High-precision intelligent quality inspection system for diamond grains

Abstract

The invention discloses a high-precision intelligent quality inspection system for diamond grains, and relates to the technical field of defect detection and contour measurement, and the system comprises a multi-mode sensing device which is used for obtaining a two-dimensional image, a three-dimensional shape and spectral data of the grains; the module is used for receiving manufacturing process parameters and constructing a theoretical three-dimensional model; the interface is used for acquiring actual manufacturing process data; the unit is used for processing the fused measured data and comparing the fused measured data with a theoretical model to generate deviation data; the hierarchical artificial intelligence analysis engine comprises a data analysis and coordination control AI model based on multi-head attention, the AI model is coupled with an attribution analysis layer, and the engine fuses and analyzes measured data, deviation data, process parameters and actual manufacturing process data. And performing deep cross-modal association, trend analysis, manufacturing data association analysis and accurate attribution by using an attention mechanism, and finally generating a quality evaluation result containing application performance prediction and process optimization suggestions with attribution information.

Description

Technical Field

[0001] The present invention relates to the technical fields of defect detection and profile metrology, and specifically to a high-precision intelligent quality inspection system for diamond abrasive grains. Background Art

[0002] Diamond abrasive grains (or diamond particles), as superhard materials, play a crucial role in fields such as precision grinding, polishing, and cutting. The quality of the abrasive grains, including their geometric morphology (especially the sharpness of the cutting edge, the distribution and accuracy of concave and convex surfaces), surface and internal defects, and the physical and chemical properties of the material itself (such as crystal purity and internal stress), directly determines their processing efficiency, processing accuracy, and service life. Therefore, high-precision and comprehensive quality inspection of diamond abrasive grains is a key link to ensure their application performance and optimize the production process. Currently, the detection technology for diamond and its products is constantly evolving, but the comprehensive intelligent quality inspection of tiny and complex abrasive grains still faces challenges.

[0003] For example, Chinese invention patent CN118111993B discloses a visual quality inspection method and system for surface defects of diamond tools. This method mainly uses a single-channel grayscale camera to collect tool images and detects surface defects through means such as image preprocessing, texture extraction and correction, and texture projection analysis. This patent has achieved a certain degree of automation in the detection of surface defects of diamond tools, replaced some work that relies on subjective human judgment, and improved the quality inspection accuracy and efficiency in specific scenarios. However, this method mainly targets tools with relatively large sizes, uses two-dimensional grayscale vision and texture analysis technology, and its detection dimension is relatively single, limited to visible surface defects, and cannot obtain three-dimensional morphology information of the abrasive grains (especially for the complex concave and convex features commonly present on the abrasive grains), internal stress, material crystallization quality, and other information that is also crucial for the performance of the abrasive grains. In addition, its analysis method has limited ability to identify complex, subtle, or non-texture-related defect patterns.

[0004] Another example is Chinese invention patent CN110006363B, which discloses a roller profile projector. This device clamps the workpiece to be measured (such as a roller) through three-point positioning, projects its profile using a light source, and measures it on a projection receiving device, mainly used to detect the profile shape and wear condition of rollers or indenters. This patent provides a convenient and high-precision method for profile measurement of specific types of components. However, this method is based on the projection principle. Although it is effective for measuring the outer profile of specific components, it is difficult to apply to the detection of tiny and three-dimensional structure complex diamond abrasive grains. Especially when the abrasive grains have concave features (such as pits and concave surfaces), the projection method cannot accurately obtain their internal profile and depth information. At the same time, this method is also limited to geometric profile measurement and cannot provide other key quality information such as surface details and material properties.

[0005] The above designs have improved the automation level and measurement accuracy of specific detection tasks to a certain extent by using vision and texture analysis to automatically detect surface defects of diamond tools (such as CN118111993B), or by using the projection method to precisely measure the profiles of specific components (such as CN110006363B). However, there are still certain limitations. Especially when it comes to comprehensively, deeply, and intelligently evaluating the quality of a single diamond abrasive grain, for example: it is impossible to comprehensively and accurately obtain key comprehensive information such as three-dimensional topography data (especially complex concave and convex features), internal stress distribution, material crystal structure, and composition purity that have an important impact on the performance of the abrasive grain. Each detection index is usually regarded in isolation, lacking effective technical means to analyze the internal correlation between geometric features, surface defects, and material properties and their comprehensive impact on the final application performance. It is disconnected from the manufacturing process, and the existing detection results are often difficult to directly establish clear and quantifiable associations with upstream manufacturing process parameters (such as material ratios, pressing parameters, sintering curves) and actual production process data (such as furnace temperature fluctuations, pressure changes), resulting in difficulties in tracing defects and ineffective guidance for precise adjustment and optimization of the quality inspection process.

[0006] Therefore, there is an urgent need to develop a new detection system that can overcome the above limitations to achieve all-round, high-precision, and intelligent quality inspection of diamond abrasive grains from geometry to material, from appearance to internal, from monomer characteristics to process correlation, and even performance prediction, so as to meet the increasingly stringent requirements of modern precision machining for superhard abrasives and promote the intelligent upgrading of related manufacturing processes. Summary of the Invention

[0007] The purpose of the present invention is to overcome the deficiencies of the prior art and propose a high-precision intelligent quality inspection system for diamond abrasive grains to solve the above problems.

