A method for kinetic testing of micron-scale particle collisions

Abstract

The present application relates to the field of boiler heating surface anti-ash and slagging and coal-fired power plant micron-level fine particulate matter removal, and provides a kind of micron particle collision dynamics test method and system.The method comprises: the continuous frame image collected by high-speed camera is pretreated;Lightweight target detection network is used to sub-pixel level positioning of micron-level particles in the image, and the complete motion trajectory of the particles is generated;The trajectory is used to identify the collision time, calculate the particle dynamics characteristic parameters and perform rationality check;Collect the temperature and humidity data of the test environment, construct the temperature and humidity and dynamics characteristic database, and mine the influence law of temperature and humidity on dynamics characteristics;Dynamic curve and visual chart are generated based on the dynamics characteristic parameters and output.The present application realizes high-precision detection and dynamics characteristic extraction of micron-level particle impact process, can effectively support the research of boiler heating surface ash and slagging mechanism, solves the removal problem of fine particulate matter in electrostatic precipitator, and has significant engineering practical value.

Description

Technical Field

[0001] This invention relates to the field of preventing ash and slag buildup on boiler heating surfaces and removing micron-sized fine particles from coal-fired power plants, specifically to a dynamic testing method and system for micron-sized particle collisions. Background Technology

[0002] Ash and slag buildup on boiler heating surfaces is a core bottleneck restricting the safe, stable, and economical operation of equipment in industries such as thermal power generation and chemical processing. Essentially, it involves a dynamic process of collision, adhesion, and rebound between 2-50 μm micron-sized particles in flue gas and the heating surface. Accurate analysis of the particle dynamics parameters in this process (such as collision velocity, angle, adhesion / rebound characteristics, etc.) is a crucial prerequisite for revealing the mechanism of ash and slag buildup and developing targeted prevention and control strategies.

[0003] Existing particle impact visual recognition technologies suffer from three major drawbacks: First, micron-level particle imaging is susceptible to interference from light scattering caused by temperature and humidity changes, uneven reflection from walls, and high-speed motion easily leads to blurring and adhesion. Traditional threshold segmentation and contour extraction methods have large positioning errors and cannot meet the requirements for sub-pixel accuracy. Second, recognition algorithms often focus on extracting a single parameter, failing to simultaneously acquire complete dynamic features such as impact velocity, angle, and coefficient of restitution, and do not incorporate environmental parameters such as temperature and humidity. Third, real-time performance is insufficient; mainstream algorithms have low frame rates and cannot capture the instantaneous process of high-speed particle impact, resulting in the loss of key collision features.

[0004] Therefore, there is an urgent need to develop high-precision, high-real-time visual recognition technology that is compatible with temperature and humidity control platforms to fill the gaps in existing technologies. Summary of the Invention

[0005] To address the problems of existing technologies, such as susceptibility to interference in micron-level particle imaging, low positioning accuracy, incomplete feature extraction, lack of environmental parameter correlation, and insufficient real-time performance, this invention provides a dynamic testing method and system for micron-level particle collisions. This method achieves high-precision and complete dynamic analysis and environmental correlation modeling of the impact process, providing effective data support for the research and engineering optimization of boiler ash and slagging prevention mechanisms.

[0006] In a first aspect, the present invention provides a dynamic testing method for collisions of micron-sized particles, comprising the following steps:

[0007] The continuous frame images acquired by the high-speed camera are preprocessed, including adaptive median filtering, multi-scale illumination enhancement, and adaptive binarization. A lightweight target detection network is used to perform sub-pixel localization of micron-level particles in preprocessed consecutive frame images, and Kalman filtering and Hungarian algorithm are combined to generate complete particle motion trajectories. Based on the complete motion trajectory of the particle, the collision time is identified, the particle dynamic characteristic parameters are calculated, and the physical rationality is verified. Collect temperature and humidity data of the test environment, associate the dynamic characteristic parameters with the temperature and humidity parameters, construct a database of temperature and humidity characteristics and dynamic characteristics, and explore the influence of temperature and humidity characteristics on dynamic characteristics. Dynamic curves and visualization charts are generated and output based on the aforementioned dynamic characteristic parameters.

