A computer vision-based slurry agglomeration online monitoring system and method

Abstract

The present application relates to the field of artificial intelligence, in particular to a kind of computer vision-based grinding fluid agglomeration online monitoring system and method, system includes, image acquisition unit is used to receive the grinding fluid image collected by camera and output pixel data, edge gap unit is used to determine edge threshold according to background area brightness difference sorting, extract grain edge point and form edge section, generate agglomeration spacing reference according to edge section spacing sorting, determine agglomeration connecting part in combination with continuous frame gap center offset and edge direction difference, grade output unit is used to generate agglomeration risk level and alarm according to the number of connected particles, continuous frame number and gray deposit intensity.The present application, by brightness difference, edge extension direction, edge section spacing and continuous frame change joint analysis, can identify stable agglomeration connection relationship under the fluctuation of grinding fluid background, reduce the misjudgment caused by instantaneous noise, improve the timeliness of agglomeration initial monitoring and risk level output reliability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an online monitoring system and method for grinding fluid agglomeration based on computer vision. Background Technology

[0002] The field of artificial intelligence technology involves data processing methods in the production sector that utilize computer equipment to model, identify, and control outputs from image samples, visual features, and status labels. Among these, online monitoring of grinding slurry agglomeration refers to an industrial control computer receiving images of the grinding slurry captured by an industrial camera, converting the images into model input data, extracting agglomeration regions, particle boundaries, or concentration distribution features through a pre-trained recognition model, and transmitting the recognition results to a controller, alarm unit, or production monitoring terminal.

[0003] Monitoring of grinding fluid agglomeration is mainly carried out through offline sampling and detection, manual microscopic observation, or overall image grayscale statistics: operators take samples from the grinding fluid circulation pipeline or working tank at regular intervals, and then confirm the presence of particle agglomeration through microscopic images, particle size detection, or experience judgment; in online image monitoring scenarios, the brightness change of the whole image, particle area change, or single frame threshold segmentation results are mostly used as the basis for judgment, with insufficient attention paid to the local gap relationship between particle edges, the spatial offset state in continuous frames, and connection stability.

[0004] The above solution has the following technical problems:

[0005] First, there is a time interval between offline sampling and on-site grinding, and the sample status cannot fully reflect the aggregation formation process in real time. Changes in the early stage of aggregation may be missed within the sampling interval, resulting in monitoring results lagging behind actual risks.

[0006] Second, manual microscopic observation relies on the operator's experience. It is difficult to consistently judge the close proximity of particle edges, the narrowing of local gaps, and the subtle connections. The results of observations by different personnel or different batches are prone to deviation.

[0007] Third, overall grayscale statistics focus more on changes in global image brightness or particle area. Fluctuations in liquid background brightness, uneven lighting, and changes in abrasive concentration can easily mask local edge differences, making it difficult to accurately identify fine aggregate connections.

[0008] Fourth, single-frame threshold segmentation is prone to mistaking instantaneous noise, overlapping images of suspended particles, or brief illumination disturbances for clustering. It lacks verification of continuous frame position offset and edge direction consistency, which increases the probability of false alarms.

[0009] Fifth, risk assessments often focus on whether particle aggregation occurs, lacking sufficient characterization of the correlation between the number of connected particles, duration, and deposition intensity, making it difficult to generate a stable aggregate risk level output.

[0010] Therefore, solutions characterized by offline sampling, manual judgment, overall grayscale statistics, and single-frame recognition are insufficient to meet the requirements of online monitoring of grinding fluid agglomeration for real-time performance, stability, anti-interference, connection relationship location, and risk classification alarm. Summary of the Invention

[0011] The purpose of this invention is to address the shortcomings of the existing technology by proposing an online monitoring system and method for grinding fluid agglomeration based on computer vision.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: an online monitoring system for grinding fluid agglomeration based on computer vision includes:

[0013] The image acquisition unit is used to receive the grinding fluid image captured by the camera, record the pixel brightness value and spatial position, and output the image pixel data;

[0014] An edge gap unit, connected to the image acquisition unit, is used to receive the image pixel data, sort the brightness difference between adjacent pixels in the liquid background area and determine the particle edge difference threshold, identify particle edge points that exceed the particle edge difference threshold, connect the particle edge points with the same extension direction to form a particle edge segment, calculate the shortest pixel distance between adjacent particle edge segments, generate a clustering spacing reference according to the first tenth position of all the shortest pixel distances in the current frame, determine suspected closed gaps, and determine the clustering connection part according to the center position offset and edge direction difference of the suspected closed gaps in consecutive frames.

[0015] The level output unit, connected to the edge gap unit, is used to receive the agglomeration connection part, calculate the number of connected particles, the number of continuous frames and the grayscale deposition intensity of the agglomeration connection part, generate the agglomeration risk level, and send an alarm signal to the controller.

[0016] Specifically, the image acquisition unit includes a brightness sampling submodule, a coordinate calibration submodule, and a pixel encapsulation submodule;

[0017] The brightness sampling submodule is used to receive the polishing slurry image frame output by the camera, read the brightness value of each pixel in the pixel scanning order, and mark the pixels with no brightness value, brightness overflow or brightness missing as invalid pixels to obtain valid brightness data.

[0018] The coordinate calibration submodule is used to determine the spatial position corresponding to the effective brightness data based on the row and column arrangement of the camera imaging area, establish the correspondence between pixel points and pixel coordinates, and obtain coordinate calibration data.

[0019] The pixel encapsulation submodule is used to associate and encapsulate the effective brightness data and the coordinate calibration data according to the same pixel point, and output the image pixel data.

