Diamond polishing and grinding online detection system

Abstract

The invention relates to the field of solid analysis, in particular to a diamond polishing and grinding online detection system which comprises the following steps: firstly, synchronously acquiring optical sensing data and thermal sensing data of the surface of a to-be-detected diamond; then, processing the acquired sensing data to detect and identify various physical characteristics of the surface, including a surface defect state, thermally induced anomaly distribution and self-calibrated surface roughness, and tracking an evolution trend of related characteristics, so as to generate a preliminary multi-dimensional characteristic analysis result; and finally, generating a comprehensive diamond defect surface physical property online evaluation report by integrating various analysis results. The method mainly solves the problems that the multi-dimensional physical characteristics of the surface of the diamond are difficult to evaluate in real time, accurately and comprehensively and the change of the surface state is difficult to respond quickly by a traditional detection means, and aims at comprehensively evaluating the surface quality of the diamond through automatic online analysis and providing a basis for guaranteeing the quality of a final product and optimizing a grinding process.

Description

Technical Field

[0001] The present invention relates to the technical field of solid analysis, and particularly to an on-line detection system for diamond polishing and grinding. Background Art

[0002] For the precise determination and in-depth analysis of the physical properties of diamond, the prior art involves various means aimed at quantitatively characterizing its key physical parameters, such as surface topography, the size and distribution of micro-defects, optical properties, etc., so as to evaluate its detection effect. For example, through specific optical configurations, such as bright and dark field illumination, differential interference contrast, the detection and characterization capabilities of micro-scratches, pits or contaminants can be enhanced, thereby evaluating the quality based on physical parameters.

[0003] However, when applying these analysis techniques to the on-line detection system of diamond, many challenges still exist. Although traditional laboratory precision analysis instruments can provide high-resolution surface physical property information, their detection speed is slow, they have strict environmental requirements, and it is difficult to directly integrate them into high-speed production lines with vibration, coolant, and grinding dust for real-time determination of physical properties. For diamond materials with high hardness, high reflectivity or transparency, some optical detection methods may encounter problems such as low signal contrast, serious stray light interference or limited penetration depth when measuring their surface physical properties, which directly affect the accuracy and reliability of the physical property determination results. In addition, the on-line detection system needs to process a large amount of sensor data in a very short time, and accurately identify and quantify various tiny and differently shaped surface defects or parameter deviations. These techniques still have deficiencies in terms of real-time performance, robustness, and intelligence level. Therefore, at present, the quality control of most diamonds still relies on off-line sampling inspection, which not only leads to a lag in feedback on the dynamic changes of the physical properties of the material, making it difficult to achieve real-time optimization of the processing technology, but also increases the risk of waste products.

[0004] Therefore, an on-line detection system for diamond polishing and grinding is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide an on-line detection system for diamond polishing and grinding, which includes a defect data acquisition unit for synchronously acquiring optical sensing data and thermal sensing data on the surface of diamond, and performing preprocessing operations including image enhancement, filtering, and digital feature extraction to output optical characterization data; a defect characteristic analysis unit, based on the optical characterization data, performing image content interpretation through a fusion of data-driven and visual detection models to identify surface defect information, and performing image feature matching and temporal tracking to output corresponding defect evolution trend information; and a defect information integration unit, using a rule engine to generate a defect surface physical property report to achieve closed-loop feedback, realizing on-line and automated diamond surface state evaluation and process optimization.

[0006] To achieve the above object, the present invention provides the following technical solutions: A diamond polishing and grinding on-line detection system, comprising: A defect data acquisition unit, configured to synchronously acquire optical sensing data and thermal sensing data on the surface of a diamond, and perform preprocessing to output optical characterization data containing surface physical property information; A defect characteristic analysis unit, configured to detect defects on the surface of the diamond based on the optical characterization data by fusing a data-driven model and a visual detection model, output surface defect information, and detect thermal-induced anomalies in combination with the thermal sensing data to output thermal-induced anomaly information; calculate the surface roughness of the diamond surface based on the optical characterization data, and perform self-calibration of the surface roughness calculation by comparing with external reference data to output surface roughness information; and compare the surface defect information with a historical defect database to output defect evolution trend information; A defect information integration unit, configured to perform feedback decision-making using a rule engine based on the surface defect information, the thermal-induced anomaly information, the surface roughness information, and the defect evolution trend information, generate a defect surface physical property report, and output the report to a controller through an industrial communication interface.

[0007] Further, the preprocessing specifically includes: Converting the optical sensing data into a single-channel grayscale image; Applying a contrast-limited adaptive histogram equalization algorithm to the single-channel grayscale image; Using a guided filter algorithm to filter out noise from the single-channel grayscale image after contrast enhancement; And for the filtered single-channel grayscale image, extracting a region of interest through template matching technology based on normalized cross-correlation, and applying multi-level wavelet transform to extract and reconstruct low-frequency approximation components and high-frequency detail components to form the optical characterization data.

