Multi-station cooperation and deep learning chip inductor appearance detection system and method

Abstract

The invention discloses a multi-station cooperation and deep learning chip inductor appearance detection system, which is characterized in that a feeding module is arranged at a feeding starting end of the appearance detection system and is connected with a glass turntable feeding area of a multi-light-source multi-station imaging module, a main control system and a preorder material circulation link of an intelligent execution sorting module; the sorting module is used for sorting disordered to-be-detected chip inductors into an ordered state and conveying the to-be-detected chip inductors to the imaging module; the core image processing and decision-making module comprises an image preprocessing unit, a feature extraction and fusion unit, a deep learning-based target detection and classification network and a data driving parameter optimization model, and is used for completing image processing, defect identification and parameter optimization; and the intelligent execution sorting module is used for distributing the elements into corresponding bins through air valves according to the instruction of the core decision module. The method has the beneficial effects of greatly shortening the debugging time, improving the production flexibility, reducing the cost, improving the efficiency and the quality and the like.

Description

Technical Field

[0001] This invention relates to the field of industrial product quality inspection technology, and more specifically, to an intelligent detection system and method for appearance defects of chip inductors based on multi-station collaboration and deep learning classification. Background Technology

[0002] As core components of modern electronic circuits, the appearance quality of micro-components such as chip inductors and ferrite beads directly determines the reliability and production yield of electronic devices. Appearance defects such as cracks, missing corners, poor coatings, electrode oxidation, and out-of-tolerance dimensions can lead to abnormal electrical performance of components and even cause circuit failures. Therefore, accurate and efficient appearance inspection of chip inductors is a key link in quality control in the electronics manufacturing industry and a prerequisite for ensuring the quality of end products. In the past, the appearance inspection of chip inductors was mainly carried out by manual microscopes. Inspectors would observe the appearance of the components with the naked eye to judge whether there were defects. With the development of automation technology, appearance inspection equipment based on machine vision has emerged on the market. These devices achieve automated inspection by combining image acquisition with traditional algorithms. They mainly rely on preset thresholds and rules such as brightness, area, and aspect ratio to judge defects. Some devices have basic sorting functions for good and bad products, which have replaced manual operation to a certain extent and improved inspection efficiency. However, the existing detection methods still have certain drawbacks. Specifically, manual inspection is inefficient, labor-intensive, and its results are significantly affected by subjective factors such as operator experience and fatigue, leading to frequent missed detections. Traditional machine vision-based inspection equipment suffers from high error rates; traditional algorithms are sensitive to changes in lighting, slight product movement, and background interference, resulting in frequent over- and under-detection issues, failing to meet high-precision inspection requirements. Most equipment lacks accurate defect classification capabilities, only distinguishing between good and bad products, unable to accurately identify specific defect types such as cracks, scratches, and electrode oxidation, hindering subsequent process traceability and quality analysis. Furthermore, existing technologies lack adaptability and intelligence. When changing product models, engineers must manually adjust numerous optical and algorithm parameters, resulting in complex debugging processes, high operator skill requirements, and a lack of self-learning and optimization capabilities. System stability is easily affected by environmental factors; lens dust, glass scratches, and changes in lighting significantly reduce detection stability, requiring frequent manual maintenance and calibration, increasing production control costs.

[0003] Therefore, those skilled in the art are dedicated to providing a multi-station collaborative and deep learning-based chip inductor appearance inspection system and method that can effectively solve the above-mentioned technical problems. Summary of the Invention

[0004] To achieve the above objectives, this invention provides a multi-station collaborative and deep learning-based surface mount inductor appearance inspection system, comprising a feeding module, a multi-source multi-station imaging module, a core image processing and decision-making module, and an intelligent execution sorting module; this invention achieves the detection and sorting of surface mount inductor appearance defects through the collaborative operation of the feeding module, the multi-source multi-station imaging module, the core image processing and decision-making module, and the intelligent execution sorting module. The feeding module is set at the feeding start end of the appearance inspection system and is connected to the glass turntable feeding area of ​​the multi-source multi-station imaging module, the main control system and the preceding material flow link of the intelligent execution sorting module. It is used to organize the disordered chip inductors to be tested into an ordered state and transport them to the imaging module to provide a continuous and qualified material supply for subsequent multi-station image acquisition. The multi-source, multi-station imaging module includes a high-speed rotating glass turntable with at least four imaging stations evenly arranged around the circumference of the turntable. Each station is equipped with a high-speed industrial camera, which takes pictures of each element sequentially through a synchronous triggering mechanism. The core image processing and decision-making module includes an image preprocessing unit, a feature extraction and fusion unit, a deep learning-based target detection and classification network, and a data-driven parameter optimization model, which completes image processing, defect identification, and parameter optimization. The intelligent sorting module, based on the instructions of the core decision-making module, uses high-precision air valves to distribute components into the corresponding bins and integrates with the MES system to achieve closed-loop management of quality data.

