Prism detection sorting quality control method and system based on defect map and process inversion

Abstract

This invention provides a prism inspection and sorting quality control method and system based on defect maps and process inversion. The method includes: acquiring multi-angle polarization images of the prism to be inspected; simultaneously acquiring prism information and storing it to form traceable raw data; performing polarization feature decomposition on the polarization images to extract structured defect features; constructing a defect evolution map; when a defect appears in a new batch, generating candidate process deviation hypotheses and obtaining root cause analysis results; performing multi-dimensional risk classification on the prisms and generating corresponding sorting decisions; based on the root cause analysis results, generating theoretical parameter adjustment amounts by combining at least one of verified historical adjustment data and process parameter response relationships, and decomposing the theoretical parameter adjustment amounts into multi-step restricted adjustment instructions according to equipment safety constraints when the process adjustment triggering conditions are met; collecting sorting decisions, process adjustment execution records, adjusted quality feedback data, and verification results to form closed-loop verification samples, and updating based on these samples.

Description

Technical Field

[0001] This invention belongs to the field of optical component quality control technology, specifically relating to a prism inspection and sorting quality control method and system based on defect maps and process inversion. Background Technology

[0002] Prisms, as crucial optical components in optical instruments, precision lenses, and aerospace optical assemblies, directly impact the imaging quality and stability of optical systems due to their surface quality, internal stress state, film uniformity, and geometric edge integrity. During prism manufacturing, processes such as cutting, annealing, grinding, polishing, cleaning, and coating can introduce defects like scratches, bubbles, edge chipping, stress unevenness, and film anomalies. Because prisms possess multiple optical surfaces, high-reflectivity interfaces, and complex optical path structures, some defects are not easily visible under ordinary lighting conditions, especially optically sensitive defects related to polarization response, film state, or internal stress, which pose significant detection challenges.

[0003] Existing prism inspection and sorting methods mainly include manual visual inspection, ordinary machine vision inspection, defect classification or defect-process parameter mapping methods based on machine learning, and conventional grading and sorting methods. Manual visual inspection relies on the experience of the inspectors, has low inspection efficiency, a high risk of missed detections and false positives, and is difficult to reliably identify optically sensitive defects such as internal stress and coating polarization anomalies. Ordinary machine vision inspection usually acquires images of ordinary intensity, which is easily affected by prism surface reflections, background stray light, and multifaceted reflections, and has limited ability to identify defects such as fine scratches, coating anomalies, and uneven stress.

[0004] Some existing methods attempt to establish a mapping relationship between defect features and process parameters using machine learning models. However, these methods typically rely on end-to-end data fitting and lack consideration for physical constraints such as process sequence, equipment adjustability, parameter coupling, and process response lag. This results in insufficient interpretability of the output, and the parameter tuning suggestions may be difficult to directly apply to on-site process control. Furthermore, existing detection data often lacks stable correlation with batches, processes, equipment, and real-time process parameters, making it difficult to trace the root causes of defects and resulting in low efficiency in process investigation and adjustment after batch defects occur.

[0005] Furthermore, conventional grading and sorting methods often rely on defect size, quantity, or appearance category for judgment, rarely considering factors such as the optical path region where the defect is located, abnormal polarization response, and the product's reworkability. This can easily lead to reworkable or downgradeable products being mistakenly scrapped, or products with optical risks flowing into subsequent processes. Existing quality control systems also generally lack a closed-loop learning mechanism based on test results, process adjustments, quality feedback, and manual review, making it difficult to continuously optimize detection, sorting, and process adjustment capabilities in response to changes in equipment status, environmental fluctuations, and process drift.

[0006] Therefore, existing technologies still cannot simultaneously meet the needs of high-precision detection of optically sensitive defects, traceable and explainable process root causes, refined classification and sorting, and closed-loop optimization of processes in prism production. Summary of the Invention

[0007] To address the shortcomings and deficiencies of existing technologies, this invention provides a prism inspection, sorting, and quality control method and system based on defect mapping and process inversion. This method acquires multi-angle polarization images of the prisms to be inspected after the processing steps of the prism production line, and simultaneously obtains process context information such as prism identity, batch, process, equipment, and real-time process parameters, forming a traceable raw data fusion body. The polarization images are subjected to polarization feature decomposition and defect detection, extracting structured defect features containing information such as polarization response characteristics and the location of optical path sensitive areas. A defect evolution map is constructed based on the structured defect features of historical batches and process context data, and attribution hypothesis relationships between defects, processes, and process parameters are established in the map. When a defect appears in a new batch, similar defects are retrieved through the map. The model generates candidate process deviation hypotheses, and filters and ranks these hypotheses based on process physical constraints to obtain root cause analysis results. Further, it combines validated historical adjustment data and process parameter response relationships to generate theoretical parameter adjustment amounts, and generates multi-step restricted adjustment instructions based on equipment safety constraints when process adjustment trigger conditions are met. Simultaneously, it performs multi-dimensional risk classification of the prism based on structured defect characteristics and root cause analysis results, and generates sorting decisions. Finally, it collects sorting decisions, process adjustment execution records, quality feedback, and review results to form a closed-loop verification sample, used to update the defect detection model, defect evolution map, and process inversion model. This invention unifies prism polarization defect detection, process root cause tracing, graded sorting, restricted parameter adjustment, and closed-loop learning into a single quality control process, improving the traceability, interpretability, and on-site executability of prism production quality control.

[0008] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0009] To overcome the problems of insufficient defect identification accuracy, difficulty in consistently detecting optically sensitive defects, ambiguity in locating process root causes, lack of safety constraints in parameter tuning, coarse sorting and grading, and difficulty in closed-loop utilization of quality control data in existing prism production quality control processes, this invention provides a prism detection and sorting quality control method and system based on defect maps and process inversion. This scheme combines multi-angle polarization imaging, synchronous binding of process context, structured defect characterization, defect evolution maps, physically constrained process inversion, generation of theoretical parameter adjustment amounts, multi-dimensional risk grading, restricted process adjustment, and a closed-loop learning mechanism, enabling the prism quality control process to expand from simple defect identification to defect root cause tracing, product sorting and disposal, and process optimization control.

[0010] The overall technical concept of this invention is as follows: First, multi-angle polarization imaging is used to acquire polarization image information that reflects the surface morphology, film state, and stress state of a prism, and this image information is synchronously associated with the prism's identity, batch, process, equipment, and real-time process parameters; then, the detected defects are converted into structured defect features containing polarization response and optical path position information; further, structured defect features from historical batches and process context data are used to construct a defect evolution map, expressing the attribution hypothesis relationship between defects, processes, and process parameters in the map; when a defect appears in a new batch, candidate defects are obtained through map retrieval. The process deviation hypothesis is assumed, and candidate hypotheses are screened and ranked using process physical constraints to obtain root cause analysis results that can point to the process links and process parameter states. Based on this, theoretical parameter adjustment amounts are generated by combining verified historical adjustment data and process parameter response relationships, and multi-step restricted adjustment instructions are generated according to equipment safety constraints. At the same time, sorting decisions are generated based on structured defect characteristics and root cause analysis results. Finally, sorting decisions, process adjustment execution records, quality feedback and review results are incorporated into closed-loop verification samples to continuously update at least one of the defect detection model, defect evolution map and process inversion model.

[0011] Specifically, the present invention provides a prism inspection and sorting quality control method based on defect maps and process inversion, including the following contents.

[0012] First, multi-angle polarization images of the prism to be inspected are acquired. Simultaneously, the corresponding identity information, batch information, process information, equipment information, and real-time process parameters of the prism are obtained. The polarization images are then associated and stored with these information to form a traceable raw data fusion. This raw data fusion ensures that the defect detection results of a single prism are no longer isolated image results, but can be mapped to a specific batch, process, equipment, and corresponding process parameter status, providing a data foundation for subsequent defect evolution mapping and process root cause analysis.

[0013] In a preferred embodiment, an inspection station can be set up after the key processing steps in the prism production line. A programmable polarization light source array, an industrial polarization camera, and a rotating stage work together to execute multi-angle, multi-polarization imaging sequences. The polarization images are then linked and stored with real-time process parameters using timestamps and prism identification codes as indexes. This establishes temporal and identification consistency between inspection data and process data, improving the reliability of subsequent quality traceability.

