A valve tightness detection method based on water bubble machine vision

Abstract

The application provides a valve leakproofness detection method based on bubble machine vision in water, and belongs to the technical field of valve detection. The application fills a liquid bag valve with standard air pressure, places the valve in a water tank and applies micro oscillation, acquires streaming video frames by using a machine vision acquisition system, inputs the video frames into a group intelligence lightweight topology dynamic evolution model, performs sparse graph topology initialization and distributed collaborative parallel reasoning layer processing, synchronously executes background difference and wavelet decomposition pre-noise removal to remove false bright spots of liquid surface ripples, implements streaming weight compensation on micro bubbles by using a multi-scale pyramid cascade network and a spatial self-organizing evolution operator, performs momentum conservation verification and removal on non-bubble disturbance by using Navier-Stokes fluid dynamics constraints, and finally determines the liquid bag valve leakproofness according to the bubble overflow rate and position distribution, thereby solving the problem that the machine vision artificial intelligence system cannot reliably distinguish real leakage bubbles under the interference of liquid surface ripple bright spots in a micro oscillation environment.

Description

Technical Field

[0001] This invention belongs to the field of valve inspection technology, and more specifically, relates to a valve sealing test method based on machine vision of air bubbles in water. Background Technology

[0002] Leakage testing of liquid bag valves is a crucial step in quality control in fields such as medical and food processing. Traditional methods primarily rely on manual underwater bubble observation or differential pressure sensor monitoring. The former is inefficient and dependent on operator experience, while the latter lacks sensitivity to minute leaks, failing to meet the demands of high-precision, batch testing. With the development of machine vision technology, image recognition-based bubble detection solutions have been gradually introduced into valve leak testing. These solutions use cameras to capture images of bubbles in a water tank and combine them with deep learning models for automatic identification. However, in actual testing conditions, to facilitate the escape of minute leaking bubbles from valve gaps, micro-oscillations are applied to the water tank. These micro-oscillations inevitably create periodic surface ripples, which produce bright spots in the image that closely resemble bubble morphology. Traditional background subtraction or fixed threshold filtering methods cannot reliably distinguish between real bubbles and bright spots from surface ripples in dynamic interference environments, resulting in a high false leak rate in the test results. In other words, existing technologies suffer from the technical problem of bright spots from surface ripples in micro-oscillation environments preventing machine vision systems from reliably distinguishing between real leaking bubbles. Summary of the Invention

[0003] In view of this, the present invention provides a valve tightness detection method based on machine vision of bubbles in water, which can solve the technical problem in the prior art that the interference of liquid surface ripples caused by the micro-vibration environment of the water tank in the liquid bag valve tightness detection scenario causes the machine vision system to be unable to accurately distinguish between real leaking bubbles and false bright spots.

[0004] This invention is implemented as follows: This invention provides a valve tightness detection method based on machine vision for air bubbles in water, comprising the following steps:

[0005] Inflate the liquid bag valve with standard air pressure, place the liquid bag valve in the water tank, apply micro-vibration to the water tank, start the machine vision acquisition system, and acquire streaming video frames;

[0006] The streaming video frames are input into the swarm intelligence lightweight topology dynamic evolution model. After processing by sparse graph topology initialization and distributed collaborative parallel inference layer, background difference and wavelet decomposition pre-denoising are performed simultaneously to remove false bright spot interference caused by liquid surface ripples.

[0007] The feature map after pre-denoising is processed by the spatial non-maximum flow suppression operator, and combined with the dynamic evolution mechanism of optical flow field topology, multi-dimensional feature decoupling is performed on the overlapping bubble region to restore the individual features of each bubble.

[0008] A multi-scale pyramid cascade network is used to perform multi-resolution alignment of the decoupled bubble features. A spatial self-organizing evolution operator is used to perform streaming weight compensation on the micro-bubble target, and the bubble position and overflow rate are output.

[0009] The Navier-Stokes fluid dynamics constraint update algorithm verifies the fluid momentum conservation of bubble motion characteristics, eliminates non-bubble perturbation velocities, and updates the weights of the swarm intelligence lightweight topology dynamic evolution model.

[0010] The airtightness of the liquid bag valve is determined based on the bubble overflow rate and location distribution. When the overflow rate exceeds the airtightness determination threshold, the leakage detection result is output.

[0011] The standard pressure is specifically obtained through multiple rounds of gradient pressurization experiments on the same batch of liquid bag valves, combined with experimental data analysis of leakage frequency and detection sensitivity. The standard pressure range is 0.05–0.30. .

[0012] Specifically, the micro-oscillation refers to applying periodic mechanical vibration to the water tank, with a vibration frequency range of 1 to 10. The amplitude range is 0.5 to 3.0. The frame rate range for streaming video frames is 60–120. .

[0013] Specifically, the swarm intelligence lightweight topology dynamic evolution model involves the input layer receiving compressed tensors of streaming video frames, the initial network topology being a sparse graph structure, the initial connection weights between nodes being generated by the local probability function of the cuckoo search algorithm, the initial connection edges being established between node pairs with probabilities higher than the activation threshold, and the network internally using a distributed collaborative parallel inference layer instead of the traditional fully connected layer.

[0014] Specifically, the distributed collaborative parallel inference layer consists of inference layer nodes that are abstracted as independent intelligent agents with a state memory trace decay mechanism, the memory trace being based on a decay equation. decays hourly, where for The value of memory traces at any given moment. The attenuation coefficient is... The recommended range is 0.01 to 0.10. .

[0015] Specifically, the sparse graph topology initialization involves feature transmission between nodes based on flow dynamic threshold collaborative logic, local activation operators triggering dynamic node splitting, dynamic allocation of child node memory based on the spatial information entropy value of the parent node's feature graph, and automatic elimination of invalid background region nodes when the memory trace is below the elimination threshold, with a recommended elimination threshold range of 0.05 to 0.20.

[0016] Specifically, the background difference and wavelet decomposition pre-denoising is based on the state entropy increase dynamic constraint. By establishing a state entropy function on the continuous inter-frame pixel difference sequence, the low-frequency brightness disturbance caused by liquid surface ripples and the high entropy value mutation caused by bubbles are distinguished. Combined with wavelet decomposition, the low-frequency noise component of the ripples is removed from the difference image.

[0017] Specifically, the dynamic evolution mechanism of the optical flow field topology is to use the optical flow estimation method to calculate the motion vector field of each pixel between adjacent frames, construct the motion vector field into a graph topology, and identify the independent motion trajectory of each bubble individual based on the evolution law of the topology between adjacent frames.

