Resistor defect identification method and system based on visual inspection

Abstract

The invention discloses a resistor defect identification method and system based on visual inspection, and relates to the technical field of visual inspection. According to each visual inspection area, a plurality of resistors in the visual inspection area trigger area type detection; the final defect grade of the resistor is determined based on the image identification of the panoramic image of the resistor, and the batch defect identification system is determined according to the final defect grade of each resistor and the corresponding visual detection area, so that the accuracy of the batch defect identification system is improved. Determining centralized maintenance content according to each resistor combination and the corresponding resistor maintenance station; and determining a maintenance worksheet of the maintenance robot according to the centralized maintenance content, the corresponding resistors and the defect areas of the resistors, and determining a dynamic maintenance system of the maintenance robot for the plurality of resistors based on a plurality of maintenance items of the maintenance worksheet, the final defect grade of each resistor and the corresponding defect areas. And the accuracy of the dynamic maintenance system is improved.

Description

Technical Field

[0001] This invention relates to the technical field of visual inspection, and more particularly to a method and system for defect identification of resistors based on visual inspection. Background Technology

[0002] With the development of technology, resistors, as electronic components, are used in electronic devices. A conveyor line dynamically transports multiple resistors to the same visual inspection station. A camera at the visual inspection station dynamically captures images of each resistor. In the current technology, each camera captures an individual resistor separately, ignoring the capture of the remaining resistors and neglecting the control of area-based inspection. This affects the accuracy of the batch defect identification system and leads to the inaccuracy of the maintenance robot's dynamic maintenance system for multiple resistors. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and system for identifying defects in resistors based on visual inspection.

[0004] This invention provides a method for defect identification of resistors based on visual inspection, comprising: When multiple resistors are in the visual inspection station, the information combination of each resistor is determined by tracing the model information of each resistor, and the primary defect level of the resistor is determined based on the identification of the information combination of each resistor. Multiple visual inspection areas are determined based on the primary defect level and corresponding location of each resistor. Regional inspection is triggered based on each visual inspection area and multiple resistors located in that visual inspection area, and multiple images of each resistor in different directions are determined. A panoramic image of the resistor is determined based on multiple images of the resistor in different directions, the corresponding primary defect levels, and the corresponding information combinations. The final defect level of the resistor is determined based on image recognition of the panoramic image. A batch defect identification system is determined based on the final defect level of each resistor and the corresponding visual detection area. The panoramic image is a four-dimensional information body containing geometry, texture, defect semantics, and historical records. The batch defect identification system includes a quality feedback system and a system self-optimization system. Based on the identification of the batch defect identification system, the corresponding classification information is determined. Based on the classification information and the corresponding multiple resistors, multiple resistor combinations are determined. Based on each resistor combination and the corresponding resistor maintenance station, the centralized maintenance content is determined. Based on the centralized maintenance content, the corresponding resistors, and the defect areas of those resistors, a maintenance worksheet for the maintenance robot is determined. Based on the multiple maintenance items in this worksheet, the final defect level of each resistor, and the corresponding defect areas, a dynamic maintenance system for multiple resistors by the maintenance robot is determined.

[0005] This invention provides a visual inspection-based resistor defect identification system, which is applied to the aforementioned visual inspection-based resistor defect identification method.

[0006] Compared with the prior art, the beneficial effects of the present invention are: (1) When multiple resistors are in the visual inspection station, the information combination of each resistor is determined by tracing the model information of each resistor, and the primary defect level of the resistor is determined based on the identification of the information combination of each resistor; multiple visual inspection areas are determined based on the primary defect level of each resistor and the corresponding position, and regional inspection is triggered according to each visual inspection area and the multiple resistors in the visual inspection area. Multiple visual inspection areas are introduced to realize the control of regional inspection, so that multiple images of each resistor in different directions can be processed.

[0007] (2) Based on multiple images of the resistor in different directions, the corresponding primary defect level and the corresponding information combination, the panoramic image of the resistor is determined. Based on the image recognition of the panoramic image of the resistor, the final defect level of the resistor is determined. Based on the final defect level of each resistor and the corresponding visual detection area, a batch defect identification system is determined. The visual detection of the panoramic image of the resistor is introduced to control the final defect level of the resistor and improve the accuracy of the batch defect identification system.

[0008] (3) Based on the identification of the batch defect identification system, the corresponding classification information is determined. Based on the classification information and the corresponding multiple resistors, multiple resistor combinations are determined. The centralized maintenance content is determined according to each resistor combination and the corresponding resistor maintenance station. The maintenance worksheet of the maintenance robot is determined according to the centralized maintenance content, the corresponding resistor and the defect area of ​​the resistor. Based on the multiple maintenance items of the maintenance worksheet, the final defect level of each resistor and the corresponding defect area, the dynamic maintenance system of the maintenance robot for multiple resistors is determined. The centralized maintenance content is controlled, and the overall consideration of the multiple maintenance items of the maintenance worksheet, the final defect level of each resistor and the corresponding defect area is realized, which improves the accuracy of the dynamic maintenance system of the maintenance robot for multiple resistors. Attached Figure Description

[0009] Figure 1 This is a schematic flowchart of a visual inspection-based resistor defect identification method according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 of the visual inspection-based resistor defect identification method in this embodiment of the invention. Figure 3 This is a flowchart illustrating step S12 in the visual inspection-based resistor defect identification method in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the visual inspection-based resistor defect identification method in this embodiment of the invention. Figure 5 This is a flowchart illustrating step S14 of the visual inspection-based resistor defect identification method in this embodiment of the invention. Figure 6 This is a flowchart illustrating step S15 of the visual inspection-based resistor defect identification method in this embodiment of the invention. Figure 7 This is a schematic diagram of the structural composition of a visual inspection-based resistor defect identification system according to an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 7 A visual inspection-based method for resistor defect identification is proposed and applied to visual inspection scenarios. The visual inspection-based method for resistor defect identification includes: Step S11: When multiple resistors are at the visual inspection station, the information combination of each resistor is determined by tracing the model information of each resistor, and the primary defect level of the resistor is determined based on the identification of the information combination of each resistor. Step S12: Determine multiple visual inspection areas based on the primary defect level and corresponding location of each resistor; trigger regional inspection based on each visual inspection area and multiple resistors located in that visual inspection area; and determine multiple images of each resistor in different directions. Step S13: Determine the panoramic image of the resistor based on multiple images of the resistor in different directions, the corresponding primary defect level, and the corresponding information combination; determine the final defect level of the resistor based on image recognition of the panoramic image of the resistor; and determine the batch defect recognition system based on the final defect level of each resistor and the corresponding visual detection area. Step S14: Determine the corresponding classification information based on the batch defect identification system, determine multiple resistor combinations based on the classification information and the corresponding multiple resistors, and determine the centralized maintenance content according to each resistor combination and the corresponding resistor maintenance station; Step S15: Determine the maintenance worksheet for the maintenance robot based on the centralized maintenance content, the corresponding resistor, and the defect area of ​​the resistor. Based on the multiple maintenance items in the maintenance worksheet, the final defect level of each resistor, and the corresponding defect area, determine the dynamic maintenance system of the maintenance robot for multiple resistors.

[0012] refer to Figure 2In step S11, the specific steps are as follows: S111: Multiple resistors are input to the same vision inspection station via a conveyor line and are simultaneously captured by multiple cameras at the vision inspection station. At the same time, the model information of each resistor is marked, the model information of each resistor is traced, and the usage information and processing information of each resistor are output. Based on the usage information and processing information of each resistor, the information combination of the resistor is constructed. S112: In multiple resistors, the information combination of each resistor is dynamically identified, and multiple defect information is determined during the identification process. Based on the multiple defect information, the corresponding defect location and the processing history of the resistor, the primary defect level of the resistor is determined.

