. The row vectors of the two-dimensional time coding matrix correspond to the sliding time window sequence number. The column elements of the two-dimensional time coding matrix store the coupling strength coefficient of the environmental disturbance to the flow signal. The two-dimensional time coding matrix is ​​the final fusion feature.

8. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1 is characterized in that: When the fully connected layer constructs the environmental disturbance compensation coefficient matrix, it implements a steady-state and transient dual-path compensation strategy: The steady-state path outputs the zero drift compensation coefficient and the range slope compensation coefficient for basic working condition calibration. The transient path outputs the pressure shock suppression gain and the temperature recovery rate compensation coefficient to cope with sudden environmental disturbances. The environmental compensation coefficient matrix is ​​stored in a two-dimensional table with row index compensation type and column index sensor ID. The table element value is the normalized compensation weight.

9. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 1, characterized in that: The generation of the joint compensation amount adopts a dynamic weight allocation mechanism controlled by a deviation threshold: When the absolute value of the timing deviation exceeds 0.5% of the full scale, the timing deviation accounts for 70% of the compensation weight, and the Euclidean norm of the compensation matrix accounts for 30%; When the absolute value of the timing deviation is lower than the threshold, the weighted sum of the coefficients of the compensation matrix accounts for 85%, and the timing deviation is only used to adjust the matrix elements. The final joint compensation amount is the linear superposition value of the two types of compensation components.

10. The intelligent flow meter calibration system based on multi-sensor fusion according to claim 2, characterized in that: The closed-loop verification unit (3) includes dual-channel back-propagation and forced triggering rules, which are as follows: The neural network optimization channel uses the root mean square of the calibration error as the loss function and adopts the stochastic gradient descent method to update the convolution kernel weight parameters. The life model optimization channel uses the cumulative distribution of timing deviations and adjusts the Weibull function segmented parameter group through the particle swarm algorithm. When the calibration error exceeds the accuracy level requirement for three consecutive times or the cumulative running time reaches 100 hours, the dual-channel optimization program is forcibly started.

Images