[0008] Furthermore, the lightweight object detection network is a lightweight convolutional neural network based on the YOLOv8-nano object detection model, namely ShuffleNetV2.

[0009] Furthermore, the lightweight object detection network includes a feature enhancement branch with an attention mechanism to enhance the edge and texture features of micron-sized particles and suppress background interference.

[0010] Furthermore, the lightweight object detection network adds convolutional layers to the regression branch of the detection head and reduces the number of channels in the classification branch for single-class detection scenarios. At the same time, it introduces a dynamic weight calibration mechanism to dynamically adjust the classification confidence score of the predicted box based on the local grayscale and texture statistical features of the image region corresponding to the predicted box.

[0011] Furthermore, when identifying the collision moment: the position change rate of the particle between consecutive frames is calculated. When the position change rate changes abruptly and the distance between the particle center and the wall is less than 1 / 2 of the particle diameter, the current frame is determined to be a collision key frame.

[0012] Furthermore, the dynamic characteristic parameters include impact velocity, impact angle, adhesion rebound state, and recovery coefficient.

[0013] Furthermore, the method for calculating the impact velocity is as follows: trace back the trajectory data of several frames before the collision keyframe, use a polynomial to fit the motion curve and take the derivative to obtain the instantaneous velocity.

[0014] Furthermore, the acquisition frequency of the temperature and humidity data is synchronized with the image acquisition frame rate of the high-speed camera.

[0015] Secondly, the present invention provides a dynamic testing system for micron-level particle collisions, comprising a hardware module and a software module: The hardware module includes a high-speed camera, a high-brightness coaxial light source, an image acquisition card, an embedded computing terminal, a temperature and humidity sensor, and a data synchronization unit. The high-speed camera is used to acquire continuous frame images of particle motion, the temperature and humidity sensor is used to synchronously acquire temperature and humidity data of the test environment, and the data synchronization unit is used to keep the image acquisition time of the high-speed camera synchronized with the data acquisition time of the temperature and humidity sensor. The embedded computing terminal is used to provide a computing platform. The software module is a visual recognition program running on an embedded computing terminal, including an image preprocessing unit, a particle positioning and trajectory tracking unit, a dynamic feature extraction unit, a temperature and humidity correlation analysis unit, and a result output unit.

[0016] Furthermore, the image preprocessing unit is used to preprocess the consecutive frame images; Furthermore, the particle localization and trajectory tracking unit is used to call a lightweight target detection network to perform sub-pixel-level particle localization on the preprocessed image, and combine it with a trajectory tracking algorithm to generate the complete motion trajectory of the particles; Furthermore, the collision feature extraction unit is used to identify the collision moment based on the complete motion trajectory and calculate the dynamic characteristic parameters of the particles; Furthermore, the temperature and humidity correlation analysis unit is used to perform correlation analysis between the dynamic characteristic parameters and the synchronously collected temperature and humidity data, construct a database of temperature and humidity characteristics and dynamic characteristics, and explore the influence of temperature and humidity characteristics on dynamic characteristics. Furthermore, the result output unit is used to output the analysis results of the dynamic characteristic parameters and generate dynamic curves and visualization charts.

[0017] Furthermore, the high-speed camera has a maximum pixel size of 1024 × 1024 and a recording time of 4.85 seconds, a minimum pixel size of 256 × 256 and a recording time of 8.73 seconds; the temperature and humidity sensor has an accuracy of ±0.5℃ for temperature and ±1%RH for humidity.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention combines adaptive image preprocessing with an improved lightweight detection network to achieve accurate sub-pixel positioning of micron-level particles (error <0.1μm), effectively suppressing noise, adhesion and illumination changes, and achieving a detection accuracy of over 95%.