[0020] Specifically, the edge gap unit includes an edge extraction submodule, a gap determination submodule, and a connection confirmation submodule;

[0021] The edge extraction submodule is used to receive the image pixel data, sort the brightness difference of adjacent pixels in the liquid background area to determine the particle edge difference threshold, and output the pixels that exceed the particle edge difference threshold as particle edge points.

[0022] The gap determination submodule is used to receive the particle edge points, connect them in a consistent direction of extension to form the particle edge segments, calculate the shortest pixel distance between adjacent particle edge segments, and compare the shortest pixel distance with the aggregation spacing benchmark to determine suspected closed gaps.

[0023] The connection confirmation submodule is used to receive the suspected closed gap in consecutive frames, compare the center position offset and the edge direction difference, and output the clustered connection part.

[0024] Specifically, the process for setting the particle edge difference threshold is as follows:

[0025] Obtain pixel pairs with horizontal and vertical adjacency in the liquid background area, remove any pixel pair marked as invalid, calculate the brightness difference between the remaining pixel pairs, sort them from low to high according to the brightness difference, read the sorted median, and determine the sorted median as the particle edge difference threshold.

[0026] Specifically, the steps for identifying the suspected closed gap are as follows:

[0027] Obtain the endpoint pixel coordinates and extension direction of adjacent particle edge segments, calculate the shortest pixel distance between the endpoints of adjacent particle edge segments, and determine the area between edge segments whose shortest pixel distance is not greater than the aggregation spacing reference as the suspected closed gap.

[0028] Specifically, the process of determining the agglomeration connection portion is as follows:

[0029] The center position of the suspected closed gap and the edge direction of the corresponding particle edge segment are obtained in the same region of consecutive frames. The center position offset of adjacent frames and the edge direction difference of adjacent frames are calculated. When the center position offset of adjacent frames is lower than a preset position drift threshold and the edge direction difference of adjacent frames is lower than a preset direction change threshold, the corresponding same region is marked as the agglomeration connection part.

[0030] Specifically, the grade output unit includes a particle count submodule, a continuous frame counting submodule, a deposition calculation submodule, and a grade generation submodule;

[0031] The particle count submodule is used to receive the agglomeration connection portion, trace the closed contour along the particle edge segments associated on both sides of the agglomeration connection portion, count the number of non-overlapping closed contours, and obtain the number of connected particles.

[0032] The continuous frame counting submodule is used to obtain the clustered connection portion in the continuous frame whose center position offset is lower than a preset position drift threshold, count the number of consecutive occurrences, and obtain the number of continuous frames;

[0033] The deposition calculation submodule is used to calculate the difference between the pixel brightness value inside the agglomeration connection and the brightness value of the adjacent liquid background, and generate the grayscale deposition intensity.

[0034] The level generation submodule is used to receive the number of connected particles, the number of continuous frames, and the grayscale deposition intensity, output the aggregation risk level, and send an alarm signal to the controller.

[0035] Specifically, the process of calculating the grayscale deposition intensity is as follows:

[0036] The pixel brightness values ​​within the area covered by the agglomeration connection and the liquid background brightness values ​​within a preset neighborhood outside the agglomeration connection are obtained. Pixels occupied by the particle edge points are removed. The brightness values ​​of the connection area and the background area are determined respectively. The difference between the brightness values ​​of the connection area and the background area is calculated to obtain the grayscale deposition intensity.

[0037] Specifically, the steps for determining the risk level of family reunification are as follows:

[0038] The number of connected particles, the number of continuous frames, and the grayscale deposition intensity are mapped to a particle number risk value, a continuous stability risk value, and a deposition occlusion risk value. The particle number risk value, the continuous stability risk value, and the deposition occlusion risk value are weighted and synthesized to obtain a clustering comprehensive score. The clustering risk level is output according to the risk level range in which the clustering comprehensive score is located.

[0039] A computer vision-based online monitoring method for grinding fluid agglomeration includes the following steps:

[0040] S1: Receive the image of the polishing fluid captured by the camera, record the brightness value and spatial position of the pixels, and output the image pixel data;

[0041] S2: Receive the image pixel data, sort the brightness difference of adjacent pixels in the liquid background area and determine the particle edge difference threshold, determine the particle edge points that exceed the particle edge difference threshold, connect the particle edge points with the same extension direction to form a particle edge segment, calculate the shortest pixel distance between adjacent particle edge segments, generate a clustering spacing reference according to the first tenth position of all the shortest pixel distances in the current frame, determine the suspected closed gap, and determine the clustering connection part according to the center position offset and edge direction difference of the suspected closed gap in consecutive frames;

[0042] S3: Receive the agglomeration connection part, calculate the number of connected particles, the number of continuous frames, and the grayscale deposition intensity of the agglomeration connection part, generate an agglomeration risk level, and send an alarm signal to the controller.

[0043] Compared with the technology, the advantages and positive effects of the present invention are as follows:

[0044] In this invention, by recording the pixel brightness and spatial position in the polishing slurry image, an edge determination benchmark that fits the current image state is formed by sorting adjacent pixels by brightness difference in the background. Then, the particle edges are organized along a consistent extension direction and the aggregation spacing benchmark is determined by sorting the edge segment spacing. This transforms the isolated grayscale changes of local gaps into comparable spatial structural features. Combined with the center offset of continuous frames and the edge direction difference, stable connection relationships are confirmed, reducing background fluctuations and instantaneous noise interference. Furthermore, the degree of aggregation development is comprehensively characterized by the number of connected particles, the number of continuous frames, and the grayscale deposition intensity, making early aggregation identification more timely, connection positions clearer, and risk levels more suitable for triggering polishing process control. Attached Figure Description

[0045] Figure 1 This is the main flowchart of the online monitoring method for grinding fluid agglomeration of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the effect of confirming a suspected closed gap in the present invention.