[0008] The purpose of the present invention is achieved through the following technical solutions: A high-precision intelligent quality inspection system for diamond abrasive grains, comprising: A process parameter interface module for receiving manufacturing process parameters related to diamond abrasive grains; A theoretical model construction module connected to the process parameter interface module for generating a theoretical three-dimensional model of diamond abrasive grains based on manufacturing process parameters; An image acquisition module for acquiring two-dimensional images of diamond abrasive grains; A three-dimensional topography measurement module for acquiring three-dimensional surface topography data of diamond abrasive grains; A spectral analysis module for acquiring material characteristic spectral data of diamond abrasive grains; A manufacturing process data interface module for obtaining actual manufacturing process data related to the production of diamond abrasive grains from a manufacturing execution system or database; A data processing and fusion unit, connected to the image acquisition module, the three-dimensional topography measurement module, and the spectral analysis module, is used to preprocess two-dimensional images, three-dimensional surface topography data, and material property spectral data, extract measured features, and fuse the measured features into a measured multimodal data record; A model comparison unit, connected to the data processing and fusion unit and the theoretical model construction module, is used to compare the measured multimodal data record or its derived measured features with the theoretical three-dimensional model, and generate deviation data; A hierarchical artificial intelligence analysis engine, which is used to receive the measured multimodal data record, deviation data, manufacturing process parameters, and actual manufacturing process data, and perform analysis. The engine includes: A data analysis AI model, whose core is based on the multi-head self-attention mechanism, is used to deeply analyze the received data, where different attention heads parallelly focus on the geometric relationships, spectral features, defect morphologies, and cross-modal feature correlations in the data; and A coordination control AI model, whose core is based on the multi-head attention mechanism, is used to manage the data analysis AI model, monitor the overall workflow status sequence of the system, analyze the process consistency, the time trend of the cumulative evaluation results, and the correlation between the cumulative evaluation results and the actual manufacturing process data, and integrate the analysis results of the data analysis AI model; And the coordination control AI model is coupled with an attribution analysis layer, and the attribution analysis layer is used to analyze the output of the coordination control AI model to determine the input features or data sources that contribute the most to specific analysis conclusions, alerts, or suggestions; The coordination control AI model is used to generate the final quality evaluation result or feedback information for the diamond abrasive grains based on its own analysis and the attribution results of the attribution analysis layer; A result output module, connected to the hierarchical artificial intelligence analysis engine, is used to output the final quality evaluation result or feedback information.

[0009] The spectral analysis module includes a Raman spectrometer; and the system further includes a fast geometric verification unit, which is used to perform a fast geometric deviation evaluation on the diamond abrasive grains based on a preset target curve and the reflection light characteristics at the focus before the spectral analysis module performs spectral acquisition, and provide the fast geometric deviation evaluation result to the coordination control AI model.

[0010] The system further includes a surface flattening analysis unit, which is used to perform a mathematical flattening process on the concave or convex surfaces in the three-dimensional surface topography data obtained by the three-dimensional topography measurement module, compare the processing result with the theoretical flattening parameters calculated based on the preset target curve to generate flattening deviation data, and provide the flattening deviation data as part of the measured multimodal data record to the data analysis AI model.

[0011] The multi-head self-attention mechanism of the data analysis AI model is used to explicitly model and analyze the cross-modal interaction relationships among the internal geometric features, spectral features, and defect features in the measured multi-modal data record by using at least one group of attention heads.

[0012] The data analysis AI model is used to perform correlation analysis on the measured features and the deviation data generated by the model comparison unit by using at least one group of attention heads, and combine the material ratio or pressing parameter information in the manufacturing process parameters from the process parameter interface module to identify the feature change patterns caused by the parameter deviations in specific process links.

[0013] The multi-head attention mechanism of the coordination control AI model is used to perform time series trend analysis on the accumulated final quality assessment results by using at least one group of attention heads to detect quality drift.

[0014] The coordination control AI model is used to analyze the correlation between specific defect patterns or parameter drifts in the accumulated assessment results and the actual manufacturing process data obtained through the manufacturing process data interface module by using its multi-head attention mechanism and attribution analysis layer, and perform precise attribution of the main driving factors of the correlation.

[0015] The hierarchical artificial intelligence analysis engine is used to predict the application performance of diamond abrasives based on its comprehensive analysis results.

[0016] The coordination control AI model is used to generate an alarm indicating potential process problems or more operable process optimization suggestions with key influencing factors marked as part of the feedback information based on the trend analysis, correlation analysis it performs, and the precise attribution results provided by the attribution analysis layer, and trigger an update to the simulation model used by the theoretical model construction module according to a preset threshold.

[0017] The attribution analysis layer is used to calculate and output the quantitative contribution degree or importance score of each element in the manufacturing process parameters, measured features in the measured multi-modal data record, deviation data, or actual manufacturing process data to the specific analysis conclusions, alarms, or suggestions generated by the coordination control AI model by using at least one interpretable AI (XAI) technology.

[0018] The beneficial effects of the present invention are: 1. Integrated with 2D vision, high-precision 3D topography measurement, and Raman spectroscopy analysis, it can simultaneously obtain multi-dimensional information such as the macroscopic / microscopic geometric shape, surface precision topography and defects, and internal material properties (such as crystal form, stress, composition) of diamond abrasives, far exceeding the information content provided by single detection methods, forming a "holographic" portrait of the quality of a single abrasive. The high-resolution sensing technology combined with data processing algorithms ensures the accuracy of various detection data. In particular, the hierarchical AI engine's in-depth analysis of multi-modal data can identify subtle defects and complex patterns that are difficult to discover by the human eye or traditional algorithms, greatly improving the reliability and consistency of quality assessment and reducing missed detections and misjudgments.