[0005] Furthermore, the feeding module includes a vibratory feeder, a linear feeder, a guide block, and a parameter adaptive adjustment unit; The vibratory feeder is located at the starting end of the feeding module and is connected to the linear feeder and the parameter adaptive adjustment unit. It is used to receive the chip inductors to be tested, and through vibration, it organizes the disordered chip inductors into an ordered state and conveys them to the linear feeder. The linear feeder is set on the material transfer path between the vibratory feeder and the glass turntable. It is connected to the discharge end of the vibratory feeder, the guide block and the parameter adaptive adjustment unit. It is used to receive the ordered chip inductors from the vibratory feeder and to smoothly and evenly transport the components to the guide block through linear vibration, and then accurately feed them into the glass turntable of the multi-light source multi-station imaging module. The guide block is positioned between the linear feeder and the glass turntable of the multi-source multi-station imaging module. It connects to the discharge end of the linear feeder and the feeding area of ​​the glass turntable. It is used to receive the chip inductors conveyed by the linear feeder. Through its own preset guide channel structure, it guides the components to the designated bearing position on the glass turntable, ensuring that the components are stable in posture and uniform in position before entering the imaging station.

[0006] The glass turntable is used to carry the chip inductors conveyed by the guide block. The high-speed rotation drives the components to pass through four imaging stations in sequence, and the synchronous triggering device ensures that each component stays at the corresponding station. The parameter adaptive adjustment unit is connected to the vibratory feeder, linear feeder, main control system, and material feeding status monitoring component. It is used to receive the specification instructions of the chip inductor to be tested issued by the main control system, automatically call the preset combination of vibratory feeder frequency, linear slot frequency, and feeding speed parameters; at the same time, it collects material feeding status data in real time; if abnormalities such as jamming or dropping are detected, it immediately fine-tunes the feeding parameters to ensure feeding stability.

[0007] Furthermore, the four imaging stations of the multi-source multi-station imaging module are respectively configured with the following different light sources: Workstation 1: A ring-shaped white LED light source is used to vertically illuminate and collect an overall image of the top surface of the component, which is used to detect surface dirt, damage, and marking errors. Workstation 2: Uses red LED light source, low-angle sweeping light to highlight scratches, dents, and flatness defects on the electrode ends; Workstation 3: A backlit light source is used to acquire silhouette images of component outlines for dimensional measurement; Station 4: A polarized light source is used to eliminate reflective interference and clearly image the electrode coating condition, detecting defects such as oxidation and incomplete coating.

[0008] The multi-source, multi-station imaging module also includes a dynamic light source spectrum adjustment unit, which is used to switch the light source spectrum band according to the material and surface coating type of the chip inductor to be tested, and to form a dual light source combination of basic light source and spectrum-adaptive light source in conjunction with the light sources of station one, station two, station three and station four.

[0009] Furthermore, in the core image processing and decision-making module: The image preprocessing unit is used to perform filtering, denoising, enhancement, and correction operations on multiple acquired images to standardize the images. The feature extraction and fusion unit performs registration and feature layer fusion on the image information from the four workstations to form a comprehensive feature vector that includes information on the top surface, side surface, contour, and coating. Deep learning-based target detection and classification networks use improved YOLOv5, SSD, or Faster R-CNN frameworks as inputs, taking multi-station images as different channels of the same target and outputting defect category and location information.

[0010] Furthermore, the data-driven parameter optimization model can record the operator's correction data on the interactive interface, and use this data to fine-tune the deep learning model and imaging parameters, so as to realize the system's self-evolution and reduce the reliance on manual debugging by engineers. Defect categories include cracks, electrode scratches, dimensional deviations, and poor coating; The core image processing and decision-making module includes a defect severity grading unit, which establishes a three-level grading standard based on defect size, location, and impact range. The output results include defect category, location, and severity level.

[0011] Furthermore, the intelligent sorting module includes a good product bin and a defective product bin. The system counts the quantity in each bin in real time. When the quantity in the defective product bin exceeds a threshold and the yield rate is lower than a set value, a feedback form is automatically generated and a rework prompt is given. Simultaneously, the intelligent sorting module also includes a secondary verification station for defective products. A high-speed camera performs secondary imaging detection on components initially identified as defective, comparing the results with the initial detection. If they match, they are sent to the defective product bin; if they do not match, they are marked as suspicious and sent to a dedicated suspicious product bin for manual verification.

[0012] Furthermore, the deep learning network of the core image processing and decision-making module is trained with multi-station image samples, enabling it to learn both global and local features of components simultaneously. The detection rate and recognition accuracy for complex and minute defects are higher than those of traditional threshold methods. The appearance inspection system also includes a human-computer interaction interface for displaying real-time images, defect classification results, statistical charts, and parameter self-optimization options.

[0013] Furthermore, the intelligent execution sorting module also includes a defect tracing data association unit; The defect tracing data association unit is connected to the core image processing and decision-making module and the production process data module of the MES system. When the intelligent sorting module sends defective products into the corresponding silo, the defect tracing data association unit automatically captures the detection data of the defective product and matches it with the production data of the component in the MES system to generate a defect production association dataset. The association dataset can support reverse querying of the process parameters of the corresponding production batch by defect type, or querying the defect rate change trend of the products produced by the production equipment. When the defect rate of a certain production equipment is higher than the average value for a preset time, the equipment process abnormality prompt is given.