[0014] Secondly, polarization feature decomposition is performed on the acquired polarization images, and a deep learning detection model is used to identify defect regions, extracting structured defect features that include polarization response characteristics and the location of optical path sensitive areas. Polarization feature decomposition can transform the light intensity information in ordinary images into polarization response information that better reflects the optical state of the material, making anomalous regions related to surface morphology, film state, or internal stress more prominent. The location of the optical path sensitive area is used to characterize the degree of influence of defects in the optical functional areas of the prism, so that defect characterization is no longer limited to appearance, size, and location, but can reflect its potential impact on optical performance.

[0015] In a preferred embodiment, polarization eigendecomposition includes calculating the Stokes vector based on multi-polarization images, further obtaining the degree of linear polarization and the polarization angle. The polarization angle is calculated using a four-quadrant arctangent function to reduce quadrant ambiguity that may arise from ordinary arctangent calculations. When acquiring circularly polarized images, the complete degree of polarization can also be calculated to accommodate defect types that may involve circularly polarized responses.

[0016] In a preferred embodiment, a polarization feature attention module is incorporated into the neck region of the deep learning detection model. This module performs global average pooling on multiple polarization feature channels to obtain channel description vectors, generates channel importance scores via a multilayer perceptron, normalizes these scores to obtain attention weights for each channel, and finally outputs a multi-channel weighted fusion feature. This structure allows the detection model to automatically adjust the contribution ratio of different polarization feature channels based on the differences in polarization response of different defects, making it more suitable for identifying polarization-sensitive defects such as fine scratches, uneven stress, and coating abnormalities.

[0017] In a preferred embodiment, the structured defect features further include basic defect geometric parameters, defect edge morphological parameters, and cross-face defect co-occurrence relationship parameters. The basic defect geometric parameters characterize the size, shape, and location of the defect; the defect edge morphological parameters characterize the roughness or irregularity of the defect profile; and the cross-face defect co-occurrence relationship parameters describe the spatial location or pattern association of defects across different optical surfaces of the same prism. These structured features provide standardized input for subsequent similar defect retrieval, attribution of data, and risk classification.

[0018] Furthermore, a defect evolution map is constructed based on the structured defect features and process context data of historical batches. This defect evolution map contains attribution hypothesis relationships between defects, processes, and process parameters, used to characterize the potential associations between different defect patterns and process parameter states under different processing steps. When a defect occurs in a new batch, similar defect patterns are retrieved from the defect evolution map to generate candidate process deviation hypotheses, which are then filtered and ranked by process physical constraints to obtain root cause analysis results. This process avoids directly mapping defect features to process parameters in a black box manner, instead using historical correlations and process constraints to jointly infer the possible sources of defects.

[0019] In a preferred embodiment, the defect evolution map also includes equipment nodes. The attribution hypothesis is a ternary association consisting of defect nodes, process nodes, and process parameter nodes. The association weights are calculated by fusing co-occurrence probability, normalized conditional mutual information, historical parameter tuning feedback confidence, and equipment status confidence. Using a ternary association allows for the simultaneous inclusion of defect patterns, process stages, and process parameter status in the attribution analysis. Compared to a simple binary mapping between defects and parameters, this is more suitable for expressing the defect formation patterns under multi-process, multi-parameter coupling conditions. The introduction of equipment status confidence and historical parameter tuning feedback confidence ensures that the map relationship not only reflects statistical correlations but also incorporates process experience validated by production feedback.

[0020] In a preferred approach, process physical constraints include process sequence constraints, equipment adjustable range constraints, parameter coupling constraints, and response lag constraints. Candidate process deviation assumptions are first filtered through hard constraints to eliminate infeasible assumptions that do not conform to process logic, equipment capabilities, or parameter physical relationships. The remaining feasible assumptions are then comprehensively ranked, and the root cause analysis results are output. This approach ensures that the root cause analysis results not only correlate with historical data but also meet on-site process control and equipment safety requirements, reducing the probability of generating unfeasible parameter tuning suggestions.

[0021] Furthermore, based on the structural defect characteristics and root cause analysis results, the prism is subjected to multi-dimensional risk classification to generate corresponding sorting decisions. The multi-dimensional risk classification not only considers the size and quantity of defects, but also combines factors such as defect polarization response, optical path position, and the possibility of product rework and repair to comprehensively evaluate the product status, enabling the sorting results to more finely distinguish different disposal directions such as qualified release, downgraded use, rework, isolation pending judgment, or scrapping.

[0022] In a preferred approach, multidimensional risk grading includes three dimensions: defect size risk level, optical performance risk level, and product rework repairability level. The defect size risk level reflects the risk posed by the geometric size and quantity of defects; the optical performance risk level reflects the impact of the optical path region where the defect is located and the degree of polarization anomaly on optical performance; and the product rework repairability level reflects the possibility of restoring the current prism to an acceptable quality state through cleaning, repolishing, replating, annealing, or other rework methods. By combining these multiple dimensions, sorting decisions can balance product quality risk and resource utilization efficiency.

[0023] Furthermore, based on the root cause analysis results, and combined with at least one of the following: validated historical adjustment data and process parameter response relationships, theoretical parameter adjustment amounts are generated. Validated historical adjustment data reflects existing parameter adjustment experience and its feedback results under similar defect patterns, while process parameter response relationships reflect the impact trend of process parameter changes on defect risk or quality indicators. By generating theoretical parameter adjustment amounts from the above data or relationships, the root cause analysis results can be transformed from qualitative deviation judgments into quantitative adjustment targets that can be used for process control.

[0024] In a preferred embodiment, the theoretical parameter adjustment amount is generated by at least one of the historical closed-loop sample weighting method and the process sensitivity matrix method. When both the historical closed-loop sample weighting method and the process sensitivity matrix method are used simultaneously, the adjustment amounts obtained by the two methods are dynamically fused, and a direction consistency check is performed before fusion. If the adjustment amounts obtained by the two methods conflict in direction and have similar confidence levels, no automatic parameter tuning command is output, and the process is transferred to manual review. Through this mechanism, a complementarity can be formed between historical experience and process response models, and erroneous automatic parameter tuning can be avoided when the adjustment direction is uncertain.

[0025] When the process adjustment triggering conditions are met, the theoretical parameter adjustment amount is decomposed into multiple restricted adjustment commands based on equipment safety constraints. This design does not directly issue the theoretical parameter adjustment amount to the equipment all at once, but converts the theoretical adjustment target into step-by-step control commands that are limited by equipment safety boundaries, single maximum adjustment capacity, and parameter linkage relationships, thereby making the process adjustment process more stable and controllable.

[0026] In a preferred approach, multi-step constrained adjustment instructions are generated using iterative optimization of the remaining adjustment amount. Constraints include the maximum single parameter adjustment amount, parameter safety boundaries, and parameter linkage rules. After each adjustment step is completed, the remaining adjustment target is updated based on the executed adjustment amount. Subsequent quality feedback is then used to determine whether to continue with subsequent adjustment steps, terminate the adjustment early, or transfer it to manual review. This approach avoids equipment instability or secondary defects caused by sudden parameter changes and improves the on-site adaptability of process adjustments.

[0027] Finally, the system collects sorting decision-making records, process adjustment execution records, adjusted quality feedback data, and verification results to form a closed-loop verification sample. Based on this sample, at least one of the defect detection model, defect evolution map, and process inversion model is updated. The closed-loop verification sample demonstrates the actual correspondence between detection, sorting, parameter tuning, quality feedback, and verification results, enabling the system to continuously refine the model and map using validated data, rather than relying on unvalidated original records for direct updates.

[0028] In a preferred approach, sample admission rules are established to screen high-confidence closed-loop validation samples with complete process records and verified adjustment effects. Samples are dynamically normalized and weighted using multi-source evidence to generate soft labels, with the weights of each evidence source dynamically updated based on its historical validation reliability. Model and graph updates employ a combination of incremental updates and full retraining, with a performance gating mechanism implemented. These mechanisms reduce the impact of noisy samples or conflicting evidence on model updates and ensure necessary performance validation before deploying the updated model or graph.

[0029] This invention also provides a prism inspection and sorting quality control system based on defect maps and process inversion. The system includes a data acquisition module, a defect detection and structuring module, a map construction and process inversion module, a sorting and parameter tuning generation module, and a closed-loop update module.

