PCBA AI defect detection method and system

Abstract

The invention belongs to but is not limited to the technical field of PCBA defect detection, and discloses a PCBA AI defect detection item demand response method and system, which adopts multi-scale long-range fusion, context information, super-resolution and block detection to improve a target detection algorithm, and performs detection and identification of a tiny target. Identifying the fine granularity of the target by adopting a contextual information and depth feature matching method; 3D point cloud and X-ray perspective are adopted, a multi-sensor fusion detection frame is used, and invisible defect detection is carried out. According to the method, AI technologies such as deep learning are utilized, the detection precision of small defects on the PCBA board is effectively improved, and missing detection and false detection are reduced. The AI technology can automatically identify and learn defect features, and efficiency and accuracy are significantly improved. The AI model can optimize the self algorithm through continuous learning, so that the detection system can adapt to different PCBA boards and production environments, and the universality and flexibility of the detection system are enhanced.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the technical field of PCBA defect detection, and in particular relates to a PCBA AI defect detection method and system. Background Art

[0002] Printed Circuit Board Assembly (PCBA) defect detection technology has undergone significant development, both in China and internationally. As the electronics manufacturing industry moves towards higher precision, faster speed and lower cost, PCBA defect detection technology has become a key link in ensuring product quality.

[0003] Printed Circuit Board Assembly (PCBA) defect detection technology has made significant progress worldwide. With the increasing requirements of the electronics manufacturing industry for accuracy, speed and cost control, this technology has become an indispensable part of ensuring product quality. Internationally, PCBA defect detection is mainly developing in the following directions:

[0004] Automation and machine vision: By using high-resolution cameras and advanced image processing algorithms, automated inspection and machine vision systems can effectively identify common defects such as poor welding, missing parts or position deviations. More advanced systems implement three-dimensional visual inspection, providing more accurate data support for defect analysis.

[0005] Artificial Intelligence and Deep Learning: The use of artificial intelligence technology, especially deep learning models, for defect detection is becoming more and more widespread. Such technology can self-learn through a large number of training sets to identify various defect characteristics on the PCBA board surface, and its performance will continue to improve over time.

[0006] X-ray inspection technology: As a non-destructive inspection method, X-ray technology can penetrate the PCBA board surface and effectively find problems such as voids and poor contact under the soldering points. It is especially suitable for the inspection of surface mount components such as BGA and QFN that are difficult to directly observe.

[0007] The combination of automatic optical inspection (AOI) and automatic X-ray inspection (AXI): The comprehensive application of these two technologies not only enhances the comprehensiveness and accuracy of detection, but also represents the future development trend of the industry.

[0008] In China, PCBA defect detection technology has also experienced rapid growth, which is specifically reflected in the following aspects:

[0009] Technological catch-up and local innovation go hand in hand: Domestic related technologies not only keep up with the international advanced level, but also demonstrate unique innovation capabilities in software algorithms, intelligent applications, etc.

[0010] Industrial cluster advantages: As the world's largest electronic product production base, China has formed a powerful industrial chain cluster effect, which has promoted the rapid development and widespread application of testing technology and equipment, and also promoted the growth and development of domestic testing equipment manufacturers and technical service providers.

[0011] Policy support: The Chinese government has given strong support to high-tech industries, and has encouraged the research and development and marketization of PCBA testing technology through various measures such as financial subsidies and tax incentives.

[0012] Market competition and international cooperation: In addition to fierce competition in the domestic market, domestic enterprises are also actively seeking technical exchanges and cooperation with overseas counterparts, and continuously improving their independent innovation capabilities by introducing advanced technologies and absorbing and transforming them.

[0013] Through the above analysis, the problems and defects of the prior art are as follows:

[0014] Difficulty in detecting tiny targets: For example, on a 300mm×200mm PCBA board, the size of some components or defects is less than 1mm×1mm, which far exceeds the definition of "small targets" for most current target detection algorithms (the maximum side length of the target does not exceed 3% of the shortest side of the entire image or 32 pixels), posing a huge challenge to existing detection technologies.

[0015] Difficulty in identifying fine-grained differences: When a comprehensive inspection of all components and their potential defects on the entire PCBA board is required, since the appearance of each component is very similar and the visual difference between defects and normal states is minimal, extremely high requirements are placed on the resolution ability of the inspection system.