[0013] In the embodiments of this application, multiple resistors are input to the same vision inspection station via a conveyor line and are simultaneously captured by multiple cameras at the vision inspection station. At the same time, the model information of each resistor is marked, the model information of each resistor is traced, and the usage information and processing information of each resistor are output. Based on the usage information and processing information of each resistor, the information combination of the resistor is constructed, which takes into account the overall consideration of the usage information and processing information of each resistor and ensures the accuracy of the information combination of the resistor.

[0014] At this time, the resistors are constantly in motion on the conveyor line. To ensure the acquisition of clear images without motion blur, the system uses an industrial camera with a global shutter to achieve microsecond-level synchronous exposure through hardware triggering signals. The multi-camera layout (top view and left and right oblique view) is optimized through optical simulation to cover all key feature surfaces of the resistors: the top camera captures the body characters, color rings, and coating defects, while the side cameras monitor pin coplanarity, bending, or oxidation. The image processing unit uses deep learning instance segmentation algorithms such as MaskR-CNN to accurately separate each resistor from the complex background, generate pixel-level masks and unique temporary IDs (such as Temp_ID_001), and output precise coordinates and bounding boxes.

[0015] A carrier traceability strategy is adopted. The data matrix code / QR code on the pallet, tape, or tin is scanned by a high-precision code reader and used as an index key for the MES or WMS system. The system obtains the carrier's general attributes (model, batch, supplier) through the code and binds this information in batches to the temporary ID objects segmented in step 1. For example, all Temp_ID objects in pallet A are assigned the model B-103K.

[0016] Using the carrier code as the key, the system queries the central database for two types of key information: processing information (supplier, production batch, material composition, process parameters, historical quality inspection data) and usage information (target production line, customer orders, quality requirements, storage records, risk warnings). The data middleware cleans, transforms, and standardizes data from heterogeneous systems (MES / ERP / QMS) to form a unified structure. At the same time, the system deeply integrates visual data (temporary ID, coordinates, image blocks), identity information (model), and historical data (processing and usage information) to generate a structured data package—an information combination—for each resistor. This information combination includes the current visual status, historical history, and future use, serving as a digital archive for subsequent intelligent decision-making.

[0017] Specifically, when the conveyor line is paused, three 5-megapixel global shutter cameras (top and 45° left and right) simultaneously flash and take pictures; the image processing server identifies 48 resistors in the tray within 50 milliseconds, assigns temporary IDs (Temp_R_001 to Temp_R_048) and records the pixel coordinates.

[0018] The system queries the database using DM codes to obtain processing information (pure tin plating, G-type terminal forming, 150°C baking for 1 hour) and usage information (server motherboard V3 line, customer X, A-grade quality), as well as key warnings (0.5% of resistor pins in this batch have out-of-tolerance coplanarity). Taking Temp_R_023 as an example, the system constructs a complete file containing visual data (multi-view images, coordinates), identity information (model, supplier, batch), and history data (process, production line, warning). When this file is passed to S112, the algorithm automatically adopts a stricter judgment threshold due to the known risk of pin coplanarity, thus achieving context-aware detection.

[0019] Furthermore, among multiple resistors, the information combination of each resistor is dynamically identified, and multiple defect information is determined during the identification process. Based on the multiple defect information, the corresponding defect location, and the manufacturing process of the resistor, the primary defect level of the resistor is determined. This comprehensive consideration of multiple defect information, corresponding defect location, and manufacturing process of the resistor ensures the accuracy of the primary defect level of the resistor.

[0020] At this time, the system dynamically switches the algorithm strategy according to the usage information of the resistor: if the target production line is automotive electronics, a model sensitive to micro-cracks is called; if it is consumer electronics, emphasis is placed on appearance and character clarity. Similarly, processing information such as tin-lead alloy electroplating triggers a dedicated pin oxidation classifier. In addition, the system dynamically adjusts the detection parameters according to the processing history warning: if there is a risk of pin coplanarity deviation exceeding the standard, the feature weight of the pin area is strengthened through a spatial attention mechanism, and the confidence determination threshold for this defect is reduced (e.g., from 90% to 75%) to achieve strict screening.

[0021] The system identifies multiple defect types (such as surface scratches, pin bending, and body damage) through a multi-label classification network and accurately locates the defect positions using a normalized coordinate system (such as x:0.2, y:-0.4), ensuring that it is not affected by the absolute position of the resistor in the image. At the same time, a confidence score from 0 to 1 is assigned to each defect to quantify its obviousness (e.g., a deep scratch is 0.95, and a slight scratch is 0.65).

[0022] A weighted fusion model is used to synthesize the defect information and historical background to output the primary defect level. The system constructs the defect information into a multi-dimensional feature vector (such as [scratch_confidence_0.8, bend_confidence_0.4]) and dynamically assigns weights according to the processing history: the weight of defects in the core area is higher than that in the edge, the weight of performance defects is higher than that of appearance, and the weight of defects matching the warning is multiplied by a risk amplification factor (e.g., 1.5). The weighted total score is judged as P0 (normal, total score < T1), P1 (suspicious, T1 ≤ total score < T2), or P2 (obvious defect, total score ≥ T2) through a threshold interval to achieve intelligent sorting.

[0023] Specifically, for the analysis of resistor R_B008: due to the warning, the algorithm highly focuses on the pin area. The defect information output shows a surface scratch (confidence 0.88) and pin coplanarity deviation exceeding the standard (confidence 0.30). During the primary grading, the scratch, as an appearance defect, has a lower weight, while although the confidence of the pin coplanarity deviation exceeding the standard is not high, its weight is amplified to an equivalent of 0.60 due to the warning. The comprehensive score exceeds the P1 threshold but does not reach P2, and R_B008 is marked as P1 (suspicious) and needs to enter S12 for high-precision re-inspection.

[0024] Analysis of resistor R_B015: sensitive to the pin area; the defect information only outputs pin coplanarity deviation exceeding the standard (confidence 0.72); due to a high match with the warning, its weight is significantly amplified to an equivalent of 0.95; although the original confidence is medium, the comprehensive score easily exceeds the P2 threshold after considering the context, and R_B015 is marked as P2 (obvious defect) and directly sorted, saving re-inspection resources.

[0025] Analysis of resistor R_B033: Standard testing procedure was performed; defect information was output as body collapse (confidence level 0.99); body collapse is a fatal performance defect with extremely high type weight, and the 0.99 confidence level further confirms its severity; no warning is needed, the comprehensive score far exceeds the P2 threshold, and R_B033 is decisively marked as P2 (obvious defect).

[0026] refer to Figure 3 In step S12, the specific steps are as follows: S121: Mark the position of each resistor relative to the visual inspection station, determine multiple inspection contour lines based on the primary defect level of each resistor and its corresponding position, and determine the corresponding visual inspection area based on the shape of the multiple inspection contour lines and the corresponding multiple resistors, so as to mark multiple visual inspection areas. S122: In multiple visual detection areas, multiple resistors are determined based on the dynamic detection of each visual detection area to mark the multiple resistors in the visual detection area. A corresponding detection signal is determined based on each visual detection area and the multiple resistors in the visual detection area. The detection signal triggers regional detection of multiple resistors in the same visual detection area to realize multi-directional detection of each resistor and to determine multiple images of each resistor in different directions.