[0008] Further, the process of outputting surface defect information includes: The data-driven model uses a lightweight convolutional neural network; using a lightweight convolutional neural network model pre-trained for diamond surface defect types, automatically extracting and identifying surface irregular defects from the input optical characterization data, and outputting a first defect detection result, including defect category, a prediction score with a confidence level greater than a preset threshold, and bounding box coordinates; The visual detection model integrates a Canny edge detection algorithm and a Hough transform algorithm to identify and locate surface regular defects, and obtain a second defect detection result; For the first defect detection result and the second defect detection result, fusion processing is performed based on the confidence weighting factor and / or the intersection over union between bounding boxes, and the surface defect information is generated through redundant elimination and conflict resolution.

[0009] Further, the process of outputting the surface roughness information includes: Based on the optical characterization data, fast Fourier transform processing is applied to analyze the spectral energy distribution of the low-frequency approximation component and the high-frequency detail component; and according to the predetermined estimation model between the spectral energy distribution and the surface roughness, the surface roughness is estimated and the texture feature parameters are quantified to obtain the preliminary surface roughness information. Through the calibration data interface, the reference value of the baseline surface roughness for the same sample area is regularly received; and an adaptive calibration algorithm is applied to calculate the deviation amount by comparing the surface roughness with the reference value of the baseline surface roughness. Based on the deviation amount, the internal parameters of the predetermined estimation model are adjusted, and the surface roughness is recalculated to output the calibrated surface roughness information.

[0010] Further, the process of outputting the defect evolution trend information includes: The surface defect information is subjected to spatio-temporal dimensional correlation processing and multi-dimensional feature parameter matching with the defect map data stored in the historical defect database to obtain the associated defect numbers. Track the time-series quantization change data corresponding to the defect numbers in the optical sensing data, and use a preset time-series analysis model to analyze the time-series quantization change data to identify and quantify the evolution patterns corresponding to the defect numbers; wherein, the evolution patterns are parameter vectors representing the defect size change rate, the morphological transformation characteristics, and the dynamic statistical characteristics of the occurrence frequency. Based on the evolution patterns and a preset risk assessment rule set, the future development trends and potential risk levels of the corresponding defect numbers are determined to obtain the defect evolution trend information.

[0011] Further, the specific implementation process of the rule engine includes: Analyze the severity of the surface defect information and the degree of deviation of the surface roughness information from the preset target specifications, and set a processing priority queue. According to the processing priority queue, the rule engine comprehensively applies the surface defect information, the defect evolution trend information, the surface roughness information, and the thermally induced abnormal information, and uses an adjustment action rule library to generate the defect surface physical property report.

[0012] Furthermore, the defect information integration unit further includes: adopting the isolation forest algorithm and the principal component analysis reconstruction error analysis method to compare the surface defect information, the thermally induced anomaly information, and the surface roughness information with the normal working condition model to identify new surface anomaly states, which are used to trigger and update the rule engine to generate the defect surface physical property report.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Through the defect data acquisition unit, the present invention realizes the synchronous and high-fidelity acquisition of the optical sensing data and the thermal sensing data on the diamond surface, and combines the precise image preprocessing process including image enhancement, filtering, and targeted feature extraction. It not only effectively overcomes the inherent problems of insufficient information dimension and low signal-to-noise ratio in comprehensively and finely characterizing the complex surface physical properties by traditional single information sources, but also the highly standardized optical characterization data containing rich physical property information output by this unit lays a high-quality and multi-modal fusion data foundation for subsequent multi-dimensional intelligent analysis and precise defect interpretation.

[0014] 2. The defect characteristic analysis unit of the present invention not only realizes the efficient identification and accurate information output of various complex and regular defects on the diamond surface by integrating the data-driven image analysis model and the machine vision algorithm; at the same time, it integrates the analysis of thermal sensing data to identify potential thermally induced anomalies, can precisely estimate the surface roughness based on the optical image features, and realizes self-calibration by comparing with external reference data to ensure the long-term accuracy of the measurement values. In addition, by intelligently comparing the current defect information with the historical defect database and performing temporal tracking, it provides accurate quantitative indicators for dynamically grasping the real-time quality status of the diamond surface.