[0019] 2. By introducing means such as surface flattening analysis, it can conduct non-traditional and more refined shape deviation assessments on the key concave and convex curved surfaces of the abrasive, quantify subtle geometric features that affect the actual use performance such as surface waviness and local irregularities. The fast geometry verification unit can quickly judge based on optical principles in the early stage whether the overall shape of the abrasive seriously deviates from the target, effectively screening out some unqualified products, optimizing the detection process, and improving the overall efficiency.

[0020] 3. Incorporate key manufacturing process parameters (formulation, pressing, sintering) into the system and construct a theoretical 3D model based on these parameters, establishing a digital bridge between process inputs and the expected product form at the microscopic particle level. By comparing the measured data with the theoretical model predicted based on the process, the system can quantify the difference between the actual abrasives produced and the state that "should have been achieved according to the standard process". This deviation directly reflects the actual process execution effect or the accuracy of the model itself. Integrate the manufacturing process data interface to obtain the actual process data in the MES system, making the detection results no longer isolated but directly correlatable with the real fluctuations in the production process (such as temperature, pressure curve, raw material batch).

[0021] 4. The coordinated control AI model can automatically analyze the trend of accumulated quality detection data and deeply correlate it with the actual manufacturing process data obtained, searching for potential causal relationships. The core advantage lies in the coupled attribution analysis layer. Using interpretable AI technology, it can quantitatively analyze the correlations discovered by the coordinated control AI model, accurately pointing out which specific process parameter fluctuations or raw material characteristics (with high contribution scores) are most likely to cause the observed specific defect patterns or quality drifts, greatly improving the efficiency and accuracy of root cause analysis and changing from "experience-based judgment" to "data-driven".

[0022] 5. The system is no longer limited to determining whether a product meets static specification standards. Instead, it uses an AI model to predict the potential performance of abrasive grains in specific downstream applications (such as grinding efficiency, service life, machining accuracy, etc.) based on a deep understanding of the comprehensive internal characteristics of abrasive grains (geometry, defects, materials, process deviation imprints, etc.). This predictive assessment enables more refined grading of abrasive grains according to expected performance (for example, "high-efficiency type", "long-life type"), thereby achieving more optimized product matching and application strategies and maximizing product value.

[0023] 6. The system can actively monitor quality trends and promptly issue alerts about potential problems. The process optimization suggestions generated based on precise root cause analysis clearly indicate the key process steps and parameters that need attention (along with impact weights), enabling process engineers to make more targeted adjustments, significantly improving the efficiency and success rate of process improvement. The system can also trigger updates to the simulation models used in the theoretical model construction module based on the deviation between the actual and theoretical values and the results of root cause analysis, ensuring that the theoretical model keeps pace with the times and enhancing its value in guiding production.

[0024] 7. The highly automated detection, analysis, and decision-making processes significantly reduce the dependence on manual operations and subjective judgments, improve the detection speed and consistency. The AI model can adaptively adjust subsequent detection strategies (such as scanning area, resolution) based on the preliminary detection results, optimize the utilization of equipment resources, and establish a comprehensive digital file for each abrasive grain (or each batch), greatly enhancing the product quality traceability ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is the system architecture of the present invention Figure 1 ; Figure 2 is the system architecture of the present invention Figure 2 . DETAILED DESCRIPTION OF THE INVENTION

[0026] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts fall within the protection scope of the present invention.

[0027] It should be noted that the orientation concepts of "left", "right", "up", "down", "front", "rear", "inner", and "outer" in the following solutions are all relative directions, and will not be listed one by one here.

[0028] Embodiment 1: As Figure 1 and Figure 2As shown, this embodiment provides a high-precision intelligent quality inspection system for diamond grits, constituting a basic version with multimodal sensing capabilities, enhanced geometric analysis means, and intelligent attribution analysis functions.

[0029] The intelligent quality inspection system of this embodiment physically or logically includes the following core components: It includes a process parameter interface module for receiving manufacturing process parameters related to the batch of diamond grits to be inspected input by the operator or imported from an external system (such as MES). These parameters at least cover material ratios, pressing process details, and preset sintering parameters. Connected to this interface module is a theoretical model construction module that internally or invokes process simulation algorithms (based on finite element, phase field, or other physical models) and can predictively generate an idealized theoretical three-dimensional model of a single grit under these process conditions. This model reflects the expected geometric shape, size, and even possible internal states. The system also includes a manufacturing process data interface module for obtaining manufacturing process data actually recorded during the production process of this batch of grits (such as actual furnace temperature curves, pressure sensor readings, raw material batch numbers, etc.) from the factory's manufacturing execution system (MES) or related databases.

[0030] It integrates multiple sensors to comprehensively capture grit information. An image acquisition module, usually including a high-resolution camera and a customized light source, is used to obtain two-dimensional image information of the grits. A three-dimensional topography measurement module, such as a high-precision confocal microscope or a white light interferometer, is used to accurately measure the three-dimensional contour and microscopic topography of the grit surface. A spectral analysis module, specifically a Raman spectrometer in this embodiment, is used to detect material properties such as chemical composition, crystal structure, and stress state at specific points on the grits. In particular, this embodiment also includes a fast geometric verification unit, which usually includes specific optical emission and reception components and is designed to perform a fast geometric compliance check before Raman spectroscopy acquisition.