[0014] Furthermore, the defect tracing data association unit also includes a time axis calibration submodule and a multi-dimensional attribution analysis submodule; The time axis calibration submodule is connected to the end sensor of the production line and the loading sensor of the detection system. It is used to collect the transmission time of the chip inductor from the production line to the loading of the detection system in real time, shift the imaging time of the defective product detection data backward by the corresponding transmission time to generate the calibration time, and then perform correlation and matching with the production data of the MES system based on the calibration time and the material batch code. The multi-dimensional attribution analysis submodule is connected to the MES system's full-chain production database and raw material supplier quality database. It is used to retrieve related data from the above databases based on the basic inspection data of defective products, forming a three-dimensional related dataset including defects, production and supply chain. The multi-dimensional attribution analysis submodule has a built-in association rule mining model. By analyzing the three-dimensional association dataset, it automatically outputs the probability ranking of the root causes of defects. When the defect rate of a certain production equipment is higher than the system average for a preset time, or when the defect rate of a batch supplied by a certain raw material supplier is higher than the preset value for several consecutive batches, the defect tracing data association unit will not only output equipment process abnormality prompts or supplier quality warnings, but also push specific abnormal data at the same time.

[0015] A multi-station collaborative and deep learning-based method for inspecting the appearance of chip inductors includes the following steps: S1: First, the feeding module starts working. The vibratory feeder receives the disordered chip inductors to be tested, and through vibration, it organizes them into an orderly state and conveys them to the linear feeder. The linear feeder smoothly conveys the components to the guide block in a linear vibration manner. The guide block then guides the components precisely to the designated bearing position on the glass turntable of the multi-light source multi-station imaging module through a preset guide channel, so that the components are stable in posture and uniform in position. At the same time, the parameter adaptive adjustment unit of the feeding module will automatically call the preset combination of vibratory feeder frequency, straight groove frequency and feeding speed parameters according to the specifications of the components to be tested. It will also collect the feeding status data in real time. If abnormalities such as jamming or dropping are detected, the feeding parameters will be finely adjusted in time to ensure continuous and stable feeding. S2: The glass turntable rotates, causing the orderly loaded components to pass through four imaging stations in sequence. The synchronous triggering device ensures that each component stops precisely at its corresponding station and completes the imaging. Station 1 uses a ring-shaped white LED light source to vertically illuminate and collect an overall image of the top surface of the component, used to detect surface dirt, damage, and marking errors. Station 2 uses a multi-angle red LED light source to low-angle sweep and collect images of the electrode ends, highlighting scratches, dents, and flatness defects. Station 3 uses a high-uniformity backlight light source to obtain a silhouette image of the component outline, used for dimensional measurement. Station 4 uses an adjustable polarized light source to eliminate reflection interference and collect images of the electrode coating to detect defects such as oxidation and incomplete plating. S3: The core image processing and decision-making module processes the four sets of acquired images. First, the image preprocessing unit performs filtering, denoising, enhancement, and correction operations to unify image standards. Then, the feature extraction and fusion unit registers and fuses the four images to form a comprehensive feature vector that includes information on the top surface, side surface, contour, and coating. Based on a deep learning target detection and classification network with improved YOLOv5, SSD, or Faster R-CNN, the multi-station images are used as different channels of input for the same target, and the output is the defect category and location information. At the same time, the data-driven parameter optimization model records the correction data of the operator in the interactive interface. This data is used to fine-tune the deep learning model and imaging parameters to achieve system self-evolution and reduce reliance on manual debugging. S4: The intelligent execution sorting module, based on the instructions output by the core decision module, uses high-precision air valves to send components into the good product warehouse and the defective product warehouse respectively, and counts the quantity in each warehouse in real time. When the quantity in the defective product warehouse exceeds the threshold and the yield rate is lower than the set value, a quality information feedback form is automatically generated and a rework is prompted. At the same time, the defect traceability data association unit starts working. First, the time axis calibration submodule collects the transmission time of the components from production line to the detection system in real time, shifts the imaging time of defective product detection data backward by the corresponding time to generate a calibration time, and then combines the material batch code with the production data in the MES system for association and matching to generate a defective production association dataset. S5: The multi-dimensional attribution analysis submodule further retrieves the associated data from the MES system's full-link production database and raw material supplier quality database to form a three-dimensional associated dataset containing defects, production, and supply chain. After analysis by the built-in association rule mining model, it automatically outputs the probability ranking of the root causes of defects. If the defect rate of a certain production equipment is higher than the average value for a continuous preset time, or if the defect rate of a certain raw material supplier's batch is higher than the preset value for several consecutive batches, it will simultaneously push equipment process abnormality prompts or supplier quality warnings and specific abnormal data, ultimately realizing the appearance inspection of chip inductors.

[0016] The present invention has the following beneficial effects: 1. The multi-source, multi-station imaging module in this invention acquires images from multiple dimensions, including the top surface, electrode terminals, contours, and coatings, through four differentiated light sources and station designs, comprehensively capturing the appearance features of components. The core image processing and decision module performs registration and feature layer fusion on the multi-station images. Combined with improved YOLOv5, SSD, or Faster R-CNN deep learning networks, it can simultaneously learn global and local features. The accuracy of identifying complex and minute defects such as cracks and electrode scratches far exceeds that of traditional threshold methods. The filtering and denoising operations in image preprocessing are further standardized, effectively resisting the effects of changes in lighting and background interference, and solving the problem of frequent over-inspection and under-inspection in traditional equipment.

[0017] 2. In this invention, the system can not only distinguish between good and bad products, but also accurately output specific defect categories and location information such as cracks, electrode scratches, dimensional deviations, and coating defects. It intelligently executes the defect source tracing data association unit of the sorting module, and through time axis calibration and multi-dimensional attribution analysis, it associates defect detection data with MES system production data and raw material supplier quality data to form a three-dimensional associated dataset. It can reverse query the process parameters of the production batch corresponding to the defect, track the trend of equipment defect rate changes, automatically output the probability ranking of the root cause of defects, and simultaneously push equipment anomaly or supplier quality warnings.

