Panel detection data processing method and apparatus

Abstract

The embodiment of the application provides a panel detection data processing method and device, the method comprises the following steps: generating a composite defect image containing sample distribution, model error, defect mechanism and rule candidate set by performing multi-layer analysis of data layer, model layer, mechanism layer and rule layer on the data; matching the composite defect image with component images in a preset component library; screening out a candidate rule component set and a model component set, and binding them with a draft gating relationship to construct an executable gating map; solidifying the gating map by conflict processing; determining a corresponding standardized panel detection collaborative unit by compiling; and performing a standardized panel detection process according to the panel detection collaborative unit.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a panel inspection data processing method and apparatus. Background Technology

[0002] In the field of panel inspection, especially in automated optical inspection (AOI) scenarios targeting defects such as blemishes, bright spots, dark spots, line defects, stains, and scratches, the construction of detection algorithms has long faced the challenge of coordinating rule-based methods and deep learning methods.

[0003] Rule-based methods, based on deterministic criteria such as grayscale thresholds, geometric morphology, frequency domain analysis, and edge distance, are suitable for defect types with well-defined mechanisms and stable characteristics. However, when faced with complex texture backgrounds, diverse defect morphologies, and domain shifts caused by process variations, pure rule-based methods often struggle to cover long-tailed distributions.

[0004] Deep learning methods, by constructing detection, segmentation, classification, or anomaly detection networks, can automatically learn high-dimensional semantic features from data, exhibiting stronger generalization ability and recall performance for complex defects. However, these model-based methods suffer from problems such as weak interpretability, sensitivity to annotation quality, and confidence distortion on boundary samples. Furthermore, in scenarios with tight production line schedules or limited computing power, relying solely on models is insufficient to meet real-time requirements.

[0005] In current industrial practice, the combination of rules and models often adopts a loose, sequential approach, such as performing rule screening before model inference, or adding a small number of rules for filtering after the model output. This ad-hoc solution lacks a systematic methodology, resulting in vastly different implementations of the same detection task by different engineers. These implementations are difficult to reuse, transfer, or form a standardized audit loop. Furthermore, when false alarms occur, it is difficult to quickly pinpoint whether the cause is a rule design flaw, model confusion, or unreasonable collaboration logic.

[0006] Therefore, there is an urgent need for a panel inspection data processing method that can improve the accuracy and precision of panel inspection tasks. Summary of the Invention

[0007] To address the problems in the prior art, this application provides a panel inspection data processing method and apparatus, which can improve the precision and accuracy of panel inspection tasks.

[0008] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides a panel detection data processing method, including: Acquire panel detection data, which includes panel detection task data, panel detection sample set, and corresponding detection metadata. The panel detection data is parsed at the data layer to determine the corresponding panel detection data profile. Based on the constraints of the panel detection data profile, the existing model is parsed at the model layer to determine the corresponding model error profile. The model error profile and the preset process mechanism are combined to perform mechanism layer parsing to determine the corresponding defect mechanism profile. Regularizable criteria are extracted from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft. Based on the panel detection data profile, the model error profile, the defect mechanism profile and the rule candidate set, the corresponding composite defect profile is determined. The composite defect profile is matched with the component profiles in the preset component library to determine the corresponding candidate rule component set and candidate model component set. The candidate rule component set and the candidate model component set are bound to the gating relationship draft to determine the corresponding executable gating graph. The executable gating graph is subjected to conflict handling and solidification operations to determine the corresponding panel detection collaboration unit, and panel detection is performed according to the panel detection collaboration unit.

[0009] Further, the step of performing data layer parsing on the panel detection data to determine the corresponding panel detection data profile includes: Data layer parsing is performed on the panel detection data to statistically analyze the distribution pattern of defect samples in the panel detection sample set, including the long-tail distribution structure and the set of difficult cases, and to determine the corresponding sample distribution profile. Alignment analysis is performed on the feature distribution of samples under different acquisition conditions to identify the source of domain offset and determine the corresponding data drift report; For samples that have historically generated false alarms, cluster them according to texture similarity, lighting artifact similarity, or splicing artifact similarity to determine the corresponding model false alarm cluster dictionary; Based on the sample distribution profile, the data drift report, and the model false alarm cluster dictionary, the corresponding panel detection data profile is determined.

[0010] Furthermore, the step of performing model-level parsing on the existing model based on the constraints of the panel detection data profile to determine the corresponding model error profile includes: A weighted sampling operation is performed based on the sample distribution profile, the data drift report, and the model false alarm cluster dictionary to determine the corresponding validation sample set; Obtain the original output of the existing model on the validation sample set, and compare the original output with the preset true value label to determine the corresponding error bucket. The error bucket includes missed detection error, false alarm error, positioning offset error, boundary instability error and category confusion error. The error buckets and the false alarm cluster dictionary are associated with spatial location and feature dimension, and confidence statistics are performed on each error bucket based on the results of the association and attribution to determine the corresponding model error profile.

[0011] Furthermore, the step of combining the model error profile and the preset process mechanism to perform mechanistic layer analysis and determine the corresponding defect mechanism profile includes: Based on the model error profile, the error binning results are mapped to a preset process mechanism to determine the corresponding mechanism domain and observable channel; For each of the aforementioned mechanistic domains, at least one set of interpretable features is provided to determine the corresponding defect mechanism profile.

[0012] Furthermore, the step of extracting ruleable criteria from the defect mechanism profile and determining the corresponding rule candidate set and gating relationship draft includes: From the interpretable features of the defect mechanism profile, deterministic computational criteria are extracted to form a set of ruleable criteria. The value of each ruleable criterion in the set is evaluated to determine the corresponding rule value score. The value evaluation includes evaluating the computational complexity of each criterion, its ability to suppress false alarm clusters, and its impact on the risk of missed detection. Rule selection is performed based on the rule value score to determine the corresponding rule candidate set, which is used to provide prior criteria for the model. Based on the mechanistic domain classification of each defect in the defect mechanism profile, and combined with the long-tail distribution structure of the defects and the error confidence statistics of the model, a draft gating relationship for the defect is determined, including at least one of rule-first gating, model-first gating, parallel execution gating, and two-way mutual verification gating.

[0013] Further, the step of binding the candidate rule component set and the candidate model component set with the gating relationship draft to determine the corresponding executable gating graph includes: Obtain the draft gating relationship, including gating type, trigger condition description, and data routing template; Map the component category nodes in the gating relationship draft to specific component instances in the candidate rule component set or the candidate model component set; The triggering conditions are compiled into an executable predicate expression, which is used to determine the selection of the execution path and the triggering of downstream nodes; Based on the data routing template and the compilation result of the predicate expression, the data flow and dependencies between nodes are obtained, and the corresponding executable gating graph is determined.

[0014] Further, the conflict resolution and solidification operation on the executable gating graph to determine the corresponding panel detection coordination unit includes: Configure a conflict determination predicate set and a policy path set for at least one conflict node in the executable gating graph; The conflict determination predicate set includes at least one of category conflict predicates, existence conflict predicates, location conflict predicates, confidence conflict predicates, and morphological conflict predicates; The set of strategy paths includes at least one of the following: rule priority path, model priority path, parallel scoring fusion path, rollback anomaly detection path, and path to enter the manual review pool. The conflict determination predicate set and strategy path set are bound to the edges, nodes and data routes in the executable gating graph to determine the corresponding panel detection collaboration unit.

[0015] Secondly, this application provides a panel detection data processing apparatus, comprising: The panel detection data acquisition module is used to acquire panel detection data, which includes panel detection task data, panel detection sample set and corresponding detection metadata. The composite defect profile determination module is used to perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile; based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile; combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile; extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft; and determine the corresponding composite defect profile based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set. The detection collaboration unit determination module is used to match the composite defect profile with the component profiles in the preset component library, determine the corresponding candidate rule component set and candidate model component set, bind the candidate rule component set and the candidate model component set with the gating relationship draft, determine the corresponding executable gating graph, perform conflict handling and solidification operations on the executable gating graph, determine the corresponding panel detection collaboration unit, and perform panel detection according to the panel detection collaboration unit.

