Connector terminal semi-finished product appearance detection method and device based on binocular machine vision and AI model

Abstract

The invention belongs to the technical field of detection equipment and AI, and discloses a connector terminal semi-finished product appearance detection method and equipment based on binocular machine vision and an AI model, and the equipment comprises a binocular machine vision detection unit, a jig, a control and display unit and a cloud server which are arranged on a rack. The binocular machine vision detection unit comprises area-array camera modules in the vertical direction and the horizontal direction which are orthogonally arranged, and images of the top and the side faces of a product are synchronously captured. The jig comprises a replaceable embedded panel made of a transparent material. The industrial personal computer is internally provided with a lightweight deep learning model for preliminary defect recognition, and the cloud server is internally provided with a high-performance AI analysis program for secondary judgment. According to the invention, an edge-cloud collaborative architecture is adopted, so that the requirement on the computing resources of the industrial personal computer is reduced, the analysis speed and efficiency are improved, meanwhile, the functions of data aggregation analysis, continuous model optimization and intelligent early warning are realized, and the method is widely applicable to intelligent appearance quality detection of high-speed cable connector semi-finished products.

Description

Technical Field

[0001] This invention belongs to the field of testing equipment and AI technology, specifically relating to a method and equipment for appearance inspection of connector terminal semi-finished products based on binocular machine vision and AI models, which is particularly suitable for multi-dimensional appearance quality inspection and intelligent quality management of high-speed cable connector semi-finished products. Background Technology

[0002] With the rapid development of artificial intelligence technology, low-voltage, high-speed, high-current cable connectors are increasingly widely used in the manufacturing of equipment such as AI servers, intelligent computing centers, and high-performance computers. For example, 224Gbps cable connectors for high-speed transmission are widely used in high-speed signal transmission between chips and boards, such as the interconnection system of NVIDIA GB300 NVLink cabinets.

[0003] The pluggable cable terminals used to manufacture this connector typically have a complex structure and multiple quality inspection items, and come in various models and specifications. Taking the 224Gbps high-speed cable connector as an example, its semi-finished product adopts a biaxial cable design, including: insulated cable (including metal core wire harness), metal semi-enclosed protective shell (including positioning holes), injection molded clips, parallel terminal pins, and cable shielding layer. The parallel pins and core wire harness are connected by laser welding (solder joints). Its semi-finished product form is that the terminals have been crimped and preliminarily packaged, but it has not been assembled into a complete connector assembly.

[0004] Quality inspection of these semi-finished products involves numerous testing items, primarily including: solder joint connection status, parallel terminal pin alignment, deformation of the metal semi-enclosed protective shell, accurate positioning of the injection-molded clips, and contact between the parallel pin terminals and the core wire harness with the cable shielding layer. Semi-finished product types generally include products with solder joints and those without, as well as straight and bent semi-finished products.

[0005] Currently, existing testing equipment mainly suffers from the following problems: (1) Single detection dimension: Existing equipment can usually only collect images from a single perspective and cannot simultaneously acquire information from multiple ends of the product, resulting in incomplete detection.

[0006] (2) Poor fixture adaptability: The fixture structure of the existing equipment is fixed and can only test one type of product, which cannot meet the testing needs of connector terminal semi-finished products of various models and structures.

[0007] (3) Difficulty in deploying AI models: High-performance deep learning models usually require expensive hardware resources such as GPUs, while industrial field equipment is often limited by cost and space, making it difficult to deploy complex models, which leads to limited detection accuracy.

[0008] (4) Lack of collaborative data analysis: Most existing equipment operates as stand-alone machines, and the test data is stored in a scattered manner, making it impossible to conduct statistical analysis across devices and batches, and making it difficult to discover systemic quality problems.

[0009] (5) Difficulty in updating models: The AI ​​model of a single device needs to be updated on a machine-by-machine basis, which is costly to maintain and makes it difficult to achieve continuous optimization of the model.

[0010] For example, CN107478164A discloses a machine vision-based connector terminal inspection device, but its machine vision inspection system can only acquire images of the upper surface of the connector terminal from the vertical direction, and cannot simultaneously acquire images of the front surface of the terminal. Furthermore, its fixture structure is fixed, it can only inspect one type of product, and its overall structure is complex, with few inspected end faces, slow speed, low efficiency, and poor adaptability. CN107895362A discloses a machine vision method for quality inspection of miniature wiring terminals. This method uses traditional image processing algorithms (such as the Canny operator and Blob analysis) for quality inspection, without employing deep learning technology. Its ability to identify complex defects is limited, and it can only inspect images from a single perspective, failing to achieve multi-dimensional simultaneous inspection. Summary of the Invention

[0011] The purpose of this invention is to provide a method and equipment for visual inspection of connector terminal semi-finished products based on binocular machine vision and AI models. Through an edge-cloud collaborative computing architecture, it reduces the demand for industrial control computer computing resources while improving analysis speed and efficiency, and solves the problems of limited detection end faces, slow speed, low efficiency, poor adaptability, and lack of data collaborative analysis in existing technologies.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: A visual inspection device for connector terminal semi-finished products based on binocular machine vision and AI models, comprising: A rack, the rack comprising a desktop base panel, an internal support, and a housing; A binocular machine vision inspection unit is mounted on the frame and includes a vertical area array camera module and a horizontal area array camera module. The vertical area array camera module and the horizontal area array camera module are arranged orthogonally or obliquely. The optical focus points of the two camera modules are fixed target positions on the fixture. The effective detection area of ​​the vertical area array camera module covers the upper end face of the fixture, and the effective detection area of ​​the horizontal area array camera module covers the front end face of the fixture. The fixture, mounted on the base panel, includes a transparent fixing bracket and a detachable insert panel. The insert panel has an insert groove that matches the shape and structure of the product to be tested. During testing, the product to be tested is inserted into the insert groove and positioned so that the upper surface of the product to be tested is flush with the upper surface of the insert panel, and the front surface is flush with the front surface of the insert panel. The fixing bracket and the insert panel are detachably connected by a positioning pin and a snap-fit ​​structure. The control and display unit includes an industrial computer, an input keyboard, and a display. The industrial computer has a built-in AI model for preliminary defect identification of the acquired images. The cloud server communicates with the industrial control computer via a wired or wireless network. The cloud server has a built-in high-performance AI analysis program, which is used to receive the detection data and preliminary analysis results uploaded by the industrial control computer, make secondary judgments, and then save the data in the cloud.

