Non-contact pipeline fluid flow velocity measurement method based on neural network model

Abstract

The invention relates to the technical field of pipeline flow velocity measurement, and provides a non-contact pipeline fluid flow velocity measurement method based on a neural network model, and the method comprises the steps: collecting signal data sets at different pipeline flow velocities through a piezoelectric film sensor; performing two-dimensional Fourier transform on the signal data set, and converting the signal into a two-dimensional energy spectrogram of a frequency-wavenumber domain; the two-dimensional energy spectrogram predicts the flow velocity of the fluid through a two-dimensional neural network model. The collected signals are subjected to two-dimensional Fourier transform and converted into the two-dimensional energy spectrogram of the frequency-wavenumber domain, the signal processing capacity is effectively improved, more accurate and meaningful input data are provided for the two-dimensional neural network model, the two-dimensional neural network model is adopted for flow velocity prediction, the two-dimensional energy spectrogram serves as input, and the flow velocity prediction accuracy is improved. And features are automatically extracted and the flow velocity is predicted through a deep learning technology. According to the method, a traditional signal algorithm for processing the workpiece problem in the complex turbulence environment through the flow speed is overcome, and the measurement precision and robustness are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline flow velocity measurement, and particularly to a non-contact pipeline fluid flow velocity measurement method based on a neural network model. Background Art

[0002] In the field of pipeline flow velocity measurement, accurate measurement of the fluid flow velocity in a pipeline is crucial. Traditional measurement methods often use non-contact piezoelectric film sensors and combine array signal processing algorithms such as MUSIC for calculation. The principle of this method is to use the piezoelectric film sensor to sense the fluid pressure pulsation signal in the pipeline, and then analyze these signals through the array signal processing algorithm to deduce the flow velocity.

[0003] However, in actual application scenarios such as dredgers, this measurement method faces many challenges. When a dredger is working, there is a large vibration, which will seriously interfere with the signals obtained by the sensor. The medium in the pipeline is mud, which is very different from clear water. The high concentration, multi-components and complex rheological properties of mud lead to completely different effects on the signal propagation and attenuation compared with clear water. In this case, there are large errors when using the array signal processing algorithm. Due to vibration interference and the characteristics of the mud medium, the noise in the signal increases significantly, and the signal characteristics change, making it difficult for the algorithm to accurately identify and analyze the effective information related to the flow velocity, and even unable to effectively identify turbulence. This directly leads to inaccurate measurement results, making it difficult to meet the strict requirements for accuracy in fluid flow velocity measurement in actual projects such as dredgers, and thus affecting the efficiency and quality of engineering operations. Therefore, it is of great practical significance to develop a new method that can adapt to the complex environment of dredgers and accurately measure the fluid flow velocity in the pipeline. Summary of the Invention

[0004] The present invention provides a non-contact pipeline fluid flow velocity measurement method based on a neural network model, which uses a two-dimensional neural network model for flow velocity prediction, uses a two-dimensional energy spectrum diagram as the input, and automatically extracts features and predicts the flow velocity through deep learning technology. This method overcomes the signal algorithm for traditional flow velocity processing of workpieces in complex turbulent environments, improves the measurement accuracy and robustness, solves the problems of large errors and inability to identify turbulence in existing measurement methods in complex environments, and improves the accuracy of pipeline fluid flow velocity measurement.

[0005] To achieve the above object, the present invention provides a non-contact pipeline fluid flow velocity measurement method based on a neural network model, and the measurement method includes:

[0006] (1) Using a piezoelectric film sensor to collect a signal data set at different pipeline flow velocities;

[0007] (2) Performing filtering processing and two-dimensional Fourier transform on the signal data set to convert the signal into a two-dimensional energy spectrum diagram in the frequency-wavenumber domain;

[0008] (3) The two-dimensional energy spectrum diagram predicts the fluid flow velocity through a two-dimensional neural network model.

[0009] Further, the specific steps of step (1) are as follows:

[0010] Assume that there are N flexible film strips in the piezoelectric film sensor installed in the pipeline. Then the signal collected by the nth flexible film strip at time t is x n (t). Then the data set X i is expressed as:

[0011] X i = {x n (t) n = 1, 2,..., N; t = 0, 1 / f s , 2 / f s ,...,(f s - 1) / f s};

[0012] Among them, f s is the sampling frequency.

[0013] Among them, the N flexible film strips of the piezoelectric film sensor are evenly distributed on the outer circumference of the pipeline. The distance between adjacent flexible film strips is set according to the pipeline diameter, fluid flow velocity range and signal propagation attenuation characteristics to ensure the spatial resolution of the collected signal.

[0014] Considering the physical characteristics of the turbulence in the pipeline, the frequency and wavenumber ranges are intercepted.