[0016] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the panel detection data processing method.

[0017] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the panel detection data processing method described above.

[0018] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the panel detection data processing method described above.

[0019] As can be seen from the above technical solution, this application provides a panel inspection data processing method and apparatus. By performing multi-layer analysis of the data layer, model layer, mechanism layer and rule layer, a composite defect profile is generated, which includes sample distribution, model error, defect mechanism and rule candidate set. The composite defect profile is matched with the component profiles in the preset component library to select candidate rule component set and model component set, and bind them with the gating relationship draft to construct an executable gating graph. The gating graph is solidified by conflict handling, and the corresponding standardized panel inspection collaborative unit is compiled and determined. The standardized panel inspection process is performed according to the panel inspection collaborative unit, thereby improving the accuracy and precision of the panel inspection task. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts illustrating the panel detection data processing method in the embodiments of this application; Figure 2 This is a structural diagram of the panel detection data processing device in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0022] Figure label: Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0025] In the field of panel detection, the long-standing challenge of synergistic collaboration between rule-based methods and deep learning methods has been a persistent problem. This application provides a panel detection data processing method and apparatus. By performing multi-layered analysis of the data (data layer, model layer, mechanism layer, and rule layer), a composite defect profile is generated, including sample distribution, model error, defect mechanism, and a set of candidate rules. This composite defect profile is then matched with component profiles in a pre-defined component library to select candidate rule component sets and model component sets. These are then bound to a draft gating relationship to construct an executable gating graph. Conflict resolution and solidification of the gating graph are performed, and corresponding standardized panel detection collaborative units are compiled and determined. A standardized panel detection process is then executed based on these collaborative units, thereby improving the accuracy and precision of panel detection tasks.

[0026] To improve the accuracy and precision of panel inspection tasks, this application provides an embodiment of a panel inspection data processing method, see [link to embodiment]. Figure 1 The panel detection data processing method specifically includes the following: Step S101: Obtain panel detection data, which includes panel detection task data, panel detection sample set, and corresponding detection metadata; Optionally, in this embodiment, a background introduction is given first.

[0027] In panel detection tasks, traditional rule-based methods have advantages in interpretability, low-sample tasks, and hard constraint application, while deep learning methods excel in complex textures and high-dimensional feature modeling. In current practice, these two approaches are often pieced together in a fragmented manner: either rule screening is performed before input into the model, or a small number of rules are added after the model output. This weakly coupled approach fails to systematically express the gating relationship, conflict handling strategies, and validation loop between the two approaches. This results in highly inconsistent ad-hoc solutions for the same task being handled by different engineers, making them difficult to reuse and transfer. To coordinate rule-based and deep learning methods and standardize and streamline panel detection tasks, this method elevates the collaboration between rules and deep learning from fragmented calls to a standard process that is explicitly compiled during the build phase, resulting in a panel detection skill (unified detection unit) to form a standardized, deployable, and auditable collaborative detection capability.

[0028] In the process of developing panel detection skills, the first step is to collect all the data involved in the panel detection process for data analysis, including but not limited to the following data: Panel inspection task data defines the business objectives and technical constraints of the inspection activities, specifically including target defect types (such as murmurs, bright spots, dark spots, line defects, etc.), the customer's tolerance for false positives and false negatives, and production line cycle time requirements. The introduction of task data ensures that all subsequent analysis and orchestration work revolves around these clear business objectives.

[0029] Panel detection sample set: includes positive samples (images containing various defects and pixel-level or border-level annotations), negative samples (clean panel images without defects), and specially selected boundary samples and typical false positive samples. The composition of the sample set directly affects the effectiveness of subsequent model training, rule threshold setting, and playback verification.

[0030] Detection metadata records environmental and equipment status information during sample acquisition, such as machine number, light source type and attenuation status, exposure gain, scanning stitching strategy, and panel temperature. The core function of metadata is to help the system distinguish between "real defect signals" and "artifacts caused by changes in imaging conditions," thereby improving the robustness of detection.

[0031] By comprehensively collecting the above data, each detection sample not only carries image pixel information, but also a complete semantic description (task objective) and physical background (collection conditions). This provides a unified data entry point for subsequent mechanism analysis, data distribution statistics, model error analysis, and rule extraction, and ensures that every step of the entire detection chain can be traced back to the original input information and collection environment, meeting the auditability requirements of industrial scenarios.

[0032] Step S102: Perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile. Based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile. Combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile. Extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft. Based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set, determine the corresponding composite defect profile. Optionally, in this embodiment, this step performs statistical analysis on all the data collected in step S101 above and performs model false alarm decomposition and defect mechanism analysis on the defects, so as to summarize all panel detection data into a composite defect profile.

[0033] First, perform data layer parsing to quantify all the data, including "what types of data are there, how are they distributed, where is the noise and drift, and what is the structure of negative samples / false alarms." Subsequent thresholds, risks, computing power budgets, and training sampling all depend on these statistics.

[0034] Second, perform model layer analysis, train / evaluate baseline models or read existing models under the constraints of data profiling, and perform error decomposition (missed detection / false alarm / unstable boundary / class confusion / OOD). Model errors are the most direct "evidence of problems", which can bring the mechanism analysis from generalities to specific channels and specific false alarm clusters.

[0035] Third, perform mechanism layer analysis, combine process segment, material structure, imaging link and model error cluster to establish defect mechanism profile (mechanism domain → observable channel → interpretable feature → controllable variable). Mechanism profile provides a more stable causal explanation and interpretable feature set across batches, which is the basis for rule regularization and gating interpretability.

[0036] Fourth, perform rule-based analysis, extract ruleable criteria from interpretable features at the mechanistic level and false alarm / false alarm clusters in the model, form a rule candidate set and perform cost-benefit evaluation. The rules should "carry stable priors and low-cost suppression", and their design must be based on computable features, threshold ranges, applicable conditions and evidence package fields.

[0037] Specifically, the interaction mechanism of multi-level analysis is shown in the table below:

[0038] The complete data processing flow is as follows: S1 Input: Task metadata (product / process section / acquisition scheme / machine / optical parameters / ROI definition) + initial sample set (including positive and negative samples and annotations) + component library (rule library / model library); S2 Multi-layer parsing: Sequentially outputs DataProfile / FPClusterDict / DriftReport →ModelErrorProfile → MechanismProfile → RuleCandidateSet, and generates a traceable "structured knowledge package K"; S3 Component Matching: Calculate the matching score using the 'defect profile vector' in K and the 'component profile vector' in the component library to obtain the candidate rule component set and the candidate model component set; at the same time, instantiate the gating draft of S2 into a candidate gating graph (DAG). S4 Collaborative Orchestration: Solidifies triggering conditions (predicates), execution order, and data routing on the candidate gating graph; configures conflict determination predicates and strategy paths (rule priority / model priority / fusion / rollback / artificial pool); solidifies evidence package fields and disk placement locations; S5 compilation and delivery: Compile 'component version + gating graph + threshold / weight + strategy set + evidence package specification' into a Skill executable package (recipe file + plugin / model reference + verification script).

[0039] Online execution (runtime): Input single / batch image + metadata → Lightweight features → Gated predicate path selection → Execute rules / models by path → Fusion and conflict handling → Output detection results + evidence package (including gated logs / intermediate features / visualization).

[0040] Steps S3-S5 are specific implementations of step S103 in this embodiment, and will be described in detail in step S103.

[0041] Optionally, in this embodiment, this step details the structure of the structured knowledge package K: Data layer output: DataProfile (distribution / long tail / drift / noise) + FPClusterDict (false alarm cluster) + RiskList (risk list).

[0042] Model layer output: ModelErrorProfile (error bucket / gray area threshold / OOD statistics / interpretive heatmap summary).

[0043] Mechanism layer output: MechanismProfile (mechanistic domain / observable channel / interpretable feature / controllable variable).