[0013] The horizontal area array camera module includes a horizontal industrial camera and a horizontal light source adjustment module. The lens of the horizontal industrial camera is focused on the center of the side end face of the insert slot. The horizontal light source adjustment module is a high-angle ring light source composed of multi-ring, multi-angle LED beads, with an adjustable illumination angle of 30°-60°.

[0014] The fixed bracket also includes two symmetrically arranged left and right legs, the bottom end face of each leg being detachably connected to the base panel through screw holes; the insert panel is provided with a vertically extending elastic cantilever buckle, and the vertical surface of the leg is provided with a groove matching the buckle; after the elastic cantilever buckle of the insert panel is inserted into the groove of the leg, a detachable connection is formed between the insert panel and the leg; a cylindrical positioning pin is provided on the top end face of the leg, and a corresponding positioning hole is provided on the insert panel, which assists in guiding and positioning during the process of forming a detachable connection between the insert panel and the leg.

[0015] The lightweight deep learning model built into the industrial control computer includes a multi-scale feature extraction module and a fast classifier; the high-performance AI analysis program built into the cloud server includes a Transformer-based fine analysis model and a time-series data analysis module.

[0016] The inspection method of the connector terminal semi-finished product appearance inspection equipment based on binocular machine vision and AI model mainly includes the following steps: product placement, local image acquisition, local preliminary analysis, data upload, cloud secondary judgment, result feedback, and result output.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: 1. Edge-Cloud Collaborative Computing: This invention employs a collaborative computing architecture that combines preliminary analysis on an industrial control computer with secondary judgment by a cloud server. The industrial control computer uses a built-in lightweight model to perform rapid preliminary screening, while the cloud server utilizes a high-performance AI program for detailed analysis. This approach reduces the computational resource requirements of the industrial control computer, ensures detection accuracy, and improves overall analysis speed and efficiency.

[0018] 2. Comprehensive inspection dimensions: This invention adopts a dual vision module orthogonal or oblique layout to simultaneously acquire images of the top and sides of the product, which can detect a variety of appearance defects such as solder joint status, protective shell deformation, pin parallelism, and injection molding buckle positioning, providing comprehensive inspection dimensions.

[0019] 3. Strong adaptability of the fixture: The fixture of the present invention adopts a detachable insert panel, which can be quickly replaced through the detachable structure of buckle and positioning pin, and can be adapted to connector terminal semi-finished products of various specifications and structures.

[0020] 4. Collaborative Data Analysis: The cloud server can aggregate testing data from multiple devices, perform cross-device and cross-batch statistical analysis, promptly identify systemic quality issues, and support quality traceability.

[0021] 5. Continuous Model Optimization: The cloud server continuously optimizes the AI ​​analysis model based on accumulated big data and distributes the optimized model parameters to each industrial control computer to realize remote updates and continuous improvement of the model.

[0022] 6. Intelligent early warning function: The cloud server can monitor the detection status of each device in real time, and automatically send early warning information when an abnormality is detected, so as to realize intelligent quality management. Attached Figure Description

[0023] Figure 1 This is a three-dimensional external structural diagram of the machine vision-based general-purpose connector terminal semi-finished product appearance inspection equipment according to Embodiment 1 of the present invention. Figure 2 This is a front-view internal three-dimensional structural diagram of the machine vision-based universal connector terminal semi-finished product appearance inspection equipment according to Embodiment 1 of the present invention. Figure 3 This is a rear-view internal three-dimensional structural diagram of the machine vision-based universal connector terminal semi-finished product appearance inspection equipment according to Embodiment 1 of the present invention. Figure 4 This is a three-dimensional structural diagram of the fixture for the machine vision-based universal connector terminal semi-finished product appearance inspection equipment according to Embodiment 1 of the present invention. Figure 5 This is a three-dimensional structural diagram of the fixture for the machine vision-based universal connector terminal semi-finished product appearance inspection equipment according to Embodiment 2 of the present invention. Figure 6This is an exploded three-dimensional structural diagram of the fixture for the machine vision-based universal connector terminal semi-finished product appearance inspection equipment according to Embodiment 2 of the present invention. Figure 7 This is a three-dimensional structural diagram of the fixture in use for the machine vision-based universal connector terminal semi-finished product appearance inspection equipment according to Embodiment 2 of the present invention. Figure 8 This is a three-dimensional structural diagram of a connector terminal semi-finished product, which is based on a machine vision-based universal connector terminal semi-finished product appearance inspection device according to Embodiment 2 of the present invention.

[0024] Figure 9 This is a photograph of a semi-finished pluggable terminal of a high-speed cable connector tested according to Embodiment 1 of the present invention. Figure 10 This is a schematic diagram of the plug-in terminal structure for subsequent manufacturing of the test product according to Embodiment 1 of the present invention; Figure 11 This is a schematic diagram of the output of the solder joint inspection result of the general-purpose connector terminal semi-finished product appearance inspection equipment based on machine vision according to Embodiment 1 of the present invention; Figure 12 This is a schematic diagram of the solder joint inspection result output by the machine vision-based general-purpose connector terminal semi-finished product appearance inspection equipment according to Embodiment 1 of the present invention. Figure 13 This is a schematic diagram illustrating an application scenario of the high-speed cable connector finished product prepared by testing the product in Embodiment 1 of the present invention.

[0025] Figure 14 This is a schematic diagram of the main process of the detection method in an embodiment of the present invention; Figure 15 This is a flowchart illustrating the multi-scale weld joint area focusing detection steps in an embodiment of the present invention; Figure 16 This is a flowchart illustrating the adaptive soft tag detection steps in an embodiment of the present invention; Figure 17 This is a schematic diagram of the CPU-based efficient neural network acceleration steps in an embodiment of the present invention.