[0018] 3. The parameter adaptive adjustment unit of the feeding module can automatically call preset parameter combinations according to component specifications, monitor and fine-tune anomalies such as jamming and material dropping in real time, and ensure stable feeding without manual intervention. The data-driven parameter optimization model can record operator correction data and automatically fine-tune the deep learning model and imaging parameters, realizing system self-evolution. When changing product models, only the pre-stored scheme needs to be switched, eliminating the need for engineers to manually adjust a large number of optical and algorithm parameters, reducing the professional requirements for operators, significantly shortening debugging time, and improving production flexibility. At the same time, this system also has self-learning and optimization capabilities, and can continuously improve detection accuracy as production data accumulates.

[0019] 4. Multi-dimensional image information fusion improves the system's dependence on single imaging conditions, effectively resists interference from environmental factors such as lens dust, glass scratches, and changes in light, reduces the need for frequent manual maintenance and calibration, and the intelligent execution sorting module is integrated with the MES system to count the quantity of each warehouse in real time. When the defective products exceed the threshold or the yield rate does not meet the standard, a quality information feedback form is automatically generated and a rework prompt is given, further reducing costs.

[0020] 5. The high-speed rotating glass turntable, combined with a synchronous triggering mechanism, enables continuous and orderly multi-station imaging of components. This, along with the rapid sorting action of high-precision air valves, ensures a high inspection speed of 4000 pcs / min. The stable conveying and adaptive adjustment of the feeding module prevents production interruptions caused by material jams or drops. Furthermore, the inspection speed of 4000 pcs / min also improves efficiency and quality. Attached Figure Description

[0021] Figure 1 This is a schematic block diagram of the main control system structure in this invention.

[0022] Figure 2 This is a schematic block diagram of the workstation imaging detection logic in this invention.

[0023] Figure 3 This is a schematic flowchart of the appearance inspection process for chip inductors in this invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments: In the description of this invention, it should be noted that the terms "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "setting," and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] like Figures 1 to 3 As shown, a multi-station collaborative and deep learning chip inductor appearance inspection system includes a feeding module (high precision), a multi-source multi-station imaging module, a core image processing and decision-making module, and an intelligent execution sorting module. This invention uses a feeding module (high precision), a multi-source multi-station imaging module, a core image processing and decision-making module, and an intelligent execution sorting module to collaboratively complete the detection and sorting of appearance defects in chip inductors; The feeding module (high-precision feeding module) is set at the feeding start end of the appearance inspection system. It is connected to the glass turntable feeding area of ​​the multi-source multi-station imaging module, the main control system and the preceding material flow link of the intelligent execution sorting module. It is used to organize the disordered chip inductors to be tested into an ordered state and transport them to the imaging module to provide a continuous and qualified material supply for subsequent multi-station image acquisition. The multi-source, multi-station imaging module includes a high-speed rotating glass turntable with at least four imaging stations evenly arranged around the turntable. Each station is equipped with a high-speed industrial camera, which takes pictures of each component sequentially through a synchronous triggering mechanism. The detection speed can reach 4000pcs / min. The core image processing and decision-making module includes an image preprocessing unit, a feature extraction and fusion unit, a deep learning-based target detection and classification network, and a data-driven parameter optimization model, which completes image processing, defect identification, and parameter optimization. The intelligent sorting module, based on the instructions of the core decision-making module, uses high-precision air valves to distribute components into the corresponding bins and integrates with the MES system to achieve closed-loop management of quality data.

[0027] The feeding module (high precision) includes a vibratory feeder, a linear feeder, a guide block, and a parameter adaptive adjustment unit. The vibratory feeder is located at the starting end of the feeding module and is connected to the linear feeder and the parameter adaptive adjustment unit. It is used to receive the chip inductors to be tested, and through vibration, it organizes the disordered chip inductors into an ordered state and conveys them to the linear feeder. The linear feeder is set on the material transfer path between the vibratory feeder and the glass turntable. It is connected to the discharge end of the vibratory feeder, the guide block and the parameter adaptive adjustment unit. It is used to receive the ordered chip inductors from the vibratory feeder and to smoothly and evenly transport the components to the guide block through linear vibration, and then accurately feed them into the glass turntable of the multi-light source multi-station imaging module. The guide block is positioned between the linear feeder and the glass turntable of the multi-source multi-station imaging module. It connects to the discharge end of the linear feeder and the feeding area of ​​the glass turntable. It is used to receive the chip inductors conveyed by the linear feeder. Through its own preset guide channel structure, it guides the components to the designated bearing position on the glass turntable, ensuring that the components are stable in posture and uniform in position before entering the imaging station.

[0028] The glass turntable is used to carry the chip inductors conveyed by the guide block. The high-speed rotation drives the components to pass through four imaging stations in sequence, and the synchronous triggering device ensures that each component stays at the corresponding station. The parameter adaptive adjustment unit is connected to the vibratory feeder, linear feeder, main control system, and material feeding status monitoring components (such as jam sensors and drop detectors). It is used to receive the specification instructions (such as 0603, 0805, 1206) of the chip inductor to be tested issued by the main control system, and automatically call the preset combination of vibratory feeder frequency, linear slot frequency, and feeding speed parameters; at the same time, it collects material feeding status data in real time; if abnormalities such as jamming or dropping are detected, it immediately fine-tunes the feeding parameters to ensure feeding stability.