[0015] 3. The defect information integration unit of the present invention realizes the rapid, accurate mapping and intelligent decision-making from the complex image analysis results to the specific defect surface physical property report through its built-in rule engine. By outputting these adjustment suggestions to the device controller in real time through a standardized industrial communication interface, it constitutes an agile closed-loop feedback loop from on-line detection to process execution, which not only improves the timeliness and pertinence of process adjustment, effectively suppresses the occurrence of quality fluctuations, but also promotes the diamond polishing and grinding process to develop in the direction of higher automation, higher precision, and better resource utilization rate, thereby improving the final product qualification rate and the overall efficiency of the production line. Description of the Drawings

[0016] Figure 1 It is a schematic structural diagram of an on-line detection system for diamond polishing and grinding provided by the present invention; Figure 2 It is a schematic flow diagram of the surface defect information output process provided by the present invention; Figure 3 This is a schematic flow chart of the surface roughness information output process provided by the present invention. Specific embodiments

[0017] Next, in conjunction with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0018] Please refer to Figures 1 to 3 , the present invention provides a diamond polishing and grinding on-line detection system, and the technical solution is as follows:

[0019] Embodiment 1: In order to effectively address the challenges faced in precisely and real-time on-line monitoring of the surface quality of workpieces during the diamond polishing and grinding process, especially for the difficulties such as diverse defect morphologies, tiny feature sizes, and interference from complex texture backgrounds, Embodiment 1 of the present invention proposes and elaborates in detail a diamond polishing and grinding on-line detection system, as Figure 1 shown, including: Refer to Figure 1 the defect data acquisition unit for synchronously acquiring the optical sensing data and thermal sensing data of the diamond surface, and performing preprocessing to output optical characterization data containing surface physical property information.

[0020] Specifically, in this embodiment, a high-speed industrial camera is integrated to capture high-resolution optical sensing data of the diamond workpiece. At the same time, the system is also equipped with an infrared thermal imaging device, which can work synchronously with the optical camera to collect thermal sensing data of the same or adjacent areas, and thus directly reflect the real-time temperature distribution on the surface of the workpiece.

[0021] Furthermore, for the optical sensing data obtained from the industrial camera, the preprocessing specifically includes: First, convert the optical sensing data captured by the industrial camera (usually in RGB format) into a single-channel grayscale image. This processing strips the color information from the image, enabling the subsequent analysis process to focus more on the two physical attributes of brightness and contrast, which are closely related to surface defects (such as scratches, pockmarks, cracks) and texture features. Through grayscale processing, the data dimension and computational complexity can be effectively reduced, and the real-time performance of subsequent processing can be improved.

[0022] Next, the contrast-limited adaptive histogram equalization (CLAHE) method is applied to the converted grayscale image. The CLAHE algorithm divides the image into multiple small grid regions (tiles) and performs local histogram equalization on each region separately, while setting clipping limits to prevent excessive local contrast amplification, thereby stably enhancing the image detail contrast. In the diamond polished surface image, it can enhance the visibility of tiny scratches and pits in low-contrast areas, which is helpful for subsequent feature extraction and defect identification.

[0023] Subsequently, the guided filtering algorithm is used to filter the noise of the single-channel grayscale image after contrast enhancement after the image processed by CLAHE. The guided filtering algorithm uses the structural features of the image itself as a guiding reference, effectively suppressing high-frequency interference while retaining edge details, such as coolant reflection, ambient light fluctuations and image noise generated by the sensor itself, while suppressing noise and retaining the continuity of key features such as micro-crack edges and scratch contours.

[0024] Next, in order to focus analysis resources and improve processing efficiency, a template matching method based on normalized cross correlation (NCC) is used to dynamically extract the region of interest (ROI) from the denoised image. The similarity score is calculated by matching the defect template image preset in the system block by block, and the matching confidence is combined to determine whether to extract a specific area as a subsequent processing object. When the matching result does not reach the set threshold, the entire image is used for processing by default to ensure the integrity and robustness of the detection.

[0025] Finally, a multi-level discrete wavelet transform is applied to the extracted ROI or the entire image for feature extraction. The wavelet transform decomposes the image into low-frequency approximate components and high-frequency detail components in multiple directions. The former represents the image contour and slowly changing areas, and the latter can accurately capture information such as edges and texture mutations representing defects in the image. In this embodiment, the wavelet basis can be selected from common forms such as "db4" or "sym4" according to the texture characteristics of the diamond surface, and 2 to 4 levels of decomposition are performed to construct a feature matrix of uniform size (such as 256×256) to facilitate subsequent recognition model calls.

[0026] After preprocessing the surface characteristics of diamond, the system can extract standardized feature image information with complete structure, retained details, concentrated features and moderate data volume from the original optical image, providing a high-quality data foundation for subsequent defect identification, roughness analysis and thermal anomaly detection, improving the system's ability to recognize tiny defects and the accuracy and stability of overall detection.