[0030] The data acquisition module acquires multi-angle polarization images of the prism and simultaneously obtains and stores corresponding identity information, process information, equipment information, and real-time process parameters. The defect detection and structuring module performs polarization feature decomposition and defect detection, outputting structured defect features containing polarization response characteristics and optical path position information. The map construction and process inversion module constructs a defect evolution map containing attribution assumptions between defects, processes, and process parameters, retrieves similar defect patterns, and outputs root cause analysis results based on process physical constraints. The sorting and parameter tuning generation module performs multi-dimensional risk classification based on structured defect features and root cause analysis results, generates sorting decisions, generates theoretical parameter adjustment amounts, and outputs multi-step restricted adjustment instructions that comply with equipment safety constraints when trigger conditions are met. The closed-loop update module collects full-process data to generate closed-loop verification samples and updates at least one of the defect detection model, defect evolution map, and process inversion model.

[0031] Compared to existing technologies, this invention and its preferred embodiments synchronously associate multi-angle polarization images with identity, process, equipment, and real-time process parameters, enabling defect detection results to be traced back to specific processing contexts and providing a reliable data foundation for subsequent process attribution. Through polarization feature decomposition and structured defect feature extraction, defects related to polarization response, such as prism surface morphology, film state, and stress anomalies, can be expressed in a form more suitable for process analysis. By constructing a defect evolution map containing attribution assumptions between defects, processes, and process parameters, and combining this with inversion analysis based on process physical constraints, the invention avoids the pitfalls of solely relying on black-box models to output process recommendations. The system addresses issues of insufficient interpretability and engineering feasibility. By combining validated historical adjustment data with process parameter response relationships to generate theoretical parameter adjustment amounts, and further decomposing them into multi-step restricted adjustment instructions according to equipment safety constraints, process adjustments can better conform to equipment capability boundaries and production safety requirements. Through multi-dimensional risk classification and sorting decisions, a more detailed basis for product handling can be formed among defect appearance, optical effects, and rework probability. By updating the defect detection model, defect evolution map, and process inversion model through closed-loop validation samples, the quality control system can be continuously corrected and optimized as production data accumulates, improving its adaptability to process fluctuations and equipment status changes. Attached Figure Description

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0033] Figure 1 This is a flowchart illustrating the overall implementation of the embodiments of the present invention. Detailed Implementation

[0034] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0037] This invention addresses long-standing pain points in prism manufacturing quality control by constructing a complete intelligent management and control system, from accurate defect identification to closed-loop process optimization. As a core component in high-end manufacturing fields such as optical instruments, aerospace components, and precision lenses, the quality of prisms directly determines the performance of the final product. However, traditional manual visual inspection is inefficient and has a high rate of missed defects. Ordinary machine vision struggles to suppress environmental reflections and cannot identify optically sensitive defects such as uneven stress. Existing AI methods based on single mapping suffer from black-box characteristics, with output parameter tuning suggestions lacking interpretability and physical constraints, leading to recurring batch defects and significant resource waste. This solution breaks the traditional black-box mapping model between defects and process parameters, forming a full-link technical route encompassing polarization imaging detection, defect structure characterization, defect evolution map construction, physical constraint process inversion, hierarchical sorting and constrained parameter tuning, and continuous learning optimization. This achieves intelligent quality control throughout the entire prism manufacturing process.

[0038] In the defect detection stage, this solution employs multi-angle, multi-polarization state imaging technology. Through the coordination of a programmable polarization light source array and a precision rotating stage, polarization images of the prism under different orientations and illumination conditions are acquired, effectively suppressing strong reflective interference from metal and optical surfaces. The system calculates linear polarization degree, complete polarization degree, and polarization angle based on Stokes vectors, and replaces the traditional arctangent function with a four-quadrant arctangent function, resolving quadrant ambiguity in polarization angle calculation. To further enhance the detection capability of subtle defects, a polarization feature attention module is introduced into the neck of the YOLOv8 detection model. This module adaptively weights the feature contributions of light intensity, linear polarization degree, and polarization angle channels according to the defect type, significantly improving the accuracy of identifying polarization-sensitive defects such as fine scratches, coating anomalies, and internal stress. After detection, the system extracts the basic geometric information, polarization response characteristics, optical path sensitive area location, edge morphological parameters, and cross-plane co-occurrence relationships for each defect, encoding them into multi-dimensional structured nodes to lay the data foundation for subsequent root cause analysis.

[0039] To address the ambiguity in root cause localization using traditional methods, this solution innovatively constructs a heterogeneous defect evolution map. This map includes four types of nodes: defects, processes, process parameters, and equipment, as well as three types of relationships: temporal sequence, co-occurrence probability, and attribution hypothesis. The solution defines the relationship between defects and process parameters as a ternary attribution hypothesis relationship, with its weights derived from a fusion of co-occurrence probability, normalized conditional mutual information, historical parameter tuning feedback confidence, and equipment status confidence. When a new batch of defects occurs, the system searches for similar defect patterns in the map to generate candidate process deviation hypotheses. These hypotheses are then subjected to hard filtering and soft sorting based on physical constraints such as process sequence, equipment parameter range, parameter coupling, and response lag. The final output is a root cause analysis result that combines statistical relevance and physical executability. Furthermore, the solution proposes a quantitative theoretical parameter tuning generation mechanism, primarily based on a historical closed-loop sample weighting method and supplemented by a process sensitivity matrix method. This addresses the problem that traditional methods can only output qualitative deviation directions, forming a complete automatic parameter tuning logic loop.

[0040] In the product sorting and process adjustment stages, this solution establishes a three-dimensional classification model based on defect size risk, optical performance risk, and product rework repairability. It comprehensively considers the physical properties of the defect itself, its impact on the optical system, and the possibility of rework repair, classifying products into five categories: qualified products, downgradeable products, reworkable products, isolated products awaiting judgment, and scrapped products. This achieves maximum resource utilization and precise control of quality risks. To ensure the safety of process adjustments, the system does not directly issue theoretical parameter adjustment values. Instead, it uses an iterative optimization method with remaining adjustment values ​​to decompose them into multi-step adjustments that comply with equipment safety specifications. Each adjustment step strictly adheres to the maximum single adjustment value, parameter safety boundaries, and parameter linkage rules, and employs differentiated adjustment rhythms and strategies for sudden, drifting, and periodic defects. After each adjustment step is completed, the system combines subsequent detection feedback to determine whether to continue with the remaining steps. If the defect has been effectively suppressed, the process terminates early; if an anomaly occurs, it proceeds to manual review, minimizing the risks of automatic parameter adjustment.

[0041] To achieve continuous iteration of system capabilities, this solution designs a robust verification closed-loop continuous learning mechanism. Addressing the common problem of unavailable absolute attribution labels in industrial scenarios, the system dynamically normalizes and weights multi-source evidence, including expert review, DOE experiments, feedback on quality improvement after parameter tuning, and equipment anomaly logs, generating high-confidence soft labels for model training, effectively reducing the impact of noise from single human labels. The system employs a dual-cycle update strategy combining high-frequency incremental updates and periodic full retraining. The defect detection model and graph edge weights are updated daily or per shift, and the process inversion model is retrained monthly or quarterly, with constraint priorities optimized. Simultaneously, the solution establishes a strict model deployment gating mechanism. A new model is only allowed to be deployed when key indicators such as root cause localization accuracy and parameter tuning effectiveness meet preset thresholds, and the updated model performance is no less than 95% of the original version, thus avoiding performance degradation caused by model mislearning.

[0042] Overall, this solution improves defect detection accuracy through polarization imaging technology, achieves interpretable root cause localization through defect evolution maps and physical constraint inversion, optimizes production efficiency and resource utilization through quantitative safety intervention and refined sorting, and ensures the long-term stability of the system through closed-loop continuous learning.