[0031] A data processing and fusion unit is responsible for performing necessary preprocessing (such as denoising, calibration, normalization) on the raw data (2D images, 3D topography data, Raman spectroscopy data) from each sensor module, extracting preliminary measured features, and integrating the feature information from different modalities into a structured measured multi-modal data record. A model comparison unit then performs spatial registration and detailed comparison between the measured multi-modal data record (or the key geometric features extracted therefrom) and the theoretical 3D model generated by the theoretical model construction module, and calculates the deviation data between the two. This embodiment also includes a surface flattening analysis unit, which, as a dedicated software module or hardware acceleration processing unit, is used to execute a specific mathematical flattening (or unfolding) algorithm on the key concave and convex surfaces in the 3D topography data, and compare the flattened result with the theoretical flattening parameters calculated based on the preset ideal curve to generate flattening deviation data. This data is then incorporated into the measured multi-modal data record for subsequent AI analysis.

[0032] A hierarchical artificial intelligence analysis engine receives the measured multi-modal data record (already containing flattening deviation), the deviation data generated by the model comparison unit, the input manufacturing process parameters, and the actual manufacturing process data obtained from the MES. The engine internally includes: The data analysis AI model is based on the multi-head self-attention mechanism and deeply analyzes the input data. Its different attention heads concurrently focus on geometric details, spectral features, defect patterns, cross-modal correlations, etc.

[0033] The coordination control AI model is also based on the multi-head attention mechanism and is responsible for managing the data analysis AI model, monitoring the status of the entire detection process (including receiving the evaluation results of the fast geometric verification unit), analyzing the process consistency, quality trends, and the correlation with the actual manufacturing process data, and integrating the analysis results of the data analysis AI model.

[0034] The attribution analysis layer is tightly coupled with the coordination control AI model. In this embodiment, at least one interpretable AI (XAI) technology (such as SHAP values, LIME, attention map analysis, etc.) is adopted to deeply analyze the output of the coordination control AI model, calculate and output the quantitative contribution degree or importance score of each input factor (including manufacturing process parameters, measured features, deviation data, actual manufacturing process data, etc.) to the specific conclusion finally obtained (such as the determination of a certain defect, the prediction of a certain performance index, or a certain optimization suggestion).

[0035] A result output module is used to present the final quality assessment results (such as qualified / unqualified, grade, list of key parameters) generated by the coordinated control AI model (combining the results of the attribution analysis layer) and / or feedback information containing attribution information (such as potential problem warnings, key influencing factor analysis) in a human-friendly manner (such as reports, dashboards), or transmit them to downstream systems (such as automatic sorting devices).

[0036] Working process: The operator or system inputs the manufacturing process parameters of the batch to be inspected through the process parameter interface module. The theoretical model construction module generates a theoretical 3D model based on these parameters. At the same time, the manufacturing process data interface module prepares to retrieve the actual manufacturing process data related to this batch from the MES.

[0037] A single diamond abrasive enters the system, is assigned a unique ID, and is associated with the generated theoretical 3D model and the corresponding actual manufacturing process data record. The coordinated control AI model starts tracking the inspection process of this ID.

[0038] When the abrasive reaches the first integrated inspection station, the image acquisition module obtains a 2D image for rapid macroscopic defect screening and ROI identification. Immediately or simultaneously, the 3D topography measurement module obtains high-precision 3D surface topography data.

[0039] The abrasive moves to the next station. The rapid geometry verification unit is first activated, positions the optical probe at the theoretical focus based on the preset target curve, measures the characteristics of the reflected light, generates a rapid geometry deviation assessment result, and sends it to the coordinated control AI model. Subsequently, the Raman spectrometer performs spectral acquisition on one or more points of the abrasive.

[0040] The data processing and fusion unit preprocesses all the collected raw data and extracts various measured features.

[0041] The surface flattening analysis unit performs a flattening algorithm on the specified surface in the preprocessed 3D topography data, calculates the deviation from the theoretical flattening, and generates flattening deviation data.

[0042] The data processing and fusion unit fuses all the extracted measured features with the flattening deviation data into a unified measured multi-modal data record.

[0043] The model comparison unit compares the measured data record (or features) with the theoretical 3D model to generate deviation data.

[0044] The data analysis AI model receives the measured multi-modal data record (including flattening deviation) and the model comparison deviation data, uses its multi-head attention mechanism for in-depth analysis, and outputs a detailed assessment of the various characteristics (geometry, material, defects, etc.) of the abrasive.

[0045] The coordinated control AI model receives the evaluation results of the data analysis AI model, the rapid geometric deviation evaluation results, the workflow status information, as well as the associated manufacturing process parameters and actual manufacturing process data. It uses the multi-head attention mechanism for global analysis, evaluates the process consistency, conducts trend comparison (if multiple samples are processed), and analyzes the relevance to the manufacturing data.

[0046] The attribution analysis layer is activated to analyze the judgment process or preliminary conclusion of the coordinated control AI model, and uses XAI technology to calculate the contribution degree or importance score of each input factor to this judgment.

[0047] The coordinated control AI model combines its own global analysis and the precise attribution results provided by the attribution analysis layer to form the final quality evaluation conclusion (such as grade determination) and / or feedback information.