[0047] Figure 3 This is a schematic diagram illustrating the abnormal frame suppression alarm effect of the present invention;

[0048] Figure 4 This is a schematic diagram illustrating the output effect of the risk level of reunification in this invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] Example

[0051] Please see Figures 1 to 4 This embodiment provides an online monitoring system and method for grinding fluid agglomeration based on computer vision. In practical applications, for example, when the grinding fluid circulation pipeline of the grinding equipment is running continuously, a transparent observation section is installed at a stable flow position in the return or supply branch. A camera is fixed to collect images of the grinding fluid in the observation section. The industrial control computer receives image frames through the camera driver and completes pixel sorting, edge gap identification, agglomeration connection confirmation, and risk output. The controller receives alarm signals through the industrial communication interface and enters the grinding fluid supply control link. It mainly includes an image acquisition unit, an edge gap unit, and a level output unit.

[0052] The camera is an area-array industrial camera with an output image resolution of 2048×1536, a grayscale bit depth of 8 bits, and an image acquisition frequency of 60 frames / s. The effective imaging area of ​​the transparent observation section is 12.0mm×9.0mm, and the actual length of a single pixel after pixel calibration is approximately 0.00586mm. The outlet pressure of the grinding slurry circulation pump is stabilized between 0.18MPa and 0.22MPa, and the flow velocity within the observation section is controlled between 0.25m / s and 0.35m / s. The industrial control computer uses an Intel i5-class industrial processor and 8GB of memory, and the processing time for a single frame image from reception to risk level output is controlled within 25ms. The above data correspond to the actual engineering operating conditions during online monitoring of the grinding slurry, supporting camera acquisition, continuous frame buffering, and real-time alarm output.

[0053] Image acquisition unit: Connected to the camera driver output, it receives the polishing slurry image frames acquired by the camera. After the polishing slurry image frames enter the industrial control computer, they first enter the image receiving buffer. The buffer content includes frame status, pixel brightness records, pixel arrangement order, and information returned by the camera interface. The image acquisition unit organizes the pixels according to the row and column arrangement of the camera's imaging area, records the pixel brightness values ​​and spatial positions, and outputs the organized data as image pixel data.

[0054] In actual acquisition, a single frame image contains 3,145,728 pixels, each corresponding to a grayscale value record within the range of 0 to 255. The image acquisition unit writes the data of each frame into the buffer in row-major order, retaining the most recent 30 valid images in the buffer for comparison of the center position and edge direction of suspected closed gaps in subsequent consecutive frames.

[0055] When the camera driver returns a complete image frame, the image acquisition unit continues to perform brightness sampling and coordinate calibration. If the camera driver returns a dropped frame, transmission interruption, inconsistent frame header / tail, or misaligned pixel arrangement, the image acquisition unit writes the frame to the abnormal frame buffer and outputs a frame abnormality flag to the subsequent processing link. The frame abnormality flag is only used to maintain processing sequence and alarm traceability; it does not generate new particle edge points. The industrial control computer retains the processing state of the previous valid frame, and the level output unit enters the pending confirmation output phase upon receiving an abnormal frame flag, without using the abnormal frame as a basis for cluster formation.

[0056] Brightness sampling submodule: Reads the polishing slurry image frames output by the camera from the image receiving buffer, and reads the brightness value of each pixel according to the camera pixel scanning order. The brightness value is a grayscale record formed by the camera's response to the light intensity within the observation segment. The brightness sampling submodule is only responsible for reading and verifying this grayscale record.

[0057] Under stable slurry circulation, the grayscale value of the liquid background area is usually concentrated between 110 and 165, and the local grayscale variation of the edge area of ​​a single abrasive particle is usually higher than that of the smooth background area. The brightness sampling submodule marks pixels with grayscale values ​​below 5 or above 250 and abnormal continuous distribution as suspected sampling anomalies, and determines whether they have entered the invalid pixel set by combining the status of adjacent pixels.

[0058] During brightness sampling, if a pixel has no brightness record, its brightness record exceeds the camera's effective output boundary, its brightness field is missing, or its pixel record is inconsistent with the scanning order, the pixel is marked as invalid. The location of invalid pixels is still retained in the frame data, and subsequent edge extraction, gap judgment, and deposition calculation all read this mark and skip the corresponding pixel. For cases where duplicate brightness records appear at the same coordinates, the brightness sampling submodule retains the record consistent with the scanning order and writes the source of the duplication to the sampling anomaly log. The sampling anomaly log is bound to the image frame index, allowing maintenance personnel to view the camera's transmission and acquisition quality.

[0059] Effective brightness data consists of pixel brightness records, effective pixel identifiers, and abnormal pixel locations. If invalid pixels appear in a continuous area, the brightness sampling submodule adds a local failure identifier to that area, and subsequent edge gap units do not generate closed gaps in that area to avoid identifying acquisition defects as particle agglomeration.

[0060] The coordinate calibration submodule reads the effective brightness data and determines the pixel coordinates corresponding to each effective brightness record based on the row and column arrangement of the camera's imaging area. Pixel coordinates are used to describe the spatial position of pixels within the same image frame. The coordinate calibration submodule establishes the correspondence between pixels and pixel coordinates based on the row and column order, image orientation markers, and camera mounting parameters.