[0043] like Figure 1 As shown, the specific implementation steps of the embodiment of the present invention include:

[0044] Step 1: Deploy a multi-angle polarization imaging and process data synchronous acquisition system. Set up an inspection station after the key processes in the prism production line, integrating a programmable polarization light source array (which can be implemented using an LED array with a liquid crystal polarizer, switching polarization states by controlling the voltage of the liquid crystal polarizer), a high-resolution industrial camera, and a precision rotating stage. After system startup, execute a preset multi-angle, multi-polarization state imaging sequence for each prism. Simultaneously, through the device interface, collect and bind the prism's batch number, unique identification code, current process number, processing equipment identifier, and the real-time process parameter set for that batch in the corresponding process. All acquired polarization images, prism facet attitude information, and process context data are linked and stored using timestamps and identification codes, forming a raw data fusion body for subsequent traceability analysis. The specific implementation method of this step is as follows:

[0045] A separate inspection station is set up after key processes (such as grinding, polishing, and coating). This station includes a precision motorized rotating stage for fixing and rotating prisms, a programmable polarization light source array for providing multi-angle, multi-polarization illumination, and a high-resolution industrial polarization camera for image acquisition. The system communicates with the production line main control system and the process controllers of upstream processing equipment (such as grinding machines and coating machines) via a PLC or industrial Ethernet. When the prism arrives at the inspection station by the conveyor mechanism, the barcode reader first reads its identification code. This code is related to the batch number in the production management system. Current process number and processing equipment markings Automatic association. The system then triggers an imaging sequence: the rotating stage rotates according to a preset angle set:

[0046]

[0047] Rotate the prism successively for each stage orientation The polarization light source array sequentially switches to a preset set of polarization states:

[0048]

[0049] like , , , Linear polarization and left-handed and right-handed circular polarization) and the set of incident angles:

[0050]

[0051] Illumination is applied. The camera simultaneously acquires polarization images under each illumination condition. During the same time window of the imaging sequence execution Internally, the system collects the real-time process parameter set for the corresponding batch of prisms at the corresponding process through the device interface:

[0052]

[0053] For example, grinding pressure Polishing fluid flow rate Vacuum degree of coating chamber All collected data is timestamped. and identity code Perform associative encapsulation on the index. The resulting encapsulated original data is then merged. This can be formally represented as:

[0054]

[0055] This rigorous spatiotemporal and identity binding mechanism ensures that the polarization image of each defect can be traced back to the precise process context in which it was generated, laying the data foundation for the subsequent construction of a high-fidelity defect evolution map and serving as a prerequisite for realizing physically constrained process inversion.

[0056] The completeness of the imaging sequence is determined by the number of rotational attitudes (M), the number of polarization states (N), and the number of incident angles (L). In standard quantitative detection mode, it is preferable to set M ≥ 8 to cover all major optical surfaces of the prism; to set at least four linear polarization states (0°, 45°, 90°, and 135°) to calculate complete linear polarization characteristics; and to set L ≥ 3 to cover typical illumination conditions such as near-normal, oblique, and grazing incidence. The total sampling time for a single item in standard quantitative detection mode is... Keep it within 10 to 30 seconds.

[0057] The system can employ a tiered sampling strategy to address different production line cycle times and quality risk levels. In regular or high-risk batches, the standard quantitative detection mode described above is used to generate complete polarization-structured defect nodes. In low-risk batches where no batch defects have occurred in multiple consecutive batches and the equipment is in stable condition, the number of rotational attitudes or incident angles can be reduced. For example, fewer prism attitudes and a single typical incident angle can be used, but it is still preferable to retain the four linear polarization states of 0°, 45°, 90°, and 135° to ensure the calculability of DoLP and AoP. If further simplified to two polarization states, this mode is used only as a rapid pre-screening mode; after anomalies are detected during pre-screening, the standard quantitative detection mode is retried.

[0058] For high-cycle production lines, the system can also employ one or more of the following methods for cycle time adaptation: parallel acquisition by multiple cameras, multi-station splitting, pre-inspection buffer queue, and asynchronous processing of imaging tasks. Specifically, parallel acquisition by multiple cameras reduces the number of rotations of the rotating platform; multi-station splitting balances the inspection pressure at a single station; the buffer queue absorbs instantaneous cycle time fluctuations in the production line; and asynchronous processing decouples image acquisition from subsequent model inference, thereby preventing the inspection station from becoming a bottleneck in the production line.

[0059] Step 2: Perform polarization image processing and defect structured feature extraction. The acquired raw images undergo geometric correction, inter-plane registration, and polarization feature decomposition to suppress environmental reflections and background interference. For the enhanced images, a deep learning-based detection model is used to identify defect regions, and a structured feature set for each defect is further calculated. This feature set not only includes basic information such as defect type, size, and location, but also specifically extracts its polarization response characteristics, such as polarization degree anomalies, the location of the sensitive area on the optical surface, the morphological parameters of the defect edges, and co-occurrence relationships with defects on other surfaces. Finally, each defect is encoded as a structured node containing multi-dimensional attributes, laying the foundation for building a network of connections between defects. The specific implementation method of this step is as follows:

[0060] First, geometric correction is performed on the original image sequences from different viewpoints and polarization states. Based on the relative pose model of the camera and prism, a pre-calibrated intrinsic parameter matrix is ​​used. and distortion coefficient Distortion correction is performed on the image. For each prism surface, the projection relationship of its three-dimensional plane equations in the camera coordinate system is established by solving the homography matrix. Achieve geometric transformation of the image to a standard frontal plane, ensuring that the images of all surfaces have a uniform scale and orientation, laying the foundation for subsequent inter-surface registration.

[0061] Next, inter-plane registration and polarization feature decomposition are performed. Corrected images of different optical surfaces of the same prism are registered based on their known physical spatial relationships. A registration method combining feature points (such as SIFT) and known geometric constraints is used to establish pixel-level correspondences between cross-surface images.

[0062] For each registered pixel position The system calculates the Stokes vector based on the multi-polarization image after dark field correction, flat field correction, and exposure normalization: .

[0063] Where I is the total light intensity, Q and U are the linearly polarized components, and V is the circularly polarized component. For images with four linearly polarized directions, the following calculation can be performed:

[0064]

[0065]

[0066]

[0067] When using only orthogonal linear polarization pairs, the following can also be used:

[0068]

[0069] As an approximation of the total light intensity, if the system acquires a left-handed circularly polarized image I_L and a right-handed circularly polarized image I_R, then the circular polarization components can be expressed as:

[0070]

[0071] Based on the Stokes components mentioned above, calculate the degree of linear polarization:

[0072]

[0073] in To prevent extremely small positive numbers with a denominator of zero, if circular polarization information is needed, the complete degree of polarization is further calculated:

[0074]

[0075] The polarization angle is calculated using the four-quadrant arctangent function:

[0076]

[0077] In this way, the system can avoid quadrant ambiguity caused by the ordinary arctangent function when Q<0, and can distinguish between linearly polarized sensitive defects and special defects that may involve circular polarization response.

[0078] This step transforms the original intensity image into a feature image containing polarization information, which can effectively suppress unpolarized environmental reflections and background clutter, and highlight areas with abnormal polarization characteristics caused by surface morphology, stress, film, etc.

[0079] For the enhanced polarization feature image, a deep learning-based detection model is used to identify defect regions. In a preferred embodiment, the model employs an improved architecture based on the YOLOv8 detection framework, introducing a CSPDarknet-like feature extraction structure into the backbone network, or replacing the original YOLOv8 backbone network with CSPDarknet53, to enhance multi-scale defect feature extraction capabilities. A polarization feature attention module (PFA Module) is introduced in the neck region, which adaptively weights the features from the light intensity channel. Linear polarization channel Polarization angle channel The characteristics of the three channels. The loss function is:

[0080]

[0081] in For the complete crossover and union ratio loss, For binary cross-entropy classification loss, A distributed focus loss is used to improve the accuracy of bounding box regression. The model outputs the class confidence, bounding box coordinates, and a preliminary mask for each defect. The beneficial effect of introducing a polarization feature attention module is that it guides the network to pay more attention to regions with abnormal polarization responses, thereby improving the detection rate and classification accuracy of defects closely related to the optical properties of materials, such as stress inhomogeneity and minor scratches.

[0082] The polarization feature attention module is used to adaptively weight the intensity channel, linear polarization degree channel, and polarization angle channel. Let the detection network be in the... The three types of features extracted by the layer are as follows:

[0083]

[0084] First, perform global average pooling on the features of each channel to obtain the channel description vector:

[0085]

[0086] Then, importance scores for each channel are generated using a shared multilayer perceptron:

[0087]

[0088] The attention weights are obtained by using Softmax normalization:

[0089]

[0090] The final fusion features are:

[0091]

[0092] This module can dynamically adjust the contribution ratio of light intensity information and polarization information under different defect types, enabling the model to have a higher detection capability for polarization-sensitive defects such as fine scratches, uneven stress, and coating abnormalities.