[0019] (2) This invention simultaneously extracts multiple dynamic parameters such as impact velocity, angle, recovery coefficient and adhesion / rebound state, and establishes a correlation model in combination with temperature and humidity environmental data, thus realizing a complete analysis of particle impact behavior.

[0020] (3) The system of the present invention has a frame rate of not less than 200fps, which can capture the instantaneous process of high-speed impact of 1~100m / s, meet the real-time requirements of dynamic testing, and effectively avoid the loss of key information.

[0021] (4) The present invention has lightweight deployment capability, can output structured parameters and visualization results in real time, directly supports the structural optimization and operation control of boiler heating surface to prevent ash accumulation and slagging, and is expected to significantly improve equipment thermal efficiency and operation safety.

[0022] (5) This invention is widely used in technical fields such as preventing ash and slag buildup on boiler heating surfaces and removing micron-sized fine particles in coal-fired power plants. It can effectively support the research on the mechanism of ash and slag buildup on boiler heating surfaces and eliminate fine particulate matter (PM2.5) in coal-fired flue gas within traditional electrostatic precipitators. 2.5 The existence of a "penetrating window" phenomenon has significant practical engineering value. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a dynamic testing method for collisions of micron-sized particles according to the present invention.

[0025] Figure 2 This is a framework diagram of a dynamic testing system for micron-level particle collisions according to the present invention.

[0026] Figure 3 This is a schematic diagram of the dynamic testing system architecture in an embodiment of the present invention.

[0027] Figure 4 This is a flowchart of the visual recognition method in an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of the YOLOv8-nano network structure in an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] This invention provides the following technical solutions: like Figure 1 The method shown here is a dynamic testing method for collisions of micron-sized particles, which specifically includes the following steps: S1. Preprocess the continuous frame images acquired by the high-speed camera.

[0032] In a preferred embodiment of this application, the preprocessing includes adaptive median filtering, multi-scale illumination enhancement, and adaptive binarization.

[0033] S2. A lightweight target detection network is used to perform sub-pixel localization of micron-level particles in the preprocessed continuous frame images, and Kalman filtering and Hungarian algorithm are combined to generate complete particle motion trajectories.

[0034] In a preferred embodiment of this application, the lightweight object detection network is a lightweight convolutional neural network based on the YOLOv8-nano object detection model, ShuffleNetV2.

[0035] In a preferred embodiment of this application, the lightweight object detection network includes a feature enhancement branch with an attention mechanism to enhance the edge and texture features of micron-sized particles and suppress background interference.

[0036] As a preferred embodiment of this application, the lightweight object detection network adds a convolutional layer to the regression branch of the detection head and reduces the number of channels in the classification branch for single-class detection scenarios. At the same time, a dynamic weight calibration mechanism is introduced to dynamically adjust the classification confidence score of the predicted box based on the local grayscale and texture statistical features of the image region corresponding to the predicted box.

[0037] S3. Based on the complete motion trajectory of the particle, identify the collision time, calculate the particle dynamic characteristic parameters, and perform physical rationality verification.

[0038] In a preferred embodiment of this application, when identifying the collision moment: the position change rate of the particle between consecutive frames is calculated, and when the position change rate changes abruptly and the distance between the particle center and the wall is less than 1 / 2 of the particle diameter, the current frame is determined to be a collision key frame.

[0039] In a preferred embodiment of this application, the dynamic characteristic parameters include impact velocity, impact angle, adhesion rebound state, and recovery coefficient.

[0040] As a preferred embodiment of this application, the method for calculating the impact velocity is as follows: trace back the trajectory data of several frames before the collision keyframe, use a polynomial to fit the motion curve and take the derivative to obtain the instantaneous velocity.

[0041] S4. Collect temperature and humidity data of the test environment, associate the dynamic characteristic parameters with the temperature and humidity parameters, construct a database of temperature and humidity characteristics and dynamic characteristics, and explore the influence of temperature and humidity characteristics on dynamic characteristics.