[0061] After coordinate calibration, the pixel coordinates are set with the top-left pixel as the origin, with horizontal coordinates ranging from 0 to 2047 and vertical coordinates ranging from 0 to 1535. The transparent observation section calibration plate is used to correct the imaging scale during installation. The calibration error is controlled within one pixel to ensure that the shortest pixel distance between adjacent particle edge segments can be stably mapped to the actual particle spacing changes within the observation section.

[0062] The coordinate calibration data includes valid pixel coordinates, invalid pixel coordinates, image orientation status, and observation area boundary status. If the image orientation marker is inconsistent with the camera installation parameters, the coordinate calibration submodule marks the frame as a coordinate aberration frame. The pixel encapsulation submodule only retains the original received record and the reason for the aberration of the frame and does not send the frame to edge extraction. If the coordinates are complete, the coordinate calibration data enters the pixel encapsulation submodule along with the valid brightness data.

[0063] Pixel Encapsulation Submodule: This submodule associates valid brightness data and coordinate calibration data at the same pixel point to form image pixel data. The image pixel data uses pixel coordinates as the retrieval entry point and includes brightness values, pixel validity, background candidate status, frame status, and acquisition log index. When the edge gap unit reads the image pixel data, it directly obtains the corresponding brightness and validity status based on the pixel coordinates.

[0064] Before output, the pixel encapsulation submodule checks the correspondence between the brightness record and the coordinate record. Pixels with a brightness record but no coordinate record are entered into the encapsulation exception buffer, while pixels with a coordinate record but no valid brightness are marked as invalid pixels. After encapsulation, the image pixel data is written to the current frame buffer of the industrial control computer, while retaining historical valid frames needed for continuous frame processing in frame order. If no historical valid frames have been formed after the device starts, the current frame still enters edge extraction, but the continuous frame confirmation step only establishes a buffer and does not directly output the stable clustering connection part.

[0065] Edge gap unit: The edge gap unit reads image pixel data from the current frame buffer, sorts the brightness differences of adjacent pixels within the liquid background region, and obtains the particle edge difference threshold. The liquid background region consists of pixel areas within the observation segment that are not occupied by the current edge point, not occupied by invalid pixels, and whose brightness changes conform to the background stability rules. The particle edge difference threshold is generated by the background brightness difference sorting result of the current valid frame, written to the current frame rule buffer, and only serves the particle edge point extraction of the current frame.

[0066] The absolute difference formula used for the brightness difference between adjacent pixels is: ΔG = |Gi - Gj|. Here, ΔG represents the grayscale difference between adjacent pixels, Gi represents the grayscale value of the i-th pixel, and Gj represents the grayscale value of the j-th pixel that is horizontally or vertically adjacent to the i-th pixel. In engineering practice, the ΔG of adjacent pixels in liquid background areas is mostly concentrated between 0 and 12, while the ΔG of particle edge areas typically reaches 18 or higher. The edge extraction submodule determines the particle edge difference threshold based on the ΔG sorting result of the current frame's background area and identifies valid pixels exceeding this threshold as particle edge points.

[0067] The edge gap unit identifies valid pixels exceeding the particle edge difference threshold as particle edge points, and then connects them according to the extension direction of adjacent particle edge points to form particle edge segments. After the particle edge segments are formed, the edge gap unit reads the endpoints of adjacent particle edge segments, calculates the nearest distance state between the endpoints, and sorts all nearest distance states in the current frame. The distance states corresponding to the previous segment positions are written into the clustering spacing reference buffer to determine whether there are suspected closed gaps between adjacent edge segments.

[0068] The pixel distance between the endpoints of adjacent particle edge segments is calculated using the Euclidean distance formula: d = Where d represents the pixel distance between the endpoints of the two edge segments, xa and ya represent the horizontal and vertical coordinates of the endpoint of the first particle edge segment, and xb and yb represent the horizontal and vertical coordinates of the endpoint of the second particle edge segment. Based on the pixel calibration relationship in this implementation scenario, a d of 12 pixels corresponds to an actual spacing of approximately 0.070 mm, and a d of 20 pixels corresponds to an actual spacing of approximately 0.117 mm. The edge gap unit sorts all d values ​​in the current frame and selects the distance state closest to the front segment as the clustering spacing benchmark, allowing edge segments closer to closure to preferentially enter the suspected closed gap judgment.

[0069] A suspected closed gap is determined by the distance between the endpoints of the edge segment, the extension direction of the edge segment, and the pixels in the region between the edge segments. When an edge gap unit detects a suspected closed gap in the current frame, it does not immediately output an alarm but instead submits it for confirmation processing in subsequent frames. If the center position of the suspected closed gap in the same region remains within the allowable range of the position drift rule in consecutive frames, and the edge direction change remains within the allowable range of the direction change rule, then that region is marked as a clustered connection.

[0070] During on-site debugging, the allowable drift range at the center of suspected closed gaps was set to within 6 pixels, and the allowable range for edge direction change was set to within 15°. Confirmation of the agglomeration connection was initiated after maintaining at least 5 consecutive frames. This data corresponds to a sampling frequency of 60 frames / s, which can filter out the phenomenon of single-frame particles briefly passing through the observation area with the flow of the polishing fluid.

[0071] Edge extraction submodule: After receiving image pixel data, it first reads pixel pairs with horizontal and vertical adjacency relationships in the liquid background region. When any pixel is marked as invalid, that pixel pair is not included in the brightness difference sorting. The brightness differences of the remaining pixel pairs are written to the background difference sorting cache in ascending order, and the brightness difference record corresponding to the sorting center position is used as the source of the particle edge difference threshold.

