A Particle Size Detection Method for Belt Conveyor Materials Based on Optimized Watershed Algorithm

Abstract

This invention relates to the field of industrial automation inspection technology, and discloses a method for detecting the particle size of materials on a belt conveyor based on an optimized watershed algorithm. The method includes: acquiring real-time image data of the material and operating condition data of the conveyor belt speed; integrating the operating condition data and feedback data by an algorithm steady-state engine to generate control commands including algorithm path selection and parameter configuration; performing feedforward correction and adaptive preprocessing on the image according to the control commands; then performing segmentation in a dual-channel parallel processing architecture, wherein the main processing channel performs segmentation of the complete image using a specified algorithm path to generate segmentation results, and the verification channel performs segmentation of representative regions using different algorithm paths to generate internal verification signals; finally, calculating the particle size based on the segmentation results and generating feedback data to update the strategy model of the algorithm steady-state engine online. This invention solves the problems of poor adaptability to operating conditions and unstable detection accuracy of traditional methods by constructing a closed-loop self-learning system.

Description

Technical Field

[0001] This invention relates to the field of industrial automation testing technology, and in particular to a method for detecting the particle size of materials on belt conveyors based on an optimized watershed algorithm. Background Technology

[0002] In industrial production processes such as metallurgy, mining, and power, belt conveyors are key equipment for transporting bulk materials such as ores and coal. The particle size distribution of the material is an important process parameter that affects the efficiency, energy consumption, and equipment stability of downstream crushing, screening, and grinding processes. Therefore, continuous and accurate online monitoring of the particle size of the material on the conveyor belt is of great practical significance.

[0003] Compared to traditional offline manual screening or costly laser scanning inspection, machine vision-based inspection methods have become an important technological direction in this field due to their advantages such as non-contact operation, real-time performance, and rich information. These methods typically involve installing industrial cameras above the conveyor belt and processing and analyzing the captured material images to achieve quantitative statistical analysis of particle size.

[0004] However, when applying machine vision technology to such dynamic and unstructured industrial scenarios, existing technologies still face some inherent technical bottlenecks. First, the continuous movement of the conveyor belt causes material displacement during camera exposure time, resulting in geometric stretching and distortion of the acquired image. If this distortion is not corrected, it directly affects the accuracy of individual particle morphology measurements. Second, bulk materials commonly exhibit stacking and adhesion during transport. Traditional image segmentation algorithms, especially the classic watershed algorithm, are prone to producing incorrect segmentation boundaries due to noise or texture details when processing such closely contacted targets. This leads to a complete material particle being segmented into multiple fragments—an oversegmentation problem—which severely interferes with the final particle size statistics.

[0005] Industrial environments are not static; changes in lighting, material humidity, and conveyor belt speed fluctuations can all affect image features. Most existing vision inspection systems rely on a set of fixed algorithm parameters that are calibrated offline under specific conditions. When actual operating conditions deviate from the calibration state, these fixed parameters become difficult to adapt, leading to decreased detection performance and a lack of adaptability to environmental changes and robustness in maintaining long-term stable operation.

[0006] Therefore, this invention proposes a method for detecting the particle size of materials in belt conveyors based on an optimized watershed algorithm to address the shortcomings of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting the particle size of conveyor belt materials based on an optimized watershed algorithm. This method solves the problem that existing visual inspection methods, which use a fixed open-loop processing mode, are difficult to adapt to dynamic working conditions such as changes in conveyor belt speed and material adhesion in industrial settings, resulting in low detection accuracy and insufficient long-term system stability.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting the particle size of materials on a belt conveyor based on an optimized watershed algorithm, comprising the following steps:

[0009] Acquire real-time image data of the material on the conveyor belt and operating condition data characterizing the real-time linear speed of the conveyor belt;

[0010] The algorithm steady-state engine receives the operating condition data and the feedback data generated in the previous processing cycle, and generates a set of control instructions containing algorithm path selection instructions and algorithm parameter configuration based on the strategy model inside the algorithm steady-state engine.

[0011] According to the algorithm parameters configured in the control command, feedforward correction and adaptive preprocessing are performed on the real-time image data to obtain a preprocessed image;

[0012] According to the control instructions, the preprocessed image is segmented in a dual-channel parallel processing architecture. The main processing channel performs segmentation on the complete preprocessed image to generate a segmentation result for particle size quantization, and the verification channel performs an algorithm path different from that of the main processing channel on a representative region of the preprocessed image to generate an internal verification signal.

[0013] Based on the segmentation results generated by the main processing channel, the particle size distribution data of the material particles are calculated and statistically analyzed.

[0014] The algorithm generates and integrates feedback data from the current processing cycle, and updates the strategy model of the steady-state engine online based on the feedback data.

[0015] Preferably, the step of the algorithm steady-state engine receiving the operating condition data and the feedback data generated in the previous processing cycle includes:

[0016] Construct a multidimensional state vector, which includes the currently acquired real-time linear speed of the conveyor belt, as well as the mean of the material particle size distribution, the standard deviation of the material particle size distribution, the segmentation confidence level, and the dual-channel verification difference degree generated from the previous processing cycle.

[0017] Preferably, the control commands include:

[0018] Algorithm path selection instruction used to specify the image segmentation algorithm path adopted by the main processing channel;

[0019] Algorithm parameter configuration used to provide a set of numerical parameters for the feedforward correction and adaptive preprocessing, as well as the dual-channel parallel processing architecture.

[0020] Preferably, the algorithm path executed by the main processing channel is a fine segmentation path, and the steps of the fine segmentation path include:

[0021] Perform SLIC superpixel preprocessing to segment the preprocessed image into superpixel blocks;

[0022] Foreground markers are generated on the binarized image corresponding to the preprocessed image through distance transformation;

[0023] Using the foreground marker as a seed point, marker-controlled watershed segmentation is performed on the superpixel block to generate a segmentation result.