[0027] In the embodiments of this application, the position of each resistor relative to the visual inspection station is marked, multiple inspection contour lines are determined according to the primary defect level of each resistor and its corresponding position, and the corresponding visual inspection area is determined according to the shape of the multiple inspection contour lines and the corresponding multiple resistors, so as to mark multiple visual inspection areas. This takes into account the overall consideration of the shape of multiple inspection contour lines and the corresponding multiple resistors, and ensures the accuracy of the corresponding visual inspection area.

[0028] At this point, the system defines a global coordinate system based on the visual inspection station (e.g., with the upper left corner of the conveyor belt start point as the origin), and then converts the pixel coordinates obtained in step S11 into global coordinates in the physical world through pre-completed hand-eye calibration or camera calibration. The system generates a dynamic list in which each element contains the resistor's unique ID, global coordinates, and primary defect level inherited from S11, laying the foundation for subsequent spatial analysis.

[0029] The system iterates through the list generated in the previous step and applies filtering rules: completely ignoring all resistors of the P0 (normal) level, significantly reducing the amount of subsequent calculations; for the remaining P1 (suspicious) and P2 (obvious defect) resistors, the system applies spatial clustering algorithms, such as DBSCAN or hierarchical clustering; DBSCAN can automatically discover clusters based on the density of points and identify isolated noise points (i.e., isolated defective resistors), which is very suitable for this type of scenario; the output of the algorithm is one or more clusters containing spatially adjacent defective resistors, as well as P1 / P2 resistors marked as isolated points.

[0030] The system calculates the geometric boundary for each cluster. Common methods include the minimum bounding rectangle with the highest computational efficiency, the convex hull with a balance between compactness and efficiency, or the precise but complex alpha shape. To ensure that the subsequent detection equipment has sufficient motion margin and field of view, the system performs a small expansion operation on the calculated boundary (such as expanding it outward by 2-5 mm). This expansion amount is a configurable engineering parameter. The final geometric boundary with the expansion amount is the detection contour line.

[0031] Each detection contour line is instantiated as a visual detection region by defining a closed polygon in the global coordinate system; the system assigns a unique ID (such as Zone-A) to each region and establishes a reverse index to record the specific resistor IDs contained within the region; S121 outputs a structured task list, such as: [{Region ID:Zone-A,Contour:[…],Contains Resistors:[R012,R013]},…], which is sent to the execution unit of S122.

[0032] Specifically, the system has generated a list of coordinates and defect levels for five defective resistors: R_B012(P1)@(102mm,155mm), R_B013(P1)@(108mm,158mm), R_B025(P2)@(310mm,402mm), R_B026(P1)@(315mm,408mm), and R_B040(P1)@(505mm,210mm). The remaining 45 P0 level resistors have been ignored.

[0033] The system starts the DBSCAN algorithm, sets the neighborhood radius to 15mm, and the minimum number of points to 2. The algorithm finds that the distances between R_B012 and R_B013, and between R_B025 and R_B026 are all less than 15mm, forming cluster 1 and cluster 2 respectively. R_B040 is more than 15mm away from any other resistor and is marked as an isolated point.

[0034] For cluster 1 (R_B012, R_B013), the system calculates its convex hull and expands it outward by 3mm to generate an irregular pentagonal detection contour line 1; for cluster 2 (R_B025, R_B026), an expanded quadrilateral detection contour line 2 is generated similarly; for isolated point R_B040, the system generates a 10mm x 10mm square detection contour line 3 centered on it.

[0035] Contour line 1 is marked as visual detection area-A, associated with R_B012 and R_B013; contour line 2 is marked as visual detection area-B, associated with R_B025 and R_B026; contour line 3 is marked as visual detection area-C, associated with R_B040; S121 finally outputs these three precise visual detection areas with unique IDs and contents, directly guiding the maintenance robot in the next step S122 to perform high-precision multi-angle shooting, perfectly realizing intelligent focusing from surface to point.

[0036] Furthermore, in multiple visual detection areas, multiple resistors are determined based on the dynamic detection of each visual detection area to mark the multiple resistors located in that visual detection area. Based on each visual detection area and the multiple resistors located in that visual detection area, a corresponding detection signal is determined. This detection signal triggers regional detection of multiple resistors in the same visual detection area to achieve multi-directional detection of each resistor, thereby determining multiple images of each resistor in different directions. This approach considers both the visual detection areas and the multiple resistors located in that visual detection area as a whole, ensuring the accuracy of the corresponding detection signal. At the same time, the introduction of multiple visual detection areas enables control over regional detection, facilitating image processing of multiple images of each resistor in different directions.

[0037] At this point, the maintenance robot performs coarse navigation based on the global coordinates provided by S121, quickly moving the end effector (equipped with a camera) to the center of the target area. The maintenance robot then uses the high-resolution end-effector camera to perform a local fine scan, converting the identified resistor pixel coordinates into precise three-dimensional coordinates and orientations relative to the center point (TCP) of the maintenance robot's end tool through a hand-eye calibration algorithm. This process not only verifies the existence of the resistors but also updates the precise orientation of the resistors to be inspected within the area, forming a high-precision task list.

[0038] The main control system maintains a task queue, and the visual inspection areas generated by S121 are queued in sequence. When a task is ready to be executed, the system generates a structured data packet containing the area ID, the precise pose list of the target resistor, and the imaging protocol (such as Standard_3View or Advanced_3D_Scan). This signal is sent to the maintenance robot controller via an industrial bus (such as EtherCAT), providing it with all the authorization and parameters to start the inspection.

[0039] The maintenance robot controller analyzes the detection signals and uses kinematic algorithms to plan a smooth, collision-free optimal path for each resistor from each viewpoint. During execution, the maintenance robot switches viewpoints (e.g., from top view to side view), and the light source controller synchronously adjusts the lighting mode (e.g., high-angle ring light to detect scratches, low-angle strip light to highlight pins). After the posture stabilizes, a hardware trigger signal drives the camera to expose, capturing images or 3D point cloud data that are precisely bound to the viewpoint, resistor ID, and timestamp, and transmitting them to the central server in real time for analysis by S13.

[0040] Specifically, after receiving the visual inspection task for Zone-A, the maintenance robot quickly moves to a position roughly above its center. The end-effector camera performs a local fine scan to confirm the presence of two resistors, and obtains their precise poses relative to the TCP through hand-eye calibration, verifying them as R_B012 and R_B013. The main control system generates a detection signal for Zone-A: {Zone-A, Target: [{ID:'R_B012', Pose:…}, {ID:'R_B013', Pose:…}], Imaging Protocol: 'Standard_3View'}. The signal is sent to the maintenance robot controller to unlock the inspection task.

[0041] The maintenance robot executes the Standard_3View protocol. For R_B012, it moves sequentially to top-down, left 45°, and right 45° viewpoints, triggering the camera under ring or strip lighting to capture three high-resolution images. The same process is repeated for R_B013. Subsequently, when processing the visual inspection area-B, since R_B025 is a P2 level defect, the system adopts the Advanced_3D_Scan protocol. After completing the standard three-view imaging, the maintenance robot additionally activates the structured light projector to project a grating onto R_B025 and simultaneously capture images, generating a high-precision 3D point cloud file for in-depth crack analysis.

