Method for identifying flow pattern of mixed flow gas-liquid mixed delivery pump based on pressure pulsation signal

Abstract

This invention discloses a flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals. The method includes: collecting pressure pulsation signals from the pump under different gas contents and flow rates, and simultaneously recording flow pattern labels to form an original signal sample set; preprocessing the pressure pulsation signals and performing mode decomposition using variational mode decomposition to obtain eight intrinsic mode components; performing Hilbert-Huang transform on each intrinsic mode component to generate a two-dimensional time-frequency image; constructing and training a signal-image dual-channel deep neural network; and deploying the trained model on an online monitoring system for real-time flow pattern identification. This invention's flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals solves the problems of existing technologies, such as single features relying on manual design, feature distortion due to mode aliasing, and poor flow pattern identification accuracy and real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of multiphase flow mechanical technology, specifically relating to a flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals. Background Technology

[0002] In the gas-liquid two-phase flow transportation processes in industries such as oil and natural gas, the internal flow pattern of the mixed-transfer pump is a key factor affecting pump efficiency, operational stability, and equipment safety. Different flow patterns result in significant differences in pump characteristics such as pressure pulsation, noise, and efficiency. Therefore, accurate and real-time identification of the internal flow pattern of the mixed-transfer pump is of great significance for achieving optimized pump control and fault early warning.

[0003] Traditional flow pattern identification methods rely on manually designed features and a single signal processing path, combined with conventional machine learning models for discrimination. This approach has many inherent limitations and is difficult to adapt to actual industrial needs. These methods rely on manual extraction of shallow signal features, failing to fully explore the deep features of unsteady two-phase flows, and are prone to insufficient feature extraction. Furthermore, the single signal processing mode is susceptible to field interference leading to modal aliasing, resulting in feature distortion. Moreover, the models are mostly trained under fixed operating conditions, exhibiting weak generalization and transfer capabilities, and their real-time performance is insufficient for dynamic monitoring requirements. In summary, traditional identification methods cannot support the efficient and stable operation of mixed-flow pumps. Therefore, developing novel intelligent flow pattern identification technologies adapted to industrial scenarios has become an urgent need in the field of multiphase flow machinery. Summary of the Invention

[0004] The purpose of this invention is to provide a flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals, which solves the problems of single features and reliance on manual design, feature distortion caused by modal aliasing, and poor flow pattern identification accuracy and real-time performance in the prior art.

[0005] The technical solution adopted in this invention is a flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals, comprising the following steps:

[0006] Step 1: Collect pressure pulsation signals of the mixed-transport pump under different gas content and flow rate conditions, and simultaneously record the flow pattern label to form the original signal sample set; Step 2: Preprocess the pressure pulsation signal and perform mode decomposition using variational mode decomposition to obtain eight intrinsic mode components; Step 3: Perform Hilbert-Huang transform on each intrinsic mode component to generate a two-dimensional time-frequency image; Step 4: Construct and train a signal-image dual-channel deep neural network; Step 5: Deploy the trained model to the online monitoring system to identify flow patterns in real time.

[0007] The invention is further characterized by: Step 1 specifically involves: deploying a high-frequency dynamic pressure sensor at the guide vane flow channel of the gas-liquid mixing pump to collect pressure pulsation signals under different gas contents and flow rates; simultaneously, using a high-speed camera to synchronously collect flow pattern images, manually labeling the collected flow pattern images, and associating the pressure pulsation signals corresponding to the labeled flow pattern labels with the time intervals according to the time synchronization relationship; the flow pattern labels are categorized into four types: bubbly flow, condensed bubbly flow, gas bladder flow, and separated flow.