[0024] Preferably, in the dual-channel parallel processing architecture, the algorithm path executed by the verification channel is the algorithm path that is not currently selected by the main processing channel.

[0025] Preferably, the step of generating feedback data specifically includes:

[0026] The segmentation confidence score is calculated based on the average circularity of all closed contours in the segmentation results.

[0027] The dual-channel validation difference is calculated by comparing the local statistics of the main processing channel and the validation channel within the same representative region.

[0028] Preferably, the step of updating the strategy model online includes:

[0029] The instant reward value for the current processing cycle is calculated based on the segment confidence and dual-channel verification difference.

[0030] The Q-learning algorithm is applied to update the Q-values ​​in the policy model corresponding to the current state and action using the immediate reward value.

[0031] Preferably, the method further includes:

[0032] Based on the dual-channel verification difference, the motion model coefficients used in the feedforward correction step are self-tuned online.

[0033] Preferably, the feedforward correction step specifically includes:

[0034] Based on the real-time linear speed of the conveyor belt, the exposure time of the image acquisition unit, and a motion model coefficient, calculate the shearing factor in an affine transformation matrix.

[0035] Using the affine transformation matrix, a geometric transformation is performed on the real-time image data to compensate for motion distortion of the image.

[0036] Preferably, the adaptive preprocessing step includes adaptive dual denoising, which specifically involves:

[0037] Apply Gaussian and median filters to grayscale images;

[0038] The kernel standard deviation of the Gaussian filter and the kernel size of the median filter are specified by the algorithm parameter configuration generated by the algorithm steady-state engine.

[0039] In summary, the present invention has at least one of the following beneficial technical effects:

[0040] 1. This invention improves the automation level and adaptability of the detection method. By setting up a steady-state algorithm engine, the optimal algorithm path and corresponding algorithm parameters can be dynamically selected based on real-time operating data and internal feedback data. This enables the method to automatically adapt to various operating conditions such as changes in conveyor belt speed and fluctuations in material properties, reducing reliance on manual intervention and ensuring the continuity of detection.

[0041] 2. This invention improves the accuracy of material particle size detection. On the one hand, through a feedforward correction step, the motion deformation of the image is actively compensated based on the real-time linear velocity. On the other hand, in the fine segmentation path, SLIC superpixels are combined with marker-controlled watershed segmentation, and the boundary constraints provided by superpixels effectively avoid the problem of over-segmentation of adhering materials. The combination of these two aspects ensures that the segmentation contour can more accurately reflect the true geometric shape of the material, providing a reliable foundation for subsequent particle size quantization.

[0042] 3. This invention enhances the long-term stability and robustness of the detection system. By introducing a dual-channel parallel processing architecture, this method obtains a quantitative index for internal cross-validation, namely the dual-channel validation difference. The system can use this feedback to continuously optimize and self-tune the internal strategy model and feedforward correction model, enabling it to resist performance degradation caused by environmental changes or equipment wear during long-term operation and maintain the reliability of the detection results. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0044] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0045] Figure 3 This is a schematic diagram of the closed-loop control principle of the steady-state engine algorithm of the present invention;

[0046] Figure 4 This is a schematic diagram of the dual-channel parallel segmentation and verification architecture of the present invention;

[0047] Figure 5 This is a schematic diagram of the particle size distribution statistics of the present invention.

[0048] Among them, 10 is the image acquisition unit; 20 is the working condition sensing unit; 30 is the feedforward correction module; 40 is the dual-channel parallel processing module; 50 is the particle size quantization module; and 60 is the algorithm steady-state engine. Detailed Implementation

[0049] The following is in conjunction with the appendix Figure 1 - Appendix Figure 5 The present invention will be further described in detail below.

[0050] See attached document Figure 1 , Figure 1 This is a schematic diagram of a system architecture according to an embodiment of the present invention. The present invention provides a system for detecting the particle size of materials on a belt conveyor, the system comprising:

[0051] An image acquisition unit 10, such as an industrial camera, is positioned above the conveyor belt and its function is to capture image data of the material carried on the conveyor belt. The field of view of the image acquisition unit 10 covers a representative area of ​​the material.

[0052] The condition sensing unit 20, such as a rotary encoder, is configured at the drive mechanism of the conveyor belt; its function is to measure the physical motion parameters of the conveyor belt in real time and output condition data characterizing its linear velocity.

[0053] The feedforward correction module 30 receives image data from the image acquisition unit 10 and working condition data from the working condition sensing unit 20. The feedforward correction module 30 is configured to perform an affine transformation on the input image data based on the working condition data in order to correct the geometric deformation of the image caused by the movement of the conveyor belt.

[0054] The dual-channel parallel processing module 40 receives the image processed by the feedforward correction module 30. The dual-channel parallel processing module 40 integrates at least two image segmentation algorithm paths with different structures. The structure of the dual-channel parallel processing module 40 is configured with a main processing channel and a verification channel to perform parallel and asymmetric segmentation processing on the image.

[0055] The particle size quantization module 50 receives the segmentation result output from the main processing channel of the dual-channel parallel processing module 40; the particle size quantization module 50 is configured to calculate the geometric parameters of each closed contour region in the segmentation result to obtain its equivalent diameter, and further statistically analyze the particle size distribution data of the entire frame image.

[0056] The algorithm steady-state engine 60, as the central control module, has its input end connected to the working condition sensing unit 20, the dual-channel parallel processing module 40 and the particle size quantization module 50, and its output end connected to the feedforward correction module 30 and the dual-channel parallel processing module 40.