[0042] refer to Figure 4 In step S13, the specific steps are as follows: S131: Among multiple resistors, the first panoramic shape of each resistor is determined based on multiple images of each resistor in different directions and the corresponding primary defect level, and the panoramic image of the resistor is determined based on the first panoramic shape of each resistor and the corresponding primary defect level. S132: Perform image recognition on the panoramic image of each resistor, identify multiple defect features during the recognition process, and determine the final defect level of the resistor based on the combination of multiple defect features, information of each resistor, and the corresponding primary defect level. S133: Mark the orientation of each resistor, determine the first batch defect identification content based on the orientation of each resistor and the corresponding final defect level, determine the second batch defect identification content based on the orientation of each resistor and the corresponding visual inspection area, and determine the batch defect identification system based on the first batch defect identification content and the second batch defect identification content.

[0043] In the embodiments of this application, among multiple resistors, a first panoramic shape of each resistor is determined based on multiple images of each resistor in different directions and the corresponding primary defect level. A panoramic image of the resistor is determined based on the first panoramic shape of each resistor and the corresponding primary defect level, which takes into account the overall consideration of the first panoramic shape and the corresponding primary defect level of each resistor, and ensures the accuracy of the panoramic image of the resistor.

[0044] At this point, the system uses SIFT or deep learning networks to extract and match thousands of key feature points from multi-angle images. Through these matching points, the system accurately calculates the spatial pose of the camera when each image was captured, and uses the principle of triangulation to intersect rays of the same physical point in different images, thereby obtaining its true three-dimensional coordinates. The key to the intelligence of this process lies in the guidance of the S11 primary defect level: the system generates an attention heatmap, weights the feature points in the suspected defect area, increases their matching weight, and tilts computing resources to optimize the reconstruction accuracy of the area. The system outputs a dense three-dimensional point cloud or polygonal mesh, i.e., the first panoramic form, which has higher point cloud density and accuracy at the suspected defect location.

[0045] The system uses a view-dependent texture mapping algorithm to precisely wrap multiple 2D images onto a 3D mesh and fuse images under different lighting conditions to generate HDR textures, thus preserving details in both dark and bright areas. The defect information identified in S112 (such as type, location, and confidence level) is used as a semantic layer or metadata and precisely overlaid on the 3D model. This 3D model, rich in visual information, is linked with the information in S111 (model, batch, processing history, etc.) to establish a permanent index link, forming a panoramic image. The panoramic image is a four-dimensional information body containing geometry, texture, defect semantics, and historical record.

[0046] Specifically, for P1-level B resistor R_B015 (the primary defect is suspected pin bending), the system acquires three high-resolution images of its top and left and right 45° angles. During feature extraction and matching, since the primary defect points to the pin, the system pays special attention to feature matching in the pin area. In the final generated first panoramic 3D mesh model, the mesh of the resistor body is relatively sparse, while the mesh of the two pin parts is unusually dense and fine, clearly showing that the two are not on the same plane. This is the result of attention-guided reconstruction.

[0047] The system performs texture mapping and HDR fusion on the 3D model, enabling the model to not only present the geometric curvature of the pins, but also to showcase the luster of the metal surface, subtle oxidation color differences, and ceramic texture. The initial discovery by S112—suspected pin curvature, with a confidence level of 75%—is attached as metadata to the geometric data of the pins in the model. This complete panoramic image is linked with the information combination of R_B015. When the intelligent model analysis by S132 is performed, it can not only see the curved pins, but also know that they come from a specific batch of Factory A, and that this batch has a risk warning of pin coplanarity. Thus, the panoramic image of R_B015 becomes a four-dimensional information body containing geometry, texture, defect semantics, and historical records, providing unparalleled rich evidence for the final accurate judgment.

[0048] Furthermore, image recognition is performed on the panoramic images of each resistor, and multiple defect features are identified during the recognition process. Based on the combination of multiple defect features, information of each resistor, and corresponding primary defect levels, the final defect level of the resistor is determined. This approach takes into account the overall consideration of multiple defect features, information of each resistor, and corresponding primary defect levels, ensuring the accuracy of the final defect level of the resistor.

[0049] At this point, the system adopts a hybrid intelligent model architecture: the CNN branch processes the high-fidelity two-dimensional texture layer to identify surface defects (such as scratches and stains); the GNN or 3D CNN branch processes the 3D mesh or point cloud to identify geometric defects (such as pin bending and body deformation); the output features of the two branches are combined in the fusion layer and jointly analyzed through a fully connected network; the model finally outputs a set of refined and structured defect features, including defect type, accurate geometric quantization values ​​(such as out-of-tolerance value of 0.12mm and crack depth of 15µm), visual quantization values ​​(such as area of ​​0.25mm²), and confidence scores between 0 and 1.

[0050] By deeply integrating visual evidence with historical context through a weighted Bayesian network or fuzzy logic inference engine, this framework comprehensively processes three types of information: quantified defect features as primary evidence, early warning information in the combination of information as strong prior knowledge, and primary defect levels as benchmark references. If the detected defect highly matches the historical early warning, its weight will be increased by a risk amplification factor. The decision framework outputs a comprehensive defect severity score and maps it to a refined final defect level system according to preset rules: Level A (Good Product), Level B (Appearance Defect), Level C (Repairable Defect), Level D (Performance Defect), and Level F (Scrap). This process ensures the extremely high reliability of the final conclusion.

[0051] Specifically, the panoramic image of resistor B, R_B015, is fed into the hybrid intelligent model; the CNN branch analyzes its surface texture and finds no anomalies; the GNN branch analyzes its three-dimensional mesh and accurately calculates the deviation between the pin endpoint plane and the standard reference plane; the model finally outputs the structured defect feature: [{Defect type: pin coplanarity out of tolerance, out-of-tolerance value: 0.12mm, confidence level: 0.98}], which exceeds the upper limit of the 0.1mm process specification.

[0052] The decision engine integrates multi-source information: Input 1 is visual evidence of pin coplanarity exceeding tolerance with a confidence level of 98%; Input 2 is prior knowledge from its information combination regarding the risk warning of pin coplanarity in batch LOT20251015, and the two are a perfect match; Input 3 is the initial level of P1 (suspicious), which is consistent with the final discovery direction; In the Bayesian network, the combination of strong visual evidence and matching prior knowledge enables the system to confirm the defect with extremely high certainty; The system queries the specification library to map this defect to a repairable category, and R_B015 is finally classified as a C-level (repairable defect).

[0053] Therefore, the orientation of each resistor is marked, and the first-level batch defect identification content is determined based on the orientation of each resistor and its corresponding final defect level. The second-level batch defect identification content is determined based on the orientation of each resistor and its corresponding visual inspection area. A batch defect identification system is determined based on the first-level and second-level batch defect identification content, which takes into account the overall considerations of the first-level and second-level batch defect identification content, ensuring the accuracy of the batch defect identification system. At the same time, visual inspection of the panoramic image of the resistor is introduced to control the final defect level of the resistor, thereby improving the accuracy of the batch defect identification system.

[0054] At this point, the system defines the orientation of each resistor as (x, y, θ) including the center coordinates and rotation angle. These data have been accurately obtained in S11 and S12. The system creates a master data table that associates the unique ID of each resistor with its orientation, the final defect level output by S132, and the visual inspection area ID planned by S121, forming a complete and traceable dataset.

[0055] For the first-level batch defect identification, the system extracts the attitude data of all non-A-grade (B, C, D, F-grade) resistors to form a defect point set, and uses spatial statistical methods for analysis; it quantifies spatial clustering by calculating indicators such as the Moran index, and uses kernel density estimation to generate a defect hotspot map to visually display the dense defect areas; the analysis results generate the first-level batch defect identification content, including clustering conclusions, hotspot area coordinates, associated defect types and batch information, directly pointing to systemic problems in upstream processes or material handling.

