Sand pump exhaust method and system based on image recognition

Abstract

The application relates to the technical field of image processing, in particular to a sand pump exhaust method and system based on image recognition. First, real-time image sequences of a sand pump and pressure sensor data are acquired, and the real-time image sequences and the pressure sensor data are preprocessed to obtain preprocessed image sequences and pressure characteristics; then, the preprocessed image sequences and the pressure characteristics are input into a pre-trained scale adaptive pyramid network to perform bubble region semantic segmentation, so that a bubble region segmentation mask is obtained; next, geometric texture features and optical flow features extracted from a bubble image are input into a bubble trajectory prediction network to analyze the motion trajectory of bubble pixels, so that the future position, size change and motion state of the bubble are predicted; according to bubble characteristic parameters and dynamic change information, a bubble density index and a bubble motion activity index are calculated, and an exhaust control signal is output through threshold judgment; the application can effectively improve the bubble image detection accuracy, so that sand pump exhaust intelligentization is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a sand pump exhaust method and system based on image recognition. Background Technology

[0002] Traditional venting methods, whether relying on experience or simple timed venting, have significant limitations and cannot meet the demands of modern industry for precision, efficiency, and automation. While attempts have been made to analyze bubble images by acquiring internal images and employing image processing techniques such as edge detection, thresholding, and morphological operations, these methods face the following challenges in practical applications: Bubbles in images are not uniform in shape and size; they may appear as circles, ellipses, or irregular shapes of varying sizes. Furthermore, bubble images exhibit dynamic changes, resulting in a complex and unstable pixel distribution within the target region of the image. Image segmentation techniques based on fixed thresholds struggle to adapt to these dynamically changing image features, while image processing algorithms based on preset shape templates are prone to inaccurate segmentation or missed detections when dealing with irregular shapes. They fail to effectively extract image features from all bubble regions, thus reducing image detection accuracy.

[0003] To address this, a sand pump exhaust method and system based on image recognition is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a sand pump exhaust method and system based on image recognition.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A sand pump exhaust method based on image recognition, comprising: Acquire real-time image sequences of the sand pump and pressure sensor data; Preprocessing of image sequences and pressure sensor data yields preprocessed image sequences and pressure characteristics. The preprocessed image sequence and pressure features are input into a pre-trained scale-adaptive pyramid network for semantic segmentation of bubble regions, resulting in bubble region segmentation masks. The scale-adaptive pyramid network has a built-in pressure-guided attention module, which dynamically adjusts the network's attention to different regions of the image based on the generated spatial attention map, thereby achieving image region weight allocation. Geometric, textural, and optical flow features of the bubble image are extracted based on the bubble region segmentation mask. The geometric, textural, and optical flow features of the bubble image are then input into the bubble trajectory prediction network to analyze the motion trajectory of the bubble pixels and predict the future position, size changes, and motion state of the bubble. Based on bubble characteristic parameters and dynamic change information, bubble density index and bubble motion activity index are calculated. When the index exceeds the preset threshold, an exhaust control signal is output.

[0006] Furthermore, the preprocessing of the image sequence and pressure sensor data includes: An enhanced image sequence is obtained by using an adaptive histogram equalization algorithm to enhance the image sequence. The enhanced image sequence is filtered using a multi-scale Gaussian filter bank to obtain a denoised image sequence; The denoised image sequence is normalized to obtain a preprocessed image sequence; Pressure sensor data is subjected to time-frequency analysis and normalization to obtain pressure characteristics.

[0007] Furthermore, the process of inputting the preprocessed image sequence and pressure features into a pre-trained scale-adaptive pyramid network for bubble region semantic segmentation to obtain the bubble region segmentation mask includes: The encoder of the scale-adaptive pyramid network is used to process the preprocessed image sequence and extract multi-scale features; Multi-scale feature fusion is performed using the scale-adaptive pyramid module of the scale-adaptive pyramid network to obtain fused features; The pressure features and fused features are input into the pressure-guided attention module, and the generated spatial attention map is used to weight the fused features to obtain weighted features. The scale-adaptive pyramid network decoder network is used to upsample the weighted features, and the bubble region segmentation mask is output through the output layer.