[0057] The algorithm steady-state engine 60 is configured to: receive and integrate real-time operating data from the outside and feedback data from the internal processing module (including particle size statistics and internal verification signals); generate a set of control commands based on the integrated data; and send the control commands to the feedforward correction module 30 and the dual-channel parallel processing module 40 respectively, so as to dynamically and collaboratively adjust the internal algorithm parameters and working paths of the two modules.

[0058] See attached document Figure 2 , Figure 2 This is a flowchart of a method according to an embodiment of the present invention; the present invention provides a method for detecting the particle size of materials on a belt conveyor based on an optimized watershed algorithm, the method including the following steps:

[0059] S100 acquires real-time image data of the material on the conveyor belt and operating condition data characterizing the real-time linear speed of the conveyor belt.

[0060] S200, the algorithm steady-state engine receives and integrates operating condition data and feedback data generated in historical processing cycles, and generates a set of control commands that include algorithm path selection instructions and algorithm parameter configuration based on the integrated data;

[0061] S300 performs feedforward correction and adaptive preprocessing on real-time image data according to the algorithm parameters configured in the control command; among which, feedforward correction is performed based on the working condition data to compensate for motion deformation of the image.

[0062] S400, according to the algorithm path selection instruction and algorithm parameter configuration in the control instruction, performs segmentation processing on the preprocessed image under the dual-channel parallel processing architecture. The main processing channel performs segmentation on the complete image to generate the segmentation result required for particle size quantization, and the verification channel performs segmentation of the candidate algorithm path on the representative region of the image to generate the internal verification signal.

[0063] S500 calculates and statistically analyzes the particle size distribution data of the material particles based on the segmentation results generated by the main processing channel, and outputs it as the detection result of the current processing cycle.

[0064] S600 generates and integrates feedback data from the current processing cycle, including segmentation confidence calculated based on the segmentation results, dual-channel verification difference calculated based on the internal verification signal, and particle size distribution data; based on the feedback data, it performs online updates and self-tuning of the algorithm's steady-state engine's internal strategy model and feedforward correction model.

[0065] To further clarify the technical details of the method of the present invention, the specific implementation methods, working principles and collaborative relationships between modules involved in the above steps will be described in detail below.

[0066] In a specific embodiment, the real-time data sensing stage corresponding to step S100 aims to provide accurate and synchronous raw input data for subsequent processing steps. This stage may specifically include the following steps:

[0067] S101, an image acquisition unit 10 configured above the conveyor belt continuously captures images of the material carried on the conveyor belt at a preset frame rate. The image acquisition unit 10 is an industrial camera, and its lens optical axis is basically perpendicular to the plane of the conveyor belt to reduce imaging distortion. The output of this step is a series of digital images containing material information, where each frame of the image is recorded as a data structure, such as a two-dimensional matrix containing RGB color channel information. At the same time, a high-precision timestamp is attached to each frame of captured image data to mark its exact acquisition time.

[0068] S102, the motion parameters of the conveyor belt are synchronously measured by the condition sensing unit 20 configured at the conveyor belt drive mechanism; in one embodiment, the condition sensing unit 20 is a photoelectric rotary encoder; the photoelectric rotary encoder outputs a series of digital pulse signals as the main drive wheel of the conveyor belt rotates; by counting the pulse signals within a unit time interval, the real-time angular velocity of the drive wheel can be calculated, and then the real-time linear velocity of the conveyor belt can be calculated; the linear velocity... The calculation can be given by the following formula:

[0069] ;

[0070] in, The radius of the conveyor belt drive pulley; To be in the time interval The number of pulses detected by the working condition sensing unit 20; The fixed number of pulses generated per revolution of the working condition sensing unit 20; similar to image data acquisition, each calculated linear velocity value is appended with a high-precision timestamp. In order to be in The number of revolutions the wheel makes within a given time period; In order to be in The total distance the wheel travels along the tangential direction within a given time period.

[0071] S103, the image data obtained in step S101 and the linear velocity data obtained in step S102 are time-stamp aligned and synchronized; for each frame of image data, the linear velocity data with the closest timestamp is found in its timestamp neighborhood and paired with it to form a synchronized data group; this data group is transmitted to subsequent processing steps as the unified input of the current processing cycle; through this synchronization operation, it is ensured that each frame of the analyzed image accurately corresponds to its physical motion state at the moment of acquisition, providing an accurate data basis for subsequent feedforward correction and algorithm decision-making.

[0072] See attached document Figure 3 In a specific embodiment, the dynamic decision-making stage of the algorithm steady-state engine corresponding to step S200 aims to generate a set of optimal control instructions based on the current comprehensive state of the system to guide subsequent processing steps. This stage is executed by the algorithm steady-state engine 60 and may specifically include the following steps:

[0073] S201, Construct a multi-dimensional state vector for the current processing cycle. At the beginning of each processing cycle, the algorithm steady-state engine 60 integrates the real-time operating data from step S100 and the feedback data generated from the previous processing cycle (i.e., step S600) to jointly form a state vector describing the current state of the system. The state vector can be represented by the following formula:

[0074] ;

[0075] in: This is the currently acquired real-time linear speed of the conveyor belt, which is real-time operating data. This is the average particle size distribution of the material calculated in the previous processing cycle, and it belongs to feedback data. The standard deviation of the particle size distribution of the material calculated in the previous processing cycle is considered feedback data. The segment confidence scores generated for the previous processing cycle are considered feedback data. The dual-channel validation difference generated in the previous processing cycle is considered feedback data.

[0076] In practical implementation, for ease of indexing and calculation, the state vector can be... Continuous value components in (e.g.) , Discretize the data (e.g., 1, 2, 3, 4) and map it to a finite state space.