[0008] Step 2 specifically includes the following steps: Step 2.1: The collected pressure pulsation signal is processed using wavelet threshold denoising method to eliminate the sensor's own electronic noise and high-frequency interference caused by factors other than flow pulsation; Step 2.2: Perform Z-score normalization on the denoised signal, as shown below:

[0009] In the formula, This represents a standardized, denoised signal. This represents the denoised signal, μ represents the mean of the signal segment, and σ is the standard deviation. Step 2.3: Set the total number of modes to be decomposed to K=8, and the quadratic penalty factor. Convergence tolerance and the maximum number of iterations Initialize each modal function Set the current iteration count n=1; Step 2.4: For the eight modes k=1,2,…,8, update the k-th mode function in parallel in the frequency domain. The following calculation formula applies:

[0010] In the formula, For frequency variables, For the center frequency, For standardized denoised signals Fourier transform; For Lagrange multipliers; This represents the k-th modal function after the (n+1)-th iteration. This represents the i-th mode function after the nth iteration; Synchronous update of Lagrange multipliers in the frequency domain , means as follows:

[0011] In the formula, This is the noise tolerance parameter; , Let these represent the Lagrange multipliers after the nth and (n+1)th iterations, respectively; Step 2.5: Calculate the residual squared error of the sum of the normalized denoised signal and all reconstructed modes within the current iteration period. , is represented as: In the formula, This means that after each iteration, the inverse Fourier transform of all modes is performed to obtain the time-domain mode function; Convergence criteria: like or n≥ If n=n+1, then terminate the iteration and proceed to step 2.6; otherwise, let n=n+1 and return to step 2.4 to continue the loop iteration. Step 2.6: After the iteration terminates, analyze the modal function obtained from the last update. Performing an inverse Fourier transform yields the final eight eigenmode components. , k= 1, 2, ..., 8,

[0012] In the formula, Indicates the inverse Fourier transform; The final components are sorted from high to low according to their center frequencies. The highest frequency component, This is the lowest frequency component.

[0013] Step 3 specifically includes the following steps: Step 3.1: For the k-th intrinsic mode component Calculate the Hilbert transform , is represented as:

[0014] In the formula, Represents the Cauchy principal value integral; Step 3.2: Construct the k-th intrinsic mode component Analyzed signal , means as follows:

[0015] In the formula, The imaginary unit, Instantaneous amplitude, It is the instantaneous phase; Step 3.3: Based on the analytical signal Calculate the instantaneous amplitude of the k-th intrinsic mode component. With instantaneous phase , means as follows:

[0016]

[0017] For instantaneous phase The instantaneous frequency is obtained by differentiation. , means as follows:

[0018] Step 3.4: Based on instantaneous frequency With instantaneous amplitude Define the Hilbert spectrum of the k-th eigenmode component. The Hilbert spectrum characterizes the time-frequency distribution density of a signal; Step 3.5: Fuse the Hilbert spectra of all components to obtain a two-dimensional time-frequency image I.

[0019] Step 3.5 specifically includes: Step 3.5.1: Define a discrete two-dimensional time-frequency grid, where the time dimension corresponds to the signal sampling points and the frequency dimension corresponds to the effective range of the instantaneous frequencies of all intrinsic mode components; Step 3.5.2: Calculate the Hilbert spectrum of each intrinsic mode component. The instantaneous energy density is accumulated in the corresponding unit of the time-frequency network to form a comprehensive time-frequency energy matrix. , is represented as:

[0020] Step 3.5.3: Combine the time-frequency energy matrix To generate an image format suitable for convolutional neural network input, standardization is first performed, starting with energy normalization, represented as:

[0021] In the formula, This represents the normalized integrated time-frequency energy matrix. This indicates the smallest positive number that should not be taken as the logarithm of 0; Secondly, the normalized integrated time-frequency energy matrix is ​​linearly mapped onto the RGB three channels to generate a two-dimensional time-frequency image I.

[0022] The signal-image dual-channel deep neural network constructed in step 4 includes a signal channel, an image channel, a feature fusion module, and an output layer; The input to the signal channel is the intrinsic mode component. The intrinsic mode component passes sequentially through two first stacked module groups connected by residuals, a global average pooling layer, a fully connected layer, and then through Batch Normalization, Relu6, and Dropout before being fused with the output features of the image channel. The input to the image channel is a two-dimensional time-frequency image. The two-dimensional time-frequency image passes through a convolutional layer, BatchNormalization, ReLU6, a second stacked module group, a two-dimensional average pooling layer, and a fully connected layer in sequence. After passing through BatchNormalization, ReLU6, and Dropout, it is fused with the output features of the signal channel. The feature fusion module first connects the features output from the signal channel and the image channel, and then fuses them through an attention fusion mechanism. The attention fusion mechanism weights and fuses the features learned from the two channels to provide weighted fused features for the output layer to identify the manifold type. The output layer is a fully connected layer and a Softmax activation function, which outputs the probability of the manifold type, and the class with the highest probability is taken as the recognition result.