[0008] Furthermore, the process of extracting the geometric features, texture features, and optical flow features of the bubble image based on the bubble region segmentation mask includes: The centroid coordinates, area, and perimeter of the bubble region are obtained through moment calculation, and the roundness and major-minor axis ratio are calculated. A gray-level co-occurrence matrix is ​​constructed within the bubble region, and its contrast and uniformity are calculated; The optical flow estimation algorithm is used to calculate the optical flow field of bubble regions between image sequences, and the average optical flow velocity and standard deviation of each bubble region in the optical flow field are extracted as optical flow features.

[0009] Furthermore, the geometric features include area, perimeter, roundness, and aspect ratio, and the texture features include the contrast and uniformity of the gray-level co-occurrence matrix.

[0010] Furthermore, the geometric texture features and optical flow features of the bubble image are input into the bubble trajectory prediction network to analyze the motion trajectory of the bubble pixels. The process of predicting the future position, size changes, and motion state of the bubble includes: A temporal feature sequence is constructed for each tracked bubble, which contains the bubble's centroid coordinates, geometric texture features, and optical flow features at each time step; The temporal feature sequence is input into the bubble trajectory prediction network, which uses a multimodal temporal fusion network to learn the temporal motion pattern of the bubble and predict the state changes of the bubble.

[0011] Furthermore, based on bubble characteristic parameters and dynamic change information, the bubble density index and bubble motion activity index are calculated. When the index exceeds a preset threshold, the process of outputting an exhaust control signal includes: The bubble density index is defined as the ratio between the total number of detected bubbles and the area of ​​the image region; the bubble motion activity index is defined as a weighted combination of the bubble's motion speed and the rate of change of its area. A predictive dual-threshold judgment mechanism is adopted. When the predicted value of the bubble density index or the bubble movement activity index exceeds the first warning threshold, an early warning is issued; when the real-time value of the index exceeds the second exhaust threshold at the same time, an exhaust control signal is output.

[0012] A sand pump exhaust system based on image recognition includes: a data acquisition unit for acquiring real-time image sequences of the sand pump and pressure sensor data; a multimodal preprocessing unit for preprocessing the image sequences and pressure sensor data to obtain preprocessed image sequences and pressure features; a bubble segmentation unit for inputting the preprocessed image sequences and pressure features into a pre-trained scale-adaptive pyramid network for semantic segmentation of bubble regions to obtain a bubble region segmentation mask; wherein, the scale-adaptive pyramid network has a built-in pressure-guided attention module, which dynamically adjusts the network's attention to different regions of the image according to the generated spatial attention map to achieve image region weight allocation; a bubble tracking and prediction unit for extracting geometric features, texture features, and optical flow features of the bubble image according to the bubble region segmentation mask, inputting the geometric texture features and optical flow features of the bubble image into a bubble trajectory prediction network for analyzing the motion trajectory of bubble pixels to predict the future position, size change, and motion state of the bubble; and an exhaust control unit for calculating bubble density and bubble motion activity indicators according to bubble feature parameters and dynamic change information, and outputting an exhaust control signal when the indicators exceed a preset threshold.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention utilizes a scale-adaptive pyramid network to generate pixel-level bubble segmentation masks. The multi-scale feature extraction and fusion process enables the network to adapt to bubbles of different sizes and shapes, thereby overcoming the problem of missed detection of irregular bubbles by traditional edge detection and threshold segmentation. The pressure-guided attention module in the network dynamically adjusts the network's attention area to the image based on pressure features, enhances the feature response related to bubbles, avoids background noise interference, and thus improves the bubble segmentation accuracy in complex fluid environments.

[0014] 2. This invention significantly improves the accuracy of predicting future positions and sizes by constructing a temporal sequence containing centroid coordinates, geometric texture, and optical flow features, and using a bubble trajectory prediction network to learn the dynamic patterns of bubbles. The temporal analysis capability of the bubble trajectory prediction network effectively captures the complex motion behavior of bubbles, enhancing the foresight of the prediction. At the same time, the real-time output of the prediction results optimizes the response speed of exhaust decision-making, significantly improving the level of automation control of the sand pump exhaust system under dynamic operating conditions.