[0042] In a preferred embodiment of this application, the acquisition frequency of the temperature and humidity data is synchronized with the image acquisition frame rate of the high-speed camera.

[0043] S5. Generate and output dynamic curves and visualization charts based on the dynamic characteristic parameters.

[0044] like Figure 2The illustrated dynamic testing system for micron-level particle collisions includes a hardware module and a software module. The hardware module includes a high-speed camera, a high-brightness coaxial light source, an image acquisition card, an embedded computing terminal, a temperature and humidity sensor, and a data synchronization unit. The high-speed camera acquires continuous frame images of particle motion, the temperature and humidity sensor synchronously acquires temperature and humidity data of the test environment, and the data synchronization unit keeps the image acquisition time of the high-speed camera synchronized with the data acquisition time of the temperature and humidity sensor. The embedded computing terminal provides a computing platform. The software module is a visual recognition program running on the embedded computing terminal, including an image preprocessing unit, a particle positioning and trajectory tracking unit, a dynamic feature extraction unit, a temperature and humidity correlation analysis unit, and a result output unit.

[0045] In a preferred embodiment of this application, the image preprocessing unit is used to preprocess the consecutive frame images; In a preferred embodiment of this application, the particle localization and trajectory tracking unit is used to call a lightweight target detection network to perform sub-pixel-level particle localization on the preprocessed image, and combine it with a trajectory tracking algorithm to generate a complete particle motion trajectory. In a preferred embodiment of this application, the collision feature extraction unit is used to identify the collision moment based on the complete motion trajectory and calculate the dynamic characteristic parameters of the particles. In a preferred embodiment of this application, the temperature and humidity correlation analysis unit is used to perform correlation analysis between the dynamic characteristic parameters and the synchronously collected temperature and humidity data, construct a database of temperature and humidity characteristics and dynamic characteristics, and explore the influence of temperature and humidity characteristics on dynamic characteristics. In a preferred embodiment of this application, the result output unit is used to output the analysis results of the dynamic characteristic parameters and generate dynamic curves and visualization charts.

[0046] In a preferred embodiment of this application, the high-speed camera has a maximum pixel size of 1024 × 1024 and a recording time of 4.85 seconds, a minimum pixel size of 256 × 256 and a recording time of 8.73 seconds; the temperature and humidity sensor has an accuracy of ±0.5℃ for temperature and ±1%RH for humidity.

[0047] Example like Figure 3The diagram shows the architecture of the dynamic testing system. This system is compatible with a single-particle impact testing platform with controllable temperature (20~800℃, accuracy ±0.5℃) and humidity (5%~95%RH, accuracy ±1%RH). It consists of hardware modules and software modules (visual recognition program). The hardware modules include a PHOTRONFASTCAM Nova S9 high-speed camera, an LED point light source, an image acquisition card, an embedded computing terminal (NVIDIA Jetson AGX Orin), a temperature and humidity sensor (SHT35 model), and a data synchronization module. All hardware components are connected sequentially via a data bus and signal lines. The data synchronization module is connected to the camera and the temperature and humidity sensor to achieve time synchronization. Table 1 below shows the technical specifications of the PHOTRONFASTCAM Nova S9 high-speed camera.

[0048] Table 1 Technical Specifications of PHOTRONFASTCAM Nova S9 High-Speed ​​Camera

[0049] The high-speed camera is installed at an adjustable angle to the wall, and the coaxial light source faces the wall to eliminate shadows; the temperature and humidity sensor is installed 5cm away from the wall; the image acquisition card connects the camera to the embedded terminal, and the data synchronization module realizes time synchronization between the camera and the sensor.

[0050] The software module branch is a visual recognition program running on an embedded computing terminal, comprising five sub-modules: image preprocessing, particle localization and trajectory tracking, collision feature extraction, temperature and humidity correlation analysis, and result output. These sub-modules interact sequentially via data interfaces and are synchronized with the hardware module through a data synchronization module. The visual recognition program is developed using Python and PyTorch, with preprocessing utilizing the OpenCV library. An improved YOLOv8-nano model is deployed in a lightweight ONNX format on the embedded terminal.