[0018] Specifically, the multidimensional feature decoupling involves extracting features from multiple dimensions such as spatial location, movement speed, and brightness gradient of the overlapping bubble region and modeling them independently, thereby restoring the aliased multi-bubble features into their respective independent individual feature representations.

[0019] Specifically, the spatial self-organizing evolution operator adaptively adjusts the feature weight allocation based on the spatial distribution of the micro-target region in the feature map, and compensates and amplifies the feature weight of the micro-bubble target region; the streaming weight compensation predicts the weight compensation region of the current frame based on the position of the micro-bubble in the previous frame, and continuously enhances the weight of the micro-target region.

[0020] Specifically, the Navier-Stokes fluid dynamics constraint update algorithm embeds the Navier-Stokes momentum equation as a physical constraint into the loss function. It verifies whether the bubble motion satisfies the fluid momentum conservation law based on the bubble edge velocity vector, water viscous drag parameters, and buoyancy force on the bubble. Motion features that do not satisfy the conservation law are offset and removed from the gradient of the loss function.

[0021] The swarm intelligence lightweight topology dynamic evolution model also embeds a topology dynamic adjustment function. , ,in The entropy of the bubble feature space information in the current frame. The average bubble overflow rate for the current frame. Background interference intensity index, , , The weighting coefficients and .

[0022] Wherein, the topology dynamic adjustment function Specifically, when When the node splitting rate parameter drops below the low threshold, the node splitting rate parameter decreases to a low level; when When the node splitting rate parameter is in the medium range, adjust it to the medium level; when When the threshold value is higher than the high-level threshold, the node splitting rate parameter is increased to the high-level threshold. The low-level threshold is 0.30, and the high-level threshold is 0.65.

[0023] The swarm intelligence lightweight topology dynamic evolution model embeds a minimum spanning tree graph theory algorithm, which is used to reconstruct the optimal connected path of the sparse graph topology after each node split or disappearance. The minimum spanning tree algorithm is executed after constructing a weighted graph based on the feature similarity matrix between nodes in the current frame.

[0024] Wherein, the weighting coefficient The recommended range is 0.3 to 0.5. The recommended range is 0.2 to 0.4. The recommended range is 0.1 to 0.3; the airtightness judgment threshold is obtained by comparing the airtightness of liquid bag valves with liquid bag valves with different levels of micro-leakage under the same standard air pressure and micro-oscillation conditions, and then determining the optimal classification boundary after statistically analyzing the distribution range of bubble overflow rate.

[0025] This invention embeds state entropy increase dynamic constraints into a background difference model. It leverages the fundamental difference between the low-frequency, low-entropy perturbations caused by liquid surface ripples and the high-entropy abrupt changes generated by bubbles. By combining wavelet decomposition to remove low-frequency noise components from the ripples, and then using Navier-Stokes fluid dynamics constraints to verify the momentum conservation of bubble motion, non-bubble perturbations that do not satisfy the conservation law are removed from the model weight update gradient, ensuring that the convergence direction of the model parameters always aligns with the fluid dynamics behavior of real bubbles. The motion characteristics of liquid surface ripples physically do not satisfy the momentum conservation law for rising bubbles in water. Therefore, these two types of constraints filter false bright spots from both the information entropy and physical dynamics dimensions, fundamentally cutting off the contamination path of non-bubble interference on the detection results. In summary, this invention solves the technical problem that liquid surface ripple bright spot interference in micro-oscillation environments prevents machine vision systems from reliably distinguishing real leaking bubbles. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 The graph shows the variation of bubble recognition accuracy and state entropy increase constraint loss values ​​on the validation set with the number of iterations during the training process of a lightweight topological dynamic evolution model for swarm intelligence.

[0028] Figure 3 This is a scatter plot showing the distribution of background differential state entropy values ​​for each sample frame during the liquid bag valve sealing test. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0030] like Figure 1 The diagram shown is a flowchart of a valve sealing performance detection method based on machine vision for air bubbles in water, provided by the present invention. This method includes the following steps:

[0031] S01. Inflate the liquid bag valve with standard air pressure, place the liquid bag valve in the water tank, apply micro-vibration to the water tank, start the machine vision acquisition system, and acquire streaming video frames.

[0032] S02. Input the streaming video frames into the swarm intelligence lightweight topology dynamic evolution model. After sparse graph topology initialization and distributed collaborative parallel inference layer processing, perform background difference and wavelet decomposition pre-denoising simultaneously to remove false bright spot interference formed by liquid surface ripples.

[0033] S03. The feature map after pre-denoising is processed by the spatial non-maximum flow suppression operator, and combined with the dynamic evolution mechanism of optical flow field topology, multi-dimensional feature decoupling is performed on the overlapping bubble region to restore the individual features of each bubble.

[0034] S04. A multi-scale pyramid cascade network is used to perform multi-resolution alignment on the decoupled bubble features. A spatial self-organizing evolution operator is used to perform streaming weight compensation on the micro bubble target, and the bubble position and overflow rate are output.

[0035] S05, the Navier-Stokes fluid dynamics constraint update algorithm performs fluid momentum conservation verification on bubble motion characteristics, eliminates non-bubble disturbance velocities by offsetting them, and updates the weights of the swarm intelligence lightweight topology dynamic evolution model.

[0036] S06. Determine the airtightness of the liquid bag valve based on the bubble overflow rate and location distribution. Output the leakage detection result when the overflow rate exceeds the airtightness determination threshold.

[0037] The standard air pressure is the rated working air pressure of the liquid bag valve. Its specific value is obtained through multiple rounds of gradient pressurization experiments on the same batch of liquid bag valves, combined with experimental data analysis of leakage frequency and detection sensitivity. The recommended range is 0.05–0.30. .

[0038] The micro-oscillation refers to applying periodic mechanical vibration to the water tank. The frequency range of this vibration is determined by experimentally evaluating the degree of superimposed interference between liquid surface ripples and bubble movements at different frequencies, with a recommended range of 1 to 10. The amplitude range is 0.5 to 3.0. The purpose of micro-oscillation is to accelerate the overflow of micro-leaks from valve gaps, thereby improving the visibility of detecting minute leaks.

[0039] The streaming video frames are compressed tensor-form video data output frame-by-frame by the machine vision acquisition system in a streaming manner. The frame rate range is determined through experimental analysis of the bubble rising speed and the minimum detectable bubble size, with a recommended range of 60–120. .