[0048] The result output module displays or transmits the final quality evaluation results and feedback information, where the feedback information may include the description of key influencing factors (based on the attribution analysis).

[0049] By integrating three key detection methods of vision, three-dimensional topography, and Raman spectroscopy, and combining with the comparison with the theoretical model, this embodiment can conduct a comprehensive and integrated basic evaluation of the geometric shape, surface micro-quality, and basic material properties of diamond abrasive grains.

[0050] The addition of the rapid geometric verification unit enables the system to quickly predict the overall shape conformity of the abrasive grains before performing time-consuming spectral analysis and complex AI calculations, which helps to promptly eliminate seriously unqualified products and improve the efficiency of the overall detection process.

[0051] The surface flattening analysis unit provides a more detailed method for evaluating the shape deviation of concave and convex surfaces beyond the standard geometric parameters (such as the radius of curvature). It can quantify features such as surface waviness and local irregularities that are difficult to describe by traditional parameters, which is particularly beneficial for ensuring the precision forming quality of specific working surfaces of abrasive grains.

[0052] The attribution analysis layer reveals the key basis for the AI model to make judgments through XAI technology, making the "black box" decision-making process relatively transparent. The output results not only tell "what" (such as the quality grade), but also can explain "why" to a certain extent (such as jointly caused by a certain geometric parameter deviation and a certain material property abnormality), which greatly enhances the user's trust in the detection results and also provides a more clear direction for subsequent corrective measures or process adjustments.

[0053] In summary, the system demonstrated in Embodiment 1 realizes a basic intelligent quality inspection solution with comprehensive functions, enhanced analysis means, and high result credibility by integrating multiple sensors, introducing an innovative analysis unit (rapid geometric verification, surface flattening), and equipping an AI engine with attribution explanation capabilities.

[0054] Embodiment 2: As Figure 1 and Figure 2 shown, this embodiment illustrates a high-precision intelligent quality inspection system for diamond grits. The intelligent quality inspection system of this embodiment is similar to that of Embodiment 1 in terms of hardware and basic software configuration, and includes a process parameter interface module, a theoretical model construction module, an image acquisition module, a three-dimensional topography measurement module, a Raman spectrometer (as a spectral analysis module), a manufacturing process data interface module, a data processing and fusion unit, a model comparison unit, a hierarchical artificial intelligence analysis engine (including a data analysis AI model, a coordination control AI model, and a coupled attribution analysis layer), and a result output module.

[0055] The key differences and focuses are as follows: This embodiment clearly includes a surface flattening analysis unit, which is used to perform specific mathematical processing on three-dimensional topography data, generate flattening deviation data, and provide this data as one of the key inputs to the data analysis AI model for more refined shape and surface texture analysis.

[0056] Deep analysis configuration of the data analysis AI model: The data analysis AI model in this embodiment is specifically configured and trained to perform more complex analysis tasks: Its internal multi-head self-attention mechanism is optimized, including at least one or a group of attention heads, which are specifically used to explicitly model and quantify the interaction relationships between different sources of information in the measured multi-modal data records. For example, the model can analyze how a specific crystal orientation (possibly from diffraction data, if integrated; or indirectly from topography or Raman stress) affects the wear behavior of a specific grinding edge (by correlating geometric features with simulated wear data), or the co-occurrence probability of internal stress (from Raman) and surface microcracks (from three-dimensional topography or vision).

[0057] The data analysis AI model uses at least one or a group of attention heads, which are specifically responsible for deeply correlating the measured features (such as a specific type of surface defect or geometric parameter anomaly) with the deviation data generated by the model comparison unit (i.e., the difference between "actual" and "theoretical prediction"). Furthermore, it will also combine the material ratio or pressing parameter information obtained through the process parameter interface module to try to identify which specific measured features or deviation patterns are likely to be the feature "fingerprints" caused by parameter deviations in upstream specific process links (such as uneven ingredient mixing, pressing pressure fluctuations).

[0058] As a whole, the hierarchical artificial intelligence analysis engine is trained and configured to achieve application performance prediction. This means that based on the in-depth understanding of all information such as abrasive geometry, defects, material properties, cross-modal associations, and process deviation characteristics by the data analysis AI model, and through the integration and judgment of the coordinated control AI model, the system can output the predicted values of the potential performance indicators of the abrasive in a specific application scenario (for example, when grinding a specific material) (such as the predicted grinding efficiency index, wear life interval, or the expected surface roughness of the workpiece after processing).

[0059] Working process: Similar to Embodiment 1, the system receives process parameters, constructs a theoretical model, and obtains actual manufacturing data. After the abrasive is identified, two-dimensional image acquisition, three-dimensional topography measurement, and Raman spectroscopy acquisition are successively completed (or at an integrated station) (this embodiment may not necessarily include a rapid geometry verification unit, and its focus is on the backend analysis).

[0060] Preprocess the collected raw data and extract basic measured features.

[0061] The surface flattening analysis unit performs a flattening operation on the three-dimensional topography data and calculates the flattening deviation data.

[0062] The data processing and fusion unit fuses all measured features (including flattening deviation data) into a measured multi-modal data record.

[0063] The model comparison unit compares the measured data with the theoretical three-dimensional model to generate deviation data.

[0064] The data analysis AI model receives the measured multi-modal data record containing the flattening deviation, the model comparison deviation data, and the associated material ratio / compression parameter information.