[0015] Further, the two-dimensional neural network model includes multiple convolutional layers, pooling layers and fully connected layers; the convolutional layer performs a convolution operation on the two-dimensional energy spectrum diagram E(f, k) through the convolution kernel W ij The pooling layer uses max pooling or average pooling to reduce the dimension of the feature map; the fully connected layer unfolds and connects the pooled feature maps and finally outputs the predicted flow velocity.

[0016] Further, the construction process of the two-dimensional neural network model is as follows:

[0017] 1) By adjusting the mud pump speed, use the piezoelectric film sensor to collect the signals x min collected at the minimum pipeline flow velocity v max and the maximum pipeline flow velocity v n . Taking the experimental platform as an example, v min = 2 m / s, v max = 7 m / s, and using an electromagnetic flowmeter as the calibration object, label the corresponding flow velocity for the collected signals;

[0018] 2) For the collected signals x n(t) is subjected to filtering. Assume that the impulse response of the adopted filter is h(n), and the filtered signal y n (t) is:

[0019]

[0020] where M is the length of the filter; among them, the filtering frequency range is preset according to the signal frequency characteristics corresponding to the fluid flow velocity, and the frequency components related to turbulence are retained.

[0021] 3) Perform two-dimensional Fourier transform on the filtered signal y n (t) to obtain the two-dimensional energy spectrum diagram E(f,k) = |F y (f,k)|^2, where F y (f,k) is the result of the two-dimensional Fourier transform of the filtered signal y n (t);

[0022] 4) Construct a two-dimensional neural network model with the two-dimensional energy spectrum diagram E(f,k) as the input and the flow velocity v as the output: The two-dimensional neural network model includes multiple convolutional layers, pooling layers and fully connected layers; the convolutional layer performs a convolutional operation on the two-dimensional energy spectrum diagram E(f,k) through the convolutional kernel W ij The pooling layer uses max pooling or average pooling to reduce the dimension of the feature map; the fully connected layer unfolds and connects the pooled feature maps, and finally outputs the predicted flow velocity;

[0023] Among them, the convolutional operation formula is:

[0024]

[0025] Among them, is the input feature map of the (l - 1)-th layer, is the result after convolution of the l-th layer, is the bias; M represents the size of the convolutional kernel in the height (vertical) direction (i.e., the number of rows of the convolutional kernel), and N represents the size of the convolutional kernel in the width (horizontal) direction;

[0026] 5) Reasonably set the size, number of layers, and stride of the convolutional kernel of the model according to the actual situation and optimization; for example, the size of the convolutional kernel can be set to 3×3 or 5×5, the stride can be set to 1 or 2, and the number of layers is selected between 3 and 10 according to the model complexity and performance requirements.

[0027] 6) Use the dataset with flow velocity labels to train the constructed two-dimensional neural network model, and use the loss function to measure the difference between the predicted flow velocity and the true flow velocity; for example, the mean square error loss function L:

[0028]

[0029] Among them, P represents the number of training samples, is the predicted value of the sample, v i is the actual value of the sample;

[0030] 7) Compare the flow velocity results output by the model with the measurement results of the electromagnetic flowmeter, calculate the error, and evaluate the accuracy of the model;

[0031] 8) Continuously adjust the convolution kernel size, number of layers, and stride of the two-dimensional neural network model through the backpropagation algorithm to make the model achieve better performance. The optimizer combined with the backpropagation algorithm is the Adam optimizer or the Stochastic Gradient Descent (SGD) optimizer, and the convolution kernel parameters and the weights of the fully connected layer are updated iteratively.

[0032] The present invention has the following beneficial effects:

[0033] (1) The present invention realizes non-contact measurement of the flow velocity of pipeline fluid through a piezoelectric film sensor, avoiding problems such as mechanical loss and resistance influence that may be brought by traditional contact measurement methods; and can work stably in a high-vibration environment, and is particularly suitable for scenarios with increased vibration such as dredgers.

[0034] (2) The present invention performs two-dimensional Fourier transform on the collected signals, converts them into two-dimensional energy spectrograms in the frequency-wavenumber domain, and arranges the signals based on the physical characteristics of turbulence to extract effective features; effectively improves the signal processing ability and provides more accurate and meaningful input data for the two-dimensional neural network model;

[0035] (3) The present invention uses a two-dimensional neural network model for flow velocity prediction, uses the two-dimensional energy spectrogram as input, and automatically extracts features and predicts the flow velocity through deep learning technology; overcomes the signal algorithms of traditional flow velocity processing for workpieces in complex turbulent environments, and improves the measurement accuracy and robustness. Description of the Drawings

[0036] Figure 1 is the measurement flow chart of the fluid flow velocity in the present invention.