[0040] The swarm intelligence lightweight topology dynamic evolution model is based on deep learning. Its specific structure is as follows: the input layer receives compressed tensors from streaming video frames; the initial network topology is a highly sparse graph structure; the initial connection weights between nodes are generated by the local probability function of the Cuckoo Search algorithm; the Cuckoo Search algorithm calculates the activation probability of connections between nodes based on the bubble distribution density map of the current frame; nodes with probabilities higher than the activation threshold establish initial connection edges; internally, a distributed collaborative parallel inference layer replaces the traditional fully connected layer; each inference layer node is abstracted as an independent intelligent agent with a state memory trace decay mechanism; after each inference, the node updates its own state memory trace, and the memory trace is determined according to the decay equation. decays hourly, where for The value of memory traces at any given moment. The attenuation coefficient is... The values ​​are obtained by comparing different The bubble tracking accuracy of the model on the validation set was determined through experimental comparison and analysis, and the recommended range is 0.01–0.10. Inter-node feature propagation is based on flow dynamic threshold collaborative logic. When the bubble feature flow rate in a certain region increases, the local activation operator triggers dynamic node splitting, splitting into child nodes to improve representation accuracy. The memory allocation of child nodes is dynamically divided proportionally based on the spatial information entropy value of the parent node's feature map, with regions with higher information entropy allocating more memory. Nodes in invalid background regions are automatically eliminated when their memory traces fall below the elimination threshold. The elimination threshold is determined by statistically analyzing the false activation rate of nodes in multiple batches of background interference experiments, with a recommended range of 0.05–0.20. Inter-layer data looping adopts a low-latency gated routing mechanism. The gated routing dynamically allocates the parallel computing memory share of each inference layer based on the bubble feature confidence score output by the previous layer. Low-degree feature flows are preferentially assigned to deeper inference nodes for secondary verification; the loss function incorporates state entropy increase dynamic constraints, using the total number of nodes in the current inference layer and the total number of connections between nodes as computational complexity penalty terms, which participate in iterative optimization together with bubble recognition accuracy, enabling the model to automatically converge to a lightweight topology while maintaining recognition accuracy; the network embeds a minimum spanning tree graph theory algorithm to reconstruct the optimal connected path of the sparse graph topology after each node split or disappearance, ensuring the global shortest feature transmission path between nodes in the distributed inference layer. The minimum spanning tree algorithm is executed after constructing a weighted graph based on the feature similarity matrix between nodes in the current frame; the output layer outputs the bubble position coordinate set and the overflow rate of each bubble in the current frame.

[0041] The steps for establishing the training dataset of the swarm intelligence lightweight topology dynamic evolution model specifically include: collecting multiple batches of videos of water tank bubbles under different air pressures and different micro-oscillation frequencies for liquid bag valves, and labeling the bubble positions, individual boundaries, overflow rates, and liquid surface ripple interference areas; performing streaming compression tensor processing on the collected videos to construct a training sample set containing typical scenarios such as bubble overlap, liquid surface interference, and microbubbles; and dividing the training set, validation set, and test set in an 8:1:1 ratio.

[0042] The specific steps for training the swarm intelligence lightweight topology dynamic evolution model include: initializing the sparse graph topology connection weights using the cuckoo search algorithm; inputting the training sample set into the distributed collaborative parallel inference layer in batches; calculating the loss value based on the loss function with embedded state entropy increase dynamic constraints; performing backpropagation to update the connection weights between nodes; synchronously executing the node dynamic splitting and extinction mechanism to update the topology structure; and performing the minimum spanning tree algorithm to reconstruct the connected paths after each topology change; iterating the training until the bubble recognition accuracy on the validation set converges, and then saving the model parameters.

[0043] The technical effects of the swarm intelligence lightweight topology dynamic evolution model are as follows: the network replaces the traditional dense fully connected structure with a sparse graph topology, so that the limited inter-node connection resources are concentrated in the bubble feature-dense region, fundamentally reducing the GPU memory and memory usage in edge computing scenarios; the dynamic splitting and demise mechanism enables the network topology to evolve in real time with bubble traffic, avoiding redundant computation in low-traffic frames of the traditional fixed architecture; the state entropy increase dynamic constraint incorporates the topology scale into the optimization objective, enabling the model to automatically find the optimal balance between recognition accuracy and computational complexity; the minimum spanning tree algorithm ensures that each node maintains optimal connectivity after dynamic topology changes, preventing feature transmission path breakage due to node demise, thereby achieving low-latency continuous and stable inference in high-concurrency streaming data scenarios.

[0044] Wherein, the topology dynamic adjustment function The node splitting rate parameter used to adjust the lightweight topology dynamic evolution model of swarm intelligence is calculated as follows: ,in The entropy of the bubble feature space information in the current frame. Average bubble overflow rate of the current frame (unit) ), Background interference intensity index, , , The weighting coefficients are determined through regression analysis of experimental data on the topology adjustment response speed of each parameter in different frame scenarios. Recommended ranges are as follows: , , The sum of the three is normalized to make (The units of measurement are dimensionless after normalization.) Dimensions and , , After normalization, they are all dimensionless; when When the node splitting rate parameter is reduced to a low level, the number of new nodes is reduced to maintain a lightweight topology; when At that time, the node splitting rate parameter was adjusted to the medium level, responding to changes in bubble characteristics at a conventional splitting frequency; when At that time, the node splitting rate parameter is increased to a high level to accelerate the generation of child nodes in order to cope with high-density bubble flow scenarios.

[0045] The background difference model is based on the state entropy increase dynamic constraint. It establishes a state entropy function on the pixel difference sequence between consecutive frames, distinguishes between low-frequency brightness disturbances caused by periodic ripples on the liquid surface and high-entropy value mutations caused by bubbles, and removes the low-frequency noise components of the ripples from the difference image by combining wavelet decomposition, outputting a clean feature map after pre-denoising. The state entropy increase dynamic constraint treats the pixel set of adjacent frame difference images as a thermodynamic state system, uses the information entropy of the pixel difference distribution as the state entropy measure, and constrains the background difference model to retain only pixel regions with significant entropy increase as effective foreground output.

[0046] The spatial non-maximum streaming suppression operator is a streaming frame-by-frame non-maximum suppression operator that performs streaming suppression on detected bubble candidate regions in the spatial domain based on the feature response intensity. For multiple candidate regions at the same spatial location, only the one with the strongest feature response is retained, thereby eliminating multiple detections caused by overlapping bubbles. The optical flow field topology dynamic evolution mechanism is to use optical flow estimation methods to calculate the motion vector field of each pixel between adjacent frames, construct the motion vector field into a graph topology, and identify the independent motion trajectory of each bubble individual based on the evolution law of the topology between adjacent frames. The multi-dimensional feature decoupling is to extract features from multiple dimensions such as spatial position, motion speed, and brightness gradient of the overlapping bubble region and model them independently, thereby restoring the aliased multi-bubble features into their own independent individual feature representations.