[0026] In the picture: 1. Frame; 11. Base panel; 12. Internal support; 13. Housing; 2. Fixture; 21. Insert panel; 22. Fixing bracket; 23. Insert slot; 24. Positioning pin; 25. Buckle; 26. Slot; 27. Screw hole; 28. Support leg; 3. Detection unit; 31. Vertical area array camera module; 32. Horizontal area array camera module; 33. Vertical light source adjustment module; 34. Horizontal light source adjustment module; 4. Control and display unit; 41. Industrial computer; 42. Input keyboard; 43. Display; 44. Focusing motor controller; 45. Power adapter; 46. Light source controller; 5. Connector terminal semi-finished products; 51. Barcode area. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Basic Implementation See appendix Figure 1-17 The connector terminal semi-finished product appearance inspection equipment based on binocular machine vision and AI model provided in this embodiment includes: A rack, the rack comprising a desktop base panel, an internal support, and a housing; A binocular machine vision inspection unit (hereinafter referred to as the inspection unit) is mounted on the frame and includes a vertical area array camera module and a horizontal area array camera module. The vertical area array camera module and the horizontal area array camera module are arranged orthogonally or obliquely (orthogonally or obliquely in the vertical and horizontal spatial directions). The optical focal points of the two camera modules are fixed target positions on the fixture. The effective detection area of ​​the vertical area array camera module covers the upper surface of the fixture, and the effective detection area of ​​the horizontal area array camera module covers the front surface of the fixture. Preferably, the effective detection area of ​​the vertical area array camera module should cover the entire upper surface of the fixture's insert panel, and the effective detection area of ​​the horizontal area array camera module should cover the entire front surface of the fixture's insert panel. The fixture, mounted on the base panel, includes a transparent fixing bracket and a detachable insert panel. The insert panel has an insert groove that matches the shape and structure of the product to be tested. During testing, the product is inserted into the insert groove and positioned so that the upper surface and front surface of the product are flush with the upper surface and front surface of the insert panel. The fixing bracket and the insert panel are detachably connected via a positioning pin and a snap-fit ​​structure. A pressure sensor is also installed in the insert groove. When the connector terminal semi-finished product is fed into the insert groove, the sensor sends a trigger signal to the industrial control computer to start the test. The insert panel is made of anti-static acrylic or transparent PC material, preferably with a surface resistivity of 10 ppm. 8 -10¹ 0The embedded panel is made of Ω-sized antistatic acrylic. It features a quick-change structure, allowing for replacement of a single panel in less than 60 seconds. The mounting bracket is compatible with various panel sizes, enabling rapid switching between different product models.

[0029] The control and display unit includes an industrial computer, an input keyboard, and a display. The industrial computer has a built-in AI model, including a lightweight deep learning model, a multi-scale solder joint region focusing module, and an adaptive soft labeling strategy module, for preliminary defect identification of the acquired images. In a specific embodiment, the feature fusion of the multi-scale solder joint region focusing module can adopt an attention fusion method, assigning attention weights to feature maps of different scales (1×1 scale weight 0.4, 0.75×0.75 scale weight 0.3, 0.5×0.5 scale weight 0.2, 0.25×0.25 scale weight 0.1), and obtaining the fused feature map after weighted summation. The cloud server communicates with the industrial control computer via a wired or wireless network. The cloud server has a built-in high-performance AI analysis program, which is used to receive the detection data and preliminary analysis results uploaded by the industrial control computer, make secondary judgments, and then save the data in the cloud.

[0030] More specific testing equipment includes a binocular machine vision inspection unit 3, a control and display unit 4, and a fixture 2 mounted on the frame 1; The frame 1 includes a desktop base panel 11, an internal support 12, and multiple housings 13. A fixture 2 is disposed outside the housings 13. The fixture 2 includes a removable, easily replaceable transparent insert panel 21 made of anti-static acrylic material with a surface resistivity of 10⁻⁶ Ω·cm. 8 -10¹ 0 Ω can effectively prevent the connector terminal semi-finished product 5 under test from attracting impurities due to static electricity, and can prevent damage to electronic components; the insert panel 21 is provided with 3 through-holes with arc-shaped cross sections, which match the shape and structure of the connector terminal semi-finished product 5 under test. The inserts are used to insert the connector terminal semi-finished product 5 under test into the inserts 23 during testing and to position the connector terminal semi-finished product 5 under test, so that the upper end face of the connector terminal semi-finished product 5 under test is flush with the upper end face of the insert panel 21 and the front end face of the connector terminal semi-finished product 5 under test is flush with the front end face of the insert panel 21. The embedding slot 23 of the embedding panel 21 is adapted to the shape and size of the connector terminal semi-finished product 5 to be tested. After the connector terminal semi-finished product 5 is fed into the embedding slot 23, the front end face of the connector terminal semi-finished product 5 to be tested faces the effective detection area of ​​the horizontal area array camera module 32, and the upper end face faces the effective detection area of ​​the vertical area array camera module 31. Different structures or specifications of connector terminal semi-finished products 5 to be tested correspond to different embedding slots 23 of the fixture 2. A pressure sensor is also provided in the embedding slot 23. When the connector terminal semi-finished product 5 is fed into the embedding slot 23, the sensor sends a trigger signal to the industrial control computer 41 to start the test.

[0031] The fixture 2 also includes a transparent fixing bracket 22, which includes two symmetrically arranged, identical left and right legs 28. The overall outline of each leg 28 is an inverted T-shape, and four screw holes 27 are evenly distributed along the rectangular outline on the bottom end face, which are detachably connected to the base panel 11. Correspondingly, the base panel 11 of the frame 1 has pre-drilled screw holes 27 that are matched one by one with the four screw holes 27. During installation, the bottom end face of the fixing bracket 22 is placed against the corresponding area of ​​the base table, so that the screw holes 27 are precisely aligned. The fixing bracket 22 is then screwed and tightened onto the base table using bolts. This connection structure ensures that the contact gap between the fixing bracket 22 and the base table is ≤0.1mm, effectively ensuring the installation stability of the fixture 2 and avoiding interference of equipment operation vibration on the focusing accuracy of the binocular machine vision inspection unit 3.

[0032] The embedded panel 21 is detachably connected to the fixed bracket 22 via a "buckle 25 + positioning pin 24" structure, enabling quick replacement of the embedded panel 21. The replacement time for a single embedded panel 21 is ≤10 seconds. On the top end face of the support leg 28 of the fixed bracket 22, and on the side that is in contact with the bottom end face of the embedded panel 21, 2-4 cylindrical positioning pins 24 are symmetrically arranged in the horizontal direction. Correspondingly, pin holes matching the size of the positioning pins 24 are opened on the bottom end face of the embedded panel 21. The diameter of the positioning pin 24 is 6mm±0.002mm, the length is 8mm, the inner diameter of the pin hole is 6mm±0.003mm, and the depth is 10mm, ensuring that the positioning pin 24 is fully embedded. When installing the embedded panel 21, the pin hole first precisely engages with the positioning pin 24, directly limiting the horizontal displacement of the embedded panel 21, ensuring that the repeatability of the positioning after replacement is ≤0.01mm, which is compatible with the focusing requirements of the detection unit 3.