[0015] 3. By defining bubble density and motion activity indices, this invention enables quantitative assessment of bubble distribution and motion state, significantly improving the scientific rigor and accuracy of exhaust decision-making. The predictive dual-threshold judgment mechanism, combined with early warning and real-time control, effectively reduces the risk of misoperation and enhances the robustness and reliability of the exhaust system under complex operating conditions. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of a sand pump exhaust method based on image recognition according to the present invention; Figure 2 This is a schematic diagram of the process for obtaining the bubble region segmentation mask according to the present invention; Figure 3 This is a schematic diagram of the structure of a sand pump exhaust system based on image recognition according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1 to 3 This invention provides a sand pump exhaust method and system based on image recognition, the technical solution of which is as follows: Example 1: To improve the automation of its sand pump exhaust system, a company used an image recognition-based sand pump exhaust method proposed in this invention. The flowchart of this method is shown below. Figure 1 As shown, it specifically includes: Acquire real-time image sequences of the sand pump and pressure sensor data; Furthermore, by using an external high-speed camera to continuously acquire images through the sand pump's observation window, an image sequence is obtained; the image sequence is synchronized with the pressure data acquired by the pressure sensor.

[0019] Preprocessing of image sequences and pressure sensor data yields preprocessed image sequences and pressure characteristics. Furthermore, the preprocessing of the image sequence and pressure sensor data includes: An enhanced image sequence is obtained by using an adaptive histogram equalization algorithm to enhance the image sequence. The enhanced image sequence is filtered using a multi-scale Gaussian filter bank to obtain a denoised image sequence; The denoised image sequence is normalized to obtain a preprocessed image sequence; Pressure sensor data is subjected to time-frequency analysis and normalization to obtain pressure characteristics; Furthermore, the adaptive histogram equalization process is implemented using the CLAHE algorithm, with the shearing limit parameter adjustable within the range of 2.0-4.0 and the grid size set to 8×8. This process enhances the local contrast of the image in blocks, avoiding over-enhancement caused by global processing. Furthermore, the multi-scale Gaussian filter bank contains three filter kernels of different scales, with standard deviations set to 0.8, 1.5, and 2.5, respectively. It performs parallel filtering on the enhanced image sequence and removes noise through a self-weighted fusion method. The weights are dynamically adjusted according to image quality assessment metrics (PSNR, SSIM, etc.). Furthermore, the image normalization adopts a local normalization method, which normalizes the pixel values ​​within each 8×8 pixel block to the range of [0,1]. Furthermore, the time-frequency analysis of the pressure data employed a sliding window fast Fourier transform with a window length of 128 sampling points, an overlap rate of 75%, and a frequency range of interest of 0-25Hz. The amplitude and phase information of the main frequency components were extracted as pressure features.

[0020] By using an adaptive histogram equalization algorithm, the contrast and detail of the sand pump image sequence can be effectively enhanced; the introduction of a multi-scale Gaussian filter bank effectively suppresses image noise, enhances the robustness of subsequent bubble segmentation, and adapts to dynamic changes in the fluid environment; the normalization processing of image and pressure data optimizes the input conditions for multimodal data fusion, significantly improving the processing accuracy and real-time performance of the scale-adaptive pyramid network.