[0051] Based on the above dynamic testing system, such as Figure 4 As shown, the specific steps of the visual recognition method are as follows: S1. Image preprocessing.

[0052] The visual recognition program receives consecutive frame images captured by a high-speed camera and performs three levels of preprocessing: S11, Adaptive Median Filtering: Dynamically adjusts the 3×3~7×7 filter window according to the regional noise intensity to remove background and particle imaging noise.

[0053] S12, Multi-scale Retinex (classic image enhancement algorithm) enhancement: Corrects uneven lighting caused by temperature and humidity, and improves the contrast between particles and background.

[0054] S13, OTSU (Maximum Inter-Class Variance) Adaptive Binarization: Combines local gray-level mean to dynamically adjust the threshold, solving the problem of particle adhesion and segmentation failure.

[0055] S2, Particle positioning and trajectory tracking.

[0056] S21. Particle Detection: The improved YOLOv8-nano network is used. The original C2f backbone module of YOLOv8-nano is removed and replaced with the lightweight bottleneck module of ShuffleNetV2 (Lightweight Convolutional Neural Network). While keeping the overall number of network stages unchanged, the sampling rhythm of the original model is adapted and the number of parameters is reduced by 50%. A micron-level particle feature enhancement branch is added to the detection head to extract edge texture features, achieving sub-pixel-level localization of 2~50μm particles (error <0.1μm).

[0057] S22. Trajectory Tracking: Kalman filter predicts the particle's position in the next frame, and Hungarian algorithm is used to complete inter-frame matching, generating complete trajectories before, during and after the impact, thus avoiding target loss due to high-speed motion.

[0058] like Figure 5 The diagram shown is a schematic of the improved YOLOv8-nano network structure. This network structure is a hierarchical block diagram, with the following layers from the input layer to the output layer: The lightweight target detection model is based on the open-source YOLOv8-nano target detection model from Ultralytics, with appropriate improvements. Addressing core issues such as "missed detection of small targets, limited lightweight deployment, and insufficient positioning accuracy" in the dynamics test scenario of micron-sized particles impacting wall surfaces, the network hierarchy structure has been redesigned. Specific improvements are as follows: First layer: Input layer (adapted for high-speed camera, high-resolution micron-sized particle image preprocessing) The YOLOv8-nano input layer only implements fixed-size pixel normalization and batch dimensionality expansion, which fails to adapt to high-resolution micron-sized images. This can lead to issues such as forced scaling and pixel range deviation, resulting in the loss of small particle details. To address this, for preprocessed 1024 × 1024 × 3 color images, an adaptive size verification module is added to avoid feature distortion caused by forced scaling. Furthermore, an optimized pixel normalization strategy is adopted, using mean 0.485 and variance 0.229 normalization parameters adapted to micron-sized scenes, replacing the default COCO dataset normalization parameters. This ensures lossless input of high-resolution micron-sized images, avoids decreased detection accuracy due to pixel range deviation and forced size adjustment, and improves small particle detail retention by 15%, laying the foundation for subsequent feature extraction.

[0059] Second layer: ShuffleNetV2 backbone network (lightweight + multi-scale feature hierarchical extraction) The YOLOv8-nano backbone network consists of four Stage C2f modules (bottlenecks in the cross-stage part of the second convolution) (with 3, 6, 6, and 3 blocks respectively). It is based on ordinary convolution + residual connection design, with parameters accounting for 60% of the total parameters of the model. However, it does not retain enough shallow features and cannot adapt to the weak feature extraction requirements of micron-sized particles. In the lightweight model, all C2f modules were replaced with lightweight ShuffleNetV2 modules. The channel splitting ratio within each module was adjusted to 1:1 (feature branches and direct connection branches each accounted for 50%), and the number of blocks in each stage was optimized to 2, 4, 4, and 2, achieving a lightweight effect with half the parameters. Furthermore, the feature extraction objectives for each stage were clearly defined: Stage 1 (2 blocks) achieved initial feature extraction; Stage 2 (4 blocks) focused on micron-level particle detail features (edges, pixel-level textures); Stage 3 (4 blocks) extracted particle contour features; and Stage 4 (2 blocks) enhanced particle semantic features to facilitate differentiation between particles and the background. The original SPPF spatial pyramid pooling module was retained to ensure multi-scale contextual information aggregation, covering the feature extraction requirements of full-size particles from 2 to 50 μm. The modified model satisfies the feature extraction requirements of full-size particles while achieving lightweight design.