[0033] On both long sides of the insert panel 21, i.e. the sides that contact the fixed bracket 22, there is a buckle 25. The buckle 25 is an elastic cantilever integrally formed on the side of the insert panel 21, and the material is the same as that of the insert panel 21. The end of the cantilever has an outward protruding "clip head". Correspondingly, on the vertical surface of the support leg 28, there is a slot 26 that matches the buckle 25. The buckle 25 of the insert panel 21 is detachably connected to the slot 26, so that the insert panel 21 is detachably mounted on the fixed bracket 22. When testing different connector terminal semi-finished products 5, it can be easily replaced with a suitable insert panel 21. The groove 26 on the support leg 28 is 4mm deep, and a guide slope is provided at the opening of the groove 26 to facilitate the insertion of the buckle 25. Align the pin hole of the insert panel 21 with the positioning pin 24 of the fixing bracket 22 and insert it. Push the insert panel 21 gently until it fits. After the cantilever of the buckle 25 is deformed by compression, the buckle head automatically engages with the groove 26 of the fixing bracket 22 to complete the fixation. Press the elastic cantilever of the buckle 25 with your finger to disengage the buckle head from the groove 26, and at the same time pull the insert panel 21 upward to remove it quickly.

[0034] The binocular machine vision inspection unit 3 includes a vertical area array camera module 31 and a horizontal area array camera module 32. The effective detection area of ​​the vertical area array camera module 31 covers the upper surface of the entire fixture 2, and the effective detection area of ​​the horizontal area array camera module 32 covers the front surface of the entire fixture 2. The vertical area array camera module 31 and the horizontal area array camera module 32 of the binocular machine vision inspection unit 3 are respectively mounted on the internal support 12 and partially covered by the housing 13, with their lenses exposed and facing the fixture 2. The vertical area array camera module 31 includes a vertical industrial camera and a vertical light source adjustment module 33, and the horizontal area array camera module 32 includes a horizontal industrial camera and a horizontal light source adjustment module 34. Both the vertical and horizontal industrial cameras are electrically connected to the focusing motor controller 44, so that the lens focus point of the vertical industrial camera is aligned with the center of the top end face of the insert slot 23, and the lens focus point of the horizontal industrial camera is aligned with the center of the side end face of the insert slot 23. The vertical light source adjustment module 33 and the horizontal light source adjustment module 34 are both high-angle ring light sources composed of multiple rings, multiple angles, and multiple LED beads. They are electrically connected to the light source controller 46, which adjusts the illumination angle of the ring light source. The vertical light source adjustment module 33 works with the vertical industrial camera in the vertical direction to optimize the illumination angle to 45°-60° for the barcode area 51. The horizontal light source adjustment module 34 works with the horizontal industrial camera in the horizontal direction to provide uniform side illumination, highlighting scratches, dents, and other defects on the side of the connector terminal semi-finished product 5.

[0035] The control and display unit 4 includes an industrial computer 41, an input keyboard 42, a display 43, a power adapter 45, a light source controller 46, and a focus motor controller 44. The industrial computer 41 is mounted on an internal bracket 12 and enclosed by a housing 13. The input keyboard 42 and the display 43 are located in the middle of the internal bracket 12. The power adapter 45, the light source controller 46, and the focus motor controller 44 are all located inside the housing 13 and are electrically connected to the industrial computer 41, the vertical area array camera module 31, and the horizontal area array camera module 32. The industrial computer 41 uses an NVIDIA Jetson Xavier NX development board, with a built-in six-core ARM CPU and a 384-core Volta GPU. Its interfaces include two USB 3.0 Type-A ports for connecting machine vision inspection modules, one HDMI 2.0 port for connecting the display 43, one Gigabit Ethernet RJ45 port for data transmission and remote debugging, and one M.2 Key M... The system includes an interface for expanding the storage module to store detection data, and an RS485 interface for connecting the focusing motor controller 44 and the light source controller 46. The industrial computer 41 is pre-installed with TensorFlow and PyTorch deep learning frameworks (conventional software), supporting the training and inference of various neural network models. The industrial computer 41 is electrically connected to the detection unit 3, the positioning sensor of the fixture 2, the focusing motor controller 44, the power adapter 45, the light source controller 46, and the display 43 via the interface. After receiving the positioning sensor signal, the industrial computer 41 synchronously controls the detection unit 3 to start image acquisition, simultaneously identifies the barcode area 51 to read barcode information, and binds and stores the detection data with the barcode information. The industrial computer 41 pre-stores a neural network detection model (conventional software) adapted to linear products with solder joints, capable of detecting defects such as solder joint connection status, parallel terminal pin parallelism, metal semi-enclosed protective shell deformation, and injection molding buckle 25 positioning.

[0036] The focusing motor controller 44 is used to control the autofocus of both vertical and horizontal industrial cameras. The controller uses an STM32F4 series microcontroller as its core control chip, operates at 24V DC, and has a rated operating current ≤1A. Its interface includes one RS485 communication interface for receiving focusing commands from the industrial computer 41, a drive interface for connecting the drive motor of the motion module, and a power input interface using a DC 5.5mm × 2.1mm connector. The focusing motor controller 44 can output subdivided drive signals to achieve focusing stroke control with an accuracy of 0.005mm, a response delay ≤5ms, and an image acquisition frame rate ≥30fps, ensuring clear images from both the vertical and horizontal industrial cameras.

[0037] The power adapter 45 uses an AC-DC switching power supply with an input voltage of AC 100V-240V and a dual-channel DC output. The main channel is 24V DC / 5A, powering the industrial computer 41 and the detection unit 3, while the auxiliary channel is 12V DC / 5A, powering the focusing motor controller 44 and the light source controller 46. The output interface uses a multi-channel DC terminal block, with the main channel interface being a 5.5mm×2.1mm DC plug and the auxiliary channel interface being a 4-pin Phoenix terminal. The power adapter 45 has overvoltage protection, overcurrent protection, and short-circuit protection functions, meeting the industrial-grade power supply stability requirements.

[0038] The light source controller 46 is connected to the vertical light source adjustment module 33 and the horizontal light source adjustment module 34 respectively, controlling the brightness and switching of the two light source adjustment modules. The light source controller 46 adopts a constant current drive circuit with a working voltage of 24V DC, which can accurately control the brightness and illumination angle of the LED light source to provide the best lighting effect. The interface types include one RS485 communication interface for receiving brightness adjustment commands from the industrial control computer 41, and two 12V DC output interfaces with 2-pin terminals, which are connected to the vertical area array camera module 31 and the horizontal area array camera module 32 respectively.