[0016] Detection barriers of invisible defects: Some defects, such as voids inside BGA solder joints, cannot be directly observed in conventional images, which increases the difficulty of detection. Summary of the invention

[0017] In view of the problems existing in the prior art, the present invention provides a PCBAAI defect detection method and system.

[0018] The present invention is implemented as follows: a PCBAAI defect detection method comprises the following steps:

[0019] Step 1: Use multi-scale block detection to improve the target detection algorithm and segment the ROI to be detected;

[0020] Step 2: Use long-range fusion and context information to detect and locate small targets in the ROI to be inspected;

[0021] Step 3: Use deep feature matching and super-resolution recognition algorithms to identify fine-grained differences and defect types of the target;

[0022] Step 4: Use 3D point cloud and X-ray perspective, and use a multi-sensor fusion detection framework to perform invisible defect detection.

[0023] Furthermore, the block detection uses SWin (ShiftWindows) method to extract block features.

[0024] Furthermore, the multi-scale long-range fusion: uses long-range fusion across multiple scales in Backbone to directly fuse higher-resolution positioning information with deeper semantic information; the network form of Backbone refers to the long-range fusion method of DLA (Deep Layer Aggregation).

[0025] Furthermore, the context information: feature extraction of context information, including local and long-distance associations; using SPP in Backbone to enhance local relationships, and using a Non-Local attention mechanism to obtain long-distance associations.

[0026] Furthermore, the deep feature matching: the template image and the image to be tested are simultaneously input into Backbone, and the deep feature matching method based on Cross-Attention in LightGlue is adopted. After the registration, the deep image difference is performed to enhance the fine-grained difference.

[0027] Furthermore, the super-resolution: SR (Super-Resolution) module is introduced into the Neck part to assist in amplifying the effective pixel quantity of the tiny target.

[0028] Furthermore, the multi-sensor fusion detection framework: uses three types of data to train three independent detection models respectively; each sensor modality is separately encoded in its own coordinates using a modality-specific feature encoder; in the fusion stage, all weights outside the fusion detection Head are frozen, and the fusion detection Head is trained with fewer complete samples.

[0029] Another object of the present invention is to provide a PCBAAI defect detection system, comprising:

[0030] The small target detection module uses multi-scale long-range fusion, context information, super-resolution and block detection to improve the target detection algorithm and detect and identify small targets;

[0031] The fine-grained difference recognition module uses context information and deep feature matching to identify the fine-grained nature of the target;

[0032] The invisible defect detection module uses 3D point cloud and X-ray perspective, and uses a multi-sensor fusion detection framework to perform invisible defect detection.

[0033] Another object of the present invention is to provide an information data processing terminal, which includes the PCBAAI defect detection system.

[0034] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0035] First, the present invention uses multi-scale long-range fusion, context information, super-resolution and block detection to improve the target detection algorithm to adapt to the detection and recognition of small targets. The method of context information and deep feature matching is used to adapt to the fine-grained recognition of targets. 3D point cloud and X-ray perspective are used, and a multi-sensor fusion detection framework is used.

[0036] Improve detection accuracy and efficiency: Using AI technologies such as deep learning can effectively improve the detection accuracy of small defects on PCBA boards and reduce missed detection and false detection. AI technology can automatically identify and learn defect characteristics, and has significantly improved efficiency and accuracy compared to manual detection and traditional automation technology.

[0037] Strengthen adaptive learning capabilities: The AI ​​model can optimize its own algorithm through continuous learning, so that the detection system can adapt to different PCBA boards and production environments, enhancing the universality and flexibility of the detection system.

[0038] Reduce costs: Although the initial investment in the project is high, AI-driven automated testing can significantly reduce long-term operating costs and reduce cost losses and material waste caused by manual testing.

[0039] Improve production efficiency: Fast and accurate inspection can significantly increase production line throughput and reduce downtime caused by defects, thereby improving the efficiency and output of the entire production process.

[0040] feasibility:

[0041] Mature technical foundation: With the continuous advancement of computing power and algorithms, the application of AI in the field of defect detection has become increasingly mature, and the threshold and risk of engineering deployment are gradually decreasing.

[0042] Easy to integrate and expand: The AI ​​detection system can be easily integrated with existing production lines and information systems, supports remote monitoring and operation, has good scalability, and can adapt to the needs of enterprise expansion and technological upgrading.