[0037] Figure 2 is the construction flow chart of the two-dimensional neural network model in the present invention. Detailed Embodiments

[0038] The technical solutions of the present invention are further described in detail below in conjunction with specific embodiments, but these embodiments do not limit the present invention. Any similar structures and their similar changes adopted in the present invention shall be included in the protection scope of the present invention. The commas in the present invention all represent the relationship of "and", and the English letters in the present invention are case-sensitive.

[0039] Such asFigure 1 As shown in the figure, the present invention provides a non-contact pipeline fluid flow velocity measurement method based on a neural network model, and the measurement method includes:

[0040] S1. Using a piezoelectric thin film sensor to collect a signal data set under different pipeline flow velocities; the specific steps are as follows:

[0041] Assume that there are N flexible film strips in the piezoelectric thin film sensor installed in the pipeline, then the signal collected by the nth flexible film strip at time t is x n (t), then the data set X i is expressed as:

[0042] X i ={x n (t)|n = 1, 2, …, N; t = 0, 1 / f s , 2 / f s , …, (f s -1) / f s};

[0043] where f s is the sampling frequency.

[0044] Among them, the N flexible film strips of the piezoelectric thin film sensor are evenly distributed on the outer circumference of the pipeline, and the distance between adjacent flexible film strips is set according to the pipeline diameter, fluid flow velocity range and signal propagation attenuation characteristics to ensure the spatial resolution of the collected signal.

[0045] S2. Filter and perform two-dimensional Fourier transform on the signal data set to convert the signal into a two-dimensional energy spectrum diagram in the frequency-wavenumber domain; considering the physical characteristics of turbulence in the pipeline, intercept the frequency and wavenumber ranges, and perform two-dimensional Fourier transform on the intercepted ranges.

[0046] S3. Predict the fluid flow velocity through the two-dimensional energy spectrum diagram by a two-dimensional neural network model.

[0047] The two-dimensional neural network model includes multiple convolutional layers, pooling layers and fully connected layers; the convolutional layer performs a convolution operation on the two-dimensional energy spectrum diagram E(f, k) through a convolution kernel W ij , and the pooling layer uses max pooling or average pooling to reduce the dimension of the feature map; the fully connected layer unfolds and connects the pooled feature maps, and finally outputs the predicted flow velocity.

[0048] As Figure 2 shown, the construction process of the two-dimensional neural network model is as follows:

[0049] S31. By adjusting the rotational speed of the mud pump, using the piezoelectric thin film sensor to collect the signals x collected at the minimum pipeline flow velocity v min and the maximum pipeline flow velocity v max collected atn (t), taking the experimental platform as an example v min = 2m / s, v max = 7m / s, and taking the electromagnetic flowmeter as the calibration object, attaching the corresponding flow velocity labels to the collected signals;

[0050] S32, filtering the collected signal x n (t), assuming that the impulse response of the adopted filter is h(n), the filtered signal y n (t) is:

[0051]

[0052] Among them, M is the length of the filter; among them, the filtering frequency range is preset according to the signal frequency characteristics corresponding to the fluid flow velocity, and the frequency components related to turbulence are retained.

[0053] S33, performing two-dimensional Fourier transform on the filtered signal y n (t) to obtain the two-dimensional energy spectrum diagram E(f,k) = |F y (f,k)|^2, where F y (f,k) is the two-dimensional Fourier transform result of the filtered signal y n (t);

[0054] S34, constructing a two-dimensional neural network model with the two-dimensional energy spectrum diagram E(f,k) as the input and the flow velocity v as the output: The two-dimensional neural network model includes multiple convolutional layers, pooling layers and fully connected layers; the convolutional layer performs convolutional operations on the two-dimensional energy spectrum diagram E(f,k) through the convolutional kernel W ij Pooling layers use max pooling or average pooling to reduce the dimension of the feature map; the fully connected layer unfolds and connects the pooled feature maps, and finally outputs the predicted flow velocity;

[0055] Among them, the convolution operation formula is:

[0056]

[0057] Among them, is the input feature map of the l-1 layer, is the result after convolution of the l-th layer, is the bias; M represents the size of the convolutional kernel in the height (vertical) direction (i.e., the number of rows of the convolutional kernel), and N represents the size of the convolutional kernel in the width (horizontal) direction

[0058] S35. Reasonably set the convolution kernel size, number of layers, and stride of the model according to the actual situation and optimization. For example, the convolution kernel size can be set to 3×3 or 5×5, the stride can be set to 1 or 2, and the number of layers is selected between 3 and 10 according to the model complexity and performance requirements.

[0059] S36. Use the dataset with flow velocity labels to train the constructed two-dimensional neural network model, and use the loss function to measure the difference between the predicted flow velocity and the actual flow velocity. For example, the mean square error loss function L:

[0060]

[0061] where P represents the number of training samples, is the predicted value of the sample, v i is the actual value of the sample;

[0062] S37. Compare the flow velocity result output by the model with the measurement result of the electromagnetic flowmeter, calculate the error, and evaluate the accuracy of the model.