[0039] The frame 1 features a tabletop design, providing operators with a flat and spacious work surface, facilitating the integrated installation of various equipment components. It also allows operators easy access to the connector terminal semi-finished products 5, enhancing the convenience of testing operations. The frame 1 is constructed from aluminum alloy profiles with an anodized surface treatment, ensuring both load-bearing stability and ease of assembly, handling, and subsequent structural adjustments. The anodizing treatment forms a dense oxide film on the surface of the frame 1, improving rust and corrosion resistance, enhancing surface wear resistance, preventing scratches during daily operation, and resulting in a clean and easy-to-clean appearance, extending the overall service life of the equipment. The frame 1 is equipped with adjustable feet at the bottom, with an adjustment range of 0-50mm, allowing for flexible leveling based on the flatness of the ground. This prevents vibrations caused by tilting during operation, ensuring the focusing accuracy of the testing unit 3, and adapting to different workstation heights, improving the equipment's site adaptability.

[0040] The AI ​​model built into the industrial control computer includes a multi-scale solder joint region focusing module. This module performs multi-scale downsampling, central region enhancement, and local texture sharpening on the original solder joint image, enabling the network to obtain high-resolution feature representation of the solder joint region at an early stage. By performing feature reconstruction and weighted fusion on local windows of different scales, the model can simultaneously focus on the overall shape of the solder joint and the local melting texture, thereby enhancing its ability to perceive the features of the solder joint under various lighting conditions, including minor deformations, irregular edges, and small shapes.

[0041] The AI ​​model built into the industrial control computer includes an adaptive soft labeling strategy module. The industrial control computer generates multi-peak flexible labels based on the degree of melting of the solder joints, enabling explicit modeling of the category transition relationship between "normal solder joints - slightly melted solder joints - severely melted solder joints". The Gaussian weight distribution of the adaptive soft labeling strategy module is calculated using a Gaussian function with a mean μ=0 and a variance σ=0.1-0.3. The vector dimension of the flexible label is consistent with the number of defect categories. For samples in the transition zone, the label weights are smoothly distributed among adjacent categories according to a Gaussian function. For example, the label vector of slightly melted solder joints is [0.15, 0.70, 0.15]. For samples in the slightly melted state, the label distribution is smoothly distributed among adjacent categories in a Gaussian manner, enabling the network to learn the gradual boundary characteristics between categories during training, thereby significantly improving the classification accuracy of boundary samples and samples with large morphological differences within a category.

[0042] The cloud server is a virtual server on a cloud platform, with a built-in deep learning classification model that matches the AI ​​model in the industrial control computer. It supports parallel processing of detection data from multiple industrial control computers. The cloud server has a computing power of no less than 32 cores and 64GB of memory, and a network bandwidth of no less than 100Mbps. The AI ​​model of the cloud server is updated synchronously with the AI ​​model of the industrial control computer. The cloud server can also perform in-depth analysis of the stored detection data to generate quality statistical reports and defect trend analysis reports, providing data support for quality optimization of the production line.

[0043] The connector terminal semi-finished product appearance inspection method based on binocular machine vision and AI model provided in this embodiment uses the aforementioned inspection equipment and includes the following steps: S1. Product Placement: Place the semi-finished connector terminal to be tested into the insert groove of the fixture. After the pressure sensor in the insert groove detects that the product is in place, it sends a trigger signal to the industrial control computer. S2. Local Image Acquisition: After receiving the trigger signal, the industrial control computer synchronously controls the vertical area array camera module and the horizontal area array camera module to acquire images, respectively obtaining the top and front surface images of the product, and transmitting them to the industrial control computer via the Gige protocol. S3. Local Preliminary Analysis: After the industrial control computer preprocesses the acquired images, it inputs the built-in lightweight deep learning model to perform preliminary defect identification and generate local preliminary analysis results. The preliminary analysis results include defect candidate regions, preliminary classification confidence, and feature vectors. Preprocessing includes denoising, enhancement, and normalization of the acquired images; The specific process of local preliminary analysis includes: extracting features from the input image through a multi-scale feature extraction module, and then generating preliminary judgment results through a fast classifier; the lightweight deep learning model adopts MobileNetV3 or ShuffleNet architecture, with a model size ≤20MB and inference time ≤50ms; Alternatively, the preprocessed image can be input into an improved deep learning classification model in an industrial control computer. This model introduces a multi-scale weld point region focusing module before the backbone network and adopts an adaptive soft label strategy in the classification stage to perform preliminary defect detection in the region to be detected and generate preliminary analysis results. S4. Data Upload: The industrial control computer uploads the raw image data, preliminary analysis results, and product information to the cloud server with the built-in AI program via wired or wireless network. The data upload adopts a tiered upload strategy: for qualified products with a preliminary analysis confidence level ≥90%, only product information and a summary of analysis results are uploaded; for products with a preliminary analysis confidence level <90% or identified as defective, complete original image data and detailed analysis results are uploaded. The communication between the industrial control computer and the cloud server adopts the MQTT protocol or HTTP / HTTPS protocol, the data transmission adopts JSON format, and the AES-256 encryption algorithm is used to ensure data security. S5. Secondary cloud-based assessment: After receiving the uploaded data, the cloud server uses its built-in high-performance AI analysis program to perform a secondary assessment. The secondary assessment includes: fine-grained defect classification based on the Transformer model, historical trend analysis based on time-series data, and batch comparative analysis with similar products; the secondary assessment results are generated, and the detection data is saved to the cloud database. The specific process of the cloud-based secondary judgment includes: S5.1 Fine Defect Classification: The cloud server inputs the received image data into a fine analysis model based on Transformer to perform fine classification on the initially identified defect candidate regions. The fine analysis model adopts the SwingTransformer or DeiT architecture. S5.2 Historical Trend Analysis: The cloud server queries the historical testing data of the product based on the product barcode information, performs time-series analysis, and determines whether the defect is sporadic or has a worsening trend. S5.3 Batch Comparison Analysis: The cloud server statistically compares the test results of the current product with the batch test data of similar products to identify whether there are batch quality problems.