[0047] The multi-scale pyramid cascaded network is a network structure that extracts features from the input feature map at multiple resolution scales and fuses them step by step to align the features of bubbles of different sizes. The spatial self-organizing evolution operator is an operator that adaptively adjusts the feature weight allocation based on the spatial distribution of the micro-target region in the feature map, compensating and amplifying the feature weights of the micro-bubble target region to improve the detection rate of micro-bubbles. The streaming weight compensation is a mechanism that continuously enhances the weights of the micro-target region by predicting the weight compensation region of the current frame based on the position of the micro-bubbles in the previous frame during streaming frame-by-frame inference.

[0048] The Navier-Stokes fluid dynamics constraint update algorithm uses the Navier-Stokes momentum equation from fluid mechanics as the physical constraint embedded in the loss function of the swarm intelligence lightweight topology dynamic evolution model. During each backpropagation, it verifies whether the bubble motion satisfies the fluid momentum conservation law based on the bubble edge velocity vector extracted by machine vision in the current frame, the viscous drag parameters of the surrounding water, and the buoyancy force on the bubble. Motion features that do not satisfy the conservation law are judged as non-bubble disturbances caused by water tank oscillations and are offset from the loss function gradient, retaining only the motion features of real leaking bubbles that satisfy the conservation law for weight update; where the water viscosity... The hysteresis parameters were determined by performing particle image velocimetry experiments on the flow field of the test tank at different micro-oscillation frequencies and fitting the viscosity coefficient. The technical effect of the algorithm is that by directly constraining the physical motion law of bubbles to the model weight update process, the convergence direction of the network parameters is always consistent with the fluid dynamics behavior of real bubbles. This fundamentally suppresses the contamination of network parameters by pseudo-features caused by non-bubble factors such as liquid surface ripples and water oscillations. The model naturally obtains robustness to fluid disturbances without relying on additional data augmentation, ensuring high-precision continuous tracking of real leaking bubble flow in a continuous micro-oscillation environment.

[0049] The airtightness determination threshold is the upper limit for determining the number or rate of bubble overflow per unit time. Its specific value is obtained by conducting a comparative experiment on a liquid bag valve with known good airtightness and a liquid bag valve with known micro-leakage of different levels under the same standard air pressure and micro-oscillation conditions, statistically analyzing the distribution range of bubble overflow rate of the two types of samples and determining the optimal classification boundary.

[0050] Optionally, the present invention also provides a computer-based implementation method to form a valve sealing performance detection system based on underwater bubble machine vision. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the steps in the above method and work in conjunction with the machine vision acquisition system and the water tank device that applies micro-vibration to realize the automatic detection and result output of the sealing performance of the liquid bag valve.

[0051] The specific implementation of step S01 is as follows: Connect the valve of the liquid bag to be tested to the inflation device, and fill the valve with the rated working air pressure as the standard air pressure. The standard air pressure is determined by conducting multiple rounds of gradient pressurization experiments on the same batch of liquid bag valves, and the recommended range is 0.05 to 0.30. After inflation, completely immerse the liquid bag valve in a sealed water tank filled with clean water, ensuring that all potential leakage points of the valve are covered by water. Then, activate the periodic mechanical vibration device at the bottom of the water tank, applying a frequency ranging from 1 to 10. The amplitude range is 0.5 to 3.0. The micro-oscillation, in its physical function, applies periodic pressure differential disturbance to the gas at the valve gap, causing tiny leak bubbles to escape from the gap more rapidly, thus improving the visibility of bubble formation under micro-leakage conditions. Simultaneously, a machine vision acquisition system is activated, operating at 60–120... The frame rate is used to stream video of the water tank, and the streamed video frames are output in compressed tensor form for subsequent model processing.

[0052] The specific implementation of step S02 is as follows: The compressed tensor of the streaming video frames is input frame by frame into the input layer of the swarm intelligence lightweight topology dynamic evolution model. The initial network topology is a highly sparse graph structure. The initial connection weights between nodes are generated by the local probability function of the Cuckoo Search algorithm. The Cuckoo Search algorithm calculates the activation probability of the connections between each node based on the bubble distribution density map of the current frame. Node pairs with probabilities higher than the activation threshold establish initial connection edges, so that the initial topology resources are concentrated in the dense bubble feature region. During the distributed collaborative parallel inference layer processing, background subtraction and wavelet decomposition pre-denoising are performed simultaneously. The background subtraction model is based on the dynamic constraint of increasing state entropy. It regards the pixel set of the difference map of adjacent frames as a thermodynamic state system. The information entropy of the pixel difference distribution is used as the state entropy measure. The low-frequency brightness disturbance caused by liquid surface ripples corresponds to a low entropy value, and the sudden brightness jump caused by real bubbles corresponds to a high entropy value. The model only retains the pixel region with significant entropy increase as the effective foreground output. Subsequently, wavelet decomposition is combined to perform multi-scale decomposition on the difference image. The low-frequency noise component of the ripple is removed from each scale component and reconstructed, outputting a clean feature map after pre-denoising, thus eliminating the interference of false bright spots on the liquid surface ripple from an information theory perspective.

[0053] The specific implementation of step S03 is as follows: The feature map after pre-denoising enters the spatial domain non-maximum streaming suppression operator. The operator performs streaming frame-by-frame non-maximum suppression on the detected bubble candidate regions in the spatial domain based on the feature response intensity. For multiple candidate regions at the same spatial location, only the one with the strongest feature response is retained, eliminating multiple detections caused by overlapping bubbles. Subsequently, the dynamic evolution mechanism of the optical flow field topology is initiated. The motion vector field of each pixel between adjacent frames is calculated using the optical flow estimation method. The motion vector field is constructed as a graph topology structure. Based on the evolution law of the topology structure between adjacent frames, the independent motion trajectory of each bubble individual is identified, distinguishing overlapping bubbles from the motion dimension. On this basis, multi-dimensional feature decoupling is performed on the overlapping bubble regions. Features are extracted and independently modeled from three dimensions: spatial position, motion velocity, and brightness gradient. The aliased multi-bubble features are restored to their respective independent individual feature representations, providing accurate single-unit feature input for subsequent scale alignment and rate calculation.

[0054] The specific implementation of step S04 is as follows: The decoupled individual bubble features are input into a multi-scale pyramid cascade network. The network extracts features from the input feature map at multiple resolution scales and fuses them step-by-step, achieving multi-resolution alignment of features for bubbles of different sizes. This solves the problem of mismatched representation between large and small bubbles in the same feature layer. After multi-resolution alignment, the spatial self-organizing evolution operator adaptively adjusts the feature weight allocation based on the spatial distribution of the small target region in the feature map, compensating and amplifying the feature weights of the small bubble target region. The compensation mechanism adopts a streaming weight compensation method, that is, predicting the weight compensation region of the current frame based on the position of the small bubble in the previous frame, continuously enhancing the weights of the small target region to prevent small bubbles from being missed during inference due to weak feature response. Finally, the output is the set of position coordinates of each bubble in the current frame and the corresponding overflow rate.