[0027] refer to Figure 1A defect characteristic analysis unit is configured to detect defects on the diamond surface based on the optical characterization data by fusing a data-driven model and a vision detection model, output surface defect information, and detect thermal anomalies in combination with the thermal sensing data to output thermal anomaly information; calculate the surface roughness of the diamond surface based on the optical characterization data, and perform self-calibration of the surface roughness calculation by comparing with external reference data to output surface roughness information; and compare the surface defect information with a historical defect database to output defect evolution trend information.

[0028] Among them, this unit combines the input thermal sensing data or its derived features to analyze whether there are thermal anomalies caused by factors such as processing friction, internal stress of materials, or coating problems. For example, the local temperature exceeds the normal operating condition threshold (such as >50 °C) and the temperature gradient is abnormal (such as >10 °C / mm), and corresponding thermal anomaly information is output.

[0029] Furthermore, Figure 2 is a schematic flowchart of the surface defect information output process provided by the present invention. As Figure 2 shown, the process of outputting surface defect information includes: The data-driven model adopts a lightweight convolutional neural network (CNN), such as MobileNetV3 or EfficientNet-Lite. This model is pre-trained in a supervised learning manner based on a large amount of diamond surface image data labeled with defect categories such as microcracks, pits, pockmarks, contamination spots, and burn marks and their bounding box positions. Data augmentation techniques such as rotation, mirroring, and scale transformation are introduced during the training process to improve the robustness of the model under different lighting, viewing angles, and texture changes.

[0030] In the online detection stage, a lightweight convolutional neural network model pre-trained for diamond surface defect types is used to automatically extract and identify surface irregular defects from the input optical characterization data, and output a structured first defect detection result, including: defect category label, confidence score greater than a preset threshold, and defect bounding box coordinate information.

[0031] The vision detection model integrates the Canny edge detection algorithm and the Hough transform method to identify and locate surface regular defects, such as straight scratches caused by abrasive scratches or circular arc defects caused by chipping. Among them, the Canny algorithm extracts the image edge structure, and the Hough transform is used to identify geometric elements that conform to the parameter model from the edge map, outputting information such as the start and end point coordinates, length, and angle of the scratch, or the center coordinates, radius, start, and end angles of the circular arc, thereby forming a second defect detection result.

[0032] To improve the integrity and accuracy of the recognition results, the first defect detection result and the second defect detection result are fused based on the confidence weighting factor and / or the intersection over union (IoU) between bounding boxes, and the surface defect information is generated through redundant elimination and conflict resolution, which specifically includes: Weighted fusion: The results are fused based on the confidence weight factor of each detection source and / or the intersection over union (IoU) of the bounding boxes. For example, a relatively high initial weight can be assigned to the irregular defects recognized by the CNN model; while for regular defects, the fusion weight is adjusted according to traditional features such as geometric structure integrity and edge strength.

[0033] Redundant elimination: Detection boxes with a high IoU overlap and low confidence among multiple detection results are eliminated to avoid duplicate counting; for the same type of defect information that is repeatedly recognized, they are merged and the one with a higher confidence is retained.

[0034] Conflict resolution: When the same area is recognized as different types of defects by two models, the system determines the most reasonable defect classification and description method based on the preset priority rules, the matching results of typical morphological features, or by introducing context features for auxiliary discrimination.

[0035] The surface defect information should include the unique identification ID, type classification, geometric position coordinates, size parameters (such as length, area or estimated depth value) of each defect, and optional morphological description features.

[0036] By fusing the data-driven model and the machine vision detection algorithm, and combining their respective advantages in the detection of irregular and regular defects, the comprehensiveness, accuracy and stability of the diamond surface defect detection are improved. With the intelligent fusion processing mechanism, not only the complementary advantages are realized, breaking through the limitations of a single detection method in defect type coverage and recognition accuracy, but also through redundant elimination and conflict resolution, it ensures that the finally output surface defect information is more reliable and complete, providing better input for subsequent quality assessment and process feedback, thus guiding production more effectively and reducing the risk of misjudgment.

[0037] Furthermore, Figure 3 This is a schematic flow chart of the surface roughness information output process provided by the present invention. As Figure 3 shown, the process of outputting the surface roughness information includes: First, based on the optical characterization data, fast Fourier transform (FFT) processing is applied to analyze the spectral energy distribution of the low-frequency approximation component and the high-frequency detail component. The spectral energy distribution includes but is not limited to statistical indicators such as the proportion of band energy, the main frequency position, and the spectral width.

[0038] Subsequently, based on the spectral energy distribution and the surface roughness (arithmetic mean roughness A predetermined estimation model between them estimates the surface roughness and quantifies the texture feature parameters to obtain the preliminary surface roughness information. Among them, the texture feature parameters include, but are not limited to, indicators such as texture directionality, periodicity, contrast, and uniformity. The surface roughness and the quantified texture feature parameters together constitute the preliminary surface roughness information.