[0044] S6. Result Feedback: The cloud server returns the secondary judgment result and confidence score to the industrial control computer; S7. Result Output: The industrial control computer integrates the local preliminary analysis results and the cloud-based secondary judgment results to generate the final inspection report. The inspection results are bound and stored with the product barcode information, and the inspection images, defect marks and judgment results are displayed on the monitor. The specific rules for generating the final test report are as follows: when the local preliminary analysis result is consistent with the cloud secondary judgment result, the result is directly used as the final test result; when the two results are inconsistent, the result with higher confidence level shall prevail; if the confidence levels of both are lower than the set threshold, it is marked as requiring manual review.

[0045] The detection method also includes an offline detection mode: when the network connection between the industrial control computer and the cloud server is interrupted, the industrial control computer only relies on the local lightweight deep learning model to complete the detection and temporarily stores the detection data in the local storage; after the network is restored, the temporarily stored data is automatically uploaded to the cloud server for supplementary analysis.

[0046] The cloud server also performs the following data management functions: The test data is categorized and stored according to product model, production batch, and test time; quality statistical reports are generated regularly, including defect type distribution, defect rate trend, and equipment test efficiency indicators; the AI ​​analysis model is continuously optimized based on the accumulated big data, and the optimized model parameters are sent to each industrial control computer for local model updates.

[0047] The cloud server is deployed in a distributed architecture, including a load balancer, multiple AI analysis nodes and a data storage cluster; the AI ​​analysis nodes dynamically allocate detection tasks according to the load, and the processing capacity of a single analysis node is ≥100 items / minute.

[0048] The detection method also includes a real-time early warning function: when the cloud server detects the following situations, it automatically sends early warning information to the management personnel: the defect rate of a single device continuously exceeds the set threshold; the defect rate of the same type of product in the same batch exceeds the set threshold; the equipment detection efficiency is lower than the set threshold or communication is abnormal.

[0049] The industrial control computer and the cloud server adopt an edge-cloud collaborative computing architecture. The load distribution strategy of the collaborative computing is dynamically adjusted according to network bandwidth, cloud load, and detection timeliness requirements, specifically including: High-bandwidth, low-latency scenarios: The industrial control computer uploads the original image, and the cloud server performs a complete analysis; Low-bandwidth scenarios: The industrial control computer performs complete local analysis and only uploads the analysis results and key feature data; Balanced scenario: The industrial control computer performs the initial analysis, and the cloud server performs the secondary fine analysis.

[0050] The following describes several more specific embodiments.

[0051] Example 1 This embodiment uses a semi-finished linear 224Gbps high-speed cable connector terminal as the testing object to explain in detail the specific application of the equipment of the present invention. See also... Figure 9 and Figure 12 The product to be tested is a pluggable terminal semi-finished product of the NVIDIA GB300 NVLink cabinet high-speed transmission cable. Its structure includes: insulated cable (including metal core wire harness), metal semi-enclosed protective shell (including positioning holes), injection molded buckle, parallel terminal pins, and cable shielding layer. The semi-finished product has completed terminal crimping and preliminary packaging, but has not been assembled into a complete connector assembly.

[0052] The main structure of the testing equipment in this embodiment is the same as that in the basic embodiment. Its testing items include: (1) Solder connection status: whether the laser welding point is complete, without false welding, and without air holes; (2) Parallelism of parallel terminal pins: whether multiple pins are kept parallel and whether the spacing deviation is within ±0.05mm; (3) Deformation of metal semi-enclosed protective shell: whether the protective shell has dents, scratches or deformation; (4) Positioning accuracy of injection molding buckle: whether the buckle position meets the design specifications; (5) Insulation between parallel pins and core wire harness: whether the pins are in contact with the cable shielding layer.

[0053] I. System Architecture The detection system in this embodiment adopts an edge-cloud collaborative architecture, which includes two parts: an on-site detection layer and a cloud analysis layer.

[0054] 1. On-site inspection layer: including racks, binocular machine vision inspection units, fixtures and industrial control computers.

[0055] The frame is a tabletop structure made of aluminum alloy profiles, with an anodized surface and adjustable feet at the bottom (adjustment range 0-50mm).

[0056] The binocular machine vision inspection unit includes a vertically arranged area array camera module and a horizontally arranged area array camera module, both orthogonally positioned. The vertical area array camera module includes a 5-megapixel industrial camera (Basler acA2500-14uc) equipped with a 35mm fixed-focus industrial lens, a resolution of 2592×1944 pixels, and a frame rate ≥30fps. The vertical light source adjustment module uses a high-angle ring light source composed of 36 white LEDs (6500K color temperature), with an adjustable illumination angle of 45°-60°. The horizontal area array camera module has the same configuration as the vertical module, and its horizontal light source adjustment module uses blue LEDs (470nm wavelength). The spatial position deviation of the focus points of the two camera modules is ≤0.02mm.

[0057] The fixture includes a fixed bracket and a detachable insert panel. The fixed bracket is connected to the frame via four M6 screw holes, with a fitting gap ≤0.1mm. The insert panel is connected to the fixed bracket via a "positioning pin + snap-fit" structure. The positioning pin diameter is 6mm±0.002mm, the pin hole inner diameter is 6mm±0.003mm, the repeatability is ≤0.01mm, and the replacement time for a single piece is ≤10 seconds. The insert slot dimensions are 25mm (L) × 8mm (W) × 5mm (H), with a tolerance of ±0.02mm, and the inner wall has a 0.5mm thick silicone buffer layer. The insert panel is made of anti-static acrylic material with a surface resistivity of 10. 8 -10¹ 0 Ω. A pressure sensor (FSR402) is installed in the insert slot, with a detection threshold of 0.5N.

[0058] The industrial PC uses an NVIDIA Jetson Xavier NX development board, featuring a built-in six-core ARM Carmel CPU (1.9GHz), a 384-core Volta GPU (1.37GHz), 8GB LPDDR4x memory, and 16GB eMMC storage. Interfaces include: 2 x USB 3.0, 1 x HDMI 2.0, 1 x Gigabit Ethernet RJ45, 1 x M.2 Key M, and 1 x RS485. It comes pre-installed with TensorFlow 2.8 and PyTorch 1.10.

[0059] 2. Cloud analytics layer: This includes cloud server clusters, deployed using a distributed architecture.

[0060] Cloud server hardware configuration: It adopts an Alibaba Cloud ECS instance (ecs.gn7i-c8g1.2xlarge), equipped with an 8-core Intel Xeon Ice Lake processor, 31GB of memory, 1×NVIDIA A10 GPU (24GB of video memory), and 500GB of SSD storage.