[0093] After obtaining the defect regions, the structured feature set for each defect is further calculated. The basic geometric features include: based on the length of the minimum bounding rectangle. ,width Aspect Ratio ,area Centroid coordinates Edge morphological parameters are calculated by pre-calculating the Fourier descriptor of the defect profile. Low-frequency coefficients To characterize roughness and irregularity. Polarization response characteristics are a key supplement, used to calculate the average polarization anomaly within the defect region:

[0094]

[0095] and polarization angle perturbation That is, within the region The standard deviation of the defect. To relate to the process context, the location of the sensitive area of ​​the defect on the optical surface needs to be calculated: the standard frontal plane is divided into the core optical path area, the transition area, and the edge area. The minimum Euclidean distance from the defect centroid coordinates to the center of each area is calculated and mapped to the corresponding region code. The cross-surface co-occurrence relationship feature is generated by searching for defects on other surfaces of the same prism at their corresponding spatial projection positions, forming a binary vector. ,in This represents the total number of facets of the prism. Indicates the first The surface has defects at the associated location.

[0096] Ultimately, each defect is encoded as a structured node. Its attribute vector is:

[0097]

[0098]

[0099] All features are normalized before splicing. This structured encoding not only contains rich physical and optical properties, but also has a unified format, which can be directly used as the basic data unit for constructing nodes in the subsequent defect evolution map, providing accurate and multidimensional input for in-depth analysis of the spatiotemporal and attributional relationships between defects.

[0100] Step 3: Construct a defect evolution map and drive process inversion analysis under physical constraints. Using historical batch data, a defect evolution map is established with defect nodes, process nodes, process parameter nodes, and equipment nodes as its core, characterizing the temporal sequence, co-occurrence probability, and attribution hypothesis relationships between nodes. When a defect appears in a new batch, the system first searches the map for similar defect patterns and their historically associated process deviation hypotheses. Subsequently, the physical constraint process inversion model is activated. This model compares and filters data-driven candidate deviation parameters with the inherent sequential constraints of processes, the physical adjustable range of equipment parameters, the coupling relationships between parameters, and the hysteresis characteristics of adjustment responses. Finally, root cause analysis results with process interpretability are output, including the most likely dominant abnormal process, specific suspected parameter groups, parameter deviation directions, and adjustment priorities, and quantitative theoretical parameter adjustment amounts are generated.

[0101] The specific implementation method for this step is as follows:

[0102] First, based on the structured defect status coding and process context data of historical batches, a heterogeneous defect evolution map is constructed. .

[0103] Among them, the node set It includes four types of nodes: defective nodes Its attribute is a defect status code; process node Associated process names and sequences; process parameter nodes Includes parameter name, nominal value, and unit; device node. Record the equipment number and status. The edge set includes process time sequence edges, defect and process co-occurrence edges, equipment and process association edges, and attribution hypothesis relationships between defects, processes, and process parameters.

[0104] The attribution hypothesis relation can be represented as a ternary relation or a hyperedge:

[0105]

[0106] Indicates a defect node At the process node Next and process parameter nodes There is a potential attribution relationship. The strength of this attribution relationship is not directly equivalent to a strict causal relationship, but is determined by co-occurrence probability, conditional mutual information, process time sequence, and historical parameter tuning feedback.

[0107] The conditional mutual information component can be represented as:

[0108]

[0109] In the formula, x, y, and z correspond to the discrete or binned values ​​of process parameter status, defect status, and process status, respectively. The system integrates this conditional mutual information with the historical defect improvement confidence, process sequence constraints, and equipment status confidence to form the attribution hypothesis weights:

[0110]

[0111] in, For normalized conditional mutual information, For historical parameter tuning feedback confidence level, The confidence level is related to the device status. These are the weighting coefficients.

[0112] When constructing the graph, edge weights are calculated by statistically analyzing historical data, for example:

[0113]

[0114] in It is a defect With process Co-occurrence frequency It is a defect Total number of occurrences.

[0115] When defects appear in the new batch At that time, the system performs subgraph retrieval within the map. Calculation With all historical defect nodes Similarity:

[0116]

[0117] in The Mahalanobis distance is based on the structural properties of defects. The scaling factor is used. The top-K similar defect nodes and their associated ternary attribution hypothesis relationships are retrieved, forming a candidate set of process deviation hypotheses.

[0118]

[0119] Subsequently, a physical constraint process inversion model was initiated. This model first employs a three-layer fully connected neural network to perform preliminary scoring on candidate hypotheses. The network input is a concatenated vector of the deviations between defect features and candidate parameters, and the output is a data-driven anomaly confidence score. The network structure can be configured with a 256-dimensional input layer and 128-dimensional or 64-dimensional hidden layers, all using... Activation function, used in the output layer Functions. However, relying solely on data-driven approaches may produce results that violate process common sense; therefore, physical constraints are introduced for reordering. Define a set of constraint functions. Including process sequence constraints If the parameter Process If a process is later in time than any other process where the defect might occur, the assumption is discarded; equipment adjustable range constraints. Parameter adjustment amount Must meet:

[0120]

[0121] Parameter coupling constraints For parameter sets with physical coupling Its adjustment must meet the following requirements. ,in The equations represent the coupling relationship; the response hysteresis constraint. For parameters with inertia, such as temperature and concentration, the adjusted batches must take into account the lagging batches. Ultimately, the combined score for each candidate hypothesis is calculated. The calculation process is as follows:

[0122] Candidate attribution hypotheses obtained from map retrieval The system first calculates the anomaly confidence level using a data-driven scoring model:

[0123]

[0124] in, As a current structural feature of defects, For candidate processes, parameters, and equipment status characteristics, () is the Sigmoid function.

[0125] Subsequently, the system performs hard constraint filtering on the candidate hypotheses. Let the set of physical constraints be:

[0126]

[0127] These represent process sequence constraints, equipment adjustable range constraints, parameter coupling constraints, and response lag constraints, respectively. Define the feasibility function:

[0128]

[0129] in:

[0130]

[0131] When Feasibleh=0, the candidate hypothesis is eliminated; when Feasibleh=1, the overall ranking score is calculated:

[0132]

[0133] in, Weights for attribution assumptions in the graph. Scoring the urgency or scope of impact of parameter adjustments. For soft-constraint sets, For the degree of violation of soft constraints, These are the weighting coefficients. The system uses... The feasible candidate hypotheses are sorted in descending order, and the dominant abnormal process, suspected parameter group, parameter deviation direction and adjustment priority are output.

[0134] The above scheme combines the flexibility of a data-driven approach with the reliability of prior physical knowledge through a combination of graph retrieval, neural network scoring, and constraint reordering. , By using constraint functions for screening, it is ensured that the inverted process deviation parameters are not only statistically relevant, but also physically achievable and safely executed in engineering, thereby significantly improving the interpretability and field operability of the root cause analysis results.

[0135] After outputting the dominant abnormal process, suspected parameter group, and parameter deviation direction, the system further converts the qualitative parameter deviation judgment into quantitative theoretical parameter adjustment amounts. Let the candidate parameter set be:

[0136]

[0137] For each candidate parameter The process inversion model outputs its deviation direction. Where +1 indicates that the current parameter is higher than the target process state, and -1 indicates that the current parameter is lower than the target process state. The corresponding correction direction is defined as follows:

[0138]

[0139] That is, when the parameter is too high, the adjustment is reduced; when the parameter is too low, the adjustment is increased.

[0140] The system prioritizes using a weighted method of historical closed-loop samples to generate theoretical adjustment values. Let the set of the top K historical closed-loop samples similar to the current defect pattern be denoted as . Each sample h contains historical effective adjustment amounts. Defect pattern similarity and adjusted validity confidence For parameters The historical weighted adjustment value is:

[0141]

[0142] in To prevent extremely small positive numbers with a denominator of zero, the theoretical adjustment amount obtained by the historical sample method is:

[0143]

[0144] When the number of historical closed-loop samples is insufficient, the similarity is below the threshold, or the validity confidence is insufficient, the system uses the process sensitivity matrix method for supplementation. Let the defect risk index vector be R, including indicators such as defect size risk, optical risk, batch occurrence frequency, and polarization anomaly amplitude; let the process parameter vector be P. The sensitivity matrix is ​​estimated through DOE experiments, small-scale controlled trial runs, or historical closed-loop samples.