[0077] S202, based on an internal policy model, according to the state vector Generate control commands; the algorithm steady-state engine 60 is internally configured with a policy model, which establishes a mapping relationship from state vectors to optimal control commands; in one embodiment, the policy model can be specifically implemented as an offline calibrated multidimensional look-up table, or a Q-table that can be updated through online learning.

[0078] The algorithm's steady-state engine 60 constructs state vectors. As input, the strategy model is queried to determine the optimal action to be performed in the current state, which is the control instruction to be generated in this step.

[0079] S203, output control command; this control command consists of two parts: algorithm path selection command and algorithm parameter configuration.

[0080] The algorithm path selection instruction specifies the image segmentation algorithm path to be used by the main processing channel in the subsequent step S400, for example, selecting between a preset fine segmentation path and a high-speed threshold path.

[0081] The algorithm parameter configuration function is to provide a complete set of specific numerical parameters for subsequent steps S300 and S400. This parameter set may include: ;in, is the kernel standard deviation of the Gaussian denoising filter; This represents the kernel size of the median denoising filter; The preset number of superpixels for SLIC superpixel segmentation; The compactness factor for SLIC superpixel segmentation; These are the motion model coefficients in feedforward affine correction.

[0082] In a specific embodiment, corresponding to the feedforward correction and adaptive preprocessing stage of step S300, it receives control instructions from step S200 and processes the real-time image data acquired in step S100; this stage aims to actively compensate for the geometric distortion of the image and prepare high-quality image input for subsequent segmentation steps.

[0083] This stage can first perform a feedforward affine correction step based on real-time operating conditions; the purpose of this step is to inversely compensate for the geometric stretching deformation of the image caused by the displacement of the material during the exposure time of the image acquisition unit 10, based on the real-time motion state of the conveyor belt; this step may specifically include:

[0084] S301, construct a two-dimensional affine transformation matrix; this two-dimensional affine transformation matrix is ​​designed as a horizontal shearing transformation to correct motion deformation of the material image; affine transformation matrix It can be represented as:

[0085] ;

[0086] S302, Calculate the shearing factor in the affine transformation matrix; shearing factor The value is obtained from the real-time linear speed of the conveyor belt in step S100. Exposure time of image acquisition unit 10 and motion model coefficients obtained from the control command in step S200. It is jointly determined; the calculation formula is as follows:

[0087] ;

[0088] in, This refers to the real-time linear speed of the conveyor belt. The fixed exposure time parameter for the image acquisition unit 10; This is a motion model coefficient dynamically configured by the algorithm steady-state engine 60. The existence of this coefficient is used to adapt to the different visual deformation effects produced by materials with different physical properties (e.g., mainly rolling particles and mainly translational blocks) during motion. The value is updated and self-tuned online by the algorithm steady-state engine 60 based on historical feedback data.

[0089] S303, Perform an affine transformation on the real-time image data; use the affine transformation matrix constructed in step S301. The raw image data obtained from step S100 (denoted as...) Perform a geometric transformation operation to obtain the corrected image data (denoted as ). For the specific implementation of this affine transformation operation, those skilled in the art can use standard image transformation functions from computer vision libraries, which are well-known techniques in the field and will not be elaborated here.

[0090] After this step, the output corrected image data It is geometrically closer to the actual shape of the material.

[0091] Following the feedforward affine correction step, this stage continues to process the corrected image data. A series of adaptive preprocessing operations are performed; these operations aim to further improve image quality and extract key features for subsequent segmentation tasks; these operations are configured by control instructions issued by the algorithmic steady-state engine 60, and may specifically include:

[0092] S304 performs grayscale space conversion; converts the feedforward corrected image data... The image is converted from RGB color space to YUV color space. YUV space was chosen because it separates the luminance information (Y channel) from the chrominance information (U and V channels). The Y channel is then extracted as the grayscale image for subsequent processing. This grayscale image can more stably reflect the light and dark contrast of materials and reduce interference caused by differences in material color or changes in light color temperature.

[0093] The S305 performs adaptive dual noise reduction; to cope with the mixed noise that may exist in the industrial field, it adopts a noise reduction strategy that combines Gaussian filtering and median filtering.

[0094] First, for grayscale images A Gaussian filter is applied to smooth the image and suppress subtle Gaussian noise; the standard deviation of the kernel function of the Gaussian filter used is... It is not a fixed value, but is specified by the algorithm parameter configuration generated by the algorithm steady-state engine 60 in step S200.

[0095] Subsequently, a median filter is applied to the Gaussian-filtered image; this step is mainly used to eliminate impulse noise or salt-and-pepper noise caused by dust, water droplets, etc. in the image; the kernel size of the median filter... Similarly, the algorithm's steady-state engine 60 provides this dynamically in its algorithm parameter configuration.

[0096] In this way, the intensity and type of denoising can be adaptively adjusted based on the algorithm's steady-state engine 60's judgment of the current operating condition; this step outputs a denoised grayscale image. .

[0097] S306 performs adaptive binarization processing on the denoised grayscale image. Applying the Otsu method (maximum inter-class variance method) to generate binarized images The Otsu method analyzes the grayscale histogram of an image to automatically calculate a global threshold that can distinguish the foreground (material particles) from the background (conveyor belt) to the greatest extent. Since this threshold is dynamically calculated based on the grayscale distribution of each frame of the image, this binarization step can adapt to changes in the overall brightness of the image caused by factors such as changes in light intensity and changes in the reflective properties of materials. The specific mathematical implementation of the Otsu method is a well-known technology in this field and will not be described in detail here.

[0098] The final output of this step is a binarized image. The areas of the material particles are clearly marked and then conveyed to step S400 for core segmentation processing.