[0055] The specific implementation of step S05 is as follows: The bubble motion features output from step S04 are input into the Navier-Stokes fluid dynamics constraint update algorithm. The algorithm uses the Navier-Stokes momentum equation as the physical constraint embedded in the loss function of the swarm intelligence lightweight topology dynamic evolution model. During each backpropagation, based on the bubble edge velocity vector extracted by machine vision in the current frame, the viscous drag parameters of the surrounding water, and the buoyancy force on the bubble, it verifies whether the bubble motion satisfies the fluid momentum conservation law. The viscous drag parameters of the water are determined by performing particle image velocimetry experiments on the flow field of the test tank at different micro-oscillation frequencies and fitting the viscosity coefficient. Motion features that do not satisfy the conservation law are judged as non-bubble disturbances caused by tank oscillation, and their corresponding gradient contributions to the loss function are offset and discarded. Only the real leakage bubble motion features that satisfy the conservation law are retained to participate in the network weight update, so that the convergence direction of the model parameters is always consistent with the fluid dynamic behavior of the real bubbles. Meanwhile, the dynamic splitting and extinction mechanism of nodes continuously updates the topology of the sparse graph, and after each topology change, the minimum spanning tree graph theory algorithm is executed to construct a weighted graph based on the feature similarity matrix between nodes in the current frame and reconstruct the optimal connected path to ensure that the feature transmission path between each node is minimized globally.

[0056] The specific implementation of step S06 is as follows: Based on the set of bubble position coordinates and overflow rate output in step S04, the number and rate distribution of bubbles overflowing from different parts of the liquid bag valve per unit time are statistically analyzed and compared with a pre-determined airtightness judgment threshold determined through comparative experiments. The airtightness judgment threshold is obtained by conducting comparative experiments on liquid bag valves with good airtightness and liquid bag valves with different levels of micro-leakage under the same standard air pressure and micro-oscillation conditions, statistically analyzing the bubble overflow rate distribution range of the two types of samples, and determining the optimal classification boundary. When the detected bubble overflow rate exceeds the airtightness judgment threshold, the system outputs the leakage detection result and marks the leakage location coordinates; when the overflow rate is lower than the threshold, the liquid bag valve is judged to be airtight.

[0057] It should be noted that the key technologies of this invention include: a background difference pre-denoising technique based on state entropy increase dynamic constraints, the core of which lies in utilizing the information theory-based essential difference between low-entropy disturbances of liquid surface ripples and high-entropy mutations of bubbles, fundamentally separating interference signals from valid signals at the feature extraction stage, so that the subsequent inference layer obtains pure bubble feature input, breaking through the limitation of high false detection rate of traditional background difference methods in dynamic water surface environments; and a physical constraint loss function technique based on the Navier-Stokes momentum equation, the core of which lies in directly encoding the hydrodynamic law of bubble rising motion into the model's weight update path, so that the network parameters naturally exclude non-bubble disturbance features that do not satisfy the physical conservation law during training and inference, and obtain robustness to fluid interference without relying on additional data augmentation; the above two technologies form a defense-in-depth collaborative mechanism in this invention, the former constructing the first filter at the feature extraction level, and the latter constructing the second constraint at the parameter optimization level. The two mechanisms complement each other, enabling the model to continuously and accurately track real leaking bubbles in the dual interference environment of liquid surface ripples and water oscillations, and achieve reliable determination of airtightness.

[0058] It should be noted that in the mass production scenario of liquid bag valve inspection, multiple bubbles may simultaneously overflow densely in the water tank, and some bubbles may approach each other during their ascent, forming an overlapping appearance. This makes it difficult for machine vision systems to accurately separate the overlapping bubbles into independent individuals and calculate their respective overflow rates and locations. The reason for this technical problem is that when bubbles move upward in the water driven by buoyancy, the liquid film between adjacent bubbles will cause the boundaries of the two bubbles to merge in the image under the action of surface tension, forming a single connected region. Traditional target detection methods identify the merged region as a single target, thus losing the independent position and rate information of each bubble, resulting in a significant decrease in the accuracy of leak location positioning and leak rate calculation. The usual solution to the above technical problem is to apply fixed morphological segmentation or ellipse fitting separation to the overlapping region. However, these methods rely on prior assumptions about bubble morphology. When bubbles deform due to water flow disturbances caused by micro-oscillations, the segmentation results of the fixed morphological model show serious deviations and cannot maintain temporal consistency in streaming frame-by-frame inference, further introducing cross-frame tracking errors. This invention effectively solves this technical problem. It first eliminates multiple candidate detections at the same location using a spatial non-maximum flow suppression operator. Then, it utilizes the dynamic evolution mechanism of the optical flow field topology to separate the independent motion trajectories of overlapping bubbles at the motion vector level. Optical flow estimation does not rely on static morphological assumptions about the bubbles but distinguishes individual bubbles based on the differences in the actual motion directions of pixels between adjacent frames. Even if bubble boundaries overlap spatially, as long as there are slight differences in the motion direction or velocity of each bubble, the optical flow field topology can distinguish them as independent trajectory nodes at the graph level. Furthermore, by combining multi-dimensional feature decoupling, the overlapping region is independently modeled from three dimensions: spatial location, motion velocity, and brightness gradient. This allows the individual features of each bubble to be separated in the feature space, providing accurate individual feature inputs for subsequent overflow rate calculation, thus achieving reliable separation and independent tracking of overlapping bubbles.

[0059] Specifically, the principle of this invention is:

[0060] The core reason why this invention can solve the above-mentioned technical problems is that liquid surface ripples and real bubbles have essential differences in two dimensions: information entropy characteristics and fluid dynamic characteristics. The solution of this invention utilizes the differences in these two dimensions to construct a dual filtering mechanism.

[0061] First, surface ripples are low-frequency periodic brightness disturbances formed at the interface of water bodies driven by periodic mechanical vibrations. Their pixel differences exhibit regular low-amplitude oscillations in the time domain, corresponding to low and periodic information entropy. In contrast, when real bubbles overflow from valve gaps, they cause sudden high-amplitude jumps in brightness in local image regions, with significantly higher information entropy than ripple disturbances. The background difference model treats the pixel set of adjacent frame difference images as a thermodynamic state system, using the information entropy of pixel difference distribution as a measure of state entropy. Only pixel regions with significant entropy increases are retained as effective foreground output. From an information theory perspective, this separates low-entropy ripple disturbances from high-entropy bubble features. Furthermore, wavelet decomposition is used to further remove low-frequency noise components of the ripples from the difference image, outputting a clean pre-denoised feature map, thereby eliminating the interference of surface ripples during the feature extraction stage.