[0043] Overall, the development and application of PCBA defect detection technology not only optimizes the production process of the electronics manufacturing industry, improves product quality and output, but also brings significant economic benefits and competitive advantages to enterprises. In the long run, the promotion of this technology will promote technological progress and structural optimization of the entire industry, and strongly support the strategic goals of intelligent manufacturing and industrial automation.

[0044] Second, the expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0045] Wide industry applicability: PCBA is the core component of a variety of electronic products, covering smart phones, computers, industrial control equipment, medical instruments and other fields. Therefore, AI defect detection technology has broad application prospects in these industries.

[0046] Improved quality control: For fields that require extremely high reliability, such as aerospace, military equipment, medical equipment, etc., AI technology can significantly improve product quality control standards and ensure product safety and reliability.

[0047] Third, the technical solution of the present invention significantly improves the accuracy, efficiency and detection capability of PCBA (printed circuit board assembly) defect detection by introducing multiple advanced technologies such as multi-scale block detection, long-range fusion, deep feature matching, super-resolution recognition and multi-sensor fusion, and especially solves several problems in the prior art. The following is a detailed description of the technical problems solved by the solution in industrial applications and the technological progress achieved.

[0048] 1. Technical problems existing in the existing technology

[0049] a. Difficulty in detecting small targets and complex backgrounds

[0050] Existing PCBA defect detection methods are usually difficult to accurately detect small-scale defects when facing tiny targets and complex backgrounds. Traditional image processing and target detection algorithms lack the ability to efficiently segment and detect tiny defects in complex backgrounds, which can easily lead to missed detections and false detections, especially for the detection of tiny components and solder joints on PCB boards.

[0051] b. Inaccurate classification of fine-grained defects

[0052] Existing inspection technologies have difficulty accurately distinguishing similar but slightly different types of defects. For example, slight cracks on solder joints and cold solder joints are often difficult to distinguish correctly in low-resolution images. Due to the limited ability of traditional inspection methods to process image details, the classification and identification of subtle defects are not accurate enough.

[0053] c. Limitations of non-visual defect detection

[0054] In the PCBA manufacturing process, many defects, such as internal voids and cracks in solder joints, internal damage to components, etc., cannot be directly observed by conventional optical inspection methods. Traditional inspection systems mostly rely on surface visual inspection, which makes it difficult to detect internal or hidden structural defects.

[0055] d. Lack of multi-sensor fusion technology

[0056] Traditional inspection technologies often only use single sensor technology, such as 2D imaging, and cannot effectively combine multi-sensor data such as X-rays and 3D point clouds. This limitation makes it impossible for the system to achieve multi-angle and high-precision inspection of PCBA defects, making it difficult to detect internal defects.

[0057] 2. Significant technological progress

[0058] a. Improve the accuracy of small target detection

[0059] The present invention effectively solves the problem of small targets being difficult to detect in complex backgrounds in the prior art by adopting a multi-scale block detection algorithm and long-range fusion technology. Multi-scale block detection can perform multi-level segmentation of the target area, and combined with long-range fusion technology, local details are combined with global context information, thereby improving the detection accuracy of small targets, and is particularly suitable for the detection of complex components and small defects on PCB boards.

[0060] b. Enhanced fine-grained defect recognition capabilities

[0061] By introducing deep feature matching and super-resolution recognition algorithms, the present invention significantly enhances the ability to classify subtle defects. Deep feature matching uses neural networks to learn the deep features of different defect types, so that similar defects such as slight cracks in solder joints and cold solder joints can be accurately distinguished. Super-resolution recognition further improves the resolution of the image, obtains more detailed information, ensures accurate identification of defect types at a fine-grained level, and solves the problem of insufficient ability to identify subtle differences in the prior art.

[0062] c. Achieve efficient detection of invisible defects

[0063] Through the multi-sensor fusion of 3D point cloud and X-ray perspective technology, the present invention overcomes the limitation of traditional detection technology that it cannot detect internal and hidden defects. 3D point cloud builds an accurate three-dimensional model for the PCB board, and X-ray perspective can penetrate the surface and capture internal structural information. Through multi-sensor fusion, the system can detect invisible defects from multiple angles and dimensions, greatly improving the coverage of PCBA detection and solving the problem of difficult detection of internal defects.

[0064] d. Innovative application of multi-sensor fusion technology

[0065] The multi-sensor fusion detection framework proposed in this invention fuses multiple data such as optical images, 3D point clouds, and X-ray perspectives to achieve multi-angle detection of PCBAs. This innovative detection framework greatly improves the reliability and comprehensiveness of detection, avoids the limitations of single sensor technology, and enables the system to maintain efficient and accurate detection capabilities when facing different types of defects.