[0065] The attention head group A (cross-modal association) analyzes the mutual influence between features such as geometry, spectroscopy, and defects, and establishes an internal relationship model.

[0066] The attention head group B (process deviation association) analyzes the relationship between the measured features and the theoretical deviation, and attempts to associate it with the input formula / compression parameters to find evidence of process influence.

[0067] The attention head group C (comprehensive evaluation) combines all information to comprehensively evaluate the internal quality attributes of the abrasive.

[0068] The coordinated control AI model receives the in-depth analysis report of the data analysis AI model, combines the process monitoring information and the global database knowledge (if any), and may trigger performance prediction calculations according to the preset tasks.

[0069] The hierarchical artificial intelligence analysis engine (mainly executed by the data analysis AI model and coordinated by the control AI model for integration) uses a pre-trained performance prediction model to calculate the predicted performance indicators under specific applications based on the comprehensive evaluation results of the abrasive particle state obtained from in-depth analysis.

[0070] The attribution analysis layer can still perform attribution analysis on the evaluation results (including performance prediction) finally formed by the coordinated control AI model to increase the interpretability of the results.

[0071] The result output module displays the final quality evaluation results. In this embodiment, the results include the predicted application performance indicators, as well as possible quality grades and a summary of key features.

[0072] By explicitly modeling and analyzing the interaction relationships between cross-modal features (geometry, material, defects, etc.), the system can go beyond surface parameters, understand the deeper combination of key factors that determine the internal quality and performance of abrasive particles, and provide a more essential basis for distinguishing between "good" and "bad" abrasive particles.

[0073] Relate the deviation between the measured features and the "theoretical expected state", and further combine the input process parameter (formula, pressing) information, enabling the AI model to more effectively identify potential upstream process problems corresponding to specific defects or abnormal features, providing more powerful clues for fault diagnosis and targeted improvement in the production process.

[0074] To achieve application-oriented predictive quality assessment, the system no longer simply determines "conforming / non-conforming" to specifications, but can predict the possible performance of abrasive particles in actual use (efficiency, lifespan, processing effect, etc.). This predictive assessment directly links the goal of quality control with the actual needs of downstream applications, achieving a true sense of "applicability" screening.

[0075] Based on the performance prediction results, the abrasive particles can be more refinedly graded (e.g., "high-efficiency type", "long-life type", "precision machining type"), enabling different grades of products to be more accurately matched to the most suitable application scenarios, thereby maximizing product value and improving the stability and consistency of end applications.

[0076] By integrating the results of surface flattening analysis into AI in-depth analysis, the system can more comprehensively consider the potential impact of subtle shape deviations and surface textures on performance, improving the comprehensiveness of analysis and the accuracy of prediction.

[0077] In summary, Example 2 demonstrates the great potential of the present invention in transcending traditional quality inspection and achieving predictive and intelligent quality assessment by strengthening the in-depth analysis ability of the data analysis AI model (especially cross-modal association and process deviation association analysis) and finally realizing the prediction of application performance.

[0078] Example 3: As Figure 1 and Figure 2 shown, this example illustrates a high-precision intelligent quality inspection system for diamond diamond grains. This system not only has the multi-modal detection and in-depth analysis capabilities described in Example 1 and Example 2 (its spectral analysis module is specifically a Raman spectrometer and includes a fast geometric verification unit), but also focuses on implementing the use of a coordinated control AI model and its coupled attribution analysis layer to perform full-process process correlation analysis, intelligent trend monitoring, accurate defect traceability, and actionable process optimization feedback, thereby constructing a complete intelligent manufacturing closed-loop.

[0079] The intelligent quality inspection system of this example is basically the same as that of Example 1 in terms of core hardware configuration, including a process parameter interface module, a theoretical model construction module, an image acquisition module, a three-dimensional topography measurement module, a Raman spectrometer (as the spectral analysis module), a manufacturing process data interface module (its role is extremely crucial in this example), a data processing and fusion unit, a model comparison unit, a fast geometric verification unit, a hierarchical artificial intelligence analysis engine (including a data analysis AI model, a coordinated control AI model, and an attribution analysis layer), and a result output module.

[0080] The key and focus of this example lie in the functional configuration and data utilization of the coordinated control AI model and its attribution analysis layer (AttributionAnalysis Layer): The coordinated control AI model is configured with a dedicated attention head group for continuously monitoring and analyzing the time series data of the accumulated final quality assessment results (stored in the database) to automatically detect whether there are statistically significant drifts, mutations, or periodic fluctuations in the key indicators of product quality.

[0081] Data association and accurate attribution. The coordinated control AI model uses its attention mechanism and is deeply coupled with the attribution analysis layer to specifically analyze the complex correlation between the quality drift or specific defect patterns (from the accumulated detection results) found by trend analysis and the actual manufacturing process data obtained through the manufacturing process data interface module (for example, raw material information of different batches, the measured pressure curve of the press, the actual temperature records of each temperature zone of the sintering furnace, the atmosphere flow rate, etc.). The attribution analysis layer plays a key role at this time to analyze the identified strong correlations and accurately locate which specific one or more actual process parameters or operation events (such as "the temperature in the third temperature zone is 10 degrees higher", "using a certain batch of binder from a certain supplier", "the holding pressure time is shortened by 2 seconds") are the main driving factors causing the quality problem.