[0044] Rule layer output: RuleCandidateSet (rule profile / threshold range / applicable conditions / cost / evidence fields) in: The data layer determines whether the threshold is valid (distribution / drift / noise) and which are long-tail risk points, providing a baseline constraint for all subsequent analysis; The model layer determines "on which clusters the model fails, where the confidence gray area is, and whether OOD / rollback is needed," and systematically evaluates the model's failure modes by utilizing the distribution characteristics revealed in the data profile. The mechanism layer determines "which channels / features are interpretable and stable" and combines the failure modes reflected in the model error profile with the pre-set process mechanism knowledge to upgrade the defects from "name labels" to "causally interpretable physical descriptions". The rule layer determines "which priors can be implemented at low cost and whether they can significantly reduce false alarm clusters or improve recall," transforming the interpretable features in the mechanism profile into a series of executable and low-cost logical criteria.

[0045] Optionally, in this embodiment, this step integrates the intermediate products obtained from all the above-mentioned hierarchical analysis to obtain a composite defect profile, including a mechanism profile, a data profile, a model error profile, and a rule candidate profile, with risk and threshold baselines. Key constraints such as data 'long tail / drift / false alarm clusters / gray area confidence' are included in the selection, so as to support subsequent automatic matching and gating orchestration.

[0046] Step S103: Match the composite defect profile with the component profiles in the preset component library to determine the corresponding candidate rule component set and candidate model component set. Bind the candidate rule component set and the candidate model component set to the gating relationship draft to determine the corresponding executable gating graph. Perform conflict handling and solidification operations on the executable gating graph to determine the corresponding panel detection collaboration unit and perform panel detection according to the panel detection collaboration unit.

[0047] Optionally, in this embodiment, after obtaining the composite defect profile, the system matches it with the profiles corresponding to each component in the preset component library.

[0048] The composite defect profile encompasses the defect's mechanistic characteristics, data distribution characteristics, model error characteristics, and rule candidate characteristics; the component profile describes the applicable conditions, computational cost, stability, and traceability of each rule component or model component. By calculating the matching score between the two, the system automatically selects the most suitable candidate rule component set and candidate model component set for the current task.

[0049] The selected candidate component set is then bound and instantiated with the previously generated gating relationship draft. The gating relationship draft only contains high-level gating intents (such as rule-first or parallel verification), while the binding operation replaces the abstract nodes in the draft with concrete component implementations and compiles the triggering conditions into executable predicate expressions. This process combines data routing, execution order, and latency budget constraints, ultimately generating a directed acyclic graph, i.e., an executable gating graph. This graph precisely defines the invocation timing, data flow, and path selection logic for each component.

[0050] The executable gating graph undergoes a standardized conflict handling process. This includes configuring specific processing strategies for each conflict resolution path (e.g., when rules and model outputs are inconsistent), such as rule priority, model priority, parallel scoring fusion, rollback anomaly detection, or manual review. The applicable conditions and trigger thresholds for each strategy are clearly defined. After standardization, all components, gating logic, conflict strategies, and evidence package specifications are encapsulated into a standardized panel detection collaborative unit. During panel detection, this collaborative unit receives input data, automatically selects a path according to the gating graph, outputs detection results along with a complete and auditable evidence package, thereby achieving structured, reusable, and traceable detection through rule-based and deep learning collaboration.

[0051] Optionally, in this embodiment, this step achieves the goal of clearly recording the role of each rule component and model component during the construction period, whether it is a pre-gating, parallel bypass, or post-verification, and writing it into the deployable description file, which can avoid engineers being unable to understand the original collaborative relationship after replacement.

[0052] A specific example is provided for illustration: Example 1: Selection of Example for Rule Component and Model Component.

[0053] For Mura detection, the rule components preferably include low-frequency uniformity analysis, regional brightness difference rules, and edge transition rules; the model components preferably include a region segmentation network or anomaly detection network.

[0054] For bright spot and dark spot tasks, the rule component can emphasize brightness threshold and area constraints, while the model component can use a small object detection network.

[0055] For line defect tasks, the rule component can emphasize directional consistency and fine length constraints, while the model component can use a fine line segmentation network.

[0056] This modular approach allows different tasks to share the same basic components, but their orchestration logic remains differentiated.

[0057] Example 2: Gating and Conflict Handling Example.

[0058] If a rule has a strong response but the model has low confidence, the system can prioritize retaining the rule or send it to the manual review pool based on the task risk. If the model has high confidence but the rule response is weak, the model result can be output first, but supplementary rule verification can be triggered at the same time to prevent misjudgment caused by model overfitting.

[0059] When both the rules and the model are strong, the system directly outputs a highly reliable conclusion and records the basis for the fusion. When both are weak, the process reverts to anomaly detection or template difference branch and includes the sample in the sample pool to be expanded.

[0060] This implementation transforms the rules and models from simple sequential connections into a configurable collaborative decision network.

[0061] This example demonstrates how this embodiment filters relevant component sets through defect profiling and constructs panel detection collaborative units by defining gating relationships to achieve standardized panel detection.

[0062] As can be seen from the above description, the panel inspection data processing method provided in this application can generate a composite defect profile containing sample distribution, model error, defect mechanism, and rule candidate set by performing multi-layer analysis of data, model, mechanism, and rule layers on the data. The composite defect profile is then matched with component profiles in a preset component library to filter out candidate rule component sets and model component sets, which are then bound to a gating relationship draft to construct an executable gating graph. The gating graph is then solidified by conflict handling, and the corresponding standardized panel inspection collaborative unit is compiled and determined. The standardized panel inspection process is then performed according to the panel inspection collaborative unit, thereby improving the accuracy and precision of the panel inspection task.

[0063] In one embodiment of the panel detection data processing method of this application, it may further include the following: Step S201: Perform data layer parsing on the panel detection data, statistically analyze the distribution pattern of defect samples in the panel detection sample set, including the long-tail distribution structure and the set of difficult cases, and determine the corresponding sample distribution profile; Step S202: Perform alignment analysis on the sample feature distribution under different acquisition conditions, identify the source of domain offset, and determine the corresponding data drift report; Step S203: For samples that have historically generated model false alarms, cluster them according to texture similarity, lighting artifact similarity, or splicing artifact similarity to determine the corresponding model false alarm cluster dictionary; Step S204: Determine the corresponding panel detection data profile based on the sample distribution profile, the data drift report, and the model false alarm cluster dictionary.

[0064] Optionally, in this embodiment, this step is the process of parsing the data layer to obtain the panel detection data profile.

[0065] Specifically, data layer parsing includes: 1. Panel inspection data sample distribution statistics: The data is statistically distributed according to dimensions such as defect type, level, location area, size scale, brightness difference, frequency bandwidth, and directionality, and outputs a long-tail structure and difficult case set.

[0066] The system focuses on identifying two key structural types: first, long-tailed distribution structures, where certain defect categories have extremely few samples and are located at the tail end of the statistical distribution; and second, difficult-case sets, which are samples that are easily confused with other defects or are on the decision boundary. Based on this statistical information, the system generates a sample distribution profile, which records the proportion of different defect categories, the characteristics of difficult cases, and their distribution locations, providing data for subsequent component selection.

[0067] 2. Data drift analysis: Perform alignment analysis on feature distributions under different batches, different equipment, and different optical conditions to identify the source of domain offset and label transferable and non-transferable features.

[0068] In actual production, panel inspection data may come from different machines, batches, or optical parameter settings. These differences can cause the same defect to exhibit different characteristic distributions under different conditions. The system calculates the degree of difference in characteristic distributions under each condition to identify which differences are normal batch fluctuations and which are domain shifts requiring attention. The analysis results are output in the form of a data drift report, which indicates the feature dimensions with severe drift and their possible sources.

[0069] 3. Negative sample structuring: False alarm samples are clustered according to similarity in texture, lighting artifacts, splicing artifacts, and grayscale quantization noise to form false alarm clusters.