[0023] Each first stacked module group of the signal channel contains three first stacked modules; the first stacked module consists of a one-dimensional pooling layer, a one-dimensional convolutional layer, and a connection layer. The workflow of each first stacked module is as follows: After the intrinsic mode components are input, they are learned in two parts. One part learns micro-detail features by passing through a one-dimensional pooling layer with a 1×3 kernel and a one-dimensional convolutional layer with a 1×1 kernel. The other part learns macro-features by passing through a 1×1 convolutional layer and then passing through three one-dimensional convolutions with kernels of 1×3, 1×5, and 1×7 in parallel. The outputs of the two branches are connected in the feature channel and then passed through BatchNormalization, ReLU6, and Dropout to obtain the signal channel output features.

[0024] The second stacked module group includes three second stacked modules, each consisting of different two-dimensional convolutional layers with residual connections. The workflow of each second stacked module is as follows: First, it goes through a 1×1 two-dimensional convolutional layer to increase the dimensionality. After the dimensionality increase, it goes through Batch Normalization and ReLU6, and then performs multi-scale learning in parallel through a 3×3 two-dimensional convolutional layer and a 5×5 two-dimensional convolutional layer. The outputs of the two branches go through a 1×1 one-dimensional convolutional layer for dimensionality reduction, and then through Batch Normalization to output the corresponding features.

[0025] The beneficial effects of this invention are: This invention presents a flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals. By constructing a signal-image dual-channel deep learning fusion architecture, it achieves high-precision and high-efficiency intelligent identification of the gas-liquid two-phase flow pattern within the pump. This invention employs variational mode decomposition, which effectively avoids mode aliasing compared to traditional methods. It also combines Hilbert-Huang transform to extract features from both time series and time-frequency image dimensions, utilizing dual-channel deep learning to achieve feature fusion, overcoming the problems of single features and reliance on manual design in traditional methods. Furthermore, the identification model of this invention can be deployed in online systems for real-time flow pattern identification, while possessing good structural flexibility to adapt to different operating conditions, providing reliable technical support for intelligent monitoring and optimized control of mixed-flow pumps. Attached Figure Description

[0026] Figure 1 This is a flowchart of the flow pattern identification method for a mixed-flow gas-liquid pump based on pressure pulsation signals according to the present invention. Figure 2 This is a structural diagram of the first stacked module in the signal-image dual-channel deep neural network constructed in this invention; Figure 3 This is a structural diagram of the second stacked module in the signal-image dual-channel deep neural network constructed in this invention; Figure 4 This is a structural diagram of the attention mechanism in the signal-image dual-channel deep neural network constructed in this invention; Figure 5 This is a layout diagram for pressure pulsation signal acquisition according to an embodiment of the present invention; Figure 6 These are example diagrams of four flow patterns captured by a high-speed camera in an embodiment of the present invention; Figure 7 This is a diagram showing the results of obtaining eight intrinsic mode components using variational mode decomposition in an embodiment of the present invention. Figure 8 This is a two-dimensional time-frequency diagram of the intrinsic mode components after Hilbert-Huang transform according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the manifold recognition results according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0028] Example 1 This embodiment provides a flow pattern identification method for mixed-flow gas-liquid pumps based on pressure pulsation signals, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect pressure pulsation signals of the mixed-transport pump under different gas content and flow rate conditions, and simultaneously record the flow pattern label to form the original signal sample set; Step 2: Preprocess the pressure pulsation signal and perform mode decomposition using Variational Mode Decomposition (VMD) to obtain eight intrinsic mode components; Step 3: Perform Hilbert-Huang transform on each intrinsic mode component to generate a two-dimensional time-frequency image that can characterize the time-frequency energy distribution of the sample; Step 4: Construct and train a signal-image dual-channel deep neural network; Step 5: Deploy the trained model to the online monitoring system to identify flow patterns in real time.