[0066] The present invention solves many technical problems existing in existing PCBA detection technology in industrial applications, especially in the detection of small targets, fine-grained defect classification, invisible defect detection and multi-sensor fusion. This innovative detection method not only improves the detection accuracy and efficiency, but also expands the scope and types of defect detection, greatly improving the reliability and accuracy of PCBA production quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of a PCBAAI defect detection method provided by an embodiment of the present invention;

[0068] Figure 2 It is a schematic diagram of a long-range fusion method of DLA provided in an embodiment of the present invention;

[0069] Figure 3 is a schematic diagram of the Non-Local attention mechanism provided by an embodiment of the present invention;

[0070] Figure 4 is a schematic diagram of a GAN-based SR module provided by an embodiment of the present invention;

[0071] Figure 5 is a schematic diagram of the ShiftWindows method provided by an embodiment of the present invention;

[0072] Figure 6 is a schematic diagram of LightGlue provided by an embodiment of the present invention;

[0073] Figure 7 It is a schematic diagram of the FusionAOI framework provided by an embodiment of the present invention;

[0074] Figure 8 It is a structural diagram of a PCBAAI defect detection system provided by an embodiment of the present invention.

[0075] Fig. 9 It is an experimental effect diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0077] like Figure 1 As shown, a PCBAAI defect detection project demand response method provided by an embodiment of the present invention includes the following steps:

[0078] S101, using multi-scale block detection to improve the target detection algorithm and segment the ROI to be detected;

[0079] S102, using long-range fusion and context information to detect and locate small targets in the ROI to be inspected;

[0080] S103, using deep feature matching and super-resolution recognition algorithms to identify fine-grained differences in targets and types of defects;

[0081] S104 uses 3D point cloud and X-ray perspective, and uses a multi-sensor fusion detection framework to perform invisible defect detection.

[0082] The PCBAAI defect detection method uses a variety of advanced technologies and sensor fusion to achieve high-precision defect detection, especially for the detection of tiny targets and invisible defects. Its working principle includes four main steps, each of which is based on the latest machine vision, deep learning and sensor technology. The following is a detailed description of the working principle:

[0083] 1. Multi-scale block detection improves the target detection algorithm to segment the ROI to be detected

[0084] In the first step, the system uses a multi-scale block detection algorithm, which can perform global and local refined analysis of PCBA (printed circuit board assembly). By performing multi-scale segmentation on the image to be inspected and extracting feature information of different resolutions, the system can automatically identify potential detection areas (ROI, Region of Interest). In this process, the improved target detection algorithm combines the multi-layer feature extraction capabilities of modern convolutional neural networks (CNN) to ensure efficient and accurate segmentation of the target area to be detected in complex backgrounds. Block detection enables the system to cope with targets of different scales and ensure detection accuracy, especially in PCBA, where the detection area includes very small components or solder joints.

[0085] 2. Long-range fusion and context information for small target detection and positioning in the ROI to be detected

[0086] In the second step, the system further improves the detection accuracy of small targets through the comprehensive application of long-range fusion and contextual information. The long-range fusion algorithm can extract key information across a large area, combine the locally detected information with the global information, and ensure that the small target is not disturbed by the complex background. At the same time, the introduction of contextual information enables the system to use the characteristics of the surrounding area to help locate the target, which is very important for the layout of components on PCB boards because there is a complex network of relationships between components. By combining long-range fusion and contextual information, the system can more accurately detect small defect targets, such as poor soldering and component damage.

[0087] 3. Deep feature matching and super-resolution recognition algorithms identify fine-grained differences in targets and defect types

[0088] In the third step, the system applies Deep Feature Matching and Super-Resolution Recognition technology to identify subtle defects on PCBA and fine-grained differences of targets. Deep Feature Matching uses neural networks to learn the deep features of different defect types, so that it can accurately distinguish defects with high similarity, such as tiny cracks and cold solder joints in solder joints. Super-resolution recognition enhances the image resolution in the ROI to be inspected by enhancing the details of low-resolution images to obtain clearer images, further ensuring that fine-grained differences in targets can be clearly identified, thereby improving the ability to recognize complex defects.