[0021] The preprocessed image sequence and pressure features are input into a pre-trained scale-adaptive pyramid network for semantic segmentation of bubble regions, resulting in bubble region segmentation masks. The scale-adaptive pyramid network has a built-in pressure-guided attention module, which dynamically adjusts the network's attention to different regions of the image based on the generated spatial attention map, thereby achieving image region weight allocation. Furthermore, the preprocessed image sequence and pressure features are input into a pre-trained scale-adaptive pyramid network for bubble region semantic segmentation, and the process of obtaining the bubble region segmentation mask is as follows: Figure 2 As shown, it specifically includes: The encoder of the scale-adaptive pyramid network is used to process the preprocessed image sequence and extract multi-scale features; Multi-scale feature fusion is performed using the scale-adaptive pyramid module of the scale-adaptive pyramid network to obtain fused features; The pressure features and fused features are input into the pressure-guided attention module, and the generated spatial attention map is used to weight the fused features to obtain weighted features. The weighted features are upsampled using the decoder network of the scale-adaptive pyramid network, and a bubble region segmentation mask is output through the output layer. Furthermore, the encoder of the scale-adaptive pyramid network adopts a lightweight CNN architecture, which contains four convolutional blocks, each containing two 3×3 convolutional layers with 32, 64, 128, and 256 channels respectively, and uses batch normalization and ReLU activation function. Furthermore, the scale-adaptive pyramid module employs three pooling layers of different scales, denoted as the first pooling layer, the second pooling layer, and the third pooling layer, with corresponding pooling scales of 2×2, 4×4, and 8×8, respectively. Each branch is followed by a 1×1 convolution for channel compression. Finally, bilinear interpolation is used to unify the feature map size and channel concatenation is performed to obtain the fused features.

[0022] Furthermore, the process of obtaining weighted features by the pressure-guided attention module is as follows: the 1-dimensional pressure features are expanded into a 2-dimensional spatial attention map through a fully connected layer, the attention weights are normalized to the range [0,1] through the sigmoid function, and then multiplied element-wise with the fused features to generate pressure-guided weighted features; Furthermore, the decoder adopts a convolutional structure, which includes three upsampling layers. Each layer is upsampled by 2x bilinear interpolation and then followed by a 3×3 convolution. Finally, a single-channel segmentation result is output through a 1×1 convolution. The activation function uses the sigmoid function to ensure that the output value is in the range [0,1].

[0023] Furthermore, the network training uses a combination of Focal Loss and Dice Loss loss functions with weights of 0.4 and 0.6 respectively. The initial learning rate is set to 1e-3, and cosine annealing learning rate scheduling is used. The batch size is 8, and the total number of training epochs is 150. Furthermore, the stress-guided attention module is designed with an adaptive weight adjustment mechanism, the implementation of which includes: using a sliding window statistical method to perform variance analysis on the stress data of the past 100 time steps to calculate the stress fluctuation intensity index; using the stress fluctuation intensity index to adaptively weight the attention weights in the spatial attention map calculation process to obtain the adaptive weight factor; and using the exponential moving average method to smooth the adaptive weight factor.

[0024] Through an adaptive weight adjustment mechanism, the attention mechanism can more accurately focus on bubble-related regions under complex working conditions with drastic pressure fluctuations, thereby improving segmentation accuracy.

[0025] By utilizing a scale-adaptive pyramid network to generate pixel-level bubble segmentation masks, the multi-scale feature extraction and fusion process enables the network to adapt to bubbles of different sizes and shapes. The pressure-guided attention module dynamically adjusts the network's attention area to the image based on pressure features, enhancing the feature response related to bubbles, thereby improving the bubble segmentation accuracy in complex fluid environments.

[0026] Geometric, textural, and optical flow features of the bubble image are extracted based on the bubble region segmentation mask. The geometric, textural, and optical flow features of the bubble image are then input into the bubble trajectory prediction network to analyze the motion trajectory of the bubble pixels and predict the future position, size changes, and motion state of the bubble. Furthermore, geometric features include area, perimeter, roundness, and aspect ratio, while texture features include the contrast and uniformity of the gray-level co-occurrence matrix. Furthermore, the process of extracting the geometric features, texture features, and optical flow features of the bubble image based on the bubble region segmentation mask includes: The centroid coordinates, area, and perimeter of the bubble region are obtained through moment calculation, and the roundness and major-minor axis ratio are calculated. A gray-level co-occurrence matrix is ​​constructed within the bubble region, and its contrast and uniformity are calculated; The optical flow estimation algorithm is used to calculate the optical flow field of bubble regions between image sequences, and the average optical flow velocity and standard deviation of each bubble region in the optical flow field are extracted as optical flow features. Furthermore, bubble regions in an image sequence can be extracted based on the bubble region segmentation mask; Furthermore, the connected domain analysis of the bubble region adopts the 8-neighborhood connection method to filter out noise regions with an area of ​​less than 50 pixels, and calculates the geometric features of the effective bubble region through moments. Furthermore, in addition to basic geometric features, hierarchical extraction of multi-scale geometric features can be performed. This process includes: using wavelet transform to decompose the bubble contour into three scale levels, with decomposition scale factors set to 2, 4, and 8 respectively, to obtain contour representations at different detail levels; using box counting to perform fractal analysis on the bubble contour, with box size increasing from 2 pixels to 32 pixels and a step size of 2 pixels, and obtaining the fractal dimension by fitting the slope of the log-log graph using the least squares method; calculating the ratio of the bubble region area to its convex hull area, which reflects the degree of concavity and convexity of the bubble shape; and combining basic geometric features, fractal features, and convex hull features to construct multi-scale geometric features. By introducing fractal dimension and convex hull ratio features, the ability to identify irregular bubbles is enhanced; the hierarchical geometric feature extraction method can more comprehensively describe the morphological changes of bubbles and improve the accuracy of bubble classification.