[0072] When the particle edge difference threshold is written to the current frame buffer, the source frame, background region identifier, and the validity status of pixels participating in the sorting are recorded simultaneously. The edge extraction submodule then reads the valid pixels within the current frame and outputs pixels with brightness differences exceeding the particle edge difference threshold as particle edge points. Valid pixels that do not meet the edge determination state are retained as background candidate pixels, and the subsequent deposition calculation submodule reads the brightness of adjacent liquid background pixels from the background candidate pixels.

[0073] When the background area is obscured by bubbles, sudden changes in lighting, or invalid pixels, preventing the sorting cache from generating a valid read result, the edge extraction submodule reads the particle edge difference threshold of the previous valid frame and appends a threshold inheritance flag to the current frame output. If there is no previous valid threshold during the startup phase, the current frame is marked as an edge-to-confirmation frame, and the gap determination submodule only saves the pixel state of that frame, without generating any suspected closed gaps. Both the threshold inheritance flag and the edge-to-confirmation flag are entered into the industrial control computer log for easy tracing of the reasons for alarm non-output or delayed confirmation.

[0074] Gap Determination Submodule: After receiving particle edge points, it connects particle edge segments according to pixel adjacency and local extension direction. The local extension direction is determined by the coordinate orientation of adjacent particle edge points and the continuous state of brightness difference changes. When the orientation of a candidate edge point is consistent with that of the current edge segment, the candidate edge point is included in the same particle edge segment; when the orientation changes abruptly, an invalid pixel block occurs, or the edge point connection relationship is interrupted, the current particle edge segment ends, and the endpoints of the edge segment are recorded.

[0075] The gap determination submodule reads the endpoint pixel coordinates and edge extension direction of adjacent particle edge segments to form endpoint distance states. These endpoint distance states are written to the current frame interval buffer and sorted according to their order. The states corresponding to the first segment positions in the current frame interval buffer are written to the clustering gap reference buffer. When the endpoint distance states of adjacent particle edge segments fall within the allowable range of the clustering gap reference, and the extension directions of the two edge segments converge, the region between the edge segments is identified as a suspected closed gap.

[0076] A suspected closed gap includes the gap center location, associated particle edge segment, endpoint location, edge direction, region pixel set, and frame index. If the gap region covers invalid pixels, the gap determination submodule adds a local failure flag, and the connection confirmation submodule delays confirmation of the gap in consecutive frames. If the edge segment is located at the boundary of the observation area, the gap determination submodule marks it as a boundary edge segment. The boundary edge segment can participate in particle contour tracking, but does not trigger the output of a suspected closed gap independently.

[0077] The connection confirmation submodule reads suspected closed gaps from the current frame and historical valid frames, and searches for the same region based on the center position, the direction of the associated edge segments, and the pixel coverage relationship of the area. The center position is used to mark the spatial location of the suspected closed gap in the observation area, and the edge direction is used to record the orientation of the particle edge segments on both sides of the suspected closed gap. The center position offset rules and edge direction change rules are derived from camera installation calibration, flow stability records of the observation section, and statistical results of historical valid frames.

[0078] After a suspected closed gap in the current frame successfully matches a suspected closed gap in a historical valid frame, the connection confirmation submodule compares the center position offset state and the edge direction change state. When the center position offset enters the position drift rule allowable state, and the edge direction change enters the direction change rule allowable state, the corresponding area is marked as a clustered connection. When the center position offset manifests as a significant drift with the flow direction, or when the edge direction change cannot maintain the same connection relationship, the suspected closed gap is marked as an instantaneous gap, and subsequent level output units do not read this area.

[0079] When there are insufficient consecutive frames, missing image frames, or interrupted frame order, the connection confirmation submodule does not use abnormal frames for matching. If there are no historical valid frames before the current valid frame, only a suspected closed gap buffer is established. If some historical valid frames exist but the continuous link is incomplete, the connection confirmation submodule outputs a pending review status. This pending review status, once entered into the level output unit, is only used to maintain the alarm link and does not directly increase the risk level. After the historical buffer is fully restored, the connection confirmation submodule re-outputs the clustered connection section according to the consecutive valid frame confirmation rules.

[0080] The graded output unit receives the clustered connection portion and reads the particle edge segments, continuous frame records, and regional brightness records associated with that connection portion, generating the number of connected particles, the number of consecutive frames, and the grayscale deposition intensity. The number of connected particles comes from the tracking results of the closed contours on both sides of the clustered connection portion; the number of consecutive frames comes from the maintenance state of the same region of the clustered connection portion in consecutive valid frames; and the grayscale deposition intensity comes from the difference between the brightness inside the clustered connection portion and the brightness of the adjacent liquid background.

[0081] The graded output unit writes the number of connected particles, the number of sustained frames, and the grayscale deposition intensity into the risk generation cache, respectively. The risk generation cache stores mapping rules for particle number risk values, sustained stability risk values, and deposition occlusion risk values. These mapping rules are derived from the polishing slurry process control configuration, camera calibration results, and historical valid image records, and are maintained by the industrial control computer's configuration management program.

[0082] The risk generation cache uses a percentage-based risk recording system. The risk value corresponding to the number of connection particles is categorized as low risk (0-30), medium risk (31-70), and high risk (71-100); the risk value corresponding to the number of sustained frames is categorized as short-term occurrence (0-35), stable existence (36-75), and persistent existence (76-100); and the risk value corresponding to the grayscale deposition intensity is categorized as slight deposition (0-30), significant deposition (31-70), and severe deposition (71-100).