[0063] S38. Continuously adjust the convolution kernel size, number of layers, and stride of the two-dimensional neural network model through the backpropagation algorithm to make the model achieve better performance. The optimizer combined with the backpropagation algorithm is the Adam optimizer or the stochastic gradient descent (SGD) optimizer, and the convolution kernel parameters and the weights of the fully connected layer are updated iteratively.

[0064] The present invention performs two-dimensional Fourier transform on the collected signal, converts it into a two-dimensional energy spectrum diagram in the frequency-wavenumber domain, organizes the signal based on the physical characteristics of turbulence, and extracts effective features; effectively improves the signal processing ability and provides more accurate and meaningful input data for the two-dimensional neural network model; uses the two-dimensional neural network model to predict the flow velocity, uses the two-dimensional energy spectrum diagram as the input, and automatically extracts features and predicts the flow velocity through deep learning technology; overcomes the signal algorithms of traditional flow velocity processing for workpieces in complex turbulent environments, and improves the measurement accuracy and robustness.

[0065] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

Claims

1. A non-contact pipeline fluid flow velocity measurement method based on a neural network model, characterized in that The measurement method includes: (1) Using a piezoelectric thin film sensor to collect a signal dataset under the pipeline flow velocity; (2) Filtering and performing two-dimensional Fourier transform on the signal dataset to convert the signal into a two-dimensional energy spectrum diagram in the frequency-wavenumber domain; (3) Predicting the fluid flow velocity through the two-dimensional energy spectrum diagram by a two-dimensional neural network model.

2. The non-contact pipeline fluid flow velocity measurement method based on a neural network model according to claim 1, wherein The specific steps of step (1) are: Assume that the outer wall of the pipeline is seamlessly attached to the piezoelectric film sensor. The piezoelectric film sensor has N flexible film strips, and the signal collected by the nth flexible film strip at time t is x n (t), then the data set X i is expressed as: X i = {x n (t)|n = 1, 2, …, N; t = 0, 1 / f s , 2 / f s , …, (f s - 1) / f s}; where f s is the sampling frequency.

3. The non-contact pipeline fluid flow velocity measurement method based on a neural network model according to claim 1, wherein The two-dimensional neural network model includes multiple convolutional layers, pooling layers, and fully connected layers; the convolutional layer performs a convolution operation on the two-dimensional energy spectrum diagram E(f,k) through the convolution kernel W ij The pooling layer uses max pooling or average pooling to reduce the dimension of the feature map; the fully connected layer unfolds and connects the pooled feature maps, and finally outputs the predicted flow velocity.

4. A non-contact pipeline fluid flow velocity measurement method based on a neural network model according to claim 1, characterized in that, The construction process of the two-dimensional neural network model is: 1) By adjusting the rotational speed of the sludge pump, the piezoelectric film sensor is used to collect the minimum pipeline flow velocity v min and the maximum pipeline flow velocity v max of the signal x n (t) collected under such conditions, and the electromagnetic flowmeter is used as the calibration object to label the corresponding flow velocity tags for the collected signal; 2) Filter the acquired signal x n (t) to obtain the filtered signal y n (t); 3) Perform a two-dimensional Fourier transform on the filtered signal y n (t) to obtain the two-dimensional energy spectrum diagram E(f,k) = |F y (f,k)|^2 in the frequency-wavenumber domain, where F y (f,k) is the result of the two-dimensional Fourier transform of the filtered signal y n (t); 4) Construct a two-dimensional neural network model with the two-dimensional energy spectrum diagram E(f,k) as the input and the flow velocity v as the output: The two-dimensional neural network model includes multiple convolutional layers, pooling layers, and fully connected layers; the convolutional layer performs a convolution operation on the two-dimensional energy spectrum diagram E(f,k) through the convolution kernel W ij to perform a convolution operation on the two-dimensional energy spectrum diagram E(f,k). The pooling layer uses max pooling or average pooling to reduce the dimension of the feature map; the fully connected layer unfolds and connects the pooled feature maps and finally outputs the predicted flow velocity; 5) Reasonably set the convolution kernel size, number of layers, and stride of the model according to the actual situation and optimization; 6) Use the dataset with flow velocity labels to train the constructed two-dimensional neural network model, and use a loss function to measure the difference between the predicted flow velocity and the true flow velocity; 7) Compare the flow velocity result output by the model with the measurement result of the electromagnetic flowmeter, calculate the error, and evaluate the accuracy of the model; 8) Continuously adjust the convolution kernel size, number of layers, and stride of the two-dimensional neural network model through the backpropagation algorithm to make the model achieve better performance.

Images