[0070] The system focuses on handling historical model false positives. It performs cluster analysis on past model false positives based on feature similarity: texture similarity clustering groups false positives with similar background and defect textures into one category; lighting artifact similarity clustering groups false positives caused by uneven lighting, reflections, etc., into another category; and stitching artifact similarity clustering groups false positives caused by artifacts at the stitching points of the scan into another category. The clustering results form a model false positive cluster dictionary, with each cluster recording typical samples, common features, and triggering conditions.

[0071] 4. Annotation consistency and noise assessment: Statistically determine the consistency rate among multiple annotators, locate classes with ambiguous boundaries and classes with inconsistent definitions, and use them as candidates for manual review pool in the conflict strategy.

[0072] By fusing the above outputs, a unified panel detection data profile is generated. It can be understood that the panel detection data profile output in this step quantifies the data thresholds / risks / long tails / drifts, which are used to drive subsequent gating thresholds, artificial pool conditions, backoff strategies, and training sampling. At the same time, the false alarm cluster dictionary transforms false alarms from 'scattered points' into 'structured clusters', guiding the triggering predicates of rule suppression, model retraining, and conflict strategies.

[0073] Examples are provided for illustration: The data layer parsing input is: labeled samples (positive / negative), level, position (by panel partition / ROI), scale / contrast / frequency band and other feature statistics + collected metadata (machine / batch / exposure / light source life / spanning method / temperature, etc.); The data layer parsing output consists of distributed objects that can be directly consumed by the system, including: ① category / level proportion; ② long-tail difficult case set; ③ location distribution (which areas are more prone to false alarms / false negatives); ④ scale / contrast / frequency band / direction distribution; ⑤ cross-batch drift index (distribution difference); ⑥ annotation noise index (multi-annotation consistency rate / boundary ambiguity class). The impact of the panel detection data profile obtained after parsing on subsequent steps is as follows: 1. Impact on component selection, based on the analysis results: Sufficient and stable sample size → Tendency to choose supervised segmentation / detection models as the primary approach; Severe long-tail defects or frequent new types of defects → Tendency to introduce anomaly detection models or rule-based fallbacks; In certain locations / operating conditions where false alarm clusters are concentrated, the system tends to select rule components that can suppress the cluster (such as edge masking and transition width constraints).

[0074] 2. Impact on gating and threshold, based on the analysis results: Different gating thresholds (regionalized thresholds) are set based on the prior probabilities of different regions / operating conditions to reduce systematic false alarms; The decision to enable two-way mutual verification gating and to enter the manual review pool is based on the long tail and risk level. The decision to enable the "rollback anomaly detection / degradation strategy" is based on the drift metric, and a trigger threshold is set (e.g., distribution variability exceeding the limit).

[0075] 3. Driving training and iteration (driving model layer closure): Training sampling: Reweight / resample according to the distribution to avoid the model only learning the head category; Difficult example mining: Long-tail difficult examples and false positive clusters are used as a continuous feedback set, and the actual benefits of gating / rules / models are evaluated periodically through replay. Calibration: Use 'gray area samples' from the distribution profile to perform confidence calibration and stabilize the online threshold. It is understandable that this data layer parsing, as the foundation of all parsing, has the following coupling points with subsequent parsing: Data → Mechanism: Drift and clustering results suggest possible changes in optical / process controllable variables (e.g., low-frequency non-uniformity caused by light source attenuation); the mechanism layer supplements the "sensitive variable - observation channel" mapping accordingly.

[0076] Data → Model: Distribution determines the model training strategy (sampling / augmentation / domain adaptation); drift metrics determine whether to introduce OOD detection and model rollback.

[0077] Data → Rules: The false positive cluster dictionary provides 'targets' for rules. The rule threshold range is derived from distribution statistics (rather than experience) and can be adaptively updated with each batch (recorded in the evidence package).

[0078] Through step S204, this embodiment successfully integrates sample distribution patterns, domain offset information, and false alarm cluster structure to form a structured data object that can be directly consumed by downstream components through matching and gating configuration.

[0079] In one embodiment of the panel detection data processing method of this application, it may further include the following: Step S301: Perform weighted sampling based on the sample distribution profile, the data drift report, and the model false alarm cluster dictionary to determine the corresponding validation sample set; Step S302: Obtain the original output of the existing model on the validation sample set, and compare the original output with the preset true value label to determine the corresponding error bucket. The error bucket includes missed detection error, false alarm error, positioning offset error, boundary instability error and category confusion error. Step S303: Associate the spatial location and feature dimension of the error buckets and the false alarm cluster dictionary, and perform confidence statistics on each error bucket based on the results of the association attribution to determine the corresponding model error profile.

[0080] Optionally, in this embodiment, this step is the process of model layer parsing to obtain a model error profile.

[0081] Specifically, model layer parsing requires decomposing the detection errors generated by existing models to obtain error types, confidence features, and corresponding interpretability heatmaps, which can be used to guide rule design and gating strategy configuration.

[0082] First, the data obtained through data layer parsing provides a sample distribution profile that shows the proportion of various defective samples in the entire dataset, including common head categories and rare tail categories. The data drift report identifies groups where the sample feature distribution has significantly shifted under different batches, machines, or lighting conditions. The model false positive cluster dictionary records the sample types and their common characteristics that have historically generated frequent misjudgments by the existing model.

[0083] Based on these three inputs, the system performs a weighted sampling operation. For long-tail category samples that account for a very small proportion in the sample distribution profile, a higher sampling weight is assigned, forcing the validation set to include a sufficient number of hard examples. For sample groups marked as having strong drift in the data drift report, the probability of their selection is also increased to evaluate the model's performance on out-of-domain data. For typical false alarm samples recorded in the false alarm cluster dictionary, it is ensured that they enter the validation set at a high proportion.

[0084] Through the above operations, a validation sample set is extracted, making it a challenging set that focuses on the model's weaknesses. This set can more sensitively reflect the model's true performance in edge scenarios, long-tail categories, and domain offset conditions, providing a high-quality data foundation for subsequent error analysis.

[0085] Next, a specific error quantification analysis is performed to obtain all the original output results of the existing model on the validation sample set, including detection boxes, segmentation masks, class labels, and confidence scores. Subsequently, these model outputs are precisely compared with preset ground truth labels that have been manually verified, on a sample-by-sample, pixel-by-pixel, or region-by-region basis.

[0086] The resulting deviations are finely decomposed into five specialized error buckets, including: The missed detection error bucket records the actual defect samples that the model failed to detect; The false alarm error bucket records spurious defect samples in the model's erroneous output; The positioning offset error bucket records samples where the model detects defects but the position deviates from the true value by a preset threshold. The boundary instability error bucket records samples where the defect boundary output of the model has a low degree of agreement with the true boundary, which usually manifests as jagged edges of the segmentation mask or excessive shrinkage and expansion. The category confusion error bucket records samples where the model detects the location of a defect but incorrectly identifies the type of the defect.

[0087] By using a binning design for model errors, model failure modes are refined from the general metric of "accuracy" into actionable engineering problems. For example, there is often a trade-off between missed detections and false positives; localization offset may point to problems with the model's receptive field design; boundary instability may indicate insufficient edge constraints in the loss function; and class confusion may reflect excessively high feature similarity between classes. This fine-grained decomposition provides precise targets for subsequent attribution analysis and targeted optimization.

[0088] Next, the samples in each error bucket are correlated with the false alarm cluster dictionary obtained from the data layer through a two-way attribution of spatial location and feature dimension.

[0089] The correlation of spatial location refers to analyzing the distribution pattern of error samples on the panel, such as whether they are concentrated in edge areas, corner areas, or near specific traces. The correlation of feature dimensions refers to analyzing the clustering characteristics of error samples in the feature space, such as whether they have similar texture patterns, frequency domain responses, or gradient distributions. Through this correlation attribution, the system can identify the dominant causes of each error bucket. For example, a certain type of missed detection error mainly occurs in areas with poor low-frequency uniformity, while a certain type of false alarm error mainly originates from background patterns similar to scratch textures.