[0056] For the second batch of defect identification, in the detection coverage analysis, the system evaluates the screening capability of S11 by calculating the recall rate (number of covered defect resistors / total number of defect resistors) and records the missed cases to optimize the algorithm; in the detection efficiency analysis, the system calculates the precision rate of each visual detection area (number of confirmed defects in the area / total number of detections in the area), identifies the resource waste caused by false alarms of S11, and triggers the tuning of relevant algorithm thresholds.

[0057] Specifically, the system generates a data table for 48 B resistors, associating each resistor's ID, orientation, final defect level, and corresponding visual inspection area ID. For example, R_B012 has an orientation of (102, 155, 3°), a final defect level of C, and belongs to Zone-A. This data table is shown in Table 1. Table 1 Data Table Resistor ID attitude Final Defect Level Visual inspection area ID R_B008 (101,152,5°) Grade A - R_B012 (102,155,3°) Class C Zone-A R_B013 (108,158,2°) Class C Zone-A R_B025 (310,402,355°) Class D Zone-B R_B026 (315,408,0°) Class D Zone-B R_B040 (505,210,90°) Grade B Zone-C R_B041 (508,215,88°) Grade A Zone-C R_B042 (512,218,91°) Grade A Zone-C The system extracts the orientation of all non-A-grade resistors for spatial analysis; the Moran's index is calculated to be 0.65 (p<0.01), indicating significant spatial clustering; the heat map generated by kernel density estimation shows two obvious defect hotspots in the upper left and lower right corners of the tray; the system generates a report indicating that the upper left cluster mainly consists of C-grade defects, and the lower right cluster mainly consists of D-grade defects, both originating from supplier A's batch LOT20251015, triggering a batch quality warning. At this point, the system generates the first batch defect identification content: two significant defect clusters are detected; the upper left cluster mainly consists of C-grade (repairable) defects, and the lower right cluster mainly consists of D-grade (performance) defects; combining the information, the resistors in both clusters originate from supplier A's batch LOT20251015; triggering a batch quality warning.

[0058] Recall evaluation showed that all five non-Class A resistors were successfully covered by the area planned by S121, with a recall rate of 100%, proving that S11 screening had no missed detections. In precision evaluation, Zone-A and Zone-B both had a precision rate of 100%, while Zone-C detected three resistors but only confirmed one defect, with a precision rate of 67%. The system generated a second batch of defect identification content: "The overall recall rate of the detection system is 100%, which is excellent; the precision rate of Zone-C is 67%, which indicates some false positives; it is recommended to optimize the judgment threshold for images similar to R_B040 in the S11 algorithm to improve detection efficiency." S133 outputs a comprehensive batch defect identification system, which includes two core parts: a quality feedback system that automatically sends alarms to the quality management system (QMS), isolates supplier A's batch LOT20251015, and suggests tracing its incoming material storage and SMT loading processes to locate the link that caused the physical damage; and a system self-optimization system that adds false alarm cases from Zone-C to the retraining dataset of the S11 algorithm model and automatically adjusts relevant thresholds to achieve self-learning and continuous improvement of the system.

[0059] refer to Figure 5 In step S14, the specific steps are as follows: S141: Real-time monitoring of batch defect identification system, dynamic identification of batch defect identification system, determination of multiple classification features during the identification process, determination of corresponding classification information based on multiple classification features, corresponding resistors and the final defect level of the resistor; S142: Based on the identification of the classification information, multiple classification items are determined. Multiple resistor combinations are determined according to the multiple classification items, the corresponding classification models and the corresponding resistors. The multiple resistor combinations are transferred to the corresponding resistor maintenance station. At this time, each resistor combination is centrally maintained at the resistor maintenance station, and the corresponding central maintenance content is output.

[0060] In the embodiments of this application, a batch defect identification system is monitored in real time, and the batch defect identification system is dynamically identified. During the identification process, multiple classification features are determined. Based on the multiple classification features, the corresponding resistors, and the final defect level of the resistors, the corresponding classification information is determined. This approach takes into account the overall consideration of multiple classification features, the corresponding resistors, and the final defect level of the resistors, ensuring the accuracy of the corresponding classification information.

[0061] At this point, after the analysis is completed, a structured event is sent to the message queue or event bus, and S141, as a subscriber, will immediately receive and trigger it. After receiving the event, the parsing engine decodes the batch defect identification system report, associates the results of quality clustering analysis and detection process effectiveness analysis with all resistance data of the current batch in memory, and constructs a complete, multi-dimensional analysis context.

[0062] Refined and standardized classification features are extracted from complex analysis reports. These features are divided into four categories: quality features from S132 (final defect level, type, quantification value), source features from S111 (supplier ID, production batch), spatial clustering features from the first level of analysis in S133 (whether it is located in a clustering area, clustering area ID, dominant defect type), and system performance features from the second level of analysis in S133 (whether it is a missed case in S11, accuracy of the detection area).

[0063] A comprehensive judgment is made using a rule base driven by business logic. The rule base consists of complex IF-THEN statements that support multiple conditional logic combinations. For example, rule 1: if the level is C and the type is pin coplanarity out of tolerance, it is classified as repairable - pin straightening. Rule 2: if it is located in a cluster area and the dominant defect is body damage, it is classified as batch isolation - high risk. The rule engine loads the feature vector of each resistor and performs high-speed matching. Once the condition is met, the conclusion is triggered, and finally a structured classification information object is generated as the direct input of S142.

[0064] Specifically, after S133 completes the analysis and publishes the report, S141 is immediately triggered; the parsing engine decodes the report, identifies two defect clusters and the low accuracy of Zone-C, and associates this information with the final grade and source batch of all resistors; the system extracts features for key resistors: R_B012's features are {Grade: 'C', Type: 'Pin Coplanarity Exceeds Tolerance', Batch: 'LOT20251015', Located in Cluster: True, ID: 'Hotspot-01'}; R_B025's features are {Grade: 'D', Type: 'Body Damage', Batch: 'LOT20251015', Located in Cluster: True, ID: 'Hotspot-02'}; R_B040's features are {Grade: 'B', Type: 'Surface Scratches', Region Accuracy: 0.67}.

[0065] The decision engine matches each resistor according to a specific rule: R_B012 is a feature-matching rule, and its classification information is determined as {Resistor ID:'R_B012', Classification information:'Repairable - Pin Straightening', Priority:'Medium'}; R_B025 matches another rule, with classification information as {Resistor ID:'R_B025', Classification information:'Scrap - Performance Defect', Priority:'High'}; R_B040 triggers a more complex rule because its level is B but the accuracy of its region is low, and it is classified as {Resistor ID:'R_B040', Classification information:'Manual Re-inspection - Suspected False Alarm', Priority:'Low'}.

[0066] Furthermore, based on the identification of this classification information, multiple classification items are determined. Based on the multiple classification items, the corresponding classification models, and the corresponding resistors, multiple resistor combinations are determined and transferred to the corresponding resistor maintenance station. At this time, each resistor combination is centrally maintained at the resistor maintenance station, and the corresponding central maintenance content is output. This takes into account the overall consideration of multiple classification items, corresponding classification models, and corresponding resistors, ensuring the accuracy of multiple resistor combinations.

[0067] At this point, the system groups the resistor IDs using the classification information field as the key. Each aggregated group is instantiated as an independent classification item and assigned a unique, machine-readable classification model (e.g., PROC-C-001 represents a standard pin straightening procedure). The system generates a structured work order for each classification item, containing the work order ID, project name, classification model, list of resistors to be processed, priority, and estimated working hours.