[0060] The third layer: Feature fusion layer (reconstructing the FPN+PAN structure (feature pyramid network + path aggregation network), lightweight feature complementarity) The Neck section of YOLOv8-nano uses an FPN+PAN structure, with feature fusion based on the C2f module. It has 128 / 256 channels, and the 1×1 convolution before upsampling easily causes feature loss for small targets, resulting in low fusion efficiency. The improved model introduces the Efficient Rep lightweight fusion module (a reparameterizable convolution module) to replace all C2f modules, reducing the computational cost of the fusion stage; it also eliminates the 1×1 convolution layer before upsampling, reducing feature information loss. Furthermore, when FPN fuses from top to bottom, it prioritizes the transmission of deep semantic features, while PAN fuses from bottom to top, strengthening shallow detail features, achieving multi-size feature complementarity. Each feature map simultaneously possesses granular semantic discrimination and detail localization capabilities.

[0061] Fourth layer: Micrometer-level feature enhancement branch (adds attention module to resist background interference) YOLOv8-nano lacks a feature enhancement branch, making it susceptible to false positives / false negatives due to background noise such as wall reflections masking particle features. The modified version adds a lightweight ECA channel attention module (high-efficiency channel attention module) after outputting feature maps at three scales in the feature fusion layer. This module constructs a micrometer-level feature enhancement branch—adaptively weighting the fused features at each scale to strengthen key particle features (grayscale, edges) while preserving original fused feature information, avoiding noise amplification caused by over-enhancement. This significantly improves feature robustness and adapts to particle detection needs in complex wall-facing scenarios.

[0062] Fifth layer: Detection head (classification + regression, enhancing positioning accuracy) YOLOv8-nano employs an anchor-free decoupled detection head with a simple regression branch structure (only 2 convolutional layers), a classification branch adaptable to 80 target classes, redundant channels, and positioning accuracy insufficient for micron-level scenes (≤0.1μm). The improved version retains the anchor-free decoupled detection head architecture, adding a 3×3 convolutional layer + BN + SiLU structure to the regression branch to enhance bounding box coordinate fitting capabilities. For single-class detection of micron-sized particles, the number of channels in the classification branch is reduced from 80 to 1, minimizing unnecessary computation. A dynamic weight calibration mechanism is introduced to adaptively adjust the classification score weights based on the grayscale difference between particles and the background, and it adapts to sub-pixel-level coordinate detection logic, working together to achieve highly accurate anchor-free detection. Sixth layer: Output layer (decoding results adapted for tracking scenarios) The YOLOv8-nano output layer only implements prediction tensor decoding and is not adapted for multi-target tracking scenarios. The output noisy bounding boxes easily lead to frequent tracking ID switching. Building upon prediction tensor decoding, a multi-scale prediction decoding module (adapting to three scale feature map outputs) and a confidence-adaptive filtering module are added to make the dynamic threshold adjustable. After decoding the particle coordinates, noisy bounding boxes are filtered out, and a standardized tensor adapted for StrongSORT multi-target tracking is output. The ID switching rate of StrongSORT tracking is reduced, and the output particle coordinates can be directly used for trajectory tracking and dynamic parameter (impact velocity, angle) calculation.

[0063] S3. Collision feature extraction.