[0089] 4. 3D point cloud and X-ray multi-sensor fusion for invisible defect detection

[0090] In the last step, the system detects invisible defects through multi-sensor fusion (Fusion Framework for Detection). In PCBA inspection, some defects such as internal cracks, solder joint voids, hidden component damage, etc. cannot be directly observed through surface imaging technology, so this method uses 3D point cloud data and X-ray perspective imaging technology. 3D point cloud technology can construct the three-dimensional structure of the PCB board and provide spatial position information of components and solder joints; X-ray perspective technology can penetrate the PCB board and detect internal hidden structures and defects. By fusing the data of these sensors, the system can analyze the internal and external structures of the PCBA from multiple angles, achieve efficient detection of invisible defects, and ensure product quality and reliability.

[0091] This multi-step, multi-technology inspection method can comprehensively cover visible and invisible defects on PCBA and is suitable for complex industrial inspection scenarios.

[0092] like Figure 2 As shown in the figure, the improved target detection algorithm in small target detection has been verified in projects such as drone aerial photography.

[0093] Multiscale Long-range Aggregation

[0094] Since the target scale is small, the simple positioning and semantic fusion mentioned above cannot obtain enough information directly on Neck. It is necessary to use long-range fusion across multiple scales in Backbone to directly fuse higher-resolution positioning information with deeper semantic information. The network form of Backbone refers to the long-range fusion method of DLA (Deep Layer Aggregation), such as Figure 3 shown.

[0095] Context Information

[0096] Due to the small number of input pixels, the amount of information and effective features contained in a single target are not enough to be directly and effectively detected. It is necessary to enhance the feature extraction of contextual information, including local and long-distance associations. In Backbone, SPP is used to enhance local relations, and the Transformer module containing the Non-Local Attention Mechanism and the MLP Multi-layer Perceptron is used to obtain long-distance associations, such as Figure 3 shown.

[0097] Super-Resolution

[0098] The small target is far below the general definition, so it is necessary to introduce the SR (Super-Resolution) module in the Neck part to assist and enlarge the effective pixel volume of the small target. Considering the computing power limitation of the project, the more effective diffusion model (DiffusionModel) is not used, but the SR module based on GAN (Generative Adversarial Network) is used in the shallow layer, such as Figure 4 shown.

[0099] Patch Detection

[0100] Since the target scale is small, the Swin (ShiftWindows) method is also used for block feature extraction. At the same time, considering the memory limitation of the project, no sliding window change and cross-block fusion are performed in the Backbone and Neck stages, and cross-block fusion is only performed in the input layer of the Head. For example Figure 5 shown.

[0101] The difference recognition algorithm in fine-grained difference recognition has been applied in projects such as intelligent paper marking and tile defect detection of the project team.

[0102] Deep Feature Matching

[0103] Due to the small features and the lack of semantic information, the traditional matching algorithm is insufficient. Therefore, the template image and the image to be tested are input into Backbone at the same time, and the deep feature matching method based on Cross-Attention in LightGlue is adopted, such as Figure 6 After the registration, depth map subtraction is performed to enhance the fine-grained differences.

[0104] Invisible defects

[0105] 3DAOI detection technology measures the entire point cloud to obtain height, volume and contour parameters, and can more comprehensively detect information such as component shape defects, position offsets and PCB bending.

[0106] X-rays have short wavelengths, high energy, and strong penetrating power. The penetrating power of X-rays on different materials and their different absorption properties can distinguish materials with different densities. Therefore, if the inspected object has internal fractures, different thicknesses, or changes in shape, the resulting image will also be different. It is suitable for internal defect detection. The multi-sensor fusion framework has been applied in logistics projects.

[0107] Multi-sensor fusion detection framework

[0108] like Figure 7 As shown in the figure, in order to effectively utilize the data, three types of data are used to train three independent detection models. Each sensor modality is encoded separately in its own coordinates using a modality-specific feature encoder. In the fusion stage, the heads of the three models are removed and their respective embedding layers are left for fusion, followed by the target detection head. All weights except the fusion detection head are frozen, and the fusion detection head is trained with fewer complete samples.

[0109] Taking into account the computing power limitations of the project, the fusion method does not use cross-attention-based cross-correlation fusion, but a lightweight information fusion method using cascade and MLP.