[0068] After receiving the work order, the main control system queries the physical coordinates of each resistor and plans the optimal picking and placing path for the automated material handling system (such as the SCARA maintenance robot). The path planning algorithm comprehensively considers the shortest time, collision avoidance and maintenance robot constraints to ensure high efficiency and safety. After receiving the instruction, the maintenance robot controller picks up the resistors from the tray and places them on the conveyor belt or fixture leading to the target maintenance station in sequence.

[0069] When the resistor reaches the designated maintenance station, the PLC identifies the work order and obtains the classification model through the sensor, and then calls the preset control program to drive the actuator. For example, the pin straightening station will control the fixture and servo probe to perform precise straightening according to the preset displacement-force curve. During the execution, the torque, displacement and other sensors provide real-time feedback data, and the PLC performs closed-loop control based on this to ensure the consistency and reliability of the operation.

[0070] After maintenance is completed, the integrated testing unit at the workstation immediately performs a rapid re-inspection of the resistor. For straightened resistors, the system will remeasure the coplanarity of their pins. For scrapped products, the integrity of the markings will be visually confirmed. Regardless of the result, the system will generate a detailed centralized maintenance report, which includes work order traceability information, key parameters before and after maintenance, process parameters, and final result judgment. This report is permanently bound to the resistor's unique ID, enabling full lifecycle data traceability from defect discovery to maintenance completion.

[0071] Specifically, the system found that R_B012 and R_B013 were both classified as repairable - pin straightening, aggregated them into classification item-01, assigned classification model PROC-C-001, and generated work order-01, which included a resistor list and priority; similarly, R_B025 and R_B026 were aggregated into scrap items, and work order-02 was generated.

[0072] The main control system instructs the SCARA maintenance robot to execute work order-01; the maintenance robot queries the coordinates of R_B012 and R_B013, plans the optimal Z-shaped picking path, and sequentially picks up the two resistors and precisely places them on the conveyor belt leading to maintenance station-01 (pin straightening machine); when the resistors arrive at maintenance station-01, the PLC reads the PROC-C-001 instruction; the program starts: the fixture fixes R_B012, the displacement sensor positions the pin, and the servo probe completes the straightening with precisely controlled force and speed. The entire process is fed back in real time by the force sensor, realizing closed-loop control.

[0073] After straightening, the 3D camera at the workstation re-inspects R_B012 and calculates that its pin coplanarity has decreased from 0.12mm before maintenance to 0.02mm. The system generates a detailed maintenance report, recording the work order ID, parameters before and after maintenance, process parameters, and success determination. This report is permanently bound to the ID of R_B012, and its status is updated from C-level to A-level to A-level, which can be re-entered into the good product warehouse, realizing unmanned, high-precision, and fully traceable intelligent maintenance.

[0074] refer to Figure 6 In step S15, the specific steps are as follows: S151: Collect centralized maintenance content, determine multiple maintenance dimensions based on the identification of centralized maintenance content, and determine the first level of maintenance content based on multiple maintenance dimensions and corresponding resistors; S152: Determine the second level of maintenance content based on multiple maintenance dimensions and the corresponding defect areas of the resistors. Determine the maintenance worksheet of the maintenance robot based on the first and second level of maintenance content. At this time, the maintenance robot performs dynamic maintenance on the resistors at the resistor maintenance station. S153: Inspect the maintenance worksheet and identify multiple maintenance items during the inspection process. Determine the corresponding dynamic maintenance framework based on the multiple maintenance items and the final defect level of each resistor. Determine the dynamic maintenance system of the maintenance robot for multiple resistors based on the dynamic maintenance framework and the defect area of ​​each resistor. Trigger the maintenance robot to perform targeted maintenance actions on each resistor based on the dynamic maintenance system.

[0075] In the embodiments of this application, centralized maintenance content is collected, multiple maintenance dimensions are determined based on the identification of centralized maintenance content, and the first level of maintenance content is determined according to the multiple maintenance dimensions and the corresponding resistance. This approach takes into account the overall consideration of multiple maintenance dimensions and the corresponding resistance, ensuring the accuracy of the first level of maintenance content.

[0076] At this point, the system collects all centralized maintenance content reports through API or database queries, and performs field-by-field parsing based on the predefined data model; the parsing engine extracts key information entities such as core identifiers, maintenance types, quantification parameters, process parameters and result status, and loads them into the memory data structure, indexed by resistor ID or work order ID.

[0077] The system inputs the parsed maintenance type into the ontology for semantic analysis and reasoning, thereby determining a set of standardized maintenance dimensions that describe the essence of the maintenance task. These dimensions include: operation dimensions that define the type of action (such as straightening and welding), target dimensions that define the desired state change (such as geometric restoration), precision dimensions that define the level of fineness (such as micrometer level), and constraint dimensions that define physical limitations (such as maximum applied force and maximum temperature).

[0078] The system binds multiple maintenance dimensions to the ID of the resistor to be processed, forming the first layer of maintenance content. It is not a specific sequence of actions, but a descriptive strategy document that provides principles and boundaries for S152 and S153. Its structure is usually a set of key-value pairs that clearly defines the maintenance strategy for the specific resistor, such as operation, objective, accuracy and specific constraint parameters.

[0079] Specifically, the system collects the maintenance report of R_B012, and the parsing engine extracts key information: {Resistor ID:'R_B012', Maintenance type:'Pin straightening', Pre-maintenance parameters:'Coplanarity 0.12mm', Process parameters:'Peak force 2.0N', Result:'Success'}. The system inputs the pin straightening data into the maintenance ontology for querying. The ontology returns the associated standardized dimensions: the operation dimension is straightening, the target dimension is geometric restoration, the accuracy dimension is micrometer level, and the constraint dimension includes safety thresholds such as {Maximum applied force:'3N', Maximum plastic deformation:'0.2mm'}.

[0080] The system associates the instance of R_B012 with the aforementioned dimensions, generating the first layer of maintenance content for it. This policy document clarifies that a geometric restoration and straightening operation with micron-level precision must be performed, and strictly stipulates constraints such as the upper limit of force, the upper limit of temperature, and prohibition of applying pressure to the ceramic body. This constitution-level policy document will guide the path planning of S152 and the dynamic adjustment of S153, ensuring that all maintenance operations are carried out within a safe and controllable framework.

[0081] Furthermore, the second level of maintenance content is determined based on multiple maintenance dimensions and the corresponding defect areas of the resistors. The maintenance worksheet of the maintenance robot is determined based on the first and second level of maintenance content. At this time, the maintenance robot performs dynamic maintenance on the resistors at the resistor maintenance station, which takes into account the overall consideration of the first and second level of maintenance content and ensures the accuracy of the maintenance worksheet of the maintenance robot.

[0082] At this point, a high-precision vision system (such as structured light or laser profilometer) at the maintenance station will perform a rapid 3D scan. This process is not blind, but rather uses the panoramic image from S131 as prior knowledge to perform high-resolution scanning near the marked defect area, achieving sub-pixel-level precise positioning. After acquiring accurate 3D data, the system combines the maintenance dimensions from S151 to generate a set of quantified, spatialized second-level maintenance content, including contact point coordinates, tool posture, force direction, target displacement / force, motion control mode, and process parameters.

[0083] The high-level strategy of S151 (first layer) is integrated and verified with the specific action parameters of S152-1 (second layer) to generate a program that the maintenance robot can directly execute. The first layer plays the role of safety boundary here. For example, if the calculated applied force exceeds its constraint limit, the system will trigger an alarm or adjust the parameters. After verification, the system calls a motion planning algorithm (such as RRT*) to plan a collision-free path, taking into account the dynamics of the maintenance robot and the tooling model. Finally, a time-sequential maintenance worksheet is generated, which includes motion instructions, I / O control instructions, force control instructions and waiting instructions, for the maintenance robot controller to fully parse.