[0029] The four imaging stations of the multi-source multi-station imaging module are each equipped with the following different light sources: Workstation 1: A ring-shaped white LED light source is used to vertically illuminate and collect an overall image of the top surface of the component, which is used to detect surface dirt, damage, and marking errors. Workstation 2: Uses red LED light source (multi-angle red LED light source), low-angle sweeping light to highlight scratches, dents, and flatness defects of the electrode ends; Workstation 3: Using a backlight source (high uniformity backlight source), obtain the silhouette image of the component outline for dimensional measurement; Station 4: Uses a polarized light source (adjustable) to eliminate reflective interference and clearly image the electrode coating condition, detecting defects such as oxidation and incomplete plating; The multi-source, multi-station imaging module also includes a dynamic light source spectrum adjustment unit, used to switch the light source spectrum band according to the material (such as ceramic substrate, metal electrode) and surface coating type of the chip inductor under test, forming a dual-light source combination of a basic light source and a spectrally adapted light source in conjunction with the light sources at stations one, two, three, and four. Through the above settings, this invention further improves the imaging contrast of defects in different materials.

[0030] The core image processing and decision-making module includes: The image preprocessing unit is used to perform filtering, denoising, enhancement, and correction operations on multiple acquired images to standardize the images. The feature extraction and fusion unit performs registration and feature layer fusion on the image information from the four workstations to form a comprehensive feature vector that includes information on the top surface, side surface, contour, and coating. Deep learning-based target detection and classification networks use improved YOLOv5, SSD, or Faster R-CNN frameworks as inputs, taking multi-station images as different channels of the same target and outputting defect category and location information.

[0031] The data-driven parameter optimization model can record the correction data of the operator on the interactive interface. This data is used to fine-tune the deep learning model and imaging parameters, enabling the system to self-evolve and reducing the reliance on manual debugging by engineers. Defect categories include cracks, electrode scratches, size out of tolerance, and coating defects. The core image processing and decision-making module includes a defect severity grading unit, which establishes a three-level grading standard based on defect size, location, and impact range. The output includes defect category, location, and severity level. The three-level grading standard includes minor defects that do not affect electrical performance; moderate defects that may affect stability; and severe defects that directly cause functional failure. The intelligent sorting module includes a good product bin and a defective product bin, which can be further expanded into multiple dedicated bins categorized by defect type. The system counts the quantity in each bin in real time. When the quantity in the defective product bin exceeds a threshold and the yield rate is lower than a set value, a feedback form (quality information feedback form) is automatically generated and rework is prompted. The intelligent sorting module also includes a secondary defect verification station. A high-speed camera performs secondary imaging detection on components initially identified as defective, comparing the results with the initial detection. If they match, they are sent to the defective product bin; otherwise, they are marked as suspicious and sent to a dedicated suspicious product bin for manual verification.

[0032] The deep learning network of the core image processing and decision-making module is trained with multi-station image samples and can learn the global and local features of the components simultaneously. It has self-learning and optimization capabilities and can continuously improve the detection accuracy as production data accumulates. The detection rate and recognition accuracy of complex defects and minor defects are higher than those of traditional threshold methods. The appearance inspection system also includes a human-machine interface for displaying real-time images, defect classification results, statistical charts and parameter self-optimization options.

[0033] The intelligent execution sorting module also includes a defect tracing data association unit; The defect tracing data association unit is connected to the core image processing and decision-making module and the production process data module of the MES system, such as material batch, production equipment number, and processing time. When the intelligent execution sorting module sends defective products into the corresponding silo, the defect tracing data association unit automatically captures the detection data of the defective product, including but not limited to defect category, severity, imaging time, and detection station number, and matches it with the production data of the component in the MES system (including batch and time, etc.) to generate a defect production association dataset. The association dataset can support reverse querying of the process parameters of the corresponding production batch by defect type, such as electrode plating thickness and baking temperature, or querying the defect rate change trend of the products produced by the production equipment. When the defect rate of a certain production equipment is continuously higher than the average value within a preset time, an equipment process abnormality prompt is issued.

[0034] The defect source tracing data association unit also includes a time axis calibration submodule and a multi-dimensional attribution analysis submodule; The time axis calibration submodule is connected to the end sensor of the production line and the loading sensor of the detection system. It is used to collect the transmission time of the chip inductor from the production line to the loading of the detection system in real time, shift the imaging time of the defective product detection data backward by the corresponding transmission time to generate the calibration time, and then perform correlation and matching with the production data of the MES system based on the calibration time and the material batch code. The multi-dimensional attribution analysis submodule is connected to the MES system's full-chain production database (which stores raw material coating thickness inspection reports, production station temperature and humidity curves, equipment maintenance records, and operator training levels) and raw material supplier quality database (which stores coating adhesion test results and raw material batch sampling data). It is used to retrieve related data from the above databases based on the basic inspection data of defective products (defect type, severity, imaging time, and inspection station number) to form a three-dimensional related dataset including defects, production, and supply chain. The multi-dimensional attribution analysis submodule has a built-in association rule mining model. By analyzing the three-dimensional association dataset, it automatically outputs the probability ranking of the root cause of defects (e.g., electrode oxidation defects, raw material coating problems 65%, baking temperature abnormalities 30%, other factors 5%). When the defect rate of a certain production equipment is continuously higher than the system average within a preset time, or when the defect rate of a batch supplied by a certain raw material supplier is higher than the preset value for several consecutive batches, the defect tracing data association unit will not only output equipment process abnormality prompts or supplier quality warnings, but also simultaneously push specific abnormal data.