[0062] Secondly, even if some ripples and bright spots pass the initial denoising stage, the Navier-Stokes fluid dynamics constraint update algorithm provides a second filtering mechanism. Real bubbles in water are driven upward by buoyancy, and the relationship between their edge velocity vector and the viscous drag of the surrounding water and the buoyancy acting on the bubble strictly satisfies the Navier-Stokes momentum conservation equation. However, the bright spots generated by surface ripples in the image move horizontally, which does not satisfy the vertical upward momentum conservation law. Therefore, they are judged as non-bubble perturbations and are offset from the loss function gradient, ensuring that the model weight update is driven only by the motion characteristics of real bubbles.

[0063] The two mechanisms are logically complementary: pre-denoising suppresses interference at the feature extraction level, while physical constraints prevent interference from contaminating network parameters at the model optimization level. This enables the swarm intelligence lightweight topology dynamic evolution model to maintain high-precision tracking of real leaking bubbles in a continuous micro-oscillation environment, thereby reliably achieving airtightness determination.

[0064] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0065] The specific implementation of step S01 is as follows: Inflate the liquid bag valve with standard air pressure. The standard air pressure is taken as the rated working air pressure of the same batch of liquid bag valves, and the recommended range is 0.05–0.30. The specific values ​​were determined through multiple rounds of gradient pressurization experiments combined with data on leakage frequency and detection sensitivity. After placing the valve in the water tank, periodic mechanical vibration was applied to the tank, with a recommended vibration frequency range of 1–10. The recommended amplitude range is 0.5–3.0. The micro-oscillation accelerates the overflow of micro-leaking bubbles from the valve gaps. The machine vision acquisition system is activated, operating at a speed of 60–120... The frame rate is used to capture streaming video frames, and each frame is output as a compressed tensor.

[0066] The specific implementation of step S02 is as follows: The compressed tensor of the streaming video frames is input into the swarm intelligence lightweight topology dynamic evolution model. After processing by sparse graph topology initialization and distributed collaborative parallel inference layer, background difference pre-denoising and wavelet decomposition denoising are performed simultaneously. The background difference model establishes a state entropy function for the sequence of pixel differences between consecutive frames based on the state entropy increase dynamic constraint. Let the first... Frame and the The set of pixel differences of the frame is The information entropy (i.e., state entropy) of its pixel difference distribution. The formula is expressed as follows:

[0067] ;

[0068] In the formula, For the first The first frame in the difference graph The probability density of the interval of pixel difference. For the first The state entropy of the frame difference graph It is the natural logarithm. This is the index for the pixel difference range. When... Exceeding the entropy increase threshold If the specified pixel area is selected as the valid foreground, it is considered a low-frequency disturbance of the liquid surface ripples and is therefore removed. The boundary between the state entropy distribution of liquid surface ripples and real bubbles was determined by statistically analyzing multiple batches of background interference experiments. Wavelet decomposition further separated the low-frequency noise components of the ripples from the difference map, outputting a clean feature map after pre-denoising.

[0069] The specific implementation of step S03 is as follows: The feature map after pre-denoising is processed by a spatial non-maximum streaming suppression operator. For multiple candidate regions at the same spatial location, only the one with the strongest feature response is retained, eliminating multiple detections of overlapping bubbles. Subsequently, optical flow estimation is used to calculate the motion vector field of each pixel between adjacent frames. The motion vector field is constructed into a graph topology structure, and the independent motion trajectory of each bubble is identified according to the topological evolution law. For overlapping bubble regions, features are extracted from multiple dimensions such as spatial position, motion speed, and brightness gradient, and modeled independently to restore the features of each bubble.

[0070] The specific implementation of step S04 is as follows: the decoupled bubble features are extracted and fused step by step at multiple resolution scales through a multi-scale pyramid cascade network to achieve alignment of bubble features of different sizes. The spatial self-organizing evolution operator adaptively adjusts the feature weights according to the spatial distribution of the micro-target region, performs streaming weight compensation on the micro-bubble region, predicts the compensation region of the current frame based on the micro-bubble position of the previous frame, continuously enhances the micro-target weights, and outputs the bubble position coordinate set and the overflow rate of each bubble.

[0071] The specific implementation of step S05 is as follows: The Navier-Stokes fluid dynamics constraint update algorithm embeds the Navier-Stokes momentum equation as the physical constraint into the loss function. Let the formula for the resultant force constraint acting on the bubble moving in water be expressed as follows:

[0072] ;

[0073] In the formula, Water density ( ), The velocity vector of the bubble ( ), For time ( ), For water pressure field ( ), The dynamic viscosity coefficient of water ( ), For the Laplace operator, The volume force term of buoyancy on the bubble ( ), For the fluid momentum conservation error term ( ), all dimensions are This ensures that both sides of the equation are consistent. The viscosity coefficient of the flow field was obtained by fitting it at different micro-oscillation frequencies through particle image velocimetry experiments. Motion features that do not satisfy the law of conservation of momentum are identified as non-bubble disturbances and are removed from the loss function gradient, retaining only the motion features of real leaking bubbles for weight updates.

[0074] The specific implementation of the swarm intelligence lightweight topology dynamic evolution model is as follows: The input layer receives compressed tensors of streaming video frames, and the initial network topology is a highly sparse graph structure. The initial connection weights between nodes are generated by the local probability function of the cuckoo search algorithm, and the activation probability of the connections between each node is calculated based on the bubble distribution density map of the current frame. The formula is expressed as follows:

[0075] ;

[0076] In the formula, For nodes With nodes The activation probability of the connection between them (dimensionless). For nodes With nodes Bubble distribution density in the corresponding region (dimensionless normalized value). For nodes The sum of the connection densities with all adjacent nodes, , For node indexing, here For nodes The traversal indexes of all adjacent nodes, and the indexes in step S02 The meanings are different, so they are distinguished by the subscript area. Nodes above the activation threshold establish initial connection edges. The decay formula for the state memory traces of each inference layer node is expressed as follows:

[0077] ;