[0145]

[0146] It should be noted that the process sensitivity matrix J is not limited to being obtained through full-factor DOE experiments. In actual production lines, the system can preferentially utilize historical closed-loop samples or small-scale safety trial data for local linear estimation. Let the set of historical closed-loop samples be... Each sample records the parameter adjustment amount and corresponding defect risk changes The local sensitivity matrix can then be estimated using weighted ridge regression:

[0147]

[0148] in, The sample weights can be determined jointly by the defect pattern similarity and the adjusted validity confidence. This is the regularization coefficient. For processes with a large number of parameters, the local sensitivity matrix can be estimated only for a subset of candidate parameters to reduce experimental and computational costs. DOE experiments are mainly used for initial calibration or key parameter calibration, and are not required for every parameter tuning.

[0149] Given the current risk vector and target risk vector ,have:

[0150]

[0151] The theoretical adjustment obtained by the sensitivity matrix method is:

[0152]

[0153] Its constraints are:

[0154]

[0155]

[0156] Ultimately, the system determines the confidence level based on historical samples. The two adjustment amounts are then combined:

[0157]

[0158] Among them, when there are sufficient historical closed-loop samples and the similarity is high, Take the larger value; when historical samples are insufficient, Take the smaller value. Instead of being directly issued to the equipment for execution, it serves as the target adjustment amount for subsequent restricted process adjustment modules.

[0159] To ensure the safety of automatic parameter adjustment, a direction consistency check must be performed before the two adjustment values ​​are merged.

[0160] Before fusion, the system first verifies the directional consistency of the adjustment values ​​obtained from the historical sample method and the sensitivity matrix method. For any candidate adjustment value... If its direction is the same as the correction direction If they are inconsistent, then directional projection is performed:

[0161]

[0162] If the adjustment amount after projection is zero, or if the historical sample method and the sensitivity matrix method show significant directional conflict in the unprojected state, then the processing should be based on the confidence levels of the two methods. Let the confidence level of the historical sample be... The confidence level of the sensitivity matrix is When the difference between the two is greater than the preset threshold, the method with higher confidence level shall prevail; when the confidence levels of the two are close and their directions conflict, the system shall not output an automatic parameter adjustment instruction, but shall mark the parameter as requiring manual review.

[0163] Step 4: Implement intelligent sorting and restricted process adjustment based on multi-dimensional risk assessment. Each prism is classified in three dimensions based on the size level of the defect, the optical risk level of its location in the optical path, and the product's reworkability level. Sorting decisions are then made accordingly, classifying products into qualified products, downgradeable products, reworkable products, products awaiting rework, products awaiting further judgment, and scrapped products. For batch defects triggering process adjustments, the system does not directly use the theoretical values ​​given by the inversion model. Instead, based on the equipment's safe operating procedures, the adjustment instructions are broken down into multiple restricted steps. The system combines the safety boundaries of parameters, the maximum permissible adjustment amount per step, and parameter linkage rules to generate a step-by-step adjustment plan, employing differentiated adjustment strategies for sudden, drifting, and periodic defect patterns.

[0164] The system establishes a three-dimensional classification model based on defect size risk, optical performance risk, and product rework repairability. Based on these multi-dimensional defect characteristics, a three-dimensional risk classification model is constructed. The three-dimensional classification vector is represented as follows:

[0165]

[0166] in, For defect size levels, It is classified as an optical risk level. This represents the product's reworkability rating. Each rating ranges from 1 to 5, with higher values ​​indicating higher risk or greater difficulty in repair.

[0167] First, calculate the defect size risk score. The system projects the defect area. Maximum length and number of defects Normalization is performed:

[0168]

[0169]

[0170]

[0171] Norm(・) is a normalization function determined based on the statistical range of historical samples or the upper and lower limits of process standards. The defect size risk score is:

[0172]

[0173] in:

[0174]

[0175] The defect size grade is:

[0176]

[0177] Secondly, the optical risk level is calculated. The system determines the optical risk score based on the sensitivity weight of the optical path region where the defect is located, the amplitude of polarization anomalies, and the polarization angle perturbation. Let the normalized optical path sensitivity be... The normalized linear polarization anomaly is The normalized polarization angle perturbation is ,but:

[0178]

[0179] in:

[0180]

[0181] The cross term is used to characterize the synergistic risk amplification effect when significant polarization anomalies occur in the high-light path sensitive region.

[0182] The optical risk level is:

[0183]

[0184] Next, calculate the product rework repairability rating. Product rework repairability is used to determine whether the current prism can be restored to an acceptable quality state through cleaning, repolishing, replating, annealing, or other rework methods. Product rework reliability score. ∈[0,1] is determined by the repairability of the defect type, the defect size, the region where the defect is located, the available rework process capability, and the historical rework success rate:

[0185]

[0186] in, Indicates the rework feasibility corresponding to the defect type. Indicates the rework feasibility corresponding to the defect size. This indicates the impact of the defect location on the rework success rate. This indicates the success rate of rework for similar defects in the past. These are the weighting coefficients.

[0187] in, It can be obtained through table lookup, historical statistics, and normalization calculations.

[0188] The feasibility of rework for a defect type can be indicated by the rework process knowledge base. For example, higher values ​​are given for contamination / attachment defects, medium values ​​for shallow scratches, and lower values ​​for severe edge chipping and internal bubble defects.

[0189] This indicates dimensional repairability, which can be derived inversely from the defect dimensional risk score:

[0190]

[0191] Alternatively, the calculation can be normalized based on the maximum defect size allowed by the rework process.

[0192] This indicates the repairability of the location, which can be determined based on the sensitivity level of the optical path. Lower values ​​are used for the core optical path area, and higher values ​​are used for the edge area.

[0193] The success rate of rework for similar defects in the past can be calculated as follows:

[0194]

[0195] in, This represents the number of times similar defects have been successfully reworked in the past. This represents the total number of reworks for similar defects in history. This is the smoothing coefficient.

[0196] The product rework repairability level is:

[0197]

[0198] when When =1, =1 indicates that it is extremely easy to rework and repair; when When =0, =5 indicates that it cannot be reworked or repaired.

[0199] At the same time, the system also calculates the process root cause repairability score. This score is used to determine the likelihood of suppressing similar defects in subsequent batches through process parameter adjustments. The score is determined by a combination of factors, including parameter adjustability, adjustment margin, constraint feasibility, historical improvement rates from parameter adjustments, and equipment status. Root cause repairability is not directly used to determine whether the current product can be reworked, but rather for decisions regarding subsequent constrained process adjustments.

[0200] Among them, only when the process root cause repairability score Higher than the preset threshold Only when the root cause location confidence (RootConf) is higher than a preset threshold will the system allow the automatic restricted parameter tuning process to begin; otherwise, only parameter tuning suggestions will be output and the process will proceed to manual review.

[0201] The system is based on three-dimensional hierarchical vectors Five sorting decisions are implemented. Assuming that higher numerical values ​​for each level indicate higher risk or greater difficulty in repair, the sorting rules include:

[0202] Qualified product: Meets the following requirements:

[0203]

[0204] Furthermore, if the defect is not located in a prohibited area, it is directly judged as a qualified product.

[0205] Degradable items: meet the following requirements:

[0206]

[0207] However, if the prism does not meet the qualified product standards, and the actual optical performance or application scenario of the prism allows for downgraded application, it is determined to be a downgradeable product. For prisms used in critical optical systems, additional random checks can be performed on transmitted wavefront error, transmittance, polarization extinction ratio, or MTF before downgrade determination.

[0208] Workpieces requiring return: must meet the following requirements:

[0209]

[0210] If there is a clear rework path, such as cleaning, polishing, replating, or annealing, then it is determined that the workpiece needs to be reworked.

[0211] Items awaiting quarantine: Items are deemed to be awaiting quarantine if they meet any of the following conditions:

[0212]

[0213] or:

[0214]

[0215] Where RootConf represents the root cause location reliability, and freq(pattern) represents the frequency of similar defect patterns occurring in the same batch. This is the threshold for batch defect warnings. These products require manual review or a dedicated process inspection.

[0216] Scrapped products: Satisfy:

[0217]

[0218] or:

[0219]

[0220] If there is no feasible rework path, it is judged as scrap.

[0221] The aforementioned thresholds are jointly calibrated based on historical quality data, customer acceptance criteria, and prism optical design requirements, and can be periodically calibrated with the continuous learning module.