[0061] Cloud server software architecture: Utilizing the Kubernetes container orchestration platform, multiple AI analysis Pods are deployed. The AI ​​analysis program is developed based on the PyTorch framework and includes: a fine-grained defect classification model based on Swin Transformer (88M parameters), a time-series analysis model based on LSTM, and a batch comparison analysis module based on statistical learning.

[0062] 3. Communication Connection: The industrial control computer and the cloud server are connected via gigabit Ethernet, using the MQTT over TLS 1.3 protocol for data transmission. The data format is JSON, and the AES-256-GCM encryption algorithm is used to ensure data security. Network latency is ≤20ms, and bandwidth is ≥100Mbps.

[0063] II. AI Model Architecture 1. Lightweight local model on industrial PC: Utilizing the MobileNetV3-Large architecture, with an input size of 512×512 pixels, 5.4M parameters, and a model size of approximately 5.4MB. The model includes: a multi-scale feature extraction module (using MBConv blocks with an expansion coefficient of 6), an SE attention module, and a fast classifier (fully connected layers outputting 4 classes: normal, slight melting, severe melting, and cold solder joints / porosity). Inference time is approximately 35ms, achieving a frame rate ≥25fps on Jetson Xavier NX.

[0064] The model training dataset contains 15,000 solder joint images (5,000 normal, 5,000 slightly melted, 3,000 severely melted, and 2,000 cold solder joints / porosity). Data augmentation includes random rotation (±15°), random scaling (0.9-1.1x), and random brightness adjustment (±20%). Training parameters: batch size 64, initial learning rate 0.001, Adam optimizer, training for 150 epochs.

[0065] 2. High-performance cloud-based model: Employs the Swin Transformer-Tiny architecture, with an input size of 384×384 pixels and 28M parameters. The model includes: Patch Embedding layer, 4 Swin Transformer Stages (layer number [2,2,6,2]), Layer Normalization layer, global average pooling layer, and classification head. Inference time is approximately 11ms (on an A10 GPU), with a classification accuracy of 99.5%.

[0066] 3. Time series analysis model: A two-layer LSTM architecture is adopted with a hidden layer dimension of 128. The input is the feature vector and classification results of the most recent 30 detections, and the output is the defect trend prediction (stable, deteriorating, occasional).

[0067] III. Testing Process The edge-cloud collaborative detection process in this embodiment is as follows: Step S1, Product Placement: The operator places the semi-finished linear 224Gbps high-speed cable terminal to be inspected into the insert slot of the fixture, ensuring that the barcode area is fully exposed. After the product is in place, the pressure sensor detects a pressure change (≥0.5N) and sends a trigger signal to the industrial control computer.

[0068] Step S2, Local Image Acquisition: After receiving the trigger signal, the industrial control computer sends a focusing command to the focusing motor controller via the RS485 interface (response delay 3ms), and simultaneously controls the light source controller to adjust two sets of light sources (vertical light source at 50°, horizontal light source for uniform illumination). The top-mounted color area scan camera module and the horizontal area scan camera module synchronously acquire images, both with a resolution of 2592×1944 pixels, and transmit them to the industrial control computer via a USB 3.0 interface.

[0069] Step S3, Local Preliminary Analysis: The industrial control computer preprocesses the acquired images (Gaussian denoising, histogram equalization, and size normalization to 512×512 pixels), and then inputs them into the MobileNetV3-Large model for inference. The model output includes: defect candidate region coordinates (x, y, w, h), preliminary classification results (4 classes), confidence scores for each class, and a 512-dimensional feature vector. The inference time is 35ms, and the preliminary analysis results are temporarily stored in a memory buffer.

[0070] Step S4, Data Upload: The industrial control computer executes a tiered upload strategy based on the preliminary analysis results. (1) For products with a preliminary analysis confidence level ≥ 90% and deemed qualified: only upload a summary of the analysis results in JSON format (product barcode, judgment result, confidence level, timestamp), with a data volume of approximately 200 bytes and an upload time ≤ 5ms.

[0071] (2) For products with a preliminary analysis confidence level of <90% or determined to be defective: upload complete original image data (2 images, approximately 800KB after JPEG compression) and detailed analysis results (including defect candidate regions, feature vectors, and heat maps), with a data volume of approximately 1MB and an upload time of ≤100ms.

[0072] Step S5, Secondary Cloud-Based Judgment: After receiving the uploaded data, the cloud server initiates the following analysis process: S5.1 Fine-grained Defect Classification: Image data is input into the Swing Transformer model to perform fine-grained classification of candidate defect regions. The model outputs more accurate classification results and confidence scores, and simultaneously generates an interpretable heatmap, annotating key feature regions of defects.

[0073] S5.2 Historical Trend Analysis: Based on the product barcode, query the product's inspection records for the last 30 times, extract the feature vector sequence, input it into the LSTM model, and analyze the defect trend. If a worsening trend is detected, mark the warning level.

[0074] S5.3 Batch Comparison Analysis: Statistically compare the test data of the current product with those of the same batch (the most recent 1000 pieces), calculate indicators such as defect rate and defect type distribution, and identify whether there are batch quality problems.

[0075] The total time for the second cloud-based judgment is approximately 80ms (including network transmission).

[0076] Step S6, Result Feedback: The cloud server encapsulates the secondary judgment results (fine classification results, confidence score, trend analysis conclusion, batch comparison conclusion) into JSON format and returns them to the industrial control computer. The returned data volume is approximately 500 bytes, and the return time is ≤10ms.

[0077] Step S7, Result Output: The industrial control computer integrates the preliminary local analysis results and the secondary judgment results from the cloud to generate the final test report. (1) If the results of the local and cloud are consistent: directly adopt the results and calculate the overall confidence level as the weighted average of the confidence levels of the two (local weight 0.3, cloud weight 0.7).

[0078] (2) If the results on the local machine and the cloud machine are inconsistent: compare the confidence levels of the two and take the result with the higher confidence level; if the confidence levels of both machines are below 80%, mark it as "requires manual review".

[0079] The final test results are displayed on the monitor in real time, including: product image, defect marker box, judgment result (pass / fail / requires review), overall confidence level, and trend indication. If the product is judged to be unqualified, a 2kHz alarm sound will be emitted by the speaker (lasting 500ms). At the same time, the test data (product barcode, image, analysis results, timestamp) is bound and stored in the local database and uploaded to the cloud server for persistent storage.