[0064] S31. Collision Timing Identification: The inter-frame difference method is used to calculate the rate of change of position of nanoscale particles, based on the fundamental data collected by StrongSORT (e.g., the center pixel coordinates of particles in each frame). x t , y t ) to obtain the inter-frame pixel displacement Δ S The frame time interval Δ is obtained from the time difference between adjacent frames (fps) of a high-speed camera.T The mean and standard deviation σ of the position change rate are obtained by calculating the change rate sequence of consecutive frames before the collision. When the change rate changes abruptly and the distance from the particle center to the wall is less than 1 / 2 of the particle diameter, the collision key frame is directly locked to trigger feature extraction (parameters such as the particle center coordinates, particle-wall relative position, incident velocity, impact angle, rebound angle, and recovery coefficient of the key frame at the time of collision).

[0065] S32. Calculation of dynamic parameters: After locking the collision keyframe, the impact velocity is obtained by backtracking and extracting the trajectory data of the 10 frames before the collision frame and performing polynomial fitting calculation to obtain the incident velocity. The optimal parameters are calculated using the parameters of the 10 frames before the collision frame. When the number of frames is small, the order of the fitting polynomial can be adjusted (a second-order polynomial is used for the 10 frames, the first derivative of the fitting curve is calculated for the 5-9 frames, and a first-order linear fitting is performed for the 1-4 frames assuming uniform motion). The impact angle is the angle between the motion direction and the wall normal vector. The adhesion / rebound state is determined by the displacement of the 3 frames after the collision. The recovery coefficient is the ratio of the rebound velocity to the incident velocity.

[0066] S33. Physical verification: Abnormal data is eliminated by combining particle density and wall material parameters.

[0067] S4. Temperature and humidity correlation analysis.

[0068] The data synchronization module acquires temperature and humidity data consistent with the image frame rate, constructs a "temperature and humidity-collision feature" database, and mines patterns through linear regression and cluster analysis to establish a quantitative mapping model between environmental parameters and particle behavior.

[0069] S5. Real-time output of kinetic parameters (particle size, velocity, angle, state, coefficient of restitution) is a direct calculation result, generating dynamic curves (such as temperature-adhesion probability curves), supporting export and visualization (trajectory animation, heat map). The output of the mapping model is the input of "basic kinetic parameters + environmental parameters (temperature and humidity) + auxiliary verification parameters (particle density, elastic modulus, etc.)", and the data processing is realized through a series of processes of "preprocessing + feature association + model fitting + numerical output". The output result of the mapping model consists of both structured numerical values ​​and visualization carriers (curves / heat maps).

[0070] This invention achieves high-precision visual recognition and dynamic analysis of the impact process of micron-sized particles. The system effectively overcomes the problems of imaging interference and inaccurate positioning through adaptive image processing and lightweight detection network. It can simultaneously extract multi-dimensional impact features and correlate them with environmental parameters. It has high frame rate real-time processing capability, and the output results directly support the optimization of engineering anti-ash accumulation and slag formation. It has the advantages of accuracy, completeness and practicality.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic testing method for collisions of micron-sized particles, characterized in that, Includes the following steps: The continuous frame images acquired by the high-speed camera are preprocessed, including adaptive median filtering, multi-scale illumination enhancement, and adaptive binarization. A lightweight object detection network is used to perform sub-pixel localization of micron-level particles in preprocessed consecutive frames of images. Kalman filtering and the Hungarian algorithm are combined to generate complete particle motion trajectories. The lightweight object detection network includes a feature enhancement branch with an attention mechanism to enhance the edge and texture features of micron-level particles and suppress background interference. The network adds convolutional layers to the regression branch of the detection head and reduces the number of channels in the classification branch for single-class detection scenarios. A dynamic weight calibration mechanism is also introduced to dynamically adjust the classification confidence score of the predicted bounding box based on the local grayscale and texture statistical features of the image region corresponding to the predicted bounding box. Based on the complete motion trajectory of the particle, the collision time is identified, the particle dynamic characteristic parameters are calculated, and the physical rationality is verified. Temperature and humidity data of the test environment are collected, and the dynamic characteristic parameters are correlated with the temperature and humidity data to construct a database of temperature and humidity characteristics and dynamic characteristics, and to explore the influence of temperature and humidity characteristics on dynamic characteristics; the acquisition frequency of the temperature and humidity data is kept synchronized with the image acquisition frame rate of the high-speed camera; Dynamic curves and visualization charts are generated and output based on the aforementioned dynamic characteristic parameters.