[0222] For batch defects that trigger process adjustments, the system does not directly use the theoretical adjustment amount. Instead of issuing the adjustment to the device, it is broken down into K constrained adjustment steps. The remaining theoretical adjustment amount before the t-th step is defined as:

[0223]

[0224] in, Let be the actual adjustment amount already performed in step τ. The actual adjustment amount in step t is obtained by solving the following optimization problem:

[0225]

[0226] Its constraints include:

[0227]

[0228]

[0229]

[0230] in, For parameters The maximum allowable adjustment amount per transaction. For equipment safety boundaries, This is a parameter linkage constraint. After each adjustment step is completed, the system updates the current parameter status:

[0231]

[0232] The system then uses subsequent testing feedback to determine whether to continue with the remaining steps. If the quality feedback indicates that the defect has been effectively suppressed, the system can terminate the subsequent adjustments in advance; if the defect has not been improved or a new abnormal pattern appears, the automatic parameter adjustment will be suspended and the process will be transferred to manual review or a special process investigation.

[0233] For sudden defects, the system uses fewer steps K and shorter adjustment intervals to quickly suppress anomalies; for drift defects, it uses a larger K and a medium adjustment interval, and evaluates quality feedback after each step; for periodic defects, the system checks equipment maintenance records, consumable replacement records and periodic environmental fluctuation records simultaneously before each adjustment, and uses a more conservative single-step adjustment amount.

[0234] To avoid sorting conflicts caused by multiple rules being triggered simultaneously, the system performs sorting judgments according to preset priorities. The priorities are as follows: scrap judgment, isolation pending judgment, qualified judgment, rework judgment, and downgradeable judgment. If the same prism meets multiple category conditions simultaneously, the final sorting result is output according to the category with the highest priority. For prisms used for customer-specified purposes or critical optical systems, the judgment of qualified and downgraded products must also meet the corresponding customer acceptance standards.

[0235] Preferably, the above thresholds can be obtained through expert rule initialization, customer acceptance standard mapping, historical quality data statistics, and manual review sample calibration. For new production lines or new prism models, when historical data is insufficient, an initial threshold can be set based on customer acceptance standards, optical design tolerances, and process expert experience. After accumulating a certain number of manual review samples, cluster analysis, ROC curve analysis, or cost-sensitive optimization methods can be used to calibrate the thresholds.

[0236] For example, a sorting cost function can be defined:

[0237]

[0238] in, Represents the set of thresholds for each level. For threshold-based For the sample The sorting results For manual review or customer acceptance results, The cost of missorting.

[0239] Meanwhile, each threshold preferably satisfies the following relative relationship:

[0240] The above thresholds can be designed using the following rules:

[0241]

[0242]

[0243] For the product rework repairability level, since a higher value indicates greater difficulty in repair, the following settings can be configured:

[0244] .

[0245] To ensure the consistency of sorting rules at different levels.

[0246] Step 5: Run the continuous learning and knowledge base update mechanism based on the verification loop.

[0247] The system continuously collects full-process data for new production batches, including inspection results, sorting decisions, implemented process adjustments, post-adjustment quality feedback, and final manual verification. Strict sample admission rules are established, marking only closed-loop data with complete process execution records and verified adjustment effects as high-confidence samples. Using these samples, the system periodically performs incremental updates or full retraining on the defect detection model, edge weights and association rules in the defect evolution graph, and constraint priorities in the process inversion model. Simultaneously, it monitors the model's performance before and after updates in key indicators such as defect identification consistency, root cause localization accuracy, and parameter tuning effectiveness, ensuring the continuous evolution and stable improvement of the system's knowledge.

[0248] The system establishes a data acquisition and closed-loop verification pipeline, automatically capturing full-process data for each production batch. For each prism workpiece, the data tuple is defined as follows: Where ID is the workpiece identification code. This is the set of structured defect state codes output from step two. The sorting decision to be executed in step four. This is the set of process parameter adjustment instructions actually executed. This is to provide quality feedback (i.e., new defect detection results) for the next batch or returned workpieces after adjustments. This pipeline provides final labels for manual review and confirmation (confirming defect category, root cause, and effectiveness of adjustments). It ensures data traceability in both time and logic.

[0249] To select high-confidence samples, a sample admission function is defined. .condition Complete process execution records are required, i.e. Each adjustment command is confirmed by a corresponding device execution log, and the parameter changes are within the safety adjustment boundaries defined in step four. The adjustment effect must be verified, i.e., there must be quality feedback. And it has been manually verified. This confirms that the defect has been suppressed or the root cause has been correctly located. Data D is labeled as a high-confidence sample and assigned a higher learning weight only when A(D)=1. Data that does not form a closed loop is assigned extremely low weight. Or temporarily exclude it.

[0250] The system employs a dual-cycle update strategy to iterate the core model. High-frequency incremental updates are triggered on a daily or shift basis, primarily targeting the edge weights in the defect detection model (such as a classification network) and the defect evolution graph.

[0251] For connecting defect nodes in the graph With process parameter nodes The weight of the attribution-related edges Update incrementally as follows:

[0252]

[0253] in, Let η be the set of high-confidence closed-loop samples within the current update period, and let η be the incremental learning rate. The confidence weights for sample k are... Describing defects in sample k With parameters The attribution association confidence is calculated. When there are no valid high-confidence samples in the current period, the edge weights remain unchanged to avoid invalid updates.

[0254] Periodic full retraining is performed monthly or quarterly to update the neural network scoring module and constraint priorities of the process inversion model. Since the true causal labels between defects and process parameters in real production environments are often difficult to observe directly, the system does not simply regard human judgments as absolute truth values. Instead, it integrates DOE experiments, small-scale controlled trial runs, historical parameter tuning feedback, equipment logs, and process expert review results into high-confidence soft labels.

[0255] For the nth high-confidence sample and the jth candidate process parameter, the soft label is defined as:

[0256]

[0257] in, For the nth sample, the available evidence sources may include expert review, DOE experiments, feedback on quality improvement after parameter tuning, and equipment anomaly logs. Let r be the anomalous support of the evidence source r for parameter j. To correspond to the weight of the evidence source, ε is a very small positive number to prevent the denominator from being zero.

[0258] As a further optimization, evidence source weighting The reliability of each evidence source in historical closed-loop verification samples can be dynamically determined. Let the verification consistency or accuracy of the r-th type of evidence source in the most recent L closed-loop samples be denoted as follows: Then it can be updated as follows:

[0259]

[0260] in, To update the coefficients, Let r be the consistency rate between evidence source r and the final closed-loop confirmation result within the current verification period. The corresponding weights can be normalized as follows:

[0261]

[0262] When different sources of evidence show significant differences in their support for the same parameter, the system calculates the degree of conflict of evidence:

[0263]

[0264] like If the learning weight of a sample exceeds a preset threshold, the learning weight of that sample will be reduced, or it will be transferred to a manual review queue.

[0265] The process inversion model uses a multi-label Sigmoid output:

[0266]

[0267] in, For the input features of the nth training sample, This indicates the prediction confidence level for parameter j as an outlier parameter.

[0268] The full training loss function is defined as the sum of the weighted multi-label binary cross-entropy and the constraint violation penalty term:

[0269]

[0270] Where N is the number of high-confidence samples, and M is the number of candidate process parameters. Let C be the sample confidence weights, and C be the set of physical constraints. Let λ represent the degree of violation of the c-th constraint, and λ be the constraint penalty coefficient. This training method can adapt to scenarios with multiple parameters experiencing abnormalities simultaneously and reduces the impact of noise from single artificial labels on model updates.

[0271] For different types of physical constraints This can be represented using a nonnegative penalty function. For inequality constraints... The penalty for violating this rule can be defined as:

[0272]

[0273] For parameter range constraints , can be defined as:

[0274]

[0275] For parameter coupling constraints , can be defined as:

[0276]

[0277] For process sequence constraints and response lag constraints, indicator function penalties or time difference penalties can be used. These penalty functions make the model training process tend to output candidate parameter combinations that conform to prior process knowledge.