[0082] Based on the above trend analysis, precise correlation analysis, and attribution results, the coordinated control AI model is configured to automatically generate high-level feedback information, which not only includes alerts indicating problems, but more importantly, generates process optimization suggestions with clear guiding significance and operability, annotated with key influencing factors (and their quantitative contribution degrees, from the attribution layer) (for example, "It is recommended to focus on checking the temperature control accuracy of the third temperature zone of the sintering furnace for batch XXX, as this factor has the highest contribution (0.75) to the recently observed increase in internal stress"). In addition, if the system detects a persistent and significant deviation between the theoretical model and a large number of actual measurement results, or the attribution analysis indicates that the existing model fails to accurately reflect the impact of certain process parameters, the coordinated control AI model can also automatically trigger a request to update or recalibrate the simulation model used in the theoretical model construction module according to preset rules or thresholds.

[0083] In this embodiment, the attribution analysis layer clearly adopts one or more advanced interpretable AI (XAI) technologies, such as SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), Integrated Gradients, or directly analyzing the internal attention weights of the coordinated control AI model, etc., to calculate the quantitative contribution degrees or importance scores of each input variable (covering input process parameters, measured data features, theoretical deviation data, actual manufacturing process data, etc.) for specific outputs (such as correlation intensity, alert triggering, optimization suggestion content) generated by the coordinated control AI model, and uses this score as part of the output.

[0084] Working process: Similar to the process of Embodiment 1 or 2, steps such as process parameter input, theoretical model construction, actual manufacturing data acquisition, multi-modal data collection (including rapid geometric verification), data preprocessing, feature extraction, flattening analysis (optional but data available), data fusion, and model comparison are completed to obtain measured multi-modal data records and deviation data, and the data analysis AI model completes its in-depth data analysis.

[0085] The coordinated control AI model receives the analysis results from the data analysis AI model, the results from the rapid geometric verification unit, the workflow status, deviation data, and most importantly, the actual manufacturing process data from the manufacturing process data interface and the historical cumulative detection data stored in the database.

[0086] Parallel processing (multi-head attention): Group 1 of heads: Monitor the consistency and real-time status of the current abrasive grain detection process.

[0087] Head group 2: Perform trend analysis on historical cumulative data to detect quality drift.

[0088] Head group 3: Conduct correlation analysis between the anomalies found in trend analysis or defect patterns of concern and the corresponding actual manufacturing process data.

[0089] For the strong correlations found by head group 3, coordinate and control the AI model to call the attribution analysis layer for in-depth analysis and output the precise contribution / importance scores of each associated factor.

[0090] Coordinate and control the AI model to integrate all analysis results (current abrasive particle evaluation, trend, correlation, precise attribution) to form a final evaluation of the current abrasive particle.

[0091] Based on the trend, correlation, and precise attribution results, coordinate and control the AI model to generate: Immediate alert: Such as detecting significant quality drift or strong negatively correlated process factors.

[0092] Actionable optimization suggestions: Clearly point out the problem, associated process steps / parameters, and key influencing factors supported by attribution analysis.

[0093] Model update trigger signal: Issued when preset conditions are met.

[0094] The result output module outputs the evaluation results of the current abrasive particle, and focuses on outputting feedback information such as the alerts generated by the coordinated control AI model and the optimization suggestions with attribution information.

[0095] Optionally, automatic sorting is performed. Meanwhile, the generated feedback information is recorded and can be used by production managers or automated control systems to adjust the actual production process or update the theoretical model, forming a data-driven intelligent manufacturing closed loop.

[0096] Through trend analysis of cumulative data, the system can shift from passive detection to proactive warning, detect signs early before quality problems occur on a large scale or deteriorate, and significantly improve the predictability of quality management.

[0097] Combined with actual manufacturing process data for correlation analysis, and accurately locate key influencing factors through the attribution analysis layer, greatly improving the accuracy and efficiency of defect traceability, and automating and intelligentizing part of the expert analysis process that originally required a lot of manpower and time.

[0098] The system no longer outputs vague suggestions, but optimization suggestions based on data, with clear directionality (which parameter has the greatest impact) and quantitative support (how much is the contribution), greatly increasing the possibility of the suggestions being adopted and successfully implemented, and directly empowering process improvement.

[0099] By comparing deviations and triggering model updates through feedback, it is ensured that the theoretical model based on physical simulation can be consistent with the constantly changing actual production conditions, improving the guiding value of the theoretical model in process design and optimization.

[0100] The full-process association, precise attribution, and intelligent feedback capabilities demonstrated in this embodiment are the key technical supports for realizing a truly data-driven intelligent manufacturing closed-loop (detection - analysis - decision - feedback - optimization), greatly enhancing the intelligent level and overall competitiveness of the production line.

[0101] The existence of the fast geometry verification unit provides an additional verification point for the process monitoring of the coordinated control AI model, helping to detect process anomalies or sensor problems at an early stage.

[0102] In summary, Embodiment 3 represents the application form of the present invention. It is not just a detection device, but an intelligent decision support system that is deeply integrated into the production process and has powerful capabilities of analysis, traceability, prediction, and optimization suggestions, and is an important technical practice leading to the future intelligent factory.