[0099] Reference Figure 4In a specific embodiment, corresponding to the dual-channel parallel segmentation and verification stage of step S400, it receives preprocessed image data from step S300 and control commands from step S200; the core of this stage is to perform accurate contour segmentation of material particles.

[0100] In one embodiment of the present invention, this stage adopts a dual-channel parallel processing architecture; wherein, the function of the main processing channel is to perform core segmentation operations on the complete preprocessed image according to the instructions of the algorithm steady-state engine 60, and generate the final segmentation result for particle size quantization in the subsequent step S500.

[0101] The main processing channel selects one of at least two preset algorithm paths to execute based on the algorithm path selection instruction received from step S200; in a specific embodiment, the two algorithm paths include a fine segmentation path and a high-speed threshold path.

[0102] If the algorithm path selection instruction specifies the use of a fine-grained segmentation path, the main processing channel performs the following steps:

[0103] S401, Perform SLIC superpixel preprocessing; this step converts the preprocessed image data into the CIELAB color space and divides it into several compact, uniform superpixel blocks; the purpose of this step is to elevate the image processing from pixel-level to region-level, providing structured boundary constraints for subsequent segmentation; during segmentation, the similarity distance between pixels is considered. Calculated using the following formula:

[0104] ;

[0105] in, The Euclidean distance between two pixels in the CIELAB color space; The Euclidean distance between two pixels in the image coordinate system; The preset superpixel grid spacing can be determined by the total number of pixels in the image. and the preset number of superpixels specified by the algorithmic steady-state engine 60 Joint decision, that is ; The compactness factor is specified by the algorithm's steady-state engine 60 to balance the weights of color similarity and spatial proximity in distance calculation.

[0106] S402, Generate foreground markers; this step involves generating the binarized image from step S306. Perform a distance transformation once to generate a distance transformation graph. The grayscale value of each foreground pixel in this image represents its Euclidean distance to the nearest background pixel; the foreground marker is this distance transformation map. All local maxima points in the dataset, each of which theoretically corresponds to the center of an independent material particle.

[0107] S403, Perform marker-controlled watershed segmentation; using the foreground markers generated in step S402 as seed points, perform the watershed algorithm on the SLIC superpixel map generated in step S401; this algorithm will start region growth from each marker point, and its segmentation boundary will preferentially form along the boundary of the SLIC superpixel; this method utilizes the edge information provided by the superpixel, effectively avoiding the over-segmentation problem caused by traditional watershed algorithms under unconstrained conditions. The output of this step is the final material particle segmentation contour.

[0108] If the algorithm path selection instruction specifies the use of the high-speed threshold path, then the main processing channel performs the following steps:

[0109] S404, Perform morphological-based segmentation; this path directly over the binarized image generated in step S306. The operation can be performed on the material particles; it may include applying a morphological opening operation to eliminate isolated foreground regions formed by fine dust or noise; and applying a morphological closing operation to fill small pores that may exist inside the material particles. The specific implementation of the morphological operation is well known in the art and will not be described in detail here. In some embodiments, if the material is slightly adhered, it can also be separated by combining distance transformation and watershed algorithm.

[0110] Regardless of whether it is through the fine segmentation path or the high-speed threshold path, the main processing channel eventually outputs a segmentation result image containing the closed contours of all material particles. This image is then sent to step S500 for subsequent particle size quantization processing.

[0111] In parallel with the main processing channel, this dual-channel parallel processing architecture also includes a verification channel; the verification channel operates synchronously with the main processing channel, and its function is to provide a reference segmentation result for cross-validation and generate key internal verification signals for subsequent steps S600; its specific implementation may include the following steps:

[0112] S405 identifies one or more representative regions (ROIs) in the preprocessed image as the processing objects of the validation channel. The representative region can be one or more rectangular regions with fixed or dynamically selected positions in the image. By processing only a part of the image instead of the whole, the computational load of the validation channel is significantly reduced, forming the basis for its lightweight implementation.

[0113] S406, in the representative region, an alternative algorithm path different from the main processing channel is executed; in a specific embodiment, the algorithm path executed by the verification channel is the one that the main processing channel has not currently selected; for example, if the algorithm path selection instruction in step S200 specifies that the main processing channel adopts a fine segmentation path, then the verification channel executes a high-speed threshold path in the representative region; conversely, if the main processing channel adopts a high-speed threshold path, then the verification channel executes a fine segmentation path in the representative region.

[0114] S407, Generate local processing results; After the verification channel completes the segmentation processing of the representative area, its output is not used for the final global particle size statistics, but a set of local statistical data obtained in the area, such as the particle count in the area, or the average particle size of the particles in the area.

[0115] S408, encapsulate and transmit the internal verification signal; the local statistical data generated in step S407 is encapsulated into an internal verification signal; the sole purpose of this signal is to be transmitted to step S600 as one of the data inputs for the algorithm steady-state engine 60 to perform system state evaluation and model self-tuning; in step S600, this signal will be quantitatively compared with the processing results obtained by the main processing channel in the same representative area to calculate the dual-channel verification difference, thereby providing a key basis for the closed-loop feedback control of the entire system.

[0116] See attached document Figure 5 In a specific embodiment, corresponding to the particle size quantization and statistical output stage of step S500, it receives the segmentation results output from the main processing channel in step S400; the purpose of this stage is to extract quantized geometric parameters from the segmentation results and aggregate them to generate statistical data characterizing the overall particle size distribution of the material in the current frame image, and its final output format can be as shown in the appendix. Figure 5 As shown; this stage may specifically include the following steps:

[0117] S501, traverse each successfully segmented closed contour region in the segmentation results; for the p-th closed contour region in the image, first calculate the total number of pixels it contains, and denot this total number of pixels as the pixel area. .