[0027] Furthermore, a gray-level co-occurrence matrix with a distance of 1 pixel is constructed within the bubble region, and the calculation directions include two main directions: 0° and 90°. Two texture descriptors, contrast and uniformity, are extracted as texture features. Furthermore, a sparse optical flow algorithm is used to calculate the optical flow vector of the bubble region between image sequences, and the mean and standard deviation of the optical flow amplitude are statistically analyzed as optical flow features.

[0028] Geometric features obtained through moment calculation can accurately quantify the morphological changes of bubbles; texture features extracted by constructing a gray-level co-occurrence matrix enhance the perception of the internal texture of bubbles; optical flow features extracted using an optical flow estimation algorithm can capture the dynamic motion features of bubbles, and these features provide stable input for subsequent bubble trajectory prediction and density calculation.

[0029] Furthermore, the geometric texture features and optical flow features of the bubble image are input into the bubble trajectory prediction network to analyze the motion trajectory of the bubble pixels. The process of predicting the future position, size changes, and motion state of the bubble includes: A temporal feature sequence is constructed for each tracked bubble, which contains the bubble's centroid coordinates, geometric texture features, and optical flow features at each time step; The temporal feature sequence is input into the bubble trajectory prediction network, which uses a multimodal temporal fusion network to learn the temporal motion pattern of the bubble and predict the state changes of the bubble. Furthermore, by combining the centroid coordinates of the bubbles with the grid indexing method, the bubble distribution within the neighborhood is obtained; Furthermore, the multimodal temporal fusion network architecture includes four parallel feature processing branches: the geometric feature branch uses a 3-layer MLP to process shape evolution information, with hidden layer dimensions of [64, 32, 16]; the texture feature branch uses a 1D-CNN to process temporal texture changes, with a convolutional kernel size of 3 and the number of channels of [32, 16, 8]; the motion feature branch uses a 2-layer BiLSTM to process motion trajectory information, with hidden layer dimensions of [64, 32]; and the context feature branch uses a 2-layer fully connected network to process spatial relationships, with hidden layer dimensions of [16, 8]. Furthermore, the fusion layer of the multimodal temporal fusion network architecture uses an attention mechanism to fuse the feature representations of the four branches to predict the state changes of the bubbles, including their number, position, and size. Furthermore, during training, four feature branches are trained separately by freezing the fusion layer. The geometry branch uses MSE loss with a learning rate of 1e-3, the texture branch uses cross-entropy loss, the motion branch uses trajectory regression loss, and the context branch uses neighborhood prediction loss. Then, all layers are unfrozen, and end-to-end joint training is performed using the combined loss function with the learning rate decayed to 5e-4. Dropout is added to prevent overfitting. By constructing a temporal sequence containing centroid coordinates, geometric texture, and optical flow features, and utilizing a bubble trajectory prediction network to learn the dynamic patterns of bubbles, the accuracy of future position and size predictions is significantly improved. The temporal analysis capability of the bubble trajectory prediction network effectively captures the complex motion behavior of bubbles, enhancing the foresight of predictions. At the same time, the real-time output of prediction results optimizes the response speed of exhaust decisions, significantly improving the level of automation control of the sand pump exhaust system under dynamic operating conditions.