[0278] After each model or map update, the system performs online gating evaluation on the reserved validation set and the most recent closed-loop sample set. Root cause localization accuracy is defined as:

[0279]

[0280] in, The set of outlier parameters predicted by the model. This is the set of valid parameters confirmed by closed-loop verification. The parameter tuning effectiveness rate is defined as:

[0281]

[0282] in, The number of samples to perform parameter tuning. This is for quality feedback after parameter tuning. The new model or new map is only allowed to be deployed to the production environment if the root cause localization accuracy, parameter tuning effectiveness, defect identification consistency, and performance difference before and after the update all meet the online threshold, and the performance of the updated model is not lower than 95% of the performance of the model before the update; otherwise, the old version is retained, and the updated samples are transferred to the manual review queue.

[0283] Compared with existing technologies, the solution provided in this embodiment adopts multi-angle polarization imaging technology, which can effectively suppress environmental reflection interference and accurately identify optically sensitive defects that are difficult to detect by traditional machine vision, such as uneven stress, minor scratches, and coating abnormalities, thus significantly improving the defect identification accuracy.

[0284] A defect evolution map containing a ternary attribution hypothesis was constructed. Combined with a physical constraint process inversion model, the root cause localization from data-driven to physical constraint was realized. The output root cause results have both statistical relevance and physical executability, and the efficiency of process adjustment is significantly improved.

[0285] A complete quantitative theoretical parameter tuning generation mechanism was proposed. By fusing historical closed-loop samples with the process sensitivity matrix, the problem that traditional methods can only output qualitative deviation direction was solved, and a complete automatic parameter tuning closed loop was formed.

[0286] A three-dimensional classification model based on defect size risk, optical performance risk, and product rework repairability was established, classifying products into five categories, enabling refined sorting and disposal, and effectively reducing scrap rate;

[0287] By adopting a soft labeling mechanism based on multi-source evidence fusion and a dual-cycle update strategy, the problem of unavailability of absolute causal labels in industrial scenarios is solved. At the same time, through a strict online gating mechanism, performance degradation caused by model mislearning is avoided, which helps to improve the overall yield rate of prism production in the long term.

[0288] In another implementation, the system includes a data acquisition module, a defect detection and structuring module, a map construction and process inversion module, a sorting and parameter tuning generation module, and a closed-loop update module. The data acquisition module acquires multi-angle polarization images of the prism and simultaneously obtains and stores identity information, process information, equipment information, and real-time process parameters. The defect detection and structuring module performs polarization feature decomposition and defect detection, outputting structured defect features containing polarization response characteristics and optical path position information. The map construction and process inversion module constructs a defect evolution map containing attribution assumptions between defects, processes, and process parameters, retrieves similar defect patterns, and outputs root cause analysis results in conjunction with process physical constraints. The sorting and parameter tuning generation module performs multi-dimensional risk classification based on structured defect features and root cause analysis results, generates sorting decisions, generates theoretical parameter adjustment amounts, and outputs multi-step restricted adjustment instructions that comply with equipment safety constraints when trigger conditions are met. The closed-loop update module acquires full-process data to generate closed-loop verification samples and updates at least one of the defect detection model, defect evolution map, and process inversion model.

[0289] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0290] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0291] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A prism detection sorting quality control method based on defect map and process inversion, characterized in that, include: Multi-angle polarization images of the prism to be tested are acquired, and the identity information, batch information, process information, equipment information and real-time process parameters corresponding to the prism to be tested are obtained simultaneously. The two are associated and stored to form traceable raw data. The polarization image is decomposed into polarization features, and a deep learning detection model is used to identify defect regions and extract structured defect features including polarization response characteristics and the location of the optical path sensitive area. A defect evolution map is constructed based on the structured defect features and process context data of historical batches. The defect evolution map includes the attribution hypothesis relationship between defects, processes and process parameters. When a new batch of defects occurs, similar defect patterns are retrieved from the defect evolution map to generate candidate process deviation hypotheses. After screening and sorting by process physical constraints, the root cause analysis results are obtained. Based on the structural defect characteristics and root cause analysis results, the prism is classified into multiple dimensions for risk assessment, and corresponding sorting decisions are generated. Based on the root cause analysis results, a theoretical parameter adjustment amount is generated by combining at least one of the verified historical adjustment data and process parameter response relationships. When the process adjustment triggering conditions are met, the theoretical parameter adjustment amount is decomposed into multi-step restricted adjustment instructions according to equipment safety constraints. Collect sorting decision-making, process adjustment execution records, adjusted quality feedback data and review results to form a closed-loop verification sample. Based on the closed-loop verification sample, update at least one of the defect detection model, defect evolution map and process inversion model.

2. The prism detection and sorting quality control method based on defect map and process inversion according to claim 1, characterized in that: The polarization feature decomposition includes calculating the Stokes vector based on multi-polarization images, and further obtaining the degree of linear polarization and the polarization angle; the polarization angle is calculated using the four-quadrant arctangent function; when acquiring circularly polarized images, the complete degree of polarization is further calculated.

3. The prism detection and sorting quality control method based on defect map and process inversion according to claim 1, characterized in that: The deep learning detection model has a polarization feature attention module in its neck section. The polarization feature attention module performs global average pooling on multiple polarization feature channels to obtain channel description vectors, generates channel importance scores through a multilayer perceptron, obtains the attention weights of each channel through normalization, and finally outputs multi-channel weighted fused features.

4. The prism inspection and sorting quality control method based on defect maps and process inversion according to claim 1, characterized in that: The structured defect features also include the defect's basic geometric parameters, defect edge morphological parameters, and cross-surface defect co-occurrence relationship parameters.

5. The prism inspection and sorting quality control method based on defect maps and process inversion according to claim 1, characterized in that: The defect evolution map also includes equipment nodes; the attribution hypothesis relationship is a ternary association relationship consisting of defect nodes, process nodes and process parameter nodes, and its association weight is calculated by fusing co-occurrence probability, normalized conditional mutual information, historical parameter tuning feedback confidence and equipment status confidence.

6. The prism inspection and sorting quality control method based on defect maps and process inversion according to claim 1, characterized in that: The physical constraints of the process include process sequence constraints, equipment adjustable range constraints, parameter coupling constraints, and response lag constraints. The system first filters out infeasible candidate process deviation assumptions through hard constraints, and then outputs the root cause analysis results after comprehensively ranking the remaining feasible assumptions.

7. The prism inspection and sorting quality control method based on defect maps and process inversion according to claim 1, characterized in that: The quantitative theoretical parameter adjustment amount is dynamically generated by the fusion of the historical closed-loop sample weighting method and the process sensitivity matrix method. Before fusion, the adjustment amount obtained by the two methods is checked for directional consistency. If the two directions conflict and the confidence levels are close, the automatic parameter adjustment command is not output and the process is transferred to manual review.

8. The prism inspection and sorting quality control method based on defect maps and process inversion according to claim 1, characterized in that: The multidimensional risk classification includes three dimensions: defect size risk level, optical performance risk level, and product rework repairability level. The multi-step restricted adjustment instruction is generated by iterative optimization of the remaining adjustment amount, and the constraints include the maximum single adjustment amount of the parameter, the parameter safety boundary, and the parameter linkage rules.

9. The prism inspection and sorting quality control method based on defect maps and process inversion according to claim 1, characterized in that: Establish sample admission rules to screen high-confidence closed-loop verification samples with complete process records and verified adjustment effects; the samples are generated by dynamic normalization and weighting of multi-source evidence, and the weight of each evidence source is dynamically updated according to its historical verification reliability. The model and graph are updated using a combination of incremental updates and full retraining, and an online performance gating mechanism is set up.

10. A prism inspection and sorting quality control system based on defect maps and process inversion, characterized in that, include: The data acquisition module is used to acquire polarization images of the prism and simultaneously obtain and store the corresponding identity information, process information, equipment information and real-time process parameters. The defect detection and structuring module is used to perform polarization feature decomposition and defect detection, and outputs structured defect features containing polarization response characteristics and optical path position information. The map construction and process inversion module is used to construct a defect evolution map that includes attribution assumptions between defects, processes, and process parameters, retrieve similar defect patterns, and output root cause analysis results in combination with process physical constraints. The sorting and parameter adjustment generation module is used to perform multi-dimensional risk classification and generate sorting decisions based on the characteristics of structured defects and the results of root cause analysis, as well as generate quantitative theoretical parameter adjustment amounts and output multi-step restricted adjustment instructions that comply with equipment safety constraints when trigger conditions are met. The closed-loop update module is used to collect data from the entire process to generate closed-loop verification samples and update at least one of the defect detection model, defect evolution map, and process inversion model.

Images