[0118] S502, based on pixel area Calculate the equivalent diameter of the closed contour region. The equivalent diameter is defined as the diameter of a circle with the same area as the region. In one embodiment, to obtain a diameter in physical units, the image acquisition system can be pre-calibrated using a calibration plate to obtain a conversion ratio between pixel distance and physical distance. The formula for calculating the equivalent diameter is as follows:

[0119] ;

[0120] in, For the first The equivalent diameter of a closed contour region The pixel area.

[0121] S503, perform statistical aggregation on the equivalent diameters of all identified material particles in the current frame image; assuming a total of N material particles are identified in the current frame image, a set containing N equivalent diameter values ​​will be obtained. Performing statistical calculations on the set may include calculating its arithmetic mean. with standard deviation .

[0122] Mean particle size distribution The calculation formula is: ;

[0123] Standard deviation of particle size distribution The calculation formula is: ;

[0124] S504, the calculated average particle size distribution and standard deviation The final detection result of the current processing cycle will be output; at the same time, these two statistical data will also be transmitted to step S600 as part of the feedback data to support the online learning and self-tuning of the algorithm steady-state engine 60 in subsequent processing cycles.

[0125] In a specific embodiment, the closed-loop feedback and online self-learning stage corresponding to step S600 is executed after the particle size detection result is output in each processing cycle. The purpose of this stage is to generate a set of quantitative feedback data to evaluate the system performance of the current processing cycle and to provide a basis for the self-optimization of the algorithm steady-state engine 60.

[0126] This stage may begin by generating a key feedback signal; based on the processing results of steps S400 and S500, this step generates at least two internal feedback signals, namely, segmentation confidence and dual-channel verification difference; its generation principle may specifically include:

[0127] S601, Calculate the segmentation confidence score; segmentation confidence score This is an indicator used to quantify the quality of the segmentation results from the main processing channel; in one embodiment, this indicator is calculated based on the geometry of the outlines of all segmented material particles; for the first segmented particle segment from the main processing channel... A closed contour region can be used to calculate its circularity. :

[0128] ;

[0129] in, The pixel area of ​​this region. Its outline perimeter; the degree of roundness The value range is (0, 1]. The closer the value is to 1, the closer the contour shape is to a standard circle.

[0130] After calculating the circularity of all N segmented contours in the current frame image, their arithmetic mean is used as the final segmentation confidence score for the current processing cycle. :

[0131] ;

[0132] The confidence level of the segmentation As a feedback signal, its value reflects the overall morphological regularity of the segmentation result under the current segmentation algorithm and parameter combination.

[0133] S602, Calculate the dual-channel validation difference; Dual-channel validation difference This is an indicator used to quantify the consistency between the processing results of the main processing channel and the verification channel; the calculation of this indicator is based on comparing the processing results of the two channels in the same representative region (ROI) in step S400; the calculation process is as follows:

[0134] From the complete segmentation result output by the main processing channel in step S400, extract the processing result within the same representative region used by the verification channel.

[0135] For the processing results of the main processing channel within this representative region, calculate local statistics, such as the particle count of materials within that region, denoted as... .

[0136] The internal verification signal generated by the verification channel is obtained from step S408. This signal contains local statistics of the same type calculated by the verification channel within the same representative region, denoted as... .

[0137] The normalized difference between the two is calculated using the following formula to obtain the dual-channel validation difference. :

[0138] ;

[0139] in, It is an extremely small positive number to avoid the case where the denominator is zero; this dual-channel verification of the difference. The closer the value is to 0, the more consistent the results obtained by the two different algorithm paths are in this representative region.

[0140] Finally, the segmentation confidence score generated in this step and dual-channel verification difference Together with the average particle size distribution calculated in step S503 and standard deviation Together, they form a complete feedback data set; this data set is transmitted to the algorithm steady-state engine 60 for subsequent model updates and self-tuning steps.

[0141] After the key feedback signal is generated, this stage continues with the online update step of the policy model based on reinforcement learning. The purpose of this step is to use the feedback data obtained in the current processing cycle to iteratively optimize the policy model inside the algorithm's steady-state engine 60, so that its decision-making ability can be adaptively improved. In a specific embodiment, the policy model is a Q-table, and its online update process may specifically include:

[0142] S603, Based on the feedback signals generated in steps S601 and S602, calculate the instantaneous reward value for the current processing cycle. This reward value is used to quantify the system's state. Next action The utility gained afterward; reward value It can be determined by the split confidence level and dual-channel verification difference Composed of weighted combinations:

[0143] ;

[0144] in, Current processing cycle The segmentation confidence; Current processing cycle Dual-channel verification of differences; and A pre-defined positive weighting coefficient is used to balance the importance of the morphological regularity of the segmentation results and the consistency of the dual-channel results in the reward evaluation.

[0145] S604 applies the update rules of the Q-learning algorithm to update the Q-values ​​corresponding to the current state and action in the policy model; the policy model (Q-table) stores a series of Q-values, each Q-value... This represents the state. Next action The long-term expected reward; its update formula is as follows:

[0146] ;

[0147] in: This is the updated Q value; This is the old Q value stored before this update; It is the state vector of the current processing cycle, constructed by step S201; In the current processing cycle, the algorithm steady-state engine 60 according to The selected control command (i.e., action); It is the instant reward value calculated in step S603; It is the learning rate, a preset parameter between 0 and 1, which controls the step size of each update; It is a discount factor, a preset parameter between 0 and 1, representing the importance of future rewards relative to immediate rewards; It is the state vector for the next processing cycle; This represents the next state. The optimal action can be found by querying the old Q-table, which is the Q value corresponding to the action, i.e., the estimate of the maximum future profit.