[0090] After completing the association attribution, the system performs confidence statistical analysis on the samples within each error bucket, including calculating the mean, variance, quantile distribution, and other statistical measures of the confidence level. Based on these statistical results, the system defines the gray-zone confidence threshold and the high-confidence false alarm threshold.

[0091] The final output model error profile is a structured diagnostic report, which includes the number and proportion of samples in the five error buckets, the confidence distribution characteristics of each bucket, the dominant failure mode obtained from the association attribution, and the suggested gray area threshold.

[0092] Optionally, the model error profile output in this step serves as the basis for selecting 'model-first / mutual verification / rollback' in the gating mechanism. If the model for detecting a defect has high confidence but weak rules, the gating collaborative decision will prioritize the model output and trigger supplementary rule verification.

[0093] Through step S303, this embodiment successfully constructed a model error profile, performed error correlation attribution on each basic model and existing model, and laid a data foundation for subsequent rule-based suppression of model error weaknesses and design of gating instance graphs.

[0094] In one embodiment of the panel detection data processing method of this application, it may further include the following: Step S401: Combining the model error profile, map the error binning results to the preset process mechanism to determine the corresponding mechanism domain and observable channel; Step S402: Provide at least one set of interpretable features for each of the aforementioned mechanistic domains to determine the corresponding defect mechanism profile.

[0095] Optionally, in this embodiment, this step is a mechanism layer analysis step to obtain a defect mechanism profile.

[0096] Specifically, the core task of this step is to establish a traceable correlation between the error information obtained from model-level analysis and the technological mechanisms of the physical world. By transforming the model's error information into "systematic deviations in a certain mechanism domain or a certain observable channel," the abstract model errors find specific physical roots, providing a clear entry point for subsequent interpretable feature extraction.

[0097] First, obtain the model error profile generated during the model layer parsing stage. Based on step S3, it can be seen that the model error profile includes error clusters obtained from five types of error buckets.

[0098] Each of the aforementioned error clusters is treated as an analysis unit and mapped one by one to a pre-defined process mechanism knowledge system. The pre-defined process mechanism is a structured knowledge base pre-constructed according to the panel manufacturing process and physical imaging principles, which clearly defines the observable signal channels corresponding to different physical causes. For example, for the mechanism domain of "low-frequency background inhomogeneity," its observable channels include the brightness domain (evaluating the mean deviation of the evaluation area) and the frequency domain (evaluating the energy proportion of low-frequency components); for the mechanism domain of "edge transition anomaly," its observable channels include the gradient domain (evaluating the rate of change of edge intensity) and the spatial domain (evaluating the width of the transition band).

[0099] For a typical sample within a specific error cluster, its feature distribution across multiple candidate channels is extracted and matched with standard feature templates for each mechanistic domain. If samples within an error cluster exhibit highly consistent distribution characteristics across a particular channel, and these characteristics can be reasonably explained by a specific mechanistic domain, then a strong correlation is established between the error cluster, the mechanistic domain, and the channel. For example, if a batch of false alarm samples all exhibit periodic spectral peaks in the frequency domain, they can be mapped to the "periodic structure artifact" mechanistic domain, with the observable channel being in the frequency domain.

[0100] Optionally, the mechanism domains include, but are not limited to: mechanism domains related to optical imaging, material structure, manufacturing process, and environment and equipment status.

[0101] Optionally, the observation channels include, but are not limited to, observable signals in the brightness domain, color domain, polarization domain, frequency domain, gradient domain, texture domain, and time domain.

[0102] Next, after clarifying the mechanistic domain and observable channel to which the error cluster belongs, we further refine the quantitative criteria for each mechanistic domain that can be directly implemented by the rule components, thus interpreting the features.

[0103] For each defined mechanistic domain, the system designs at least one set of characteristic indices that can be calculated using deterministic algorithms, based on their physical definitions and signal characteristics. Simultaneously, controllable variables are established, listing key variables that can significantly alter defect performance, such as exposure, gain, light source attenuation, polarizer angle, scanning and stitching strategy, panel temperature, and platform vibration.

[0104] Examples of design principles: In Example 1, taking the "low-frequency non-uniformity" mechanism domain as an example, the interpretable features provided by the system include: the grayscale standard deviation of the background image after low-pass filtering (used to quantify the overall uniformity), the maximum brightness difference between the local region and the global background (used to locate significantly deviated areas), and the cumulative energy proportion of low-frequency components in the frequency domain (used to assess the spectral concentration of non-uniformity). The calculation of these features does not depend on training data, and each step has a clear mathematical definition and physical correspondence.

[0105] Example 2: Taking the "edge transition anomaly" mechanism domain as an example again, the interpretable features provided by the system may include: the slope of the grayscale profile along the normal direction of the defect boundary (used to determine whether the transition is too gentle or steep), the contrast ratio on both sides of the edge (used to determine whether there is an optical abrupt change due to the bonding process), and the number of pixels in the width of the transition area (used to determine whether it meets the normal process specifications).

[0106] The interpretable feature set designed for each mechanism domain is integrated with the mechanism domain labels, observable channel information, and correlation mapping relationships determined in step S401, and the original model error profile, to finally form a structured defect mechanism profile.

[0107] The core value of defect mechanism profiling lies in binding together the errors made by the model (error profiling), the physical causes of the errors (mechanistic domains and channels), and the quantitative indicators that can be verified in engineering (interpretable features).

[0108] This provides direct material for the design of subsequent rule components. Each interpretable feature can be encapsulated as a rule component, and its threshold range and applicable conditions are determined by the statistical distribution of the corresponding error cluster. At the same time, it provides a computable basis for the triggering conditions of gating relationships: the predicate expression in the gating logic (such as "low-frequency uniformity score is lower than the cleaning threshold") directly references the interpretable features and their calculation methods defined in the defect mechanism profile.

[0109] Through step S402, this embodiment successfully obtains a defect mechanism profile, establishing a complete transformation link from model error to physical mechanism, and then to computable features and executable rules.

[0110] In one embodiment of the panel detection data processing method of this application, it may further include the following: Step S501: Extract deterministic computational criteria from the interpretable features of the defect mechanism profile to form a set of ruleable criteria. Evaluate the value of each ruleable criterion in the set and determine the corresponding rule value score. The value evaluation includes evaluating the computational complexity of each criterion, its ability to suppress false alarm clusters, and its impact on the risk of missed detection. Step S502: Select rules based on the rule value score to determine the corresponding rule candidate set, which is used to provide prior criteria for the model; Step S503: Based on the mechanism domain classification of each defect in the defect mechanism profile, and combined with the long-tail distribution structure of the defect and the error confidence statistics of the model, determine the draft gating relationship for the defect, including at least one of rule-first gating, model-first gating, parallel execution gating and two-way mutual verification gating.

[0111] Optionally, in this embodiment, this step is rule layer parsing to obtain a set of rule candidates and a draft of gating relationships.

[0112] Specifically, from the interpretable features described by the defect mechanism profile, we extract those criteria that can be achieved through deterministic computation, and use the criteria to select a candidate set of rules to make up for the model's prediction deficiencies.

[0113] Deterministic computation refers to directly deriving results based on fixed mathematical formulas, logical comparisons, or signal processing operators. For example, from the interpretable feature of "low-frequency uniformity," a deterministic criterion such as "calculating the local variance after low-pass filtering the image, and determining uniformity if the variance is below a certain threshold" can be extracted. From the feature of "edge transition width," a criterion such as "calculating the gray-level gradient change along the edge normal direction and counting the number of transition pixels" can be extracted. These criteria together constitute a set of regularizable criteria.

[0114] After forming the set, the system performs a multi-dimensional value assessment of each criterion, outputting a comprehensive rule value score. The assessment dimensions include three aspects: First, computational complexity assessment. Analyze the computational load required for this criterion, including filtering operations, statistical calculations, morphological processing, etc., and estimate its time consumption under production line cycle time constraints. Criteria with low complexity are more suitable as pre-screening gates, capable of quickly filtering a large number of negative samples without affecting overall throughput.