2. The dynamic testing method for collisions of micron-sized particles according to claim 1, characterized in that, The lightweight object detection network is a lightweight convolutional neural network based on the YOLOv8-nano object detection model, namely ShuffleNetV2.

3. The dynamic testing method for collisions of micron-sized particles according to claim 1, characterized in that, When identifying the collision moment: calculate the position change rate of the particle between consecutive frames. When the position change rate changes abruptly and the distance between the particle center and the wall is less than 1 / 2 of the particle diameter, determine that the current frame is a collision key frame.

4. The dynamic testing method for collisions of micron-sized particles according to claim 1, characterized in that, The dynamic characteristic parameters include impact velocity, impact angle, adhesion rebound state, and recovery coefficient.

5. The dynamic testing method for collisions of micron-sized particles according to claim 4, characterized in that, The method for calculating the impact velocity is as follows: trace back the trajectory data of several frames before the collision keyframe, use a polynomial to fit the motion curve and take the derivative to obtain the instantaneous velocity.

6. A dynamic testing system for collisions of micron-sized particles, characterized in that, Includes hardware modules and software modules: The hardware module includes a high-speed camera, a high-brightness coaxial light source, an image acquisition card, an embedded computing terminal, a temperature and humidity sensor, and a data synchronization unit. The high-speed camera is used to acquire continuous frame images of particle motion, the temperature and humidity sensor is used to synchronously acquire temperature and humidity data of the test environment, and the data synchronization unit is used to keep the image acquisition time of the high-speed camera synchronized with the data acquisition time of the temperature and humidity sensor. The embedded computing terminal is used to provide a computing platform; The software module is a visual recognition program running on an embedded computing terminal, including an image preprocessing unit, a particle positioning and trajectory tracking unit, a dynamic feature extraction unit, a temperature and humidity correlation analysis unit, and a result output unit. The image preprocessing unit is used to preprocess the consecutive frame images; The particle localization and trajectory tracking unit is used to call a lightweight object detection network to perform sub-pixel-level particle localization on the preprocessed image and generate the complete motion trajectory of the particles by combining the trajectory tracking algorithm. The lightweight object detection network includes a feature enhancement branch with an attention mechanism to enhance the edge and texture features of micron-level particles and suppress background interference. The lightweight object detection network adds a convolutional layer to the regression branch of the detection head and reduces the number of channels in the classification branch for single-class detection scenarios. At the same time, it introduces a dynamic weight calibration mechanism to dynamically adjust the classification confidence score of the predicted box based on the local grayscale and texture statistical features of the image region corresponding to the predicted box. The collision feature extraction unit is used to identify the collision moment based on the complete motion trajectory and to calculate the dynamic characteristic parameters of the particles. The temperature and humidity correlation analysis unit is used to perform correlation analysis between the dynamic characteristic parameters and the synchronously collected temperature and humidity data, construct a database of temperature and humidity characteristics and dynamic characteristics, and explore the influence of temperature and humidity characteristics on dynamic characteristics. The result output unit is used to output the analysis results of the dynamic characteristic parameters and generate dynamic curves and visualization charts.

7. A dynamic testing system for micron-scale particle collisions according to claim 6, characterized in that, The high-speed camera has a maximum pixel count of 1024 × 1024 and a recording time of 4.85 seconds, and a minimum pixel count of 256 × 256 and a recording time of 8.73 seconds; the temperature and humidity sensor has an accuracy of ±0.5℃ for temperature and ±1%RH for humidity.

Images