[0029] Example 2 Step 1 specifically involves: deploying a high-frequency dynamic pressure sensor at the guide vane flow channel of the gas-liquid mixing pump to collect pressure pulsation signals under different gas contents and flow conditions; simultaneously, using a high-speed camera to synchronously collect flow pattern images, manually labeling the collected flow pattern images, and associating the pressure pulsation signals corresponding to the time intervals of the labeled flow pattern labels with labels based on the time synchronization relationship. When annotating the images, based on the classic definition of gas-liquid two-phase flow, the specific flow patterns are labeled as four types: bubbly flow, condensed bubbly flow, gas pocket flow, and separated flow.

[0030] Example 3 The pressure pulsation signal acquired in step 1 will be further processed, specifically including: Step 2.1: The collected pressure pulsation signal is processed using wavelet threshold denoising method to eliminate the sensor's own electronic noise and high-frequency interference caused by factors other than flow pulsation; Step 2.2: Perform Z-score normalization on the denoised signal, as shown below:

[0031] In the formula, This represents a standardized, denoised signal. This represents the denoised signal, μ represents the mean of the signal segment, and σ is the standard deviation. Step 2.3: Set the total number of modes to be decomposed to K=8, and the quadratic penalty factor. Convergence tolerance and the maximum number of iterations Initialize each modal function Set the current iteration count n=1; Step 2.4: For the eight modes k=1,2,…,8, update the k-th mode function in parallel in the frequency domain. The following calculation formula applies:

[0032] In the formula, For frequency variables, For the center frequency, For standardized denoised signals Fourier transform; For Lagrange multipliers; This represents the frequency domain form of the k-th modal component in the (n+1)-th iteration. It represents the sum of the Fourier transforms of all modes except the k-th mode; Synchronous update of Lagrange multipliers in the frequency domain , means as follows:

[0033] In the formula, This is the noise tolerance parameter; , Let these represent the Lagrange multipliers after the nth and (n+1)th iterations, respectively; Step 2.5: Calculate the residual squared error of the sum of the normalized denoised signal and all reconstructed modes within the current iteration period. , is represented as: In the formula, This means that after each iteration, the inverse Fourier transform of all modes is performed to obtain the time-domain mode function; Convergence criteria: like or n≥ If n=n+1, then terminate the iteration and proceed to step 2.6; otherwise, let n=n+1 and return to step 2.4 to continue the loop iteration. Step 2.6: After the iteration terminates, analyze the modal function obtained from the last update. Performing an inverse Fourier transform yields the final eight eigenmode components. , k= 1, 2, ..., 8,

[0034] In the formula, Indicates the inverse Fourier transform; The final components are sorted from high to low according to their center frequencies. The highest frequency component, This is the lowest frequency component.

[0035] Example 4 The eight intrinsic mode components obtained in step 2 are subjected to Hilbert-Huang transform to generate a two-dimensional time-frequency image, specifically: Step 3.1: For the k-th intrinsic mode component Calculate the Hilbert transform , is represented as:

[0036] In the formula, Represents the Cauchy principal value integral; Step 3.2: Construct the k-th intrinsic mode component Analyzed signal , means as follows:

[0037] In the formula, The imaginary unit, Instantaneous amplitude, It is the instantaneous phase; Step 3.3: Based on the analyzed signal Calculate the instantaneous amplitude of the k-th intrinsic mode component. With instantaneous phase , means as follows:

[0038]

[0039] For instantaneous phase The instantaneous frequency is obtained by differentiation. , means as follows:

[0040] Step 3.4: Based on instantaneous frequency With instantaneous amplitude Define the Hilbert spectrum of the k-th eigenmode component. The Hilbert spectrum characterizes the signal's distribution density over time and frequency. Step 3.5: Fuse the Hilbert spectra of all components to obtain a two-dimensional time-frequency image I.