[0083] After outputting the clustering risk level, the risk level output unit sends an alarm signal to the controller. The alarm signal includes the clustering risk level, the location status of the clustering connection, the current frame index, anomaly flag, and rule version. Upon receiving confirmation from the controller, the risk level output unit writes the confirmation result to the alarm log. If the controller does not confirm, the communication interface returns an error, or the alarm signal transmission fails, the risk level output unit maintains the current risk output state and appends a communication confirmation flag to the next valid alarm frame output.

[0084] Particle count submodule: After receiving the clustered connection, it traces the closed contour starting from the particle edge segments associated on both sides of the clustered connection. The closed contour is formed by connecting adjacent particle edge points along the same particle boundary. The contour record includes the set of edge points, endpoint connection status, associated clustered connection, and repeated tracking identifier. The particle count submodule reads segment by segment along the edge segment connection direction until the edge segment returns to the same contour closure position or encounters boundary truncation or invalid pixel blocking.

[0085] In actual grinding slurry images, the equivalent diameter of a single normal particle profile typically ranges from 8 μm to 35 μm, corresponding to a local edge width of approximately 1.4 to 6 pixels in the image. If three or more non-overlapping closed profiles are continuously tracked on both sides of the same agglomeration junction, and there is a stable edge connection between these profiles, the particle count submodule records the agglomeration junction as a multi-particle connection state.

[0086] Non-overlapping closed contours are entered into the connected particle count record. When a contour that has already been tracked is accessed again by the edge segments on both sides of the clustered connection, the particle count submodule skips the contour based on the repeated tracking flag. Boundary truncated contours, invalid pixel blocking contours, and edge segments that do not form a closed relationship are only written into the abnormal contour record and are not entered into the connected particle count. The connected particle count and abnormal contour records are passed together to the grading generation submodule, which uses this information to determine the mapping path for the particle count risk value.

[0087] The continuous frame counting submodule reads the frame index, center position, and pending verification status of the clustered connection from the connection confirmation submodule, and counts the retention status of the clustered connection in the same region across consecutive valid frames. If the clustered connection of the current valid frame matches the clustered connection of the previous valid frame, and the center position offset is still within the allowable range of the position drift rule, the continuous frame counting submodule continues the continuous recording of that region.

[0088] Under a data acquisition rate of 60 frames / s, 5 consecutive frames correspond to a stable observation time of approximately 83ms, and 15 consecutive frames correspond to a stable observation time of approximately 250ms. The continuous frame counting submodule uses 5 consecutive frames as the starting point for confirming a stable connection and 15 consecutive frames as the basis for judging the continuous existence state, in order to distinguish between short-term particle proximity and stable aggregation connections.

[0089] When the current frame is an abnormal frame, the frame sequence is interrupted, or the clustered connection enters a pending review state, the continuous frame counting submodule does not treat this frame as a valid occurrence record, but only retains the previous valid continuous state. When a suspected closed gap is determined to be a transient gap by the connection confirmation submodule, the continuous frame counting submodule ends the continuous recording of the corresponding region. When the clustered connection is confirmed again subsequently, the continuous frame counting submodule establishes a new continuous record. The continuous record is output to the level generation submodule and written to the frame-level risk cache.

[0090] The deposition calculation submodule reads the pixel brightness values ​​within the area covered by the aggregation connection region, as well as the liquid background brightness values ​​in the outer neighborhood of the aggregation connection region. The outer neighborhood is obtained by extending the boundary of the aggregation connection region outwards to the surrounding background candidate pixels, and its range is derived from the camera observation segment calibration and the current frame background candidate state. The deposition calculation submodule excludes particle edge points, invalid pixels, and pixels with abnormal boundaries when reading the brightness.

[0091] The grayscale deposition intensity is calculated using the regional mean difference formula: Idep = |Gcavg - Gbavg|. Here, Idep represents the grayscale deposition intensity, Gcavg represents the average grayscale value of effective pixels within the area covered by the agglomeration junction, and Gbavg represents the average grayscale value of effective pixels in the liquid background neighborhood outside the agglomeration junction. In field operation, the average grayscale of the liquid background is typically between 120 and 160. For slight deposition, Idep is typically 10 to 20; for significant deposition, Idep is typically 21 to 45; and for severe deposition, Idep is typically higher than 45. The deposition calculation submodule writes Idep into the grayscale deposition intensity field and passes it to the grade generation submodule.

[0092] The remaining valid pixels within the coverage area of ​​the clustered connection form the brightness state of the connection area, while the remaining valid background pixels in the outer neighborhood form the brightness state of the background area. The deposition calculation submodule compares the difference between the brightness states of the connection area and the background area and outputs the grayscale deposition intensity. The grayscale deposition intensity serves as input to the deposition occlusion risk value and is not output as a separate alarm signal.

[0093] When the outer neighborhood is heavily occupied by particle edge points, there are insufficient background candidate pixels, or local invalid pixels affect background reading, the deposition calculation submodule reads the background brightness state of the same area in the previous valid frame and adds a background inheritance flag. If there is no inheritable background state during the startup phase, the deposition calculation submodule outputs a deposition confirmation flag. The level generation submodule retains two inputs in this frame: particle count and continuous stability, and does not use deposition occlusion paths to increase the risk level.

[0094] The rating generation submodule reads the number of connected particles, the number of consecutive frames, and the grayscale deposition intensity, and then enters the mapping paths for particle number risk value, sustained stability risk value, and deposition occlusion risk value, respectively. The particle number risk value is used to represent the particle outline state connected to the agglomeration joint, the sustained stability risk value is used to represent the retention state of the agglomeration joint in consecutive valid frames, and the deposition occlusion risk value is used to represent the occlusion state of the liquid background brightness by the deposition inside the agglomeration joint.