[0035] A multi-station collaborative and deep learning-based method for inspecting the appearance of chip inductors includes the following steps: S100: Vibratory feeder for sorting out disordered components; S200: Linear feeder conveying; S300: The guide block feeds the material to the glass turntable; S400: Calls preset parameters according to component specifications, and monitors and handles material jamming and dropping anomalies in real time; S500: The glass turntable rotates, and the components pass through four stations in sequence to complete the shooting; S600: Image processing; S700: Defect identification; S800: Data-driven model recording allows operators to correct data and fine-tune deep learning models and imaging parameters; S900: High-precision air valves distribute components to the good or bad product warehouse, count the quantity in real time, automatically generate a quality information feedback form and prompt rework when the number of defective products exceeds the threshold, link the MES system production data and generate a defect production association dataset, retrieve the production and supplier databases, output the probability ranking of the root cause of defects, and push anomaly prompts. Furthermore, in this invention, a more specific appearance inspection method includes the following steps: S1: First, the feeding module (high precision) starts working. The vibratory feeder receives the disordered chip inductors to be tested, and through vibration, it organizes them into an orderly state and conveys them to the linear feeder. The linear feeder smoothly conveys the components to the guide block in a linear vibration manner. The guide block then guides the components precisely to the designated bearing position of the glass turntable of the multi-light source multi-station imaging module through a preset guide channel, so that the components are stable in posture and uniform in position. At the same time, the parameter adaptive adjustment unit of the feeding module will automatically call the preset combination of vibratory feeder frequency, straight groove frequency and feeding speed parameters according to the specifications of the components to be tested (such as 0603, 0805, 1206). It will also collect the feeding status data in real time. If abnormalities such as jamming or dropping are detected, the feeding parameters will be finely adjusted in time to ensure continuous and stable feeding. S2: The glass turntable rotates, causing the orderly loaded components to pass through four imaging stations in sequence. The synchronous triggering device ensures that each component stops precisely at its corresponding station and completes the imaging. Station 1 uses a ring-shaped white LED light source to vertically illuminate and collect an overall image of the top surface of the component, used to detect surface dirt, damage, and marking errors. Station 2 uses a multi-angle red LED light source to low-angle sweep and collect images of the electrode ends, highlighting scratches, dents, and flatness defects. Station 3 uses a high-uniformity backlight light source to obtain a silhouette image of the component outline, used for dimensional measurement. Station 4 uses an adjustable polarized light source to eliminate reflection interference and collect images of the electrode coating to detect defects such as oxidation and incomplete plating. S3: The core image processing and decision-making module processes the four sets of acquired images. First, the image preprocessing unit performs filtering, denoising, enhancement, and correction operations to unify image standards. Then, the feature extraction and fusion unit registers and fuses the four images to form a comprehensive feature vector that includes information on the top surface, side surface, contour, and coating. Based on a deep learning target detection and classification network using improved YOLOv5, SSD, or Faster R-CNN, the multi-station images are used as different channels of input for the same target. The outputs defect categories (such as cracks, electrode scratches, dimensional deviations, coating defects, etc.) and location information. At the same time, the data-driven parameter optimization model records the correction data from the operator on the interactive interface. This data is used to fine-tune the deep learning model and imaging parameters, enabling the system to self-evolve and reduce reliance on manual debugging. S4: The intelligent execution sorting module, based on the instructions output by the core decision module, uses high-precision air valves to send components into the good product warehouse and the defective product warehouse (or a dedicated warehouse classified by defect type), and counts the quantity in each warehouse in real time. When the quantity in the defective product warehouse exceeds the threshold and the yield rate is lower than the set value, a quality information feedback form is automatically generated and a rework prompt is given. At the same time, the defect traceability data association unit starts working. First, the time axis calibration submodule collects the transmission time of the components from production line to the inspection system in real time, shifts the imaging time of the defective product inspection data backward by the corresponding time to generate a calibration time, and then combines the material batch code with the production data in the MES system (such as material batch, production equipment number, processing time) for association and matching to generate a defect production association dataset. S5: The multi-dimensional attribution analysis submodule further retrieves the associated data from the MES system's full-link production database (storing raw material coating thickness inspection reports, production station temperature and humidity curves, etc.) and raw material supplier quality database (storing coating adhesion test results, etc.) to form a three-dimensional associated dataset containing defects, production, and supply chain. After analysis by the built-in association rule mining model, it automatically outputs the probability ranking of the root causes of defects. If the defect rate corresponding to a certain production equipment is higher than the average value for a continuous preset time, or if the defect rate of a certain raw material supplier's batch is higher than the preset value for several consecutive batches, it will simultaneously push equipment process abnormality prompts or supplier quality warnings and specific abnormal data, ultimately realizing the appearance inspection of chip inductors.