[0110] like Figure 8 As shown, a PCBAAI defect detection project demand response system of a PCBAAI defect detection project demand response method provided by an embodiment of the present invention includes:

[0111] The small target detection module uses multi-scale long-range fusion, context information, super-resolution and block detection to improve the target detection algorithm and detect and identify small targets;

[0112] The fine-grained difference recognition module uses context information and deep feature matching to identify the fine-grained nature of the target;

[0113] The invisible defect detection module uses 3D point cloud and X-ray perspective, and uses a multi-sensor fusion detection framework to perform invisible defect detection.

[0114] The relevant evidence of the technical effects obtained by the embodiments of the present invention is as follows:

[0115] The PCBA image to be inspected is improved through multi-scale block detection to improve the target detection algorithm, and the ROI to be inspected is segmented. Then, long-range fusion and context information are used to complete the detection and positioning of small targets in the ROI to be inspected. In the next step, deep feature matching and super-resolution recognition algorithms are used. Finally, a multi-sensor fusion detection framework is used to complete the positioning and classification of PCBA defects such as multiple pieces, reverse sticking, tombstones, exposed copper, offset, damage, leaking solder, false soldering and invisible defects.

[0116] The experimental results are as follows Fig. 9 shown.

[0117] System accuracy comparison:

[0118] Traditional testing line This system Random inspection Sampling rate <10% Full inspection Human Resources 2 people per testing line Each person rechecks 5 lines False negatives Unquantifiable <1％ New PCB import time 2 hours 5 minutes

[0119] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0120] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A PCBAAI defect detection method, characterized in that: The following steps are involved: Step 1: Use multi-scale block detection to improve the target detection algorithm and segment the ROI to be detected; Step 2: Use long-range fusion and context information to detect and locate small targets in the ROI to be detected; Step 3: Use deep feature matching and super-resolution recognition algorithms to identify fine-grained differences and defect types of the target; Step 4: Use 3D point cloud and X-ray perspective, and use a multi-sensor fusion detection framework to perform invisible defect detection.

2. The PCBAAI defect detection method according to claim 1, characterized in that: The multi-scale long-range fusion: In Backbone, long-range fusion across multiple scales is used to directly fuse higher-resolution positioning information with deeper semantic information; the network form of Backbone refers to the long-range fusion method of DLA.

3. The PCBAAI defect detection method according to claim 1, characterized in that: The context information: feature extraction of context information, including local and long-distance associations; using SPP in Backbone to enhance local relationships, and using a Transformer module including a Non-Local attention mechanism and an MLP multi-layer perceptron to obtain long-distance associations.

4. The PCBAAI defect detection method according to claim 1, characterized in that: The super-resolution method introduces an SR module in the Neck part to assist in amplifying the effective pixel quantity of tiny targets; and uses a GAN-based SR module in the shallow layer.

5. The PCBAAI defect detection method according to claim 1, characterized in that: The block detection: adopts the Swin method to extract block features; at the same time, the sliding window is not changed and cross-block fusion is not performed in the Backbone and Neck stages, and cross-block fusion is only performed in the input layer of the Head.

6. The PCBAAI defect detection method according to claim 1, characterized in that: The deep feature matching is as follows: the template image and the image to be tested are simultaneously input into Backbone, and a deep feature matching method based on Cross-Attention in LightGlue is adopted; after registration, a deep image difference is performed to enhance fine-grained differences.

7. The PCBAAI defect detection method according to claim 1, characterized in that: The multi-sensor fusion detection framework: trains three independent detection models using three types of data respectively; each sensor modality is individually encoded in its own coordinates using a modality-specific feature encoder; In the fusion stage, the heads of the three models are removed and their respective embedding layers are left for fusion, followed by the target detection head; all weights except the fusion detection head are frozen, and the fusion detection head is trained with fewer complete samples.

8. The PCBAAI defect detection method according to claim 7, characterized in that: The fusion method uses a lightweight information fusion method of cascade and MLP.

9. A PCBAAI defect detection system according to any one of claims 1 to 8, characterized in that: include: The small target detection module uses multi-scale long-range fusion, context information, super-resolution and block detection to improve the target detection algorithm and detect and identify small targets; The fine-grained difference recognition module uses context information and deep feature matching to identify the fine-grained nature of the target; The invisible defect detection module uses 3D point cloud and X-ray perspective, and uses a multi-sensor fusion detection framework to perform invisible defect detection.

10. An information data processing terminal, characterized in that: The information data processing terminal includes the PCBAAI defect detection system described in claim 9.

Images