[0030] Based on bubble characteristic parameters and dynamic change information, bubble density index and bubble motion activity index are calculated. When the index exceeds the preset threshold, an exhaust control signal is output.

[0031] Furthermore, based on bubble characteristic parameters and dynamic change information, the bubble density index and bubble motion activity index are calculated. When the index exceeds a preset threshold, the process of outputting an exhaust control signal includes: The bubble density index is defined as the ratio between the total number of detected bubbles and the area of ​​the image region; the bubble motion activity index is defined as a weighted combination of the normalized bubble motion speed and the rate of change of area. A predictive dual-threshold judgment mechanism is adopted. When the predicted value of the bubble density index or the bubble motion activity index exceeds the first warning threshold, an early warning is issued; when the real-time value of the index exceeds the second exhaust threshold at the same time, an exhaust control signal is output. Furthermore, the weighting coefficients for bubble motion speed and area change rate in the bubble motion activity index are 0.6 and 0.4, respectively; Furthermore, the first warning thresholds are 0.15 particles / cm² and 0.32 particles / cm², respectively, and the second exhaust thresholds are 0.22 particles / cm² and 0.58 particles / cm², respectively. These thresholds can be adjusted according to actual conditions and are not unique. Furthermore, an adjustment strategy based on equipment health status is used to dynamically adjust the first warning threshold and the second exhaust threshold. The implementation process includes: establishing an equipment wear assessment model, performing trend analysis on the bubble density index and the motion activity index respectively, comparing the average value of each index with the benchmark value, and then weighting and summing the decay rates of each index to obtain the equipment performance decay rate; defining the equipment health status level according to the equipment performance decay rate, which is divided into healthy state, light wear state, moderate wear state, and heavy wear state, and the threshold between the states can be flexibly adjusted; implementing a dynamic threshold adjustment strategy, in the healthy state, the first warning threshold and the second exhaust threshold remain at their original values, in the light wear state they are adjusted to 0.9 times the original value, in the moderate wear state they are adjusted to 0.8 times the original value, and in the heavy wear state they are adjusted to 0.7 times the original value.

[0032] By using a threshold adjustment strategy based on equipment health status, real-time monitoring and intelligent response of equipment health status are achieved. The dynamic threshold adjustment strategy reduces energy consumption and improves the overall reliability and economic benefits of the sand pump system while ensuring exhaust effect.

[0033] By defining bubble density and motion activity indices, a quantitative assessment of bubble distribution and motion status is achieved, significantly improving the scientific rigor and accuracy of exhaust decision-making. The predictive dual-threshold judgment mechanism, combined with early warning and real-time control, effectively reduces the risk of misoperation and enhances the robustness and reliability of the exhaust system under complex operating conditions.

[0034] This embodiment proposes a sand pump exhaust method based on image recognition. First, real-time image sequences of the sand pump and pressure sensor data are acquired and preprocessed to obtain preprocessed image sequences and pressure features. Then, the preprocessed image sequences and pressure features are input into a pre-trained scale-adaptive pyramid network for semantic segmentation of the bubble region, obtaining a bubble region segmentation mask. Next, geometric texture features and optical flow features extracted from the bubble images are input into a bubble trajectory prediction network to analyze the motion trajectory of bubble pixels, predicting the future position, size changes, and motion state of the bubble. Based on the bubble feature parameters and dynamic change information, bubble density and bubble motion activity indices are calculated, and an exhaust control signal is output based on a threshold judgment. This invention can effectively improve the accuracy of bubble image detection, thereby realizing intelligent sand pump exhaust.

[0035] Example 2: This invention also proposes a sand pump exhaust system based on image recognition, the structure of which is as follows: Figure 3 As shown, it includes: The data acquisition unit is used to acquire real-time image sequences of the sand pump and pressure sensor data; The multimodal preprocessing unit is used to preprocess the image sequence and pressure sensor data to obtain the preprocessed image sequence and pressure features; Furthermore, the multimodal preprocessing unit includes an image preprocessing module and a signal preprocessing module; the image preprocessing module incorporates the CLAHE enhancement algorithm, the multi-scale Gaussian filtering algorithm, and the normalization algorithm; the signal preprocessing module incorporates the time-frequency analysis algorithm and the normalization algorithm.