[0084] The maintenance robot controller loads and executes the maintenance worksheet, switching from pure position control to a force / position hybrid control mode during critical contact and operation steps. In this mode, the maintenance robot maintains position control in certain degrees of freedom, while switching to force control in other degrees of freedom (such as perpendicular to the workpiece surface). The wrist-mounted six-dimensional force / torque sensor provides real-time data feedback at a frequency of several kilohertz. The controller adjusts the maintenance robot's motion through algorithms such as PID to accurately maintain the target force or track the force curve. This allows the maintenance robot to sense contact and operate compliantly, rather than blindly applying force.

[0085] Specifically, R_B012 is fixed at the maintenance station, and a laser profilometer performs a rapid scan. The system loads its panoramic image, accurately locates the defect in pin 2, and calculates that it bends 0.12mm in the +Y direction at a distance of 1.2mm from the root. Combining the straightening dimension, the system generates the second layer of maintenance content: {Contact point coordinates: (15.3, 20.1, 5.5), Tool posture: (0°, 0°, 90°), Force direction: (0, -1, 0), Target displacement: 0.12mm, Control mode: 'Force / position hybrid'}.

[0086] The system integrates and verifies the above parameters with the first constraint of S151 (maximum force < 3N) to confirm safety; the motion planner generates a collision-free path and a maintenance worksheet, instructing the maintenance robot to: ① move above the contact point; ② descend to the contact point; ③ activate force control mode to apply force along the -Y direction, monitoring in real time that the force does not exceed 3N; ④ retreat to a safe point.

[0087] When the maintenance robot executes the worksheet, the force sensor reading changes when the probe contacts the pin, and the controller immediately switches to force control mode. The maintenance robot then begins to push the pin along the -Y direction, dynamically maintaining a smooth force increase and monitoring the displacement in real time. When the displacement reaches 0.12mm, the maintenance robot stops and holds for 0.5 seconds. If the force reading approaches 3N, the maintenance robot will immediately sound an alarm and stop to prevent damage to the pin, perfectly demonstrating adaptive dynamic maintenance based on real-time feedback.

[0088] Therefore, the maintenance worksheet is inspected, and multiple maintenance items are identified during the inspection process. Based on these maintenance items and the final defect level of each resistor, a corresponding dynamic maintenance framework is determined. Based on this dynamic maintenance framework and the defect areas of each resistor, a dynamic maintenance system for the maintenance robot is established for the multiple resistors. The maintenance robot then performs targeted maintenance actions on each resistor based on this dynamic maintenance system. This approach incorporates a holistic consideration of the dynamic maintenance framework and the defect areas of each resistor, ensuring the accuracy of the dynamic maintenance system for the multiple resistors. Simultaneously, it controls the centralized maintenance content, achieving a holistic consideration of the multiple maintenance items in the maintenance worksheet, the final defect level of each resistor, and the corresponding defect areas, thus improving the accuracy of the dynamic maintenance system for the multiple resistors.

[0089] At this point, the system loads the worksheet into the high-fidelity digital twin environment of the maintenance station and runs the program once to detect collisions, reachability, singularities, and dynamic feasibility. After the pre-run, the system decomposes the macroscopic maintenance worksheet into a series of discrete, semantically clear atomic maintenance items, such as approaching the target, contact detection, and applying force. Each item represents an indivisible basic action stage, laying the foundation for subsequent refined monitoring.

[0090] A decision-making model is constructed to determine the success or failure of each atomized item and formulate corresponding strategies during execution. This framework is based on two core inputs: the atomized maintenance item (defining the current action) and the final defect level (providing the background history of the object). At the core of the framework is an expert system or reinforcement learning model containing a knowledge base consisting of context-sensitive IF-THEN-ELSE rules. For example, for a Class C straightening task, the rule definition is: if the force reading suddenly drops by more than 20% when force is applied, it is determined that the pin is broken and the operation is stopped immediately. The system loads the matching rule set according to the current defect level of the resistor, combines it with the atomized item, and forms a complete dynamic maintenance framework for a specific resistor.

[0091] The dynamic maintenance system means that when the maintenance robot executes the worksheet, each of its atomic items is under the real-time monitoring of the dynamic maintenance framework; multi-sensor data streams from force sensors, encoders and vision systems are continuously fed into the framework; once the sensor data matches the conditions in the rule base, the framework will immediately trigger targeted maintenance actions, such as parameter adjustment, action retry or immediate abort, and interrupt or overwrite the current program with high priority, while updating the resistor status and material flow path.

[0092] Specifically, after the straightening procedure of R_B012 was pre-rehearsed in the digital twin, it was decomposed into five atomic maintenance items: approaching the target, contact detection, applying force, maintaining stress, and withdrawing. The system identified R_B012 as a Class C defect, loaded a rule set for Class C straightening, and constructed a dynamic maintenance framework. The framework includes a key rule: if the force sensor reading suddenly drops by more than 20% during the force application phase, it is determined that the pin is broken and the task is immediately terminated.

[0093] The dynamic maintenance system is activated, and the maintenance robot begins execution. After completing contact detection, it enters the force application phase. When the force reading reaches 1.8N, it suddenly drops to 0.05N. The dynamic maintenance framework captures this anomaly in real time, matches the rule conditions, and immediately triggers a high-priority interrupt command. The maintenance robot instantly stops all movement and quickly retracts. At the same time, the system updates the status of R_B012 to F-level - maintenance failure (pin breakage) and automatically modifies its logistics path, directing it to the waste bin. Through S153, the system successfully transforms a potential equipment damage or defective product event into a precise and automated fault diagnosis and handling process, demonstrating true intelligence and robustness that surpasses traditional automation.

[0094] Please see Figure 7 , Figure 7 This is a schematic diagram of the structural composition of a visual inspection-based resistor defect identification system according to an embodiment of the present invention; the visual inspection-based resistor defect identification system includes: The primary defect level module 21 is used to determine the information combination of each resistor based on the traceability of the model information of each resistor when multiple resistors are in the visual inspection station, and to determine the primary defect level of the resistor based on the identification of the information combination of each resistor. The regional detection module 22 is used to determine multiple visual detection areas based on the primary defect level and corresponding position of each resistor, trigger regional detection according to each visual detection area and multiple resistors located in that visual detection area, and determine multiple images of each resistor in different directions. The batch defect identification system module 23 is used to determine the panoramic image of the resistor based on multiple images of the resistor in different directions, the corresponding primary defect level and the corresponding information combination, determine the final defect level of the resistor based on the image recognition of the panoramic image of the resistor, and determine the batch defect identification system based on the final defect level of each resistor and the corresponding visual detection area. The centralized maintenance content module 24 is used to determine the corresponding classification information based on the batch defect identification system, determine multiple resistor combinations based on the classification information and the corresponding multiple resistors, and determine the centralized maintenance content according to each resistor combination and the corresponding resistor maintenance station. The dynamic maintenance system module 25 is used to determine the maintenance worksheet of the maintenance robot based on the centralized maintenance content, the corresponding resistor and the defect area of ​​the resistor, and to determine the dynamic maintenance system of the maintenance robot for multiple resistors based on the multiple maintenance items of the maintenance worksheet, the final defect level of each resistor and the corresponding defect area.