[0036] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A multi-station collaborative and deep learning-based surface mount inductor appearance inspection system, characterized by: It includes a material feeding module, a multi-source multi-station imaging module, a core image processing and decision-making module, and an intelligent execution sorting module; The feeding module is set at the feeding start end of the appearance inspection system and is connected to the glass turntable feeding area of ​​the multi-source multi-station imaging module, the main control system and the preceding material flow link of the intelligent execution sorting module. It is used to organize the disordered chip inductors to be tested into an ordered state and transport them to the imaging module. The multi-source multi-station imaging module includes a high-speed rotating glass turntable with at least four imaging stations evenly arranged along the circumference of the turntable. Each station is equipped with an industrial camera, which takes pictures of each element sequentially through a synchronous triggering mechanism. The core image processing and decision-making module includes an image preprocessing unit, a feature extraction and fusion unit, a deep learning-based target detection and classification network, and a data-driven parameter optimization model, which completes image processing, defect identification, and parameter optimization. The intelligent sorting module, according to the instructions of the core decision-making module, uses air valves to distribute components into the corresponding hoppers.

2. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 1, characterized in that: The feeding module includes a vibratory feeder, a linear feeder, a guide block, and a parameter adaptive adjustment unit; The vibratory feeder is located at the starting end of the feeding module and is connected to the linear feeder and the parameter adaptive adjustment unit. It is used to receive the chip inductors to be tested, and through vibration, it organizes the disordered chip inductors into an ordered state and conveys them to the linear feeder. The linear feeder is set on the material transfer path between the vibratory feeder and the glass turntable. It is connected to the discharge end of the vibratory feeder, the guide block and the parameter adaptive adjustment unit. It is used to receive the ordered chip inductors from the vibratory feeder and transport the components to the guide block through linear vibration, and then feed them into the glass turntable of the multi-light source multi-station imaging module. The guide block is set between the linear feeder and the glass turntable of the multi-light source multi-station imaging module. It is connected to the discharge end of the linear feeder and the feeding area of ​​the glass turntable. It is used to receive the chip inductors conveyed by the linear feeder and guide the components to the designated bearing position of the glass turntable. The glass turntable is used to carry the chip inductors conveyed by the guide block. The high-speed rotation drives the components to pass through four imaging stations in sequence, and the synchronous triggering device ensures that each component stays at the corresponding station. The parameter adaptive adjustment unit is connected to the vibratory feeder, linear feeder, main control system, and material feeding status monitoring component. It is used to receive the specification instructions of the chip inductor to be tested issued by the main control system, and call the preset combination of vibratory feeder frequency, linear slot frequency, and feeding speed parameters; at the same time, it collects material feeding status data in real time.

3. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 1, characterized in that: The four imaging stations of the multi-source multi-station imaging module are each equipped with the following different light sources: Workstation 1: A ring-shaped white LED light source is used to vertically illuminate and collect an overall image of the top surface of the component, which is used to detect surface dirt, damage, and marking errors. Workstation 2: Uses red LED light source, low-angle sweeping light to highlight scratches, dents, and flatness defects on the electrode ends; Workstation 3: A backlit light source is used to acquire silhouette images of component outlines for dimensional measurement; Station 4: A polarized light source is used to eliminate reflective interference and clearly image the electrode coating condition, detecting oxidation and plating defects. The multi-source, multi-station imaging module also includes a dynamic light source spectrum adjustment unit, which is used to switch the light source spectrum band according to the material and surface coating type of the chip inductor to be tested, and to form a dual light source combination of basic light source and spectrum-adaptive light source in conjunction with the light sources of station one, station two, station three and station four.

4. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 1, characterized in that: The core image processing and decision-making module includes: The image preprocessing unit is used to perform filtering, denoising, enhancement, and correction operations on multiple acquired images to standardize the images. The feature extraction and fusion unit performs registration and feature layer fusion on the image information from the four workstations to form a comprehensive feature vector that includes information on the top surface, side surface, contour, and coating. Deep learning-based target detection and classification networks use improved YOLOv5, SSD, or Faster R-CNN frameworks as inputs, taking multi-station images as different channels of the same target and outputting defect category and location information.

5. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 4, characterized in that: The data-driven parameter optimization model can record the operator's correction data on the interactive interface, and use this data to fine-tune the deep learning model and imaging parameters; Defect categories include cracks, electrode scratches, dimensional deviations, and poor coating. The core image processing and decision-making module includes a defect severity grading unit, which establishes a three-level grading standard based on defect size, location, and impact range. The output results include defect category, location, and severity level.

6. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 1, characterized in that: The intelligent sorting module includes a good product bin and a defective product bin. The system counts the quantity in each bin in real time. When the quantity in the defective product bin exceeds the threshold and the yield rate is lower than the set value, a feedback form is automatically generated and a rework prompt is given. The intelligent sorting module also includes a defective product secondary verification station. A high-speed camera performs secondary imaging detection on components that are initially judged to be defective. The results are compared with the initial detection results. If they match, they are sent to the defective product bin. If they do not match, they are marked as suspicious products and sent to a dedicated suspicious product bin for manual verification.

7. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 1, characterized in that: The deep learning network of the core image processing and decision-making module is trained with multi-station image samples and can learn the global and local features of the components simultaneously. The appearance inspection system also includes a human-machine interface for displaying real-time images, defect classification results, statistical charts, and parameter self-optimization options.

8. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 1, characterized in that: The intelligent execution sorting module also includes a defect tracing data association unit; The defect tracing data association unit is connected to the core image processing and decision-making module and the production process data module of the MES system. When the intelligent sorting module sends defective products into the corresponding silo, the defect tracing data association unit automatically captures the detection data of the defective product and matches it with the production data of the component in the MES system to generate a defect production association dataset. The association dataset can support reverse querying of the process parameters of the corresponding production batch by defect type, or querying the defect rate change trend of the products produced by the production equipment. When the defect rate of a certain production equipment is higher than the average value for a preset time, the equipment process abnormality prompt is given.

9. The multi-station collaborative and deep learning-based chip inductor appearance inspection system according to claim 8, characterized in that: The defect source tracing data association unit also includes a time axis calibration submodule and a multi-dimensional attribution analysis submodule; The time axis calibration submodule is connected to the end sensor of the production line and the loading sensor of the detection system. It is used to collect the transmission time of the chip inductor from the production line to the loading of the detection system in real time, shift the imaging time of the defective product detection data backward by the corresponding transmission time to generate the calibration time, and then perform correlation and matching with the production data of the MES system based on the calibration time and the material batch code. The multi-dimensional attribution analysis submodule is connected to the MES system's full-chain production database and raw material supplier quality database. It is used to retrieve related data from the above databases based on the basic inspection data of defective products, forming a three-dimensional related dataset including defects, production and supply chain. The multi-dimensional attribution analysis submodule has a built-in association rule mining model, which automatically outputs the probability ranking of the root causes of defects by analyzing the three-dimensional association dataset. When the defect rate of a certain production equipment is higher than the system average for a continuous period of time, or when the defect rate of a batch supplied by a certain raw material supplier is higher than the preset value for several consecutive batches, the defect traceability data association unit will not only output equipment process abnormality prompts or supplier quality warnings, but also push specific abnormal data simultaneously.

10. A multi-station collaborative and deep learning-based method for appearance inspection of chip inductors, characterized in that, Includes the following steps: S1: First, the feeding module starts working. The vibratory feeder receives the disordered chip inductors to be tested, and through vibration, it organizes them into an orderly state and conveys them to the linear feeder. The linear feeder smoothly conveys the components to the guide block in a linear vibration manner. The guide block then guides the components precisely to the designated bearing position on the glass turntable of the multi-light source multi-station imaging module through a preset guide channel, so that the components are stable in posture and uniform in position. At the same time, the parameter adaptive adjustment unit of the feeding module will automatically call the preset combination of vibratory feeder frequency, straight groove frequency and feeding speed parameters according to the specifications of the components to be tested. It will also collect the feeding status data in real time. If abnormalities such as jamming or dropping are detected, the feeding parameters will be finely adjusted in time to ensure continuous and stable feeding. S2: The glass turntable rotates, causing the orderly loaded components to pass through four imaging stations in sequence. The synchronous triggering device ensures that each component stops precisely at its corresponding station and completes the imaging. Station 1 uses a ring-shaped white LED light source to vertically illuminate and collect an overall image of the top surface of the component, used to detect surface dirt, damage, and marking errors. Station 2 uses a multi-angle red LED light source to low-angle sweep and collect images of the electrode ends, highlighting scratches, dents, and flatness defects. Station 3 uses a high-uniformity backlight light source to obtain a silhouette image of the component outline, used for dimensional measurement. Station 4 uses an adjustable polarized light source to eliminate reflection interference and collect images of the electrode coating to detect defects such as oxidation and incomplete plating. S3: The core image processing and decision-making module processes the four sets of acquired images. First, the image preprocessing unit performs filtering, denoising, enhancement, and correction operations to unify image standards. Then, the feature extraction and fusion unit registers and fuses the four images to form a comprehensive feature vector that includes information on the top surface, side surface, contour, and coating. Based on a deep learning target detection and classification network with improved YOLOv5, SSD, or Faster R-CNN, the multi-station images are used as different channels of input for the same target, and the output is the defect category and location information. At the same time, the data-driven parameter optimization model records the correction data of the operator in the interactive interface. This data is used to fine-tune the deep learning model and imaging parameters to achieve system self-evolution and reduce reliance on manual debugging. S4: The intelligent execution sorting module sends components into the good product warehouse and the defective product warehouse respectively through high-precision air valves according to the instructions output by the core decision module, and counts the quantity in each warehouse in real time. When the quantity in the defective product warehouse exceeds the threshold and the yield rate is lower than the set value, a quality information feedback form is automatically generated and a rework is prompted. At the same time, the defect source traceability data association unit starts working. First, it collects the transmission time of components from production line to testing system loading in real time through the time axis calibration submodule. It shifts the imaging time of defective product detection data backward by the corresponding time to generate calibration time. Then, it combines the material batch code with the production data in the MES system for association and matching to generate defective production association dataset. S5: The multi-dimensional attribution analysis submodule further retrieves the associated data from the MES system's full-link production database and raw material supplier quality database to form a three-dimensional associated dataset containing defects, production, and supply chain. After analysis by the built-in association rule mining model, it automatically outputs the probability ranking of the root causes of defects. If the defect rate of a certain production equipment is higher than the average value for a continuous preset time, or if the defect rate of a certain raw material supplier's batch is higher than the preset value for several consecutive batches, it will simultaneously push equipment process abnormality prompts or supplier quality warnings and specific abnormal data, ultimately realizing the appearance inspection of chip inductors.

Images