[0115] Second, the suppression capability of false alarm clusters is evaluated. The criteria are compared with the dictionary of false alarm clusters to determine whether the criteria can effectively distinguish between real defects and a certain type of false alarm samples. For example, for a false alarm cluster such as "artifacts caused by illumination vignetting", if a "regional brightness difference criterion" can effectively distinguish it from real Mura, then the suppression capability score of the criterion is high.

[0116] Third, assess the impact of missed detection risk. Analyze whether using this criterion as a filtering or judgment condition will introduce new missed detections, i.e., whether it will incorrectly classify real defects as non-defects. This assessment is usually based on the distribution boundaries of defect characteristics in historical samples to ensure that the rule threshold is not excessively tightened.

[0117] Through the three-dimensional evaluation described above, each rule criterion receives a quantified value score. These scores are used for rule selection and prioritization in subsequent steps, ensuring that the rules selected for the candidate set are low-cost, high-return, and risk-controllable.

[0118] Next, the set of ruleable criteria is filtered by a scoring threshold to determine the criteria that will ultimately enter the rule candidate set. The core function of the rule candidate set is to "provide prior criteria for the model." In subsequent gating execution, these rule criteria can be executed before, after, or in parallel with model inference, providing physical constraints or semantic priors for the model's output.

[0119] For example, in Mura detection, the "low-frequency uniformity criterion" in the rule candidate set can serve as a pre-gating mechanism: if the low-frequency uniformity of the image is good, it directly outputs "no defect," skipping model inference and saving computational resources; if the low-frequency uniformity is poor, it triggers further model analysis. Similarly, the "edge transition width criterion" can serve as a post-validation mechanism: if the edge transition width of the defect region output by the model is abnormal, its confidence level is reduced or a backtracking detection is triggered. The construction of the rule candidate set makes the rules no longer fragmented, temporary logic, but rather reusable, configurable, and auditable standardized components.

[0120] Optionally, in this embodiment, this step generates a matching gating relationship draft for each type of defect based on the mechanism domain in the defect mechanism profile, combined with the long-tail distribution structure produced by the data layer and the error confidence statistics produced by the model layer.

[0121] The basic gating preference is determined based on the mechanistic domain of the defect. For example, for defects with clear mechanisms and strong interpretable features (such as bright spots and dark spots), rule-based gating is preferred, with low-cost rule screening performed first; for defects with complex shapes and relying on high-dimensional semantic features (such as cloud-like mura), model-based gating is preferred, with the model responsible for recall and rules validated later.

[0122] A long-tail distribution structure is introduced for correction. If samples of a certain type of defect are scarce and the model is difficult to train sufficiently, a rule fallback or anomaly detection fallback path is added to the gating to avoid missed detections by the model.

[0123] The model's error confidence statistics should be used to further refine the approach. If the model's confidence is artificially high in a certain false alarm cluster, a rule-based suppression path should be added to the gating draft. If the model's confidence is generally low in a certain defect category, the gating draft should reduce the model's priority weight or add a two-way verification mechanism.

[0124] Preferably, the "data constraint dimensions" that can also be selected when choosing the rule component or model component are as follows:

[0125] Through multi-dimensional integration, a draft gating relationship was finally obtained, which includes at least rule-first gating, model-first gating, parallel execution gating, and two-way mutual verification gating.

[0126] Understandably, the gating relationship draft is a structured description document that records which gating combination(s) should be used for each type of defect, the execution order between the gatings, and the conditions for triggering the switching.

[0127] Through step S503, this embodiment successfully obtained the rule candidate set and the gating relationship draft, laying the foundation for binding with specific components and instantiating into an executable gating graph in subsequent steps.

[0128] In one embodiment of the panel detection data processing method of this application, it may further include the following: Step S601: Obtain the draft gating relationship, including gating type, trigger condition description and data routing template; Step S602: Map the component category nodes in the gating relationship draft to specific component instances in the candidate rule component set or the candidate model component set; Step S603: Compile the triggering condition description into an executable predicate expression, which is used to determine the selection of the execution path and the triggering of the downstream node; Step S604: Based on the data routing template and the compilation result of the predicate expression, obtain the data flow direction and dependency relationship between each node, and determine the corresponding executable gating graph.

[0129] Optionally, in this embodiment, the gating relationship draft is used to establish triggering conditions and data routing between rules and models, and the core is to define the computation graph as a directed acyclic graph.

[0130] The gating relationship draft contains three core types of information: gating type, trigger condition description, and data routing template. The gating type specifies the collaboration mode between rule components and model components, such as rule-first, model-first, parallel execution, or bidirectional verification. The trigger condition description is a judgment criterion recorded in natural language or semi-formal language, indicating under what conditions which execution path should be activated. The data routing template defines how data is passed between components, such as whether the entire image is passed to subsequent components, only local slices of candidate regions are passed, or intermediately generated masks or feature maps are passed.

[0131] The abstract component category nodes in the draft gating relationship are replaced with actual executable concrete component instances. For example, the draft might contain a category node for "low-frequency uniformity analysis rules," and the system needs to select a specific rule plugin instance from the candidate rule component set, such as a specific version of the low-pass filter uniformity calculation module. During component instantiation, the best component is selected based on the matching score calculated in the previous component matching stage, computational cost constraints, and traceability requirements. Specifically, the matching score calculates scores for each component based on matching degree, stability, false alarm suppression capability, cost, and traceability, and forms a weighted total score.

[0132] After mapping is complete, each node is bound to an executable code unit or model service, and its version number and calling protocol are recorded.

[0133] Next, the human-readable trigger condition descriptions in the gating relationship draft will be transformed into computer-executable formal predicate expressions.

[0134] Predicate expressions take lightweight computational features, collected metadata, and the output status of upstream components as input variables, and output a Boolean or enumerated value to determine the direction of the execution path. For example, a trigger condition described as "if the low-frequency uniformity score is lower than the cleanliness threshold, skip the model and directly output non-defect" is compiled into a predicate expression by the system. The system reads the low-frequency score output by the rule component in real time, compares it with the preset threshold, and if the condition is true, guides the process to skip the subsequent model node and directly enter the output stage. Through this compilation, the originally vague condition description is transformed into a logical branch that can be quickly determined at runtime, realizing the automation and determinism of gating decisions.

[0135] Based on the compilation results of the data routing template and predicate expressions, the data flow and dependencies between nodes are established. The data routing template determines how the output of the upstream component is passed to the downstream component, such as full graph transmission, ROI slice transmission, or feature vector transmission. The compilation results of the predicate expressions determine under what conditions which edges will be activated, thus forming dynamic execution paths. The system organizes these nodes and edges into a directed acyclic graph, ensuring that there are no circular dependencies, and each path from the entry to the exit represents a possible cooperative execution sequence. The final generated executable gating graph is a complete computation graph description, which includes specific component nodes, conditional edges, data transmission specifications, and execution order constraints for each node, and can be directly loaded and executed by the runtime engine.

[0136] Through step S604, this embodiment successfully constructed an executable gating graph, which enables standardized detection of each panel detection task.

[0137] In one embodiment of the panel detection data processing method of this application, it may further include the following: Step S701: Configure a conflict determination predicate set and a policy path set for at least one conflict node in the executable gating graph; Step S702: The conflict determination predicate set includes at least one of category conflict predicates, existence conflict predicates, location conflict predicates, confidence conflict predicates, and morphological conflict predicates; Step S703: The set of strategy paths includes at least one of the following: rule priority path, model priority path, parallel scoring fusion path, rollback anomaly detection path, and path to enter the manual review pool; Step S704: Bind the conflict determination predicate set and strategy path set to the edges, nodes and data routes in the executable gating graph to determine the corresponding panel detection collaboration unit.

[0138] Optionally, in this embodiment, a conflict resolution operation is performed on the executable gating graph. During the execution of the executable gating graph, there are "conflict nodes". The characteristics of a conflict node are: the outputs of the rule component and the model component converge here, and their conclusions may be inconsistent.