[0148] S605, the new Q value calculated in step S604 The value is written back to the Q-table stored internally by the algorithm steady-state engine 60, replacing the old value. By repeating steps S603 to S605 in each processing cycle, the strategy model of the algorithm steady-state engine 60 can continuously and online learn and optimize, enabling it to make better algorithm path selection and parameter configuration decisions when facing constantly changing working conditions.

[0149] In the closed-loop feedback and online self-learning phase of this method, in addition to updating the policy model of the algorithm steady-state engine 60, a step of self-tuning the feedforward correction model can be further included. The purpose of this step is to use the consistency of the dual-channel processing results as a criterion to fine-tune the key parameters used for motion deformation compensation in step S300. This step can specifically include:

[0150] S606, Perform feedforward model self-tuning based on the verification difference; the goal of this step is to adjust the motion model coefficients used in the feedforward affine correction step S302. The adjustment is based on the dual-channel verification difference calculated in step S602, which characterizes the consistency of the path results of the two different segmentation algorithms. .

[0151] In one specific embodiment, the self-tuning process can be implemented as a proportional-integral (PI) controller; this controller verifies the difference using a dual-channel approach. As an error input, the motion model coefficients Adjustments will be made.

[0152] S607, calculate the adjustment amount of the motion model coefficients; first, define an error term for each processing cycle t. Its value is equal to the dual-channel verification difference of the current period. Simultaneously, maintain the error integral term. It is accumulated in the following way:

[0153] ;

[0154] in, This is the error integral term from the previous processing cycle; subsequently, the motion model coefficients for the current cycle are calculated according to the following formula. Adjustment amount :

[0155] ;

[0156] in, This is the proportional gain coefficient. These are the integral gain coefficients; these two coefficients are preset system parameters used to control the response speed and stability of the self-tuning process.

[0157] S608, Update and apply the motion model coefficients; Apply the adjustment calculated in step S607 to the current motion model coefficients to obtain the new coefficient values ​​to be used in the next processing cycle. :

[0158] ;

[0159] in, These are the motion model coefficient values ​​used in the current cycle; to ensure system stability, the updated coefficient values ​​can be adjusted. Amplitude limiting is applied to keep it within a preset reasonable range.

[0160] The updated motion model coefficients The data will be stored internally by the steady-state engine 60 and output to the feedforward correction module 30 as part of the algorithm parameter configuration when generating control commands in the next processing cycle. Through this self-tuning mechanism, the system can continuously fine-tune its compensation degree for material motion deformation in order to seek an optimal correction state that allows different segmentation algorithms to obtain the most consistent results, thereby improving the adaptability of the entire detection method to different material physical properties.

[0161] To illustrate the present invention in more detail, the following describes the method for detecting the particle size of conveyor belt materials based on the optimized watershed algorithm in conjunction with a specific application scenario.

[0162] This embodiment aims to perform continuous and automatic online detection of the particle size distribution of iron ore pellets carried on a conveyor belt during the sintering process of a steel plant.

[0163] To achieve this detection, system calibration is first performed. With the conveyor belt stationary, a standard calibration plate with multiple circular shapes of known diameters is placed on the plane through which the material passes, and an industrial camera captures its image. By calculating the number of pixels in the circular regions of the image, a relationship is established between the pixel area and the actual physical area (mm²). 2 The conversion relationship is used to obtain a precise pixel-to-millimeter ratio, which is then stored in the system. Within a single processing cycle, the specific implementation of this method is as follows:

[0164] First, data acquisition is performed. An industrial camera installed 1.5 meters directly above the conveyor belt continuously captures material images at a resolution of 1920x1080 pixels at a rate of 30 frames per second, with the exposure time fixed at 1 / 1000 second. Simultaneously, a rotary encoder (outputting 1024 pulses per revolution) mounted on the main drive wheel of the conveyor belt (0.5 meters in radius) detects 15 pulses within a sampling interval of 1 / 30 second. The real-time linear velocity of the conveyor belt can then be calculated using a formula. The velocity is approximately 1.37 m / s. The image frame is timestamped and paired with this velocity value as input data for the current cycle.

[0165] The steady-state engine of the algorithm performs dynamic decision-making; this engine first integrates the input data with the feedback data from the previous cycle to construct the current state vector. Based on this state vector, the algorithm's steady-state engine queries its internal Q-table policy model to determine the optimal action: select the "fine-grained path segmentation" and configure the algorithm parameter set as follows. The motion model coefficients here The initial value of 0.8 is an empirical estimate based on the physical properties of iron ore pellets (approximately spherical shape, low rolling friction).

[0166] Feedforward correction and adaptive preprocessing were performed; the shearing factor of the affine transformation was calculated based on the real-time linear velocity of 1.37 m / s, the exposure time of 1 / 1000 sec, and the motion model coefficient of 0.8. The affine transformation matrix constructed using this factor is used to perform geometric correction on the original image to compensate for the motion deformation of the pellets during exposure. The corrected image is converted into a grayscale image and then subjected to double denoising by a Gaussian filter with a standard deviation of 1.5 and a 3x3 median filter. Finally, the Otsu method is applied to the denoised image for binarization to generate a clear binary image of the material foreground and background.

[0167] The process involves dual-channel parallel segmentation. The main processing channel, according to instructions, performs a "fine-grained segmentation path" on the preprocessed image: first, the SLIC algorithm is applied to segment the image into superpixel blocks according to a preset superpixel count of 1500 and a compactness factor of 20; then, foreground markers are determined on the binary image through distance transformation; finally, using these markers as seeds, a marker-controlled watershed algorithm is executed on the SLIC superpixel image to obtain a global and accurate spherical particle segmentation contour. Simultaneously, the verification channel performs a "high-speed thresholding path" within a representative 400x400 pixel region at the center of the image, processing the binary image through morphological opening and closing operations and counting the number of particles in that region as 25. This count is then transmitted as an internal verification signal to subsequent steps.