[0039] In addition, the predetermined estimation model is modeled according to the relationship between the image frequency domain features and the external physical measurement results to improve the accuracy and generalization ability of roughness estimation. The following can be selected: A physical analysis model established based on the optical scattering theory; An empirical model based on regression training of large sample data, such as a linear regression model, a support vector regression model, or a lightweight neural network model, etc.

[0040] Through the set calibration data interface, the system regularly (periodically or on demand) receives the reference value of the benchmark surface roughness for the same sample area. This reference value is usually measured by instruments such as an atomic force microscope (AFM), a white light interferometer (WLI), or a contact profilometer, and can be imported into the system through a manual input interface, file import, or network connection. Common calibration trigger conditions include: a fixed daily cycle, batch switching, abrasive tool replacement, or after adjustment of key process parameters. The system applies an adaptive calibration algorithm. By comparing the surface roughness with the reference value of the benchmark surface roughness, it calculates the absolute error or relative error to obtain the deviation amount, which is used as the basis for model calibration.

[0041] Based on the deviation amount, adjust the internal parameters of the predetermined estimation model, and recalculate the surface roughness, and output the calibrated surface roughness information. For example: For a linear model , the coefficients a, b, and c can be corrected according to the error feedback, where is the arithmetic mean roughness, and are spectral features that can reflect roughness-related information extracted from the spectral energy distribution; For a look-up table model, the output value of the corresponding spectral entry can be adjusted; For a neural network or other non-linear model, its weight parameters or bias terms can be updated.

[0042] On the one hand, using FFT to analyze the spectral energy of image features can quickly and non-contactly obtain rich information related to the surface microtopography for online estimation On the other hand, it provides an effective way. Through the self - calibration function of regularly comparing with the high - precision physical measurement reference value and automatically adjusting the estimation model parameters, it effectively overcomes the defects of the roughness measurement method based solely on image analysis, which is prone to drift or systematic errors due to factors such as light changes, surface contamination, and insufficient model generalization ability, and provides reliable feedback data for fine - diamond process control.

[0043] Furthermore, the process of outputting the defect evolution trend information includes: First, perform spatio - temporal correlation processing and multi - dimensional feature parameter matching on the surface defect information and the defect map data stored in the historical defect database to obtain the associated defect number.

[0044] During the association process, the system comprehensively considers spatio - temporal attributes such as the proximity of the defect spatial position, the predicted potential movement trajectory, and the corresponding processing stage timestamp, and at the same time matches multi - dimensional features such as the type, size, and shape factor of the defect to identify whether it belongs to the evolution body of the same defect in the historical database, and then establish a tracking relationship and obtain the associated defect number.

[0045] Among them, the historical defect database records the defect states observed at multiple different detection time points or processing stages of the same workpiece (such as repeated detection at a fixed workstation) or workpieces of the same batch, providing a reference basis for time - series evolution analysis.

[0046] Track the time - series quantization change data corresponding to the defect number in the optical sensing data, including the geometric parameters (such as length, width, area, depth), morphological parameters (such as aspect ratio, circularity, edge complexity), and relative position parameters (such as distance from the edge or offset within the region) in consecutive detection images that change with time or processing cycle. And use a preset time - series analysis model to analyze the time - series quantization change data to identify and quantify the evolution pattern corresponding to the defect number. The model can be a traditional statistical method (such as ARIMA, adaptive exponential smoothing, Kalman filter) or a deep - learning - based sequence model (such as RNN or LSTM), which is used to identify the evolution law of the defect and extract its dynamic change characteristics.

[0047] Among them, the evolution pattern is a parameter vector representing the defect size change rate (such as linear growth, exponential growth, or steady state), morphological transformation characteristics (such as a point expanding into a line, the edge tending to be irregular), and the dynamic statistical characteristics of the occurrence frequency (such as the defect appearing periodically or the trend of the occurrence frequency increasing). This vector constitutes a mathematical representation of the defect time - series evolution behavior.

[0048] Based on the evolution pattern and a preset risk assessment rule set, determine the future development trend and potential risk level of each corresponding defect number to obtain the defect evolution trend information.

[0049] Among them, the risk assessment rule set can be established based on expert experience or historical big data regression analysis. For example: If the growth rate of the crack length exceeds the set threshold, the predicted future development trend is "rapid expansion", and the corresponding potential risk level is set to "high".

[0050] The system finally generates defect evolution trend information, which includes predictive future development trends and potential risk levels, and can be used for subsequent diamond process adjustment strategies or maintenance plan formulation. By intelligently associating the current detection information with historical data and performing time-series modeling, the system can accurately quantify the change rate, morphological evolution trend, and periodic behavior of defects, and accordingly achieve forward-looking risk assessment. Preventive measures can be taken before the defects evolve into an unacceptable state, thereby reducing the risk of batch defects, improving the reliability of diamond polishing and grinding, and providing a data basis for process optimization.