[0095] The aggregate comprehensive score uses a weighted summation formula: S = Wn × Rn + Wt × Rt + Wd × Rd. Parameter S represents the aggregate comprehensive score, Rn represents the particle number risk value, Rt represents the sustained stability risk value, Rd represents the deposition occlusion risk value, Wn represents the weight corresponding to the particle number risk value, Wt represents the weight corresponding to the sustained stability risk value, and Wd represents the weight corresponding to the deposition occlusion risk value. Wn is set to 0.35, Wt to ​​0.30, and Wd to 0.35. A normal output is triggered when S is below 40, a concern output is triggered when S is between 40 and 69, and an alarm output is triggered when S reaches 70 or above. This score corresponds to the controller's alarm level. Upon receiving an alarm, the controller triggers grinding fluid circulation filtration or fluid supply adjustment.

[0096] The risk rating generation submodule reads the weight field and rating interval field from the risk rule cache, and synthesizes the particle quantity risk value, continuous stability risk value, and deposition occlusion risk value into a comprehensive agglomeration score status. The weight field is a process configuration field, recording the participation order and influence relationship of different risk sources in the rating generation process. It originates from the grinding fluid control process configuration and is updated by the configuration management program. The rating interval field is an alarm rule field, used to map the comprehensive agglomeration score status to a normal status, a concern status, an alarm status, or a fault pending confirmation status. When the configuration is updated, the industrial control computer records the configuration source and rule version. Subsequent valid frames use the new version of the rules, while historical frame logs retain the original rule version.

[0097] After the risk level generation submodule outputs the clustering risk level, the alarm interface sends the clustering risk level, the location status of the clustering connection, the anomaly flag, and the rule version to the controller. When the controller confirms receipt, the alarm log records the confirmation status. If the controller does not confirm, the alarm interface retains the pending confirmation flag and continues to maintain the most recent valid risk level. If the communication interface failure persists, the risk level output unit outputs a communication anomaly status to the industrial control computer's status management program. The status management program saves alarm records and image frame indexes, facilitating maintenance personnel to trace the clustering identification and alarm output process frame by frame.

[0098] Alarm signals are sent via Modbus TCP or an industrial Ethernet interface. Each alarm data frame includes the risk level, the center coordinates of the agglomeration connection, the current frame number, the anomaly identifier, and the rule version number. When the controller receives a status of concern, it maintains the original liquid supply strategy and records the trend. When it receives an alarm status, it increases the activation priority of the filter branch and sends an alarm receipt back to the industrial control computer.

[0099] The computer vision-based online monitoring method for grinding fluid agglomeration, operating in conjunction with the aforementioned system, is completed within the same industrial control computer processing link. First, it receives images of the grinding fluid captured by a camera, records pixel brightness values ​​and spatial positions, forming image pixel data. Then, the edge gap processing link receives the image pixel data, sorts the brightness differences between adjacent pixels in the liquid background area, determines the particle edge difference threshold, extracts particle edge points, connects particle edge points with consistent extension directions to form particle edge segments, reads the nearest distance status between adjacent particle edge segments, generates an agglomeration spacing benchmark based on the nearest distance sorting results of the current frame, identifies suspected closed gaps, and confirms agglomeration connections through continuous frame center position offset and edge direction changes. Subsequently, the level output processing link receives the agglomeration connections, generates the number of connected particles, the number of consecutive frames, and the grayscale deposition intensity, forming an agglomeration risk level, and sends an alarm signal to the controller.

[0100] Under the above operating conditions, the system can continuously process 60 frames / second of image data while the grinding slurry is circulating, and perform frame-by-frame tracking of particle edges, suspected closed gaps, and agglomeration junctions within the observation section. When the processing time of a single frame is controlled within 25ms, the industrial control computer can complete the risk output of the current frame before the next batch of images enters the buffer. After the alarm status is acknowledged by the controller, it is written into the operation log, forming a complete closed loop from image acquisition, image processing, risk judgment to control feedback.

[0101] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A computer vision-based online monitoring system for grinding slurry agglomeration, used for interactive monitoring of grinding slurry agglomeration status between a camera, an industrial computer, and a controller, characterized in that, The system includes: The image acquisition unit is used to receive the grinding fluid image captured by the camera, record the pixel brightness value and spatial position, and output the image pixel data; An edge gap unit, connected to the image acquisition unit, is used to receive the image pixel data, sort the brightness difference between adjacent pixels in the liquid background area and determine the particle edge difference threshold, identify particle edge points that exceed the particle edge difference threshold, connect the particle edge points with the same extension direction to form a particle edge segment, calculate the shortest pixel distance between adjacent particle edge segments, generate a clustering spacing reference according to the first tenth position of all the shortest pixel distances in the current frame, determine suspected closed gaps, and determine the clustering connection part according to the center position offset and edge direction difference of the suspected closed gaps in consecutive frames. A level output unit, connected to the edge gap unit, is used to receive the agglomeration connection part, calculate the number of connected particles, the number of continuous frames and the grayscale deposition intensity of the agglomeration connection part, generate an agglomeration risk level, and send an alarm signal to the controller. The grade output unit includes a grain count submodule, a continuous frame counting submodule, a deposition calculation submodule, and a grade generation submodule; The particle count submodule is used to receive the agglomeration connection portion, trace the closed contour along the particle edge segments associated on both sides of the agglomeration connection portion, count the number of non-overlapping closed contours, and obtain the number of connected particles. The continuous frame counting submodule is used to obtain the clustered connection portion in the continuous frame whose center position offset is lower than a preset position drift threshold, count the number of consecutive occurrences, and obtain the number of continuous frames; The deposition calculation submodule is used to calculate the difference between the pixel brightness value inside the agglomeration connection and the brightness value of the adjacent liquid background, and generate the grayscale deposition intensity. The grade generation submodule is used to receive the number of connected particles, the number of continuous frames, and the grayscale deposition intensity, output the aggregation risk level, and send an alarm signal to the controller. The process of calculating the grayscale deposition intensity is as follows: The pixel brightness values ​​within the area covered by the agglomeration connection and the liquid background brightness values ​​within a preset neighborhood outside the agglomeration connection are obtained. Pixels occupied by the particle edge points are removed. The brightness values ​​of the connection area and the background area are determined respectively. The difference between the brightness values ​​of the connection area and the background area is calculated to obtain the grayscale deposition intensity.