[0080] IV. Cloud Server Data Management Functions The cloud server in this embodiment also performs the following data management functions: 1. Data Classification and Storage: Time-series database (InfluxDB) is used to store testing time-series data, object storage (MinIO) is used to store image files, and relational database (PostgreSQL) is used to store product information and quality statistics. Data is indexed by product model, production batch, and testing time, supporting fast querying.

[0081] 2. Quality Statistics Report: A quality statistics report for the previous day is automatically generated every morning at midnight, including: a pie chart of defect type distribution, a defect rate trend curve, a comparison of testing efficiency for each device, and a batch quality comparison. The report is stored in PDF format and emailed to quality management personnel.

[0082] 3. Continuous Model Optimization: Every Sunday at midnight, the cloud server performs incremental model training based on the detection data from the most recent week (approximately 50,000 records). A transfer learning strategy is employed to fine-tune the classification head while maintaining the pre-trained weights. After training, the new model's effectiveness is verified through A / B testing. If the accuracy improves by ≥0.5%, a new version of the model is automatically released. Model parameters (approximately 5.4MB) are distributed to each industrial control computer via differential updates, with an update time of ≤2 minutes.

[0083] V. Offline Detection Mode When the network connection between the industrial control computer and the cloud server is interrupted, the system automatically switches to offline detection mode: (1) The industrial control computer relies solely on the local MobileNetV3 model to complete the detection. The detection process is the same as the normal mode, but skips the data upload and cloud analysis steps.

[0084] (2) The detection data (images, analysis results) are temporarily stored in the local storage (M.2 SSD, capacity 256GB, which can store about 30,000 records).

[0085] (3) The system attempts to reconnect to the cloud server every 30 seconds. After the network is restored, it automatically uploads the temporary data in batches according to the time sequence. After the upload is completed, the local cache is cleared.

[0086] (4) The detection accuracy rate is slightly lower in offline mode (approximately 98.5% vs 99.5%), but still meets production requirements.

[0087] VI. Real-time early warning function The cloud server monitors the detection status of each device in real time. When the following situations are detected, it automatically sends an alert to the administrator via the WeChat / DingTalk API: (1) If the defect rate of 50 consecutive products on a single machine exceeds 5%, send an “Equipment Abnormal Warning” and suggest checking the equipment status.

[0088] (2) If the rate of similar defects in the same batch of products exceeds 3%, send a “batch quality warning” and suggest investigating the production process.

[0089] (3) If the equipment detection efficiency is less than 20 pieces / minute or the communication delay exceeds 500ms: send a "performance abnormality warning" and suggest checking the network or equipment load.

[0090] VII. The operational test results of the prototype in this embodiment are as follows: (1) Detection speed: The complete detection time for a single product is ≤500ms (including cloud analysis), which is about 60% faster than pure local detection (which requires the deployment of a large model and an inference time of about 200ms) and about 75% faster than pure cloud detection (which uploads the original image and has a transmission time of about 500ms).

[0091] (2) Detection accuracy: The overall detection accuracy rate is 99.5%, of which the accuracy rate of weld point defect identification is 99.5% and the accuracy of terminal parallelism detection is 0.01mm.

[0092] (3) Resource requirements: The industrial control computer only needs to run a lightweight model (20MB), and the GPU memory usage is <1GB, which greatly reduces the hardware cost.

[0093] (4) Data value: Data from multiple devices is aggregated in the cloud, supporting cross-device and cross-batch analysis, and timely detection of systemic quality problems.

[0094] (5) Operation and maintenance efficiency: The model is updated remotely and automatically, eliminating the need for operation on each machine, reducing maintenance costs by 80%.

[0095] Example 2 This embodiment uses a bent (L-shaped) high-speed cable connector terminal semi-finished product (see...). Figure 8 Using the sample as the detection object, this invention demonstrates the adaptability of the edge-cloud collaborative detection device to products with different structures.

[0096] I. Differences in Equipment Configuration The main structure of the detection equipment in this embodiment is basically the same as that in the basic embodiment, with the main difference being: 1. Dedicated Insert Panel: Utilizes an L-shaped insert groove adapted for bent terminals, with a groove depth of 6mm and a 2mm radius of curvature at the bend, perfectly conforming to the product's shape. The insert groove sidewalls are equipped with polyurethane elastic positioning ribs, providing 2-3N lateral clamping force. A 2mm diameter vent hole is located at the bottom of the insert groove for easy product placement and removal.

[0097] 2. Camera Angle Adjustment: The optical axis of the horizontal area array camera module is tilted at a 15° angle to the horizontal plane. This is precisely adjusted via the focusing motor controller to align the lens focus point with the center of the side of the bent terminal. The focusing travel is ±5mm, and the adjustment accuracy is 0.005mm.

[0098] 3. AI Model Adaptation: Deploy a dedicated analysis model for bent terminals in the cloud, adding a module for extracting geometric features of bent parts based on Swin Transformer.

[0099] II. Edge-Cloud Collaborative Detection Process The detection process in this embodiment is basically the same as that in the basic embodiment. The main difference lies in the secondary judgment stage in the cloud: Steps S1-S4 (product placement, local image acquisition, local preliminary analysis, data upload): Same as in Example 1.

[0100] Step S5, Secondary judgment in the cloud: S5.1 Detailed Analysis of Bending Areas: The cloud-based Swin Transformer model adds a feature extraction branch for bending areas, specifically analyzing the geometry, surface quality, and presence of cracks or deformation at the bend. The model was fine-tuned on a bending area detection dataset (8000 images), achieving a 98.5% accuracy rate in identifying bending defects.

[0101] S5.2 3D Shape Reconstruction: The cloud uses binocular images (top + side) for stereo matching to reconstruct the local 3D shape of the product, calculates the bending angle deviation (threshold ±2°), and determines whether it meets the specifications.

[0102] S5.3 Historical Trend Analysis: In addition to conventional time series analysis, add trend tracking of bending angle changes to identify whether there are process problems such as bending springback.

[0103] Steps S6-S7 (Result Feedback, Result Output): Same as in Example 1, the final test report adds the bending angle measurement value and the three-dimensional morphology visualization.

[0104] III. Load Balancing Strategies In this embodiment, the cloud server uses a dynamic load balancing strategy to handle detection requests from multiple devices: (1) The Kubernetes Horizontal Pod Autoscaler automatically scales up and down AI analysis Pods based on CPU utilization (threshold 70%) and the length of the pending queue (threshold 100), with a scaling response time of ≤30 seconds.

[0105] (2) The consistent hashing algorithm is used to allocate detection tasks to ensure that multiple detection requests for the same product are routed to