[0051] The defect information integration unit is used to generate a defect surface physical property report through feedback decision-making using a rule engine based on the surface defect information, the thermally induced anomaly information, the surface roughness information, and the defect evolution trend information, and output it to the controller through an industrial communication interface.

[0052] Furthermore, the specific implementation process of the rule engine includes: The rule engine first receives various output information from the defect characteristic analysis unit, mainly including: The surface defect information: including defect type, location, size, quantity, severity level assessment, etc.; The defect evolution trend information: including predictions of the future development and risks of each defect; The calibrated surface roughness information: such as the Ra value and its deviation from the target specification; And the thermally induced anomaly information: such as the maximum temperature and area range of the local overheating area.

[0053] Before processing these input information, the system first evaluates the severity of the surface defects (such as whether they are critical defects, whether the size exceeds the threshold, and whether the evolution trend shows rapid deterioration), and analyzes the deviation degree of the Ra value from the target roughness specification.

[0054] Based on these analysis results, the system sets processing priorities for various input information and constructs a dynamic priority task queue, thus forming a processing priority queue. For example, if the system detects the development of high-risk cracks or a continuous serious over-standard of the Ra value, this task will be assigned the highest priority to ensure the priority response to key process problems.

[0055] Based on the processing priority queue, the rule engine comprehensively considers the surface defect information, the defect evolution trend information, the surface roughness information, and the thermally induced anomaly information, and applies the adjustment action rule library to generate the physical property report of the defect surface.

[0056] Among them, the rule library is a structured logical set, which can be configured and maintained by process experts. It is constructed in the "IF-THEN" logical format and specifically includes: Condition part (IF): composed of the combined logic of one or more detection parameters; Action part (THEN): clearly instructs to perform specific adjustment operations on one or more processing parameters, such as adjusting the pressure of the grinding head, the spindle speed, the feed rate, the abrasive concentration, or the coolant flow rate, etc.; the adjustment forms include percentage increase or decrease, setting target values, or switching process parameter groups.

[0057] If multiple rules are triggered simultaneously, the system processes them according to the preset rule priorities, conflict resolution strategies, or meta-rules. The processing strategies include: selecting a conservative adjustment plan (to avoid over-response); weighted fusion of multiple proposed adjustment amplitudes; preferentially retaining high-risk response actions according to the risk level.

[0058] Through the above rule matching and execution process, the rule engine finally generates a specific and operable physical property report of the defect surface. These suggestions are output in a standardized data format (for example, including parameter names, target adjustment values or adjustment amplitudes, execution priorities, etc.).

[0059] By introducing the processing priority queue mechanism, the system can give priority to responding to the anomalies that have the most significant impact on the quality of diamond, improving the timeliness and effectiveness of feedback. At the same time, the configurability of the rule library facilitates experts to solidify their experience and optimization strategies into the system, and can flexibly adjust the control logic according to actual production needs, realizing a more refined, intelligent, and efficient closed-loop feedback control of the diamond polishing and grinding process, thereby improving the final product qualification rate and the overall efficiency of the production line.

[0060] Further, the defect information integration unit further includes: adopting the isolation forest algorithm and the principal component analysis reconstruction error analysis method to compare the surface defect information, the thermally induced anomaly information, and the surface roughness information with the normal working condition model to identify a new type of surface anomaly state, which is used to trigger and update the rule engine to generate the defect surface physical property report.

[0061] Based on the isolation forest algorithm and the principal component analysis reconstruction error analysis method, the online detection system's ability to identify and respond to unknown risks and new fault modes during diamond polishing and grinding is enhanced. This detection does not rely on predefined defect types or rules. By comparing with the historical normal working condition model, it can promptly and sensitively capture statistically significant abnormal states such as surface quality problems caused by sudden equipment issues or unseen diamond defects. Once such new anomalies are identified, not only can early warnings or emergency interventions be triggered in a timely manner, but these abnormal events and their related data can also provide valuable input for subsequent updating and optimizing the rule engine, enabling the system to learn from the unknown, continuously expand its knowledge base and response capabilities, improve the robustness and intelligence level of the entire production process, and avoid production losses that may be caused by the failure to promptly identify and handle new anomalies.

[0062] The present invention covers three core units: multi-modal data acquisition and preprocessing, multi-dimensional intelligent analysis and interpretation, and real-time feedback closed-loop control, realizing the full-process automatic quality monitoring and process optimization of the diamond polishing and grinding process. The system can comprehensively obtain workpiece surface state information, including optically visible defects, thermally induced anomalies, precise roughness, and defect evolution trends. Through intelligent decision-making, the system can dynamically adjust process parameters according to real-time feedback, thereby effectively improving the quality and consistency of diamond products, increasing production efficiency, and enhancing the intelligence and adaptability of the production process.