[0078] In the formula, for The memory trace value (dimensionless) at each time point. The memory trace value of the previous time step (dimensionless). The attenuation coefficient ( The recommended range is 0.01 to 0.10. Through different The bubble tracking accuracy of the model on the validation set was determined through experimental comparison and analysis under the given values. It is a natural constant. When the bubble characteristic flow rate in a certain region increases, the local activation operator triggers dynamic node splitting, and the memory allocation of child nodes is based on the spatial information entropy of the parent node's feature map. Dynamically divided according to proportion. Higher memory allocation is applied to regions with higher memory density. Memory traces of nodes in invalid background regions are below the extinction threshold. (Recommended range: 0.05–0.20) It will automatically disappear. The false activation rate of nodes was determined after statistical analysis of multiple batches of background interference experiments. The loss function incorporates a dynamic constraint of increasing state entropy, which is applied to the total number of nodes in the current inference layer. Total number of connections between nodes As a penalty for computational complexity, it is related to the accuracy of bubble recognition. They jointly participate in iterative optimization, loss function The formula is expressed as follows:

[0079] ;

[0080] In the formula, This represents the total loss value (dimensionless). The basic loss (dimensionless) for bubble detection. This is the complexity penalty coefficient (dimensionless). The accuracy bonus coefficient (dimensionless). This represents the total number of nodes in the current inference layer (a dimensionless integer). This represents the total number of connections between nodes (a dimensionless integer). The bubble recognition accuracy (dimensionless, ranging from 0 to 1). The loss error term is dimensionless, and all terms are normalized to ensure consistent dimensions on both sides of the equation. The minimum spanning tree algorithm is based on the feature similarity matrix between nodes in the current frame. After constructing a weighted graph, the optimal connected path is reconstructed. The formula is expressed as follows:

[0081] ;

[0082] In the formula, For nodes With nodes Feature similarity weights (dimensionless) between features. Given the total number of nodes in the current frame, the minimum spanning tree algorithm uses... As input, ensure that the feature propagation path between nodes is the shortest globally.

[0083] Topology dynamic adjustment function The formula is expressed as follows:

[0084] ;

[0085] In the formula, The entropy of the bubble feature space in the current frame is dimensionless. The dimensionless value of the average bubble overflow rate of the current frame after normalization (the original dimension is...). (dimensionless after normalization) The background interference intensity index (dimensionless). , , The weighting coefficients (dimensionless) satisfy the following conditions: The recommended ranges are as follows: , , The parameters were determined through regression analysis of experimental data on the topology adjustment response speed under different frame scenarios. It is a dimensionless value. When When the node splitting rate parameter drops to a low level; when When, adjust to medium setting; when At that time, the system is upgraded to a higher level to accelerate the generation of child nodes in order to cope with high-density bubble flow scenarios.

[0086] The specific implementation of step S06 is as follows: The airtightness of the liquid bag valve is determined based on the bubble overflow rate and location distribution. The airtightness determination threshold is obtained by comparing liquid bag valves known to have good airtightness and those known to have different levels of micro-leakage under the same standard air pressure and micro-oscillation conditions. The distribution range of bubble overflow rate for the two types of samples is statistically analyzed, and the optimal classification boundary is determined. When the overflow rate exceeds the airtightness determination threshold, the leakage detection result is output.

[0087] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0088] This embodiment uses the airtightness test of a batch of medical liquid bag valves as a scenario. The method described in this invention is used to determine the airtightness of the valves in batches, and to verify the method's ability to detect real leaking bubbles in a micro-vibration environment.

[0089] In this embodiment, 60 medical liquid bag valves from the same batch were selected. Among them, 45 samples were known to have passed the airtightness test, and 15 samples were known to have micro-leakage of different levels. The micro-leakage levels were divided into three categories: slight leakage, moderate leakage, and significant leakage, with 5 samples in each category. The valves to be tested were sequentially filled with 0.15... The standard air pressure, determined through gradient pressurization experiments on this batch of valves, strikes a balance between the visibility of air bubbles in micro-leaks and the structural safety of the valves. The pressurized valves are then completely immersed in a 600mm water tank. ×400 ×300 The water temperature was maintained at 20℃, and the water viscous drag parameter was determined to be 0.00102 after fitting the data through a particle image velocimetry experiment. The frequency of the water tank is 5. The amplitude is 1.5. The machine vision acquisition system captures micro-oscillations at 90 degrees. Frame rate is used for streaming video capture.

[0090] When establishing the training dataset for the swarm intelligence lightweight topology dynamic evolution model, data from this batch of valves at 0.10 were collected. 0.15 0.20 Three types of air pressure and 3 5 7 Videos of water tank bubbles under three combinations of micro-oscillation frequencies were used to annotate bubble positions, individual boundaries, overflow rates, and liquid surface ripple interference areas. A training sample set was constructed including typical scenarios such as bubble overlap, liquid surface interference, and microbubbles, and divided into training, validation, and test sets in an 8:1:1 ratio. During model training, the attenuation coefficient... Set to 0.05 The extinction threshold is set to 0.10, and the weighting coefficient is... , , The sum of the three is normalized to 1.

[0091] like Figure 2The figure shows the curve of bubble recognition accuracy on the validation set as the number of training iterations increases during the training process of the swarm intelligence lightweight topology dynamic evolution model in this embodiment, as well as the convergence trend of the corresponding state entropy increase constraint loss value. From Figure 2 It can be observed that in the early stages of training, the model has many false detections of liquid surface ripples and bright spots, resulting in low recognition accuracy. This is because the state entropy increase constraint of the background difference model has not yet fully separated low-entropy ripple disturbances from high-entropy bubble mutations. As the number of iterations increases, the loss value of the state entropy increase constraint continuously decreases and tends to stabilize. The model's ability to filter out liquid surface ripples gradually improves, and the bubble recognition accuracy increases accordingly, eventually converging to a stable level on the validation set. This trend indicates that the introduction of the state entropy increase dynamic constraint enables the model to spontaneously distinguish signal categories with different information entropy characteristics during the iteration process, rather than relying on manually set fixed filtering thresholds, thereby forming an adaptive false bright spot removal capability in dynamic interference environments.

[0092] The airtightness threshold was determined through comparative experiments. The bubble overflow rates of 45 qualified samples and 15 slightly leaking samples were statistically analyzed under the same standard pressure and micro-oscillation conditions. The overflow rate distribution ranges of the two types of samples are shown in Table 1.

[0093] Table 1. Distribution of bubble overflow rate between qualified samples and samples with minor leaks

[0094]

[0095] As shown in Table 1, there is a significant gap between the overflow rate distribution intervals of qualified samples and samples with minor leaks at each level. The optimal classification boundary was determined by statistically analyzing the intersection points of the overflow rate distributions of the two types of samples. The threshold for minor leaks was set at 0.30. .