[0095] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for defect identification of resistors based on visual inspection, characterized in that, include: When multiple resistors are in the visual inspection station, the information combination of each resistor is determined by tracing the model information of each resistor, and the primary defect level of the resistor is determined based on the identification of the information combination of each resistor. Multiple visual inspection areas are determined based on the primary defect level and corresponding location of each resistor. Regional inspection is triggered based on each visual inspection area and multiple resistors located in that visual inspection area, and multiple images of each resistor in different directions are determined. A panoramic image of the resistor is determined based on multiple images of the resistor in different directions, the corresponding primary defect level, and the corresponding information combination. The final defect level of the resistor is determined based on the image recognition of the panoramic image of the resistor. A batch defect recognition system is determined based on the final defect level of each resistor and the corresponding visual detection area. A panoramic image is a four-dimensional information body containing geometry, texture, defect semantics, and historical records; this batch defect identification system includes a quality feedback system and a system self-optimization system. Based on the identification of the batch defect identification system, the corresponding classification information is determined. Based on the classification information and the corresponding multiple resistors, multiple resistor combinations are determined. Based on each resistor combination and the corresponding resistor maintenance station, the centralized maintenance content is determined. Based on the centralized maintenance content, the corresponding resistors, and the defect areas of those resistors, a maintenance worksheet for the maintenance robot is determined. Based on the multiple maintenance items in this worksheet, the final defect level of each resistor, and the corresponding defect areas, a dynamic maintenance system for multiple resistors by the maintenance robot is determined.

2. The method for defect identification of resistors based on visual inspection according to claim 1, characterized in that, When multiple resistors are at a visual inspection station, the information combination of each resistor is determined by tracing the model information of each resistor, and the primary defect level of the resistor is determined based on the identification of the information combination of each resistor, including: Multiple resistors are fed into the same vision inspection station via a conveyor line and are simultaneously captured by multiple cameras at the vision inspection station. At the same time, the model information of each resistor is marked, the model information of each resistor is traced, and the usage information and processing information of each resistor are output. Based on the usage information and processing information of each resistor, the information combination of the resistor is constructed. In a series of resistors, the information combination of each resistor is dynamically identified, and multiple defect information is determined during the identification process. Based on the multiple defect information, the corresponding defect location and the manufacturing process of the resistor, the primary defect level of the resistor is determined.

3. The method for defect identification of resistors based on visual inspection according to claim 1, characterized in that, The process involves determining multiple visual inspection areas based on the primary defect level and corresponding location of each resistor, triggering regional detection based on each visual inspection area and multiple resistors located within that area, and determining multiple images of each resistor in different directions, including: Mark the position of each resistor relative to the visual inspection station, determine multiple inspection contour lines based on the primary defect level of each resistor and its corresponding position, and determine the corresponding visual inspection area based on the shape of the multiple inspection contour lines and the corresponding multiple resistors, so as to mark multiple visual inspection areas. In multiple visual detection areas, multiple resistors are determined based on the dynamic detection of each visual detection area to mark the multiple resistors in that visual detection area. Based on each visual detection area and the multiple resistors in that visual detection area, a corresponding detection signal is determined. The detection signal triggers regional detection of multiple resistors in the same visual detection area to achieve multi-directional detection of each resistor and to determine multiple images of each resistor in different directions.

4. The method for defect identification of resistors based on visual inspection according to claim 1, characterized in that, The process involves determining a panoramic image of the resistor based on multiple images of the resistor in different directions, the corresponding primary defect levels, and the corresponding information combinations; determining the final defect level of the resistor based on image recognition of the panoramic image; and determining a batch defect identification system based on the final defect levels of each resistor and the corresponding visual inspection areas. This includes: Among multiple resistors, the first panoramic shape of each resistor is determined based on multiple images of each resistor in different directions and the corresponding primary defect level, and the panoramic image of the resistor is determined based on the first panoramic shape of each resistor and the corresponding primary defect level.

5. The method for defect identification of resistors based on visual inspection according to claim 4, characterized in that, The process of determining a panoramic image of the resistor based on multiple images of the resistor in different directions, corresponding primary defect levels, and corresponding information combinations; determining the final defect level of the resistor based on image recognition of the panoramic image; and determining a batch defect identification system based on the final defect level of each resistor and its corresponding visual inspection area, further includes: Image recognition is performed on the panoramic images of each resistor, and multiple defect features are identified during the recognition process. The final defect level of the resistor is determined based on the combination of multiple defect features, information of each resistor, and the corresponding primary defect level. The orientation of each resistor is marked. The first batch defect identification content is determined based on the orientation of each resistor and the corresponding final defect level. The second batch defect identification content is determined based on the orientation of each resistor and the corresponding visual inspection area. The batch defect identification system is determined based on the first and second batch defect identification contents.

6. The method for defect identification of resistors based on visual inspection according to claim 1, characterized in that, The batch defect identification system identifies the corresponding classification information, and based on this classification information and the corresponding multiple resistors, determines multiple resistor combinations; based on each resistor combination and the corresponding resistor maintenance station, the centralized maintenance content is determined, including: The system monitors batch defect identification in real time, performs dynamic identification of batch defects, determines multiple classification features during the identification process, and determines the corresponding classification information based on multiple classification features, the corresponding resistor, and the final defect level of the resistor.

7. The method for defect identification of resistors based on visual inspection according to claim 6, characterized in that, The batch defect identification system identifies the corresponding classification information, and based on the classification information and the corresponding multiple resistors, determines multiple resistor combinations. The centralized maintenance content is determined based on each resistor combination and its corresponding maintenance station, and also includes: Based on the identification of this classification information, multiple classification items are determined. According to the multiple classification items, the corresponding classification models and the corresponding resistors, multiple resistor combinations are determined and transferred to the corresponding resistor maintenance station. At this time, each resistor combination is centrally maintained at the resistor maintenance station, and the corresponding central maintenance content is output.

8. The method for defect identification of resistors based on visual inspection according to claim 1, characterized in that, The maintenance worksheet for the maintenance robot is determined based on the centralized maintenance content, the corresponding resistor, and the defect area of ​​the resistor. Based on the multiple maintenance items in this worksheet, the final defect level of each resistor, and the corresponding defect area, a dynamic maintenance system for the maintenance robot for multiple resistors is determined, including: Collect centralized maintenance content, determine multiple maintenance dimensions based on the identification of centralized maintenance content, and determine the first level of maintenance content based on multiple maintenance dimensions and corresponding resistors; The second level of maintenance content is determined based on multiple maintenance dimensions and the corresponding defect areas of the resistors. The maintenance worksheet of the maintenance robot is determined based on the first and second level of maintenance content. At this time, the maintenance robot performs dynamic maintenance on the resistors at the resistor maintenance station.

9. The method for defect identification of resistors based on visual inspection according to claim 8, characterized in that, The process of determining a maintenance worksheet for the maintenance robot based on centralized maintenance content, corresponding resistors, and defect areas of those resistors, and determining a dynamic maintenance system for multiple resistors by the maintenance robot based on multiple maintenance items in the worksheet, the final defect level of each resistor, and the corresponding defect area, further includes: The maintenance worksheet is inspected, and multiple maintenance items are identified during the inspection process. Based on the multiple maintenance items and the final defect level of each resistor, a corresponding dynamic maintenance framework is determined. Based on the dynamic maintenance framework and the defect area of ​​each resistor, a dynamic maintenance system for the maintenance robot for multiple resistors is determined. Based on the dynamic maintenance system, the maintenance robot is triggered to perform targeted maintenance actions on each resistor.

10. A defect identification system for resistors based on visual inspection, characterized in that, The visual inspection-based resistor defect identification system is applied to the visual inspection-based resistor defect identification method as described in any one of claims 1-9.

Images