[0063] Embodiment 2: For further detailed description and application examples of Embodiment 1, an on-line detection system for diamond polishing and grinding includes: A defect data acquisition unit, configured to synchronously acquire optical sensing data and thermal sensing data on the diamond surface and perform preprocessing to output optical characterization data containing surface physical property information; Table 1 Comparison of Surface Defect Detection Performance

[0064] A defect characteristic analysis unit, which is used to detect defects on the diamond surface based on the optical characterization data by fusing a data-driven model and a vision detection model, output surface defect information, and detect thermal anomalies in combination with the thermal sensing data to output thermal anomaly information; calculate the surface roughness of the diamond surface based on the optical characterization data, and through comparison with external reference data, perform self-calibration of the surface roughness calculation to output surface roughness information; and compare the surface defect information with a historical defect database to output defect evolution trend information.

[0065] Specifically, in order to prove the detection effect of the present invention by fusing a data-driven model and a vision detection model, as shown in Table 1, the performance of surface defect detection is compared, and it can be seen that the defect detection effect of the present invention is the best.

[0066] A defect information integration unit, which is used to perform feedback decision-making using a rule engine based on the surface defect information, the thermal anomaly information, the surface roughness information, and the defect evolution trend information, generate a physical property report of the defective surface, and output it to the controller through an industrial communication interface.

[0067] Furthermore, the defect information integration unit further includes: adopting an isolation forest algorithm and a principal component analysis reconstruction error analysis method to compare the surface defect information, the thermal anomaly information, and the surface roughness information with a normal working condition model to identify a new type of surface abnormal state, and the new type of surface abnormal state is used to trigger and update the rule engine to generate the physical property report of the defective surface.

[0068] Specifically, the system first integrates multi-source information in the current detection cycle from an image feature analysis unit, including: surface defect information, thermal anomaly information, and surface roughness information, to construct a feature vector comprehensively representing the current surface state of the workpiece. The feature vector will be compared with a normal working condition model. The normal working condition model is constructed based on workpiece data that have been confirmed to be in a stable and high-quality state in a large number of historical samples, and can be regarded as a health baseline in a multi-dimensional feature space.

[0069] The system pre-learns a large amount of historical detection data (i.e., the set of the above-mentioned comprehensive feature vectors) that are confirmed to represent normal or high-quality production working conditions, and constructs and stores a normal working condition model or a baseline model.

[0070] This anomaly detection function adopts a technical path combining an isolation forest algorithm and a principal component analysis (PCA) reconstruction error analysis method to compare the current comprehensive feature vector with a normal working condition model in real time and calculate its anomaly score or the degree of deviation from the normal mode.

[0071] If the anomaly score calculated by the Isolation Forest algorithm exceeds the preset threshold, or the PCA reconstruction error is significantly greater than the normal fluctuation range, the system determines that there is a new surface anomaly state on the current workpiece surface. Such anomalies are usually caused by unknown or uncovered process disturbances, such as abnormal wear of the grinding tool, failure of the cooling system, material anomalies, etc., which may cause conventional rules to fail to respond effectively. The new anomaly state can also be recorded as a key event and used as input for subsequent updates and expansions of the rule engine. On the one hand, this event can be incorporated into the rule library of the rule engine through manual annotation or the online learning module, which can supplement previously uncovered scenarios; on the other hand, the identified anomaly information can also be used to adjust the risk assessment logic and optimize the generation strategy of future defect surface physical property reports. Through the above anomaly identification and response mechanism, the system can break through the limitations of the rule engine for known problems, possess the ability to identify unknown anomalies, improve the sensitivity and response speed to potential quality risks, and enhance the intelligent decision-making ability and robustness of the system under complex and dynamic working conditions.

[0072] To verify the technical advantages of the defect information integration unit of the present invention in terms of intelligence, response speed, and adaptive ability, a comparison was made with two representative baseline models, as shown in Table 2. First, in terms of the feedback decision response time, it is not much different from the response times of other baseline models and can meet the real-time requirements of the diamond online production line. Second, in terms of the accuracy of anomaly identification and feedback, the present invention still has the highest recognition rate when dealing with sudden and non-preset surface states (such as cooling failure, grinding tool detachment), while the baseline models generally cannot effectively handle unknown anomalies and have obvious deficiencies in detection ability. In terms of the quality results, the present invention can reduce the surface defect rate and the roughness fluctuation range, with the largest reduction in the standard deviation of Ra fluctuations, while the improvement of the basic model is limited. Correspondingly, the additional scrap rate caused by improper feedback is the lowest, and the present invention is significantly better than other baseline models.