[0041] Step 3.5.1: Define a discrete two-dimensional time-frequency grid, where the time dimension corresponds to the signal sampling points and the frequency dimension corresponds to the effective range of the instantaneous frequencies of all intrinsic mode components; Step 3.5.2: Calculate the Hilbert spectrum of each intrinsic mode component. The instantaneous energy density is accumulated in the corresponding unit of the time-frequency network to form a comprehensive time-frequency energy matrix. , is represented as:

[0042] Step 3.5.3: Combine the time-frequency energy matrix To generate an image format suitable for convolutional neural network input, standardization is first performed, starting with energy normalization, represented as:

[0043] In the formula, This represents the normalized integrated time-frequency energy matrix. This indicates the smallest positive number that should not be taken as the logarithm of 0; Secondly, the normalized integrated time-frequency energy matrix is ​​linearly mapped onto the RGB three channels to generate a two-dimensional time-frequency image I.

[0044] Example 5 Construct a signal-image dual-channel deep neural network, specifically including a signal channel, an image channel, a feature fusion module, and an output layer; The input to the signal channel is the intrinsic mode component. The intrinsic mode component passes sequentially through two first stacked module groups connected by residuals, a global average pooling layer, a fully connected layer, and then through Batch Normalization, Relu6, and Dropout before being fused with the output features of the image channel. Each first stacked module group of the signal channel contains three first stacked modules; the first stacked module consists of a one-dimensional pooling layer, a one-dimensional convolutional layer, and a connection layer. The structure of the first stacking module is as follows: Figure 2 As shown, the workflow of each first stacked module is as follows: After the intrinsic mode components are input, the neural network automatically learns and identifies them in two parts. One part learns microscopic details by passing through a one-dimensional pooling layer with a 1×3 kernel and a one-dimensional convolutional layer with a 1×1 kernel. The other part learns macroscopic features by passing through a 1×1 convolutional layer and then through three one-dimensional convolutions with kernels of 1×3, 1×5, and 1×7 in parallel. The outputs of the two branches are concatenated in the feature channels and then processed through Batch Normalization, ReLU6, and Dropout to normalize the fused features, perform nonlinear transformations, and prevent overfitting, thus enhancing the model's generalization ability. Finally, the signal channel output features are obtained.

[0045] The input to the image channel is a two-dimensional time-frequency image I. The two-dimensional time-frequency image I passes through a convolutional layer, BatchNormalization, ReLU6, a second stacked module group, a two-dimensional average pooling layer, and a fully connected layer in sequence. After passing through BatchNormalization, ReLU6, and Dropout, it is fused with the output features of the signal channel. The second stacked module group includes three second stacked modules, each consisting of different two-dimensional convolutional layers with residual connections. The structure of the second stack module is as follows: Figure 3As shown, the workflow of each second stacked module is as follows: First, it goes through a 1×1 two-dimensional convolutional layer to increase the dimensionality. After the dimensionality increase, it goes through Batch Normalization and ReLU6, and then performs multi-scale learning in parallel through a 3×3 two-dimensional convolutional layer and a 5×5 two-dimensional convolutional layer. The outputs of the two branches go through a 1×1 two-dimensional convolutional layer for dimensionality reduction, and then through Batch Normalization to output the corresponding features.

[0046] The feature fusion module first concatenates the features output from the signal channel and the image channel, and then fuses them using an attention fusion mechanism. This attention fusion mechanism weights and fuses the features learned from the two channels, such as... Figure 4 As shown, weighted fusion features are provided to the output layer to identify the manifold type; The output layer is a fully connected layer and a Softmax activation function, which outputs the probability of the manifold type, and the class with the highest probability is taken as the recognition result.

[0047] The trained signal-image dual-channel deep neural network is deployed to the online monitoring system of the gas-liquid mixed transport pump. The pressure pulsation signal of the mixed transport pump is collected in real time. After being processed in the same way as in the training stage, it is input into the trained signal-image dual-channel deep neural network. The highest probability category output by the network is the real-time identified flow pattern result.

[0048] Example 6 The accuracy and real-time performance of the signal-image dual-channel deep neural network for manifold recognition of the present invention were verified through the following experiments.

[0049] The test system was constructed to collect data. The entire test system mainly consists of water supply pipeline, gas supply pipeline, gas-liquid mixing pipeline, mixed flow gas-liquid mixed transport pump visualization test platform and data acquisition system.