[0036] The bubble segmentation unit is used to input the preprocessed image sequence and pressure features into the pre-trained scale-adaptive pyramid network to perform semantic segmentation of bubble regions and obtain bubble region segmentation masks. The scale-adaptive pyramid network has a built-in pressure-guided attention module, which dynamically adjusts the network’s attention to different regions of the image according to the generated spatial attention map to achieve image region weight allocation. Furthermore, the core of the bubble segmentation unit is a scale-adaptive pyramid network, which includes four main components: an encoder, a pyramid feature fusion module, a pressure-guided attention module, and a decoder. The encoder is responsible for multi-scale feature extraction, the pyramid module is responsible for cross-scale feature fusion, the attention module is responsible for pressure-guided feature weighting, and the decoder is responsible for generating pixel-level segmentation masks. Furthermore, the process of inputting the preprocessed image sequence and pressure features into a pre-trained scale-adaptive pyramid network for bubble region semantic segmentation to obtain the bubble region segmentation mask includes: The encoder of the scale-adaptive pyramid network is used to process the preprocessed image sequence and extract multi-scale features; Multi-scale feature fusion is performed using the scale-adaptive pyramid module of the scale-adaptive pyramid network to obtain fused features; The pressure features and fused features are input into the pressure-guided attention module, and the generated spatial attention map is used to weight the fused features to obtain weighted features. The scale-adaptive pyramid network decoder network is used to upsample the weighted features, and the bubble region segmentation mask is output through the output layer.

[0037] The bubble tracking prediction unit is used to extract the geometric features, texture features and optical flow features of the bubble image based on the bubble region segmentation mask. The geometric and texture features and optical flow features of the bubble image are input into the bubble trajectory prediction network to analyze the motion trajectory of the bubble pixels and predict the future position, size change and motion state of the bubble. Furthermore, the bubble tracking prediction unit includes a feature extraction module and a trajectory prediction module; the feature extraction module uses the image moment algorithm to extract geometric features, the gray-level co-occurrence matrix algorithm to extract texture features, and the Lucas-Kanade algorithm to extract optical flow features; the trajectory prediction module uses a bubble trajectory prediction network to learn temporal motion patterns.

[0038] Exhaust control unit: It is used to calculate the bubble density index and bubble motion activity index based on bubble characteristic parameters and dynamic change information. When the index exceeds the preset threshold, it outputs an exhaust control signal.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A sand pump exhaust method based on image recognition, characterized in that, include: Acquire real-time image sequences of the sand pump and pressure sensor data; Preprocessing of image sequences and pressure sensor data yields preprocessed image sequences and pressure characteristics. The preprocessed image sequence and pressure features are input into a pre-trained scale-adaptive pyramid network for bubble region semantic segmentation to obtain a bubble region segmentation mask. The scale-adaptive pyramid network has a built-in pressure-guided attention module, which dynamically adjusts the network’s attention to different regions of the image based on the generated spatial attention map to achieve image region weight allocation. Geometric, textural, and optical flow features of the bubble image are extracted based on the bubble region segmentation mask. The geometric, textural, and optical flow features of the bubble image are then input into the bubble trajectory prediction network to analyze the motion trajectory of the bubble pixels and predict the future position, size changes, and motion state of the bubble. Based on bubble characteristic parameters and dynamic change information, bubble density index and bubble motion activity index are calculated. When the index exceeds the preset threshold, an exhaust control signal is output.

2. The sand pump exhaust method based on image recognition according to claim 1, characterized in that, The preprocessing of image sequences and pressure sensor data includes: An enhanced image sequence is obtained by using an adaptive histogram equalization algorithm to enhance the image sequence. The enhanced image sequence is filtered using a multi-scale Gaussian filter bank to obtain a denoised image sequence; The denoised image sequence is normalized to obtain a preprocessed image sequence; Pressure sensor data is subjected to time-frequency analysis and normalization to obtain pressure characteristics.