[0139] This step configures two executable sets of configurations for each conflict node. The first set is the "conflict determination predicate set," which describes under what circumstances a conflict is considered to have occurred; the second set is the "policy path set," which describes which path the system should follow to continue processing once a conflict is confirmed.

[0140] First, the composition of the predicate set for conflict determination.

[0141] The conflict decision predicate set consists of multiple computable decision conditions, each used to detect a certain type of inconsistency between the rule output and the model output. Category conflict predicates determine whether the defect categories determined by the two predicates are different; existence conflict predicates determine when one predicate identifies a defect while the other identifies no defect; location conflict predicates determine whether the positional deviation of the defect region output by the two predicates exceeds a preset tolerance; confidence conflict predicates determine when the model confidence falls into the gray area while the rule feature values ​​exhibit abnormal behavior; and morphological conflict predicates determine whether the defect morphology output by the model violates the geometric or structural constraints of the rules. These predicates are all expressed in a computable form, with inputs including the rule output, model output, and intermediate feature values, and outputs being Boolean values.

[0142] Second, the composition of the strategy path set.

[0143] The rule-first path directly uses the rule's conclusion as the final output, while recording conflicting samples for subsequent analysis; the model-first path uses the model's conclusion as the final output; the parallel scoring fusion path fuses the rule's output score with the model's output confidence score according to preset weights to generate a comprehensive score as the final basis; the fallback anomaly detection path abandons the current rule and model's regular output and instead calls an anomaly detection model for secondary judgment; the path to manual review pool indicates that the system cannot automatically resolve the conflict, and sends the sample and its related evidence to the manual review queue for final judgment by a human.

[0144] Bind the above configuration to the edges, nodes, and data routes in the executable gating graph.

[0145] A conflict node is itself a node in the gating graph. Its incoming edges connect to the outputs of the rule component and the model component, and its outgoing edges connect to subsequent fusion, fallback, or output modules. During binding, the configured predicate set is written into the node's decision logic, and the policy path set is mapped to each outgoing edge of the node. Simultaneously, the data flow direction during the execution of each path needs to be clearly defined. For example, when using rule priority, the rule's output data needs to be transmitted; when using a fusion path, both the rule and model's output data need to be transmitted. After binding, the entire gating graph, conflict handling configuration, and data routing together constitute a complete and executable unit—the panel detection collaboration unit. The detection collaboration unit must undergo three types of constraint checks: latency, risk, and traceability, before finally forming an executable DAG.

[0146] This unit can be deployed and run independently. Upon receiving input data, it automatically performs gating decisions, conflict detection, and strategy selection, outputting the final detection results and complete processing logs. Furthermore, the resulting collaborative detection unit is not static: when data distribution shifts, model versions are updated, or rule thresholds are adjusted, both the draft and the gating graph can evolve in a versioned manner; the online logs and evidence packages of the gating graph will then update the draft in reverse (forming a closed loop).

[0147] Through step S704, this embodiment successfully integrates conflict handling into an executable gating diagram, forming a detection collaboration unit and achieving standardization of the panel detection process.

[0148] To improve the accuracy and precision of panel inspection tasks, this application provides an embodiment of a panel inspection data processing apparatus for implementing all or part of the panel inspection data processing method, see [link to embodiment]. Figure 2 The panel detection data processing device specifically includes the following components: The panel detection data acquisition module 10 is used to acquire panel detection data, which includes panel detection task data, panel detection sample set and corresponding detection metadata. The composite defect profile determination module 20 is used to perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile; based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile; combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile; extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft; and determine the corresponding composite defect profile based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set. The detection collaboration unit determination module 30 is used to match the composite defect profile with the component profiles in the preset component library, determine the corresponding candidate rule component set and candidate model component set, bind the candidate rule component set and the candidate model component set with the gating relationship draft, determine the corresponding executable gating graph, perform conflict handling and solidification operations on the executable gating graph, determine the corresponding panel detection collaboration unit, and perform panel detection according to the panel detection collaboration unit.

[0149] As can be seen from the above description, the panel inspection data processing device provided in this application embodiment can generate a composite defect profile containing sample distribution, model error, defect mechanism and rule candidate set by performing multi-layer analysis of data layer, model layer, mechanism layer and rule layer. The composite defect profile is matched with the component profile in the preset component library to select candidate rule component set and model component set, and bind them with gating relationship draft to construct an executable gating graph. The gating graph is solidified by conflict handling, and the corresponding standardized panel inspection collaborative unit is compiled and determined. The standardized panel inspection process is performed according to the panel inspection collaborative unit, thereby improving the accuracy and precision of panel inspection tasks.

[0150] From a hardware perspective, in order to improve the precision and accuracy of panel inspection tasks, this application provides an embodiment of an electronic device for implementing all or part of the panel inspection data processing method, wherein the electronic device specifically includes the following: The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the panel detection data processing method and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the panel detection data processing method in the present embodiment, and the contents of the embodiments of the panel detection data processing method are incorporated herein, and repeated parts will not be described again.

[0151] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0152] In practical applications, the panel detection data processing method can be partially executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing capabilities of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0153] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0154] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0155] In one embodiment, the panel detection data processing method function can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control: Step S101: Obtain panel detection data, which includes panel detection task data, panel detection sample set, and corresponding detection metadata; Step S102: Perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile. Based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile. Combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile. Extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft. Based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set, determine the corresponding composite defect profile. Step S103: Match the composite defect profile with the component profiles in the preset component library to determine the corresponding candidate rule component set and candidate model component set. Bind the candidate rule component set and the candidate model component set to the gating relationship draft to determine the corresponding executable gating graph. Perform conflict handling and solidification operations on the executable gating graph to determine the corresponding panel detection collaboration unit and perform panel detection according to the panel detection collaboration unit.

[0156] As described above, the electronic device provided in this application provides a composite defect profile by performing multi-layered analysis of data (data layer, model layer, mechanism layer, and rule layer) on the data. This composite defect profile includes sample distribution, model error, defect mechanism, and rule candidate set. The composite defect profile is then matched with component profiles in a preset component library to select candidate rule component sets and model component sets. These are then bound to a gating relationship draft to construct an executable gating graph. By handling and solidifying the gating graph through conflict resolution, the corresponding standardized panel detection collaborative unit is compiled and determined. A standardized panel detection process is then performed based on the panel detection collaborative unit, thereby improving the accuracy and precision of the panel detection task.

[0157] In another embodiment, the panel detection data processing method can be configured separately from the central processing unit 9100. For example, the panel detection data processing method can be configured as a chip connected to the central processing unit 9100, and the panel detection data processing method function can be implemented through the control of the central processing unit.

[0158] like Figure 3As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0159] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0160] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0161] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0162] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0163] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0164] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0165] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Furthermore, the audio processor 9130 is also coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored sound via the speaker 9131.

[0166] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the panel detection data processing method with a server or client as the execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the panel detection data processing method with a server or client as the execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step S101: Obtain panel detection data, which includes panel detection task data, panel detection sample set, and corresponding detection metadata; Step S102: Perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile. Based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile. Combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile. Extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft. Based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set, determine the corresponding composite defect profile. Step S103: Match the composite defect profile with the component profiles in the preset component library to determine the corresponding candidate rule component set and candidate model component set. Bind the candidate rule component set and the candidate model component set to the gating relationship draft to determine the corresponding executable gating graph. Perform conflict handling and solidification operations on the executable gating graph to determine the corresponding panel detection collaboration unit and perform panel detection according to the panel detection collaboration unit.

[0167] As described above, the computer-readable storage medium provided in this application provides a composite defect profile by performing multi-layered analysis of data, model, mechanism, and rule layers on the data. This composite defect profile includes sample distribution, model error, defect mechanism, and rule candidate set. The composite defect profile is then matched with component profiles in a preset component library to select candidate rule component sets and model component sets. These are then bound to a gating relationship draft to construct an executable gating graph. By handling and solidifying the gating graph through conflict resolution, the corresponding standardized panel detection collaborative unit is compiled and determined. A standardized panel detection process is then performed based on the panel detection collaborative unit, thereby improving the accuracy and precision of the panel detection task.