[0050] To accurately capture transient pressure fluctuations related to flow pattern, a high-frequency dynamic pressure sensor and high-speed data acquisition equipment were employed. Two M112A11 high-frequency dynamic pressure sensors from PCB (Computers Manufacturing Corporation) were used, with a range of 0.07–345 kPa, a linearity accuracy of ±1%, and a sensitivity of 14.5 mV / kPa. These sensors exhibit excellent dynamic response characteristics, and their high-frequency response capability is sufficient to capture rapid pressure pulsations caused by flow pattern transitions within the mixed-transfer pump without distortion.

[0051] Given the complexity of the internal flow within the guide vanes of the mixed-flow gas-liquid pump and its tendency to form air pockets, two sensors, designated as measuring points P1 and P2, are placed at key flow path locations within the guide vanes to comprehensively monitor pressure field changes within the flow channel. Figure 5As shown. This dual-point arrangement helps to acquire richer spatial flow field information and enhances the robustness of subsequent feature learning. The pressure pulsation signal was synchronously acquired using a USB-6229 data acquisition board from National Instruments. This board supports a synchronous sampling rate of up to 80 ks / s with a precision of 16 bits, ensuring accurate digitization of high-frequency signals.

[0052] To obtain unambiguous manifold labels corresponding to pressure signal samples, this invention employs a high-speed camera system for synchronous visualization observation. The camera used is the X213 high-speed camera from CAS World Thousand Eyes Wolf Company, with a maximum resolution of 1280×1024 pixels. In this experiment, to balance field of view and temporal resolution, the frame rate was set to 6000 frames per second.

[0053] An observation window is installed on the transparent pump casing of the mixed-flow gas-liquid pump or on a transparent pipeline section adjacent to the pump outlet to ensure that the high-speed camera's field of view can clearly capture the stable flow pattern structure inside the pump or at the outlet. A hardware trigger circuit enables the data acquisition board and the high-speed camera to be controlled by the same trigger signal, achieving synchronous startup with microsecond-level precision, ensuring that each frame of image corresponds strictly in time to each pressure signal segment.

[0054] The experiment was conducted at room temperature using air and water as the media. The liquid and gas flow rates were systematically varied to cover a typical range of 0% to 50% volumetric gas content at the pump inlet, while the pump speed was fixed at 1500 rpm to simulate various operating conditions in real-world scenarios. At each stable, preset operating point, the pressure acquisition system and high-speed camera system were simultaneously triggered. The pressure signal sampling frequency was fixed at 10240 Hz, with each continuous acquisition lasting 5 seconds, meaning each raw signal sample contained 51200 data points. The high-speed camera synchronously recorded the manifold evolution video at a frame rate of 6000 fps.

[0055] After the experiment, professionals first reviewed the high-speed video and manually interpreted and labeled the stable flow pattern intervals in the video according to the classic definition of gas-liquid two-phase flow. Then, based on the time synchronization relationship, the labeled flow pattern labels were precisely correlated with the pressure signal segments of the corresponding time intervals. Ultimately, the flow patterns were confirmed to be categorized into bubbly flow, condensing bubbly flow, gas pocket flow, and separated flow, as shown in the examples below. Figure 6 As shown.

[0056] For the pressure pulsation signals collected above , After preprocessing, variational mode decomposition (VMD) was used to perform mode decomposition, yielding eight intrinsic mode components. During mode decomposition, the total number of modes to be decomposed was set to K=8, and a quadratic penalty factor was defined. Set convergence tolerance and maximum number of iterations Initialize each modal function Set the current iteration count n=1.

[0057] When synchronously updating the Lagrange multipliers in the frequency domain, the noise tolerance parameter is taken. =0.

[0058] Ultimately, eight intrinsic mode components (IMCs) are obtained, which serve as the input to the signal channel of the signal-image dual-channel deep neural network. The eight IMCs are ordered from high to low according to their center frequencies. The highest frequency component, The lowest frequency component is shown in the example of the decomposition result. Figure 7 As shown.

[0059] A Hilbert-Huang transform is performed on the eight intrinsic mode components to generate a two-dimensional time-frequency image, which serves as the input to the image channel of a signal-image dual-channel deep neural network. An example of the generated result is shown below. Figure 8 As shown.

[0060] The intrinsic mode components of the pressure pulsation signal collected above, processed, and the two-dimensional time-frequency image were used as training data for the constructed signal-image dual-channel deep neural network. The network was trained autonomously, sequentially passing through the signal channel, image channel, feature fusion module, and classification output layer. The four-dimensional probability vector [p1, p2, p3, p4] of the network's Softmax output layer was obtained, corresponding to the predicted probabilities of bubbly flow, condensed bubbly flow, air sac flow, and separation flow, respectively. The category with the highest probability was taken as the recognition result for the current time period. To verify the effectiveness of the method of this invention, 240 test samples were selected for flow pattern recognition experiments. The results showed that the model correctly identified 238 samples, and the overall recognition accuracy reached 99%.

Claims

1. A method for identifying the flow pattern of a mixed-flow gas-liquid pump based on pressure pulsation signals, characterized in that, Includes the following steps: Step 1: Collect pressure pulsation signals of the mixed-transport pump under different gas content and flow rate conditions, and simultaneously record the flow pattern label to form the original signal sample set; Step 2: Preprocess the pressure pulsation signal and perform mode decomposition using variational mode decomposition to obtain eight intrinsic mode components; Step 3: Perform Hilbert-Huang transform on each intrinsic mode component to generate a two-dimensional time-frequency image; Step 4: Construct and train a signal-image dual-channel deep neural network; Step 5: Deploy the trained model to the online monitoring system to identify flow patterns in real time.

2. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 1, characterized in that, Step 1 specifically involves: deploying a high-frequency dynamic pressure sensor at the guide vane channel of the gas-liquid mixing pump to collect pressure pulsation signals under different gas contents and flow rates; simultaneously, using a high-speed camera to synchronously collect flow pattern images, manually labeling the collected flow pattern images, and associating the pressure pulsation signals corresponding to the labeled flow pattern labels within the same time interval according to the time synchronization relationship; the flow pattern labels are categorized into four types: bubbly flow, condensed bubbly flow, gas bladder flow, and separated flow.

3. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: The collected pressure pulsation signal is processed using wavelet threshold denoising method to eliminate the sensor's own electronic noise and high-frequency interference caused by factors other than flow pulsation; Step 2.2: Perform Z-score normalization on the denoised signal, as shown below: In the formula, This represents a standardized, denoised signal. This represents the denoised signal, μ represents the mean of the signal segment, and σ is the standard deviation. Step 2.3: Set the total number of modes to be decomposed to K=8, and the quadratic penalty factor. Convergence tolerance and the maximum number of iterations Initialize each modal function Set the current iteration count n=1; Step 2.4: For the eight modes k=1,2,…,8, update the k-th mode function in parallel in the frequency domain. The following calculation formula applies: In the formula, For frequency variables, For the center frequency, For standardized denoised signals Fourier transform; For Lagrange multipliers; This represents the frequency domain form of the k-th modal component in the (n+1)-th iteration. It represents the sum of the Fourier transforms of all modes except the k-th mode; Synchronous update of Lagrange multipliers in the frequency domain , means as follows: In the formula, This is the noise tolerance parameter; , Let these represent the Lagrange multipliers after the nth and (n+1)th iterations, respectively; Step 2.5: Calculate the squared residual error of the sum of the normalized denoised signal and all reconstructed modes within the current iteration period. , represented as: In the formula, This means that after each iteration, the inverse Fourier transform of all modes is performed to obtain the time-domain mode functions; Convergence criteria: like or n≥ If n=n+1, then terminate the iteration and proceed to step 2.6; otherwise, let n=n+1 and return to step 2.4 to continue the loop iteration. Step 2.6: After the iteration terminates, analyze the modal function obtained from the last update. Performing an inverse Fourier transform yields the final eight eigenmode components. , k= 1, 2, ..., 8, In the formula, Indicates the inverse Fourier transform; The final components are sorted from high to low according to their center frequencies. The highest frequency component, This is the lowest frequency component.

4. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1: For the k-th intrinsic mode component Calculate the Hilbert transform , represented as: In the formula, Represents the Cauchy principal value integral; Step 3.2: Construct the k-th intrinsic mode component Analyzed signal , means as follows: In the formula, The imaginary unit, Instantaneous amplitude, It is the instantaneous phase; Step 3.3: Based on the analyzed signal Calculate the instantaneous amplitude of the k-th intrinsic mode component. With instantaneous phase , means as follows: For instantaneous phase Differentiation yields the instantaneous frequency , means as follows: Step 3.4: Based on instantaneous frequency With instantaneous amplitude Define the Hilbert spectrum of the k-th eigenmode component. The Hilbert spectrum characterizes the time-frequency distribution density of the signal; Step 3.5: Fuse the Hilbert spectra of all components to obtain a two-dimensional time-frequency image I.

5. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 4, characterized in that, Step 3.5 specifically includes: Step 3.5.1: Define a discrete two-dimensional time-frequency grid, where the time dimension corresponds to the signal sampling points and the frequency dimension corresponds to the effective range of the instantaneous frequencies of all intrinsic mode components; Step 3.5.2: Calculate the Hilbert spectrum of each intrinsic mode component. The instantaneous energy density is accumulated in the corresponding unit of the time-frequency network to form a comprehensive time-frequency energy matrix. , represented as: Step 3.5.3: Combine the time-frequency energy matrix To generate an image format suitable for convolutional neural network input, standardization is performed first, followed by energy normalization, represented as: In the formula, This represents the normalized integrated time-frequency energy matrix. This indicates the smallest positive number that should not be taken as the logarithm of 0; Secondly, the normalized integrated time-frequency energy matrix is ​​linearly mapped onto the RGB three channels to generate a two-dimensional time-frequency image I.

6. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 1, characterized in that, The signal-image dual-channel deep neural network constructed in step 4 includes a signal channel, an image channel, a feature fusion module, and an output layer; The input of the signal channel is the intrinsic mode component. The intrinsic mode component passes sequentially through two first stacked module groups connected by residuals, a global average pooling layer, a fully connected layer, and then through Batch Normalization, Relu6, and Dropout before being fused with the output features of the image channel. The input to the image channel is a two-dimensional time-frequency image. The two-dimensional time-frequency image passes through a convolutional layer, BatchNormalization, ReLU6, a second stacked module group, a two-dimensional average pooling layer, and a fully connected layer in sequence. After passing through BatchNormalization, ReLU6, and Dropout, it is fused with the output features of the signal channel. The feature fusion module first connects the features output from the signal channel and the image channel, and then fuses them through an attention fusion mechanism. The attention fusion mechanism weights and fuses the features learned from the two channels to provide weighted fused features for the output layer to identify the manifold type. The output layer is a fully connected layer and a Softmax activation function, used to output the probability of the manifold type, and the category with the highest probability is taken as the recognition result.

7. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 6, characterized in that, Each first stack module group of the signal channel contains three first stack modules; The first stacked module consists of a one-dimensional pooling layer, a one-dimensional convolutional layer, and a connection layer; The workflow of each of the first stacked modules is as follows: After the intrinsic mode components are input, they are learned in two parts. One part is learned by a one-dimensional pooling layer with a pooling kernel size of 1×3 and a one-dimensional convolutional layer with a convolutional kernel size of 1×1 to learn micro-detail features. A portion of the data is processed through a 1×1 convolutional layer and then simultaneously through three one-dimensional convolutions with kernels of 1×3, 1×5, and 1×7 to learn macroscopic features. The outputs of the two branches are connected in the feature channel and then processed through BatchNormalization, Relu6, and Dropout to obtain the signal channel output features.

8. The method for flow pattern identification of a mixed-flow gas-liquid pump based on pressure pulsation signals according to claim 6, characterized in that, The second stacked module group includes three second stacked modules, each consisting of different two-dimensional convolutional layers with residual connections. The workflow of each second stacked module is as follows: First, it goes through a 1×1 two-dimensional convolutional layer for dimensionality increase. After dimensionality increase, it goes through Batch Normalization and ReLU6, and then performs multi-scale learning in parallel through a 3×3 two-dimensional convolutional layer and a 5×5 two-dimensional convolutional layer. The outputs of the two branches go through a 1×1 two-dimensional convolutional layer for dimensionality reduction, and then through Batch Normalization to output the corresponding features.

Images