[0096] The test results of 60 samples are summarized in Table 2:

[0097] Table 2 Summary of Liquid Bag Valve Tightness Test Results

[0098]

[0099] As shown in Table 2, 59 out of 60 samples were correctly judged. One qualified sample was mistakenly judged as having a slight leak due to a brief bubble-like bright spot caused by local eddies generated by micro-vibration at the bottom of the water tank. This misjudged sample was corrected in the re-inspection by increasing the airtightness judgment threshold.

[0100] like Figure 3As shown, this is a scatter plot of the distribution of the information entropy output value of the background difference state entropy increase constraint as a function of the frame number during the detection process of 60 samples in this embodiment. The horizontal axis represents the frame number, and the vertical axis represents the state entropy value of the current frame difference map. Qualified sample frames and slightly missed sample frames are distinguished by different labels. From Figure 3 It can be observed that the state entropy values ​​of qualified sample frames are concentrated in the low-entropy region, consistent with the low-frequency disturbance characteristics of liquid surface ripples. However, in micro-leak sample frames, the state entropy values ​​show significant jumps at the moment of bubble overflow, with the jump amplitude increasing with the micro-leak level. Slightly leaking samples show smaller jump amplitudes and shorter durations, while significantly leaking samples show larger jump amplitudes and longer durations. This distribution pattern indicates that the state entropy increase dynamic constraint can effectively capture the information entropy mutation characteristics of bubble overflow events without triggering false detections due to periodic disturbances of liquid surface ripples, thus achieving reliable differentiation between real bubbles and false bright spots at the physical mechanism level. Based on this, the Navier-Stokes fluid dynamics constraint further verifies the momentum conservation of candidate features passing the entropy sieve, ensuring that the features ultimately participating in the weight update all come from real leakage events that satisfy the physical laws of bubble rise. The cascading of these two constraints ensures the model's resistance to complex disturbances under micro-oscillation environments.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A valve tightness detection method based on machine vision for underwater air bubbles, characterized in that, Includes the following steps: Inflate the liquid bag valve with standard air pressure, place the liquid bag valve in the water tank, apply micro-vibration to the water tank, start the machine vision acquisition system, and acquire streaming video frames; The streaming video frames are input into the swarm intelligence lightweight topology dynamic evolution model. After processing by sparse graph topology initialization and distributed collaborative parallel inference layer, background difference and wavelet decomposition pre-denoising are performed simultaneously to remove false bright spot interference caused by liquid surface ripples. The feature map after pre-denoising is processed by the spatial non-maximum flow suppression operator, and combined with the dynamic evolution mechanism of optical flow field topology, multi-dimensional feature decoupling is performed on the overlapping bubble region to restore the individual features of each bubble. A multi-scale pyramid cascade network is used to perform multi-resolution alignment of the decoupled bubble features. A spatial self-organizing evolution operator is used to perform streaming weight compensation on the micro-bubble target, and the bubble position and overflow rate are output. The Navier-Stokes fluid dynamics constraint update algorithm verifies the fluid momentum conservation of bubble motion characteristics, eliminates non-bubble perturbation velocities, and updates the weights of the swarm intelligence lightweight topology dynamic evolution model. The airtightness of the liquid bag valve is determined based on the bubble overflow rate and location distribution. When the overflow rate exceeds the airtightness determination threshold, the leakage detection result is output.

2. The valve tightness detection method based on underwater bubble machine vision according to claim 1, characterized in that, The standard pressure is specifically obtained by conducting multiple rounds of gradient pressurization experiments on the same batch of liquid bag valves, combined with experimental data analysis of leakage frequency and detection sensitivity. The standard pressure range is 0.05 to 0.30 MPa.

3. The valve tightness detection method based on underwater bubble machine vision according to claim 2, characterized in that, The micro-oscillation specifically refers to applying periodic mechanical vibration to the water tank, with a vibration frequency range of 1 to 10 Hz, an amplitude range of 0.5 to 3.0 mm, and a streaming video frame rate range of 60 to 120 fps.

4. The valve tightness detection method based on underwater bubble machine vision according to claim 3, characterized in that, The aforementioned swarm intelligence lightweight topology dynamic evolution model specifically involves the input layer receiving compressed tensors of streaming video frames, the initial network topology being a sparse graph structure, the initial connection weights between nodes being generated by the local probability function of the cuckoo search algorithm, and the establishment of initial connection edges between node pairs with probabilities higher than the activation threshold. The network internally uses a distributed collaborative parallel inference layer to replace the traditional fully connected layer.

5. The valve tightness detection method based on underwater bubble machine vision according to claim 4, characterized in that, The distributed collaborative parallel inference layer is specifically defined as follows: each inference layer node is abstracted into an independent intelligent agent with a state memory trace decay mechanism, and the memory trace decays step by step according to the decay equation.

6. The valve tightness detection method based on underwater bubble machine vision according to claim 5, characterized in that, The sparse graph topology initialization specifically involves feature transmission between nodes based on flow dynamic threshold collaborative logic, local activation operators triggering dynamic node splitting, dynamic allocation of child node memory based on the spatial information entropy value of the parent node's feature map, and automatic elimination of invalid background region nodes when the memory trace is below the elimination threshold, with a recommended elimination threshold range of 0.05 to 0.

20.

7. The valve tightness detection method based on underwater bubble machine vision according to claim 6, characterized in that, The background difference and wavelet decomposition pre-denoising method is specifically based on the state entropy increase dynamic constraint. By establishing a state entropy function on the continuous inter-frame pixel difference sequence, the low-frequency brightness disturbance caused by liquid surface ripples and the high entropy value mutation caused by bubbles are distinguished. Combined with wavelet decomposition, the low-frequency noise component of the ripples is removed from the difference image.

8. The valve tightness detection method based on underwater bubble machine vision according to claim 7, characterized in that, The dynamic evolution mechanism of the optical flow field topology structure specifically involves using optical flow estimation methods to calculate the motion vector field of each pixel between adjacent frames, constructing the motion vector field into a graph topology structure, and identifying the independent motion trajectory of each bubble individual based on the evolution law of the topology structure between adjacent frames.

9. The valve tightness detection method based on underwater bubble machine vision according to claim 8, characterized in that, The multidimensional feature decoupling specifically involves extracting features from multiple dimensions such as spatial location, movement speed, and brightness gradient of the overlapping bubble region and modeling them independently, thereby restoring the aliased multi-bubble features into their respective independent individual feature representations.

10. The valve tightness detection method based on underwater bubble machine vision according to claim 9, characterized in that, The spatial self-organizing evolution operator specifically adjusts the feature weight allocation adaptively based on the spatial distribution of the micro-target region in the feature map, and compensates and amplifies the feature weight of the micro-bubble target region; the streaming weight compensation predicts the weight compensation region of the current frame based on the position of the micro-bubble in the previous frame, and continuously enhances the weight of the micro-target region.

Images