[0168] Embodiments of this application also provide a computer program product capable of implementing all steps of the panel detection data processing method in the above embodiments, where the execution subject is a server or a client. When this computer program / instruction is executed by a processor, it implements the steps of the panel detection data processing method. For example, the computer program / instruction implements the following steps: Step S101: Obtain panel detection data, which includes panel detection task data, panel detection sample set, and corresponding detection metadata; Step S102: Perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile. Based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile. Combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile. Extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft. Based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set, determine the corresponding composite defect profile. Step S103: Match the composite defect profile with the component profiles in the preset component library to determine the corresponding candidate rule component set and candidate model component set. Bind the candidate rule component set and the candidate model component set to the gating relationship draft to determine the corresponding executable gating graph. Perform conflict handling and solidification operations on the executable gating graph to determine the corresponding panel detection collaboration unit and perform panel detection according to the panel detection collaboration unit.

[0169] As described above, the computer program product provided in this application provides a composite defect profile by performing multi-layered analysis of data (data layer, model layer, mechanism layer, and rule layer) on the data. This composite defect profile includes sample distribution, model error, defect mechanism, and rule candidate set. The composite defect profile is then matched with component profiles in a preset component library to select candidate rule component sets and model component sets. These are then bound to a gating relationship draft to construct an executable gating graph. By handling and solidifying the gating graph through conflict resolution, the corresponding standardized panel detection collaborative unit is determined through compilation. A standardized panel detection process is then performed based on the panel detection collaborative unit, thereby improving the accuracy and precision of the panel detection task.

[0170] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0171] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0172] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0173] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0174] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A panel detection data processing method, characterized in that, The method includes: Acquire panel detection data, which includes panel detection task data, panel detection sample set, and corresponding detection metadata. The panel detection data is parsed at the data layer to determine the corresponding panel detection data profile. Based on the constraints of the panel detection data profile, the existing model is parsed at the model layer to determine the corresponding model error profile. The model error profile and the preset process mechanism are combined to perform mechanism layer parsing to determine the corresponding defect mechanism profile. Regularizable criteria are extracted from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft. Based on the panel detection data profile, the model error profile, the defect mechanism profile and the rule candidate set, the corresponding composite defect profile is determined. The composite defect profile is matched with the component profiles in the preset component library to determine the corresponding candidate rule component set and candidate model component set. The candidate rule component set and the candidate model component set are bound to the gating relationship draft to determine the corresponding executable gating graph. The executable gating graph is subjected to conflict handling and solidification operations to determine the corresponding panel detection collaboration unit, and panel detection is performed according to the panel detection collaboration unit.

2. The panel detection data processing method according to claim 1, characterized in that, The step of performing data layer parsing on the panel detection data to determine the corresponding panel detection data profile includes: Data layer parsing is performed on the panel detection data to statistically analyze the distribution pattern of defect samples in the panel detection sample set, including the long-tail distribution structure and the set of difficult cases, and to determine the corresponding sample distribution profile. Alignment analysis is performed on the feature distribution of samples under different acquisition conditions to identify the source of domain offset and determine the corresponding data drift report; For samples that have historically generated false alarms, cluster them according to texture similarity, lighting artifact similarity, or splicing artifact similarity to determine the corresponding model false alarm cluster dictionary; Based on the sample distribution profile, the data drift report, and the model false alarm cluster dictionary, the corresponding panel detection data profile is determined.

3. The panel detection data processing method according to claim 2, characterized in that, The constraint based on the panel detection data profile, performing model layer analysis on the existing model to determine the corresponding model error profile, includes: A weighted sampling operation is performed based on the sample distribution profile, the data drift report, and the model false alarm cluster dictionary to determine the corresponding validation sample set; Obtain the original output of the existing model on the validation sample set, and compare the original output with the preset true value label to determine the corresponding error bucket. The error bucket includes missed detection error, false alarm error, positioning offset error, boundary instability error and category confusion error. The error buckets and the false alarm cluster dictionary are associated with spatial location and feature dimension, and confidence statistics are performed on each error bucket based on the results of the association and attribution to determine the corresponding model error profile.

4. The panel detection data processing method according to claim 3, characterized in that, The step of combining the model error profile and the preset process mechanism to perform mechanism-level analysis and determine the corresponding defect mechanism profile includes: Based on the model error profile, the error binning results are mapped to a preset process mechanism to determine the corresponding mechanism domain and observable channel; For each of the aforementioned mechanistic domains, at least one set of interpretable features is provided to determine the corresponding defect mechanism profile.

5. The panel detection data processing method according to claim 4, characterized in that, The step of extracting ruleable criteria from the defect mechanism profile and determining the corresponding rule candidate set and gating relationship draft includes: From the interpretable features of the defect mechanism profile, deterministic computational criteria are extracted to form a set of ruleable criteria. The value of each ruleable criterion in the set is evaluated to determine the corresponding rule value score. The value evaluation includes evaluating the computational complexity of each criterion, its ability to suppress false alarm clusters, and its impact on the risk of missed detection. Rule selection is performed based on the rule value score to determine the corresponding rule candidate set, which is used to provide prior criteria for the model. Based on the mechanistic domain classification of each defect in the defect mechanism profile, and combined with the long-tail distribution structure of the defects and the error confidence statistics of the model, a draft gating relationship for the defect is determined, including at least one of rule-first gating, model-first gating, parallel execution gating, and two-way mutual verification gating.

6. The panel detection data processing method according to claim 1, characterized in that, The step of binding the candidate rule component set and the candidate model component set with the gating relationship draft to determine the corresponding executable gating graph includes: Obtain the draft gating relationship, including gating type, trigger condition description, and data routing template; Map the component category nodes in the gating relationship draft to specific component instances in the candidate rule component set or the candidate model component set; The triggering conditions are compiled into an executable predicate expression, which is used to determine the selection of the execution path and the triggering of downstream nodes; Based on the data routing template and the compilation result of the predicate expression, the data flow and dependencies between nodes are obtained, and the corresponding executable gating graph is determined.

7. The panel detection data processing method according to claim 1, characterized in that, The conflict resolution and solidification operation of the executable gating graph, determining the corresponding panel detection coordination unit, includes: Configure a conflict determination predicate set and a policy path set for at least one conflict node in the executable gating graph; The conflict determination predicate set includes at least one of category conflict predicates, existence conflict predicates, location conflict predicates, confidence conflict predicates, and morphological conflict predicates; The set of strategy paths includes at least one of the following: rule priority path, model priority path, parallel scoring fusion path, rollback anomaly detection path, and path to enter the manual review pool. The conflict determination predicate set and strategy path set are bound to the edges, nodes and data routes in the executable gating graph to determine the corresponding panel detection collaboration unit.

8. A panel detection data processing device, characterized in that, The device includes: The panel detection data acquisition module is used to acquire panel detection data, which includes panel detection task data, panel detection sample set and corresponding detection metadata. The composite defect profile determination module is used to perform data layer parsing on the panel detection data to determine the corresponding panel detection data profile; based on the constraints of the panel detection data profile, perform model layer parsing on the existing model to determine the corresponding model error profile; combine the model error profile and the preset process mechanism to perform mechanism layer parsing to determine the corresponding defect mechanism profile; extract ruleable criteria from the defect mechanism profile to determine the corresponding rule candidate set and gating relationship draft; and determine the corresponding composite defect profile based on the panel detection data profile, the model error profile, the defect mechanism profile, and the rule candidate set. The detection collaboration unit determination module is used to match the composite defect profile with the component profiles in the preset component library, determine the corresponding candidate rule component set and candidate model component set, bind the candidate rule component set and the candidate model component set with the gating relationship draft, determine the corresponding executable gating graph, perform conflict handling and solidification operations on the executable gating graph, determine the corresponding panel detection collaboration unit, and perform panel detection according to the panel detection collaboration unit.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the panel detection data processing method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the panel detection data processing method according to any one of claims 1 to 7.

Images