[0103] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, and can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. Any changes and variations made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A high-precision intelligent quality inspection system for diamond particles, characterized in that, Comprising: A process parameter interface module for receiving manufacturing process parameters related to the diamond abrasive grains; A theoretical model construction module connected to the process parameter interface module for generating a theoretical three-dimensional model of the diamond abrasive grains based on the manufacturing process parameters; An image acquisition module for acquiring two-dimensional images of the diamond abrasive grains; A three-dimensional topography measurement module for acquiring three-dimensional surface topography data of the diamond abrasive grains; A spectral analysis module for acquiring material characteristic spectral data of the diamond abrasive grains; A manufacturing process data interface module for obtaining actual manufacturing process data related to the production of the diamond abrasive grains from a manufacturing execution system or a database; A data processing and fusion unit connected to the image acquisition module, the three-dimensional topography measurement module, and the spectral analysis module for preprocessing the two-dimensional images, the three-dimensional surface topography data, and the material characteristic spectral data, extracting measured features, and fusing the measured features into a measured multimodal data record; A model comparison unit connected to the data processing and fusion unit and the theoretical model construction module for comparing the measured multimodal data record or the measured features derived therefrom with the theoretical three-dimensional model and generating deviation data; A hierarchical artificial intelligence analysis engine for receiving the measured multimodal data record, the deviation data, the manufacturing process parameters, and the actual manufacturing process data and performing analysis. The engine includes: a data analysis AI model, the core of which is based on a multi-head self-attention mechanism for performing in-depth analysis on the received data, where different attention heads concurrently focus on geometric relationships, spectral features, defect morphologies, and cross-modal feature correlations in the data; and a coordination control AI model, the core of which is based on a multi-head attention mechanism for managing the data analysis AI model and monitoring the overall workflow status sequence of the system, analyzing process consistency, the time trend of cumulative evaluation results, and the correlation between the cumulative evaluation results and the actual manufacturing process data, and integrating the analysis results of the data analysis AI model; And the coordination control AI model is coupled with an attribution analysis layer for analyzing the output of the coordination control AI model to determine the input features or data sources that contribute the most to the analysis conclusion, alert, or suggestion; The coordination control AI model is used to generate a final quality evaluation result or feedback information for the diamond abrasive grains based on its own analysis and the attribution results of the attribution analysis layer; A result output module connected to the hierarchical artificial intelligence analysis engine for outputting the final quality evaluation result or feedback information.

2. The high-precision intelligent quality inspection system for diamond grains according to claim 1, wherein: The spectral analysis module includes a Raman spectrometer; and the system further includes a fast geometry verification unit for performing a fast geometry deviation evaluation on the diamond abrasive grains based on a preset target curve and the reflection light characteristics at the focal point before the spectral analysis module performs spectral acquisition and providing the fast geometry deviation evaluation result to the coordination control AI model.

3. The high-precision intelligent quality inspection system for diamond grits according to claim 1, wherein: The system further includes a surface flattening analysis unit, which is used to perform mathematical flattening processing on the concave or convex surfaces in the three-dimensional surface topography data obtained by the three-dimensional topography measurement module, compare the processing results with the theoretical flattening parameters calculated based on a preset target curve to generate flattening deviation data, and provide the flattening deviation data as a part of the measured multi-modal data record to the data analysis AI model.

4. The high-precision intelligent quality inspection system for diamond grains according to claim 1, wherein: The multi-head self-attention mechanism of the data analysis AI model is used to explicitly model and analyze the cross-modal interaction relationships among the internal geometric features, spectral features, and defect features in the measured multi-modal data record by using at least one attention head group.

5. The high-precision intelligent quality inspection system for diamond particles according to claim 1, wherein: The data analysis AI model is used to perform correlation analysis on the measured features and the deviation data generated by the model comparison unit by using at least one attention head group, and combine the material ratio or pressing parameter information in the manufacturing process parameters from the process parameter interface module to identify the feature change patterns caused by process parameter deviations.

6. The high-precision intelligent quality inspection system for diamond grains according to claim 1, characterized in that: The multi-head attention mechanism of the coordination control AI model is used to perform time series trend analysis on the accumulated final quality assessment results by using at least one attention head group to detect quality drift.

7. The high-precision intelligent quality inspection system for diamond grains according to claim 1, characterized in that: The coordination control AI model is used to analyze the correlation between the defect patterns or parameter drifts in the accumulated assessment results and the actual manufacturing process data obtained through the manufacturing process data interface module by using its multi-head attention mechanism and the attribution analysis layer, and perform precise attribution of the main driving factors of the correlation.

8. The high-precision intelligent quality inspection system for diamond grains according to claim 1, wherein: The hierarchical artificial intelligence analysis engine is used to predict the application performance of the diamond abrasive grains based on its comprehensive analysis results.

9. The high-precision intelligent quality inspection system for diamond grains according to claim 1, characterized in that: The coordination control AI model is used to generate an alarm indicating potential process problems or more operable process optimization suggestions with key influencing factors marked as a part of the feedback information based on the trend analysis, correlation analysis it performs, and the precise attribution results provided by the attribution analysis layer, and trigger an update of the simulation model used by the theoretical model construction module according to a preset threshold.

10. The high-precision intelligent quality inspection system for diamond particles according to claim 1, characterized in that: The attribution analysis layer is used to calculate and output the quantitative contribution degrees or importance scores of each element in the manufacturing process parameters, the measured features in the measured multi-modal data record, the deviation data, or the actual manufacturing process data to the analysis conclusions, alarms, or suggestions generated by the coordination control AI model by using at least one interpretable AI technology.

Images