[0073] Table 2 Comparison of key performance data

[0074] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An on-line detection system for diamond polishing and grinding, characterized in that, Including: A defect data acquisition unit, which is used to synchronously acquire the optical sensing data and thermal sensing data on the surface of the diamond, and perform preprocessing to output optical characterization data containing surface physical property information; A defect characteristic analysis unit, which is used to detect the defects on the surface of the diamond based on the optical characterization data by fusing a data-driven model and a visual detection model, output surface defect information, and detect thermal-induced anomalies in combination with the thermal sensing data to output thermal-induced anomaly information; Calculate the surface roughness of the diamond surface based on the optical characterization data, and perform self-calibration of the surface roughness calculation by comparing with external reference data to output surface roughness information; And compare the surface defect information with a historical defect database to output defect evolution trend information; A defect information integration unit, which is used to make a feedback decision using a rule engine based on the surface defect information, the thermal-induced anomaly information, the surface roughness information, and the defect evolution trend information, generate a defect surface physical property report, and output it to a controller through an industrial communication interface.

2. The on-line detection system for diamond polishing and grinding according to claim 1, characterized in that The preprocessing specifically includes: Convert the optical sensing data into a single-channel grayscale image; Apply a contrast-limited adaptive histogram equalization algorithm to the single-channel grayscale image; Use a guided filter algorithm to filter the noise of the single-channel grayscale image after contrast enhancement; And for the filtered single-channel grayscale image, extract the region of interest through template matching technology based on normalized cross-correlation, and apply multi-level wavelet transform to extract and reconstruct the low-frequency approximation component and high-frequency detail component to form the optical characterization data.

3. The on-line detection system for diamond polishing and grinding according to claim 1, wherein The process of outputting surface defect information includes: The data-driven model uses a lightweight convolutional neural network; use a lightweight convolutional neural network model pre-trained for diamond surface defect types to automatically extract and identify surface irregular defects from the input optical characterization data, and output a first defect detection result, including defect category, prediction score with a confidence level greater than a preset threshold, and bounding box coordinates; The visual detection model integrates the Canny edge detection algorithm and the Hough transform algorithm to identify and locate surface regular defects to obtain a second defect detection result; For the first defect detection result and the second defect detection result, perform fusion processing based on a confidence-weighted factor and / or the intersection over union between bounding boxes, and generate the surface defect information through redundancy elimination and conflict resolution.

4. The on-line detection system for diamond polishing and grinding according to claim 1, characterized in that The process of outputting the surface roughness information includes: Based on the optical characterization data, apply fast Fourier transform processing to analyze the spectral energy distribution of the low-frequency approximation component and the high-frequency detail component; and estimate the surface roughness and quantify the texture feature parameters according to the predetermined estimation model between the spectral energy distribution and the surface roughness to obtain the preliminary surface roughness information; Regularly receive the reference value of the reference surface roughness for the same sample area through a calibration data interface; and apply an adaptive calibration algorithm to calculate the deviation amount by comparing the surface roughness with the reference surface roughness reference value; Based on the deviation amount, adjust the internal parameters of the predetermined estimation model, and recalculate the surface roughness, and output the calibrated surface roughness information.

5. The on-line detection system for diamond polishing and grinding according to claim 1, wherein, The process of outputting the defect evolution trend information includes: Perform spatio-temporal correlation processing and multi-dimensional feature parameter matching on the surface defect information and the defect map data stored in the historical defect database to obtain the associated defect numbers; Track the time-series quantization change data corresponding to the defect numbers in the optical sensing data, and analyze the time-series quantization change data using a preset time-series analysis model to identify and quantify the evolution patterns corresponding to the defect numbers; wherein, the evolution pattern is a parameter vector characterizing the defect size change rate, the morphology transformation characteristics, and the dynamic statistical characteristics of the occurrence frequency; Based on the evolution pattern and a preset risk assessment rule set, determine the future development trends and potential risk levels of the corresponding defect numbers to obtain the defect evolution trend information.

6. The on-line detection system for diamond polishing and grinding according to claim 1, wherein The specific implementation process of the rule engine includes: Analyze the severity of the surface defect information and the deviation degree of the surface roughness information from the preset target specifications, and set a processing priority queue; Based on the processing priority queue, the rule engine comprehensively uses the surface defect information, the defect evolution trend information, the surface roughness information, and the thermally induced anomaly information, and applies an adjustment action rule library to generate the defect surface physical property report.

7. An on-line detection system for diamond polishing and grinding according to claim 1, characterized in that, The defect information integration unit further includes: using the isolation forest algorithm and the principal component analysis reconstruction error analysis method to compare the surface defect information, the thermally induced anomaly information, and the surface roughness information with a normal working condition model to identify a new surface abnormal state, and the new surface abnormal state is used to trigger and update the rule engine to generate the defect surface physical property report.

Images