[0168] The process involves particle size quantization and statistical output. This step iterates through each closed contour in the global segmentation result generated by the main processing channel. For each contour, its pixel area is calculated, and its equivalent diameter is derived based on a pre-calibrated pixel-to-millimeter ratio. In the current frame, 512 agglomerated particles were identified. By statistically analyzing their equivalent diameters, the mean particle size distribution of the current material was calculated to be 12.8 mm, with a standard deviation of 2.3 mm. These two values ​​are output as the final detection result for this cycle.

[0169] Finally, closed-loop feedback and online self-learning are performed. This step first generates a feedback signal: by calculating the average roundness of the 512 particle contours segmented by the main channel, the segmentation confidence level for the current period is obtained. Simultaneously, the main processing channel identified 23 particles within the same 400x400 pixel representative area used by the verification channel, compared them with the 25 particles identified in the verification channel, and calculated the dual-channel verification difference. The instant reward value for the current period is calculated based on the segment confidence and validation discrepancy. Using Q-learning update rules, the corresponding entries in the Q-table of the algorithm's steady-state engine are updated by combining the current state, the executed action, the obtained reward, and the prediction of the next state. Simultaneously, a validation difference of 0.08 is used as an error input to a PI controller (whose proportional gain and integral gain are tuned to empirical values ​​during system debugging to ensure convergence stability), and the coefficients of the motion model are calculated. A small adjustment, such as -0.01, updates the original coefficient 0.8 to 0.79 for use in the feedforward correction of the next processing cycle.

[0170] This completes one processing and learning cycle. The system continuously repeats this process to achieve accurate detection of material particle size and online optimization of its own performance.

Claims

1. A method for detecting particle size of materials on a belt conveyor based on an optimized watershed algorithm, characterized in that, Includes the following steps: Acquire real-time image data of the material on the conveyor belt and operating condition data characterizing the real-time linear speed of the conveyor belt; The algorithm steady-state engine receives the operating condition data and the feedback data generated in the previous processing cycle, and generates a set of control instructions containing algorithm path selection instructions and algorithm parameter configuration based on the strategy model inside the algorithm steady-state engine. According to the algorithm parameters configured in the control command, feedforward correction and adaptive preprocessing are performed on the real-time image data to obtain a preprocessed image; According to the control instructions, the preprocessed image is segmented in a dual-channel parallel processing architecture. The main processing channel performs segmentation on the complete preprocessed image to generate a segmentation result for particle size quantization, and the verification channel performs an algorithm path different from that of the main processing channel on a representative region of the preprocessed image to generate an internal verification signal. Based on the segmentation results generated by the main processing channel, the particle size distribution data of the material particles are calculated and statistically analyzed. The algorithm generates and integrates feedback data from the current processing cycle, and updates the strategy model of the steady-state engine online based on the feedback data.

2. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, The steps by which the algorithm's steady-state engine receives the operating condition data and the feedback data generated in the previous processing cycle include: Construct a multidimensional state vector, which includes the currently acquired real-time linear speed of the conveyor belt, as well as the mean of the material particle size distribution, the standard deviation of the material particle size distribution, the segmentation confidence level, and the dual-channel verification difference degree generated from the previous processing cycle.

3. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, The control commands include: Algorithm path selection instruction used to specify the image segmentation algorithm path adopted by the main processing channel; Algorithm parameter configuration used to provide a set of numerical parameters for the feedforward correction and adaptive preprocessing, as well as the dual-channel parallel processing architecture.

4. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, The algorithm path executed by the main processing channel is a fine segmentation path, and the steps of the fine segmentation path include: Perform SLIC superpixel preprocessing to segment the preprocessed image into superpixel blocks; Foreground markers are generated on the binarized image corresponding to the preprocessed image through distance transformation; Using the foreground marker as a seed point, marker-controlled watershed segmentation is performed on the superpixel block to generate a segmentation result.

5. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, In the dual-channel parallel processing architecture, the algorithm path executed by the verification channel is the algorithm path that is not currently selected by the main processing channel.

6. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, The steps for generating feedback data specifically include: The segmentation confidence score is calculated based on the average circularity of all closed contours in the segmentation results. The dual-channel validation difference is calculated by comparing the local statistics of the main processing channel and the validation channel within the same representative region.

7. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 6, characterized in that, The steps for updating the policy model online include: The instant reward value for the current processing cycle is calculated based on the segment confidence and dual-channel verification difference. The Q-learning algorithm is applied to update the Q-values ​​in the policy model corresponding to the current state and action using the immediate reward value.

8. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 6, characterized in that, The method further includes: Based on the dual-channel verification difference, the motion model coefficients used in the feedforward correction step are self-tuned online.

9. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, The feedforward correction steps specifically include: Based on the real-time linear speed of the conveyor belt, the exposure time of the image acquisition unit, and a motion model coefficient, calculate the shearing factor in an affine transformation matrix. Using the affine transformation matrix, a geometric transformation is performed on the real-time image data to compensate for motion distortion of the image.

10. The method for detecting particle size of conveyor belt materials based on the optimized watershed algorithm according to claim 1, characterized in that, The adaptive preprocessing step includes adaptive dual denoising, which specifically involves: Apply Gaussian and median filters to grayscale images; The kernel standard deviation of the Gaussian filter and the kernel size of the median filter are specified by the algorithm parameter configuration generated by the algorithm steady-state engine.

Images