2. The computer vision-based online monitoring system for grinding fluid agglomeration according to claim 1, characterized in that, The image acquisition unit includes a brightness sampling submodule, a coordinate calibration submodule, and a pixel encapsulation submodule; The brightness sampling submodule is used to receive the polishing slurry image frame output by the camera, read the brightness value of each pixel in the pixel scanning order, and mark the pixels with no brightness value, brightness overflow or brightness missing as invalid pixels to obtain valid brightness data. The coordinate calibration submodule is used to determine the spatial position corresponding to the effective brightness data based on the row and column arrangement of the camera imaging area, establish the correspondence between pixel points and pixel coordinates, and obtain coordinate calibration data. The pixel encapsulation submodule is used to associate and encapsulate the effective brightness data and the coordinate calibration data according to the same pixel point, and output the image pixel data.

3. The computer vision-based online monitoring system for grinding fluid agglomeration according to claim 1, characterized in that, The edge gap unit includes an edge extraction submodule, a gap determination submodule, and a connection confirmation submodule; The edge extraction submodule is used to receive the image pixel data, sort the brightness difference of adjacent pixels in the liquid background area to determine the particle edge difference threshold, and output the pixels that exceed the particle edge difference threshold as particle edge points. The gap determination submodule is used to receive the particle edge points, connect them in a consistent direction of extension to form the particle edge segments, calculate the shortest pixel distance between adjacent particle edge segments, and compare the shortest pixel distance with the aggregation spacing benchmark to determine suspected closed gaps. The connection confirmation submodule is used to receive the suspected closed gap in consecutive frames, compare the center position offset and the edge direction difference, and output the clustered connection part.

4. The computer vision-based online monitoring system for grinding fluid agglomeration according to claim 3, characterized in that, The process for setting the particle edge difference threshold is as follows: Obtain pixel pairs with horizontal and vertical adjacency in the liquid background area, remove any pixel pair marked as invalid, calculate the brightness difference between the remaining pixel pairs, sort them from low to high according to the brightness difference, read the sorted median, and determine the sorted median as the particle edge difference threshold.

5. The computer vision-based online monitoring system for grinding fluid agglomeration according to claim 3, characterized in that, The steps for identifying the suspected closed gap are as follows: Obtain the endpoint pixel coordinates and extension direction of adjacent particle edge segments, calculate the shortest pixel distance between the endpoints of adjacent particle edge segments, and determine the area between edge segments whose shortest pixel distance is not greater than the aggregation spacing reference as the suspected closed gap.

6. The computer vision-based online monitoring system for grinding fluid agglomeration according to claim 3, characterized in that, The process for determining the agglomeration connection is as follows: The center position of the suspected closed gap and the edge direction of the corresponding particle edge segment are obtained in the same region of consecutive frames. The center position offset of adjacent frames and the edge direction difference of adjacent frames are calculated. When the center position offset of adjacent frames is lower than a preset position drift threshold and the edge direction difference of adjacent frames is lower than a preset direction change threshold, the corresponding same region is marked as the agglomeration connection part.

7. The computer vision-based online monitoring system for grinding fluid agglomeration according to claim 1, characterized in that, The steps for determining the risk level of family reunification are as follows: The number of connected particles, the number of continuous frames, and the grayscale deposition intensity are mapped to a particle number risk value, a continuous stability risk value, and a deposition occlusion risk value. The particle number risk value, the continuous stability risk value, and the deposition occlusion risk value are weighted and synthesized to obtain a clustering comprehensive score. The clustering risk level is output according to the risk level range in which the clustering comprehensive score is located.

8. A computer vision-based online monitoring method for grinding fluid agglomeration, characterized in that, The method is used to implement the computer vision-based online monitoring system for grinding fluid agglomeration as described in any one of claims 1-7, and includes the following steps: Receive images of the polishing slurry captured by the camera, record the brightness values ​​and spatial positions of the pixels, and output the image pixel data; The image pixel data is received, the brightness difference between adjacent pixels in the liquid background area is sorted and a particle edge difference threshold is determined, particle edge points exceeding the particle edge difference threshold are identified, and particle edge points with the same extension direction are connected to form particle edge segments. The shortest pixel distance between adjacent particle edge segments is calculated, and a clustering spacing reference is generated according to the first tenth position of all the shortest pixel distances in the current frame. Suspected closed gaps are identified, and clustering connection parts are determined according to the center position offset and edge direction difference of the suspected closed gaps in consecutive frames. The system receives the agglomeration connection portion, calculates the number of connected particles, the number of consecutive frames, and the grayscale deposition intensity of the agglomeration connection portion, generates an agglomeration risk level, and sends an alarm signal to the controller.

Images