3. The sand pump exhaust method based on image recognition according to claim 1, characterized in that, The process of inputting the preprocessed image sequence and pressure features into a pre-trained scale-adaptive pyramid network for bubble region semantic segmentation to obtain the bubble region segmentation mask includes: The encoder of the scale-adaptive pyramid network is used to process the preprocessed image sequence and extract multi-scale features; Multi-scale feature fusion is performed using the scale-adaptive pyramid module of the scale-adaptive pyramid network to obtain fused features; The pressure features and fused features are input into the pressure-guided attention module. The pressure features are expanded into a spatial attention map through a fully connected layer. The generated spatial attention map is used to multiply the fused features element-wise to obtain weighted features. The scale-adaptive pyramid network decoder network is used to upsample the weighted features, and the bubble region segmentation mask is output through the output layer.

4. The sand pump exhaust method based on image recognition according to claim 1, characterized in that, The geometric features include area, perimeter, roundness, and aspect ratio, while the texture features include the contrast and uniformity of the gray-level co-occurrence matrix.

5. The sand pump exhaust method based on image recognition according to claim 1, characterized in that, The process of extracting geometric, texture, and optical flow features of a bubble image based on a bubble region segmentation mask includes: The centroid coordinates, area, and perimeter of the bubble region are obtained through moment calculation, and the roundness and major-minor axis ratio are calculated. A gray-level co-occurrence matrix is ​​constructed within the bubble region, and its contrast and uniformity are calculated; The optical flow estimation algorithm is used to calculate the optical flow field of bubble regions between image sequences, and the average optical flow velocity and standard deviation of each bubble region in the optical flow field are extracted as optical flow features.

6. The sand pump exhaust method based on image recognition according to claim 1, characterized in that, The process of inputting the geometric texture features and optical flow features of the bubble image into the bubble trajectory prediction network to analyze the motion trajectory of bubble pixels and predict the future position, size changes, and motion state of the bubble includes: A temporal feature sequence is constructed for each tracked bubble, and the sequence contains the bubble's centroid coordinates, geometric texture features, and optical flow features at each time step; The temporal feature sequence is input into the bubble trajectory prediction network. The network uses a multimodal temporal fusion network to learn the temporal motion pattern of the bubble and predict the state changes of the bubble.

7. The sand pump exhaust method based on image recognition according to claim 1, characterized in that, Based on bubble characteristic parameters and dynamic change information, the process of calculating bubble density and bubble motion activity indices, and outputting exhaust control signals when these indices exceed preset thresholds includes: The bubble density index is defined as the ratio between the total number of detected bubbles and the area of ​​the image region; the bubble motion activity index is defined as a weighted combination of the normalized bubble motion speed and the rate of change of area. A predictive dual-threshold judgment mechanism is adopted. When the predicted values ​​of the bubble density index and / or bubble movement activity index exceed the first warning threshold, an early warning is issued; when the real-time values ​​of the indexes simultaneously exceed the second exhaust threshold, an exhaust control signal is output.

8. A sand pump exhaust system based on image recognition, characterized in that, include: The data acquisition unit is used to acquire real-time image sequences of the sand pump and pressure sensor data; The multimodal preprocessing unit is used to preprocess the image sequence and pressure sensor data to obtain the preprocessed image sequence and pressure features; The bubble segmentation unit inputs preprocessed image sequences and pressure features into a pre-trained scale-adaptive pyramid network for semantic segmentation of bubble regions, obtaining a bubble region segmentation mask. The scale-adaptive pyramid network incorporates a pressure-guided attention module, which dynamically adjusts the network's attention to different image regions based on the generated spatial attention map, achieving image region weight allocation. The bubble tracking and prediction unit extracts geometric, texture, and optical flow features from the bubble image based on the bubble region segmentation mask. It then inputs these features into a bubble trajectory prediction network to analyze the motion trajectory of bubble pixels, predicting the future position, size changes, and motion state of the bubble. The exhaust control unit calculates bubble density and bubble motion activity indices based on bubble feature parameters and dynamic change information. When these indices exceed a preset threshold, it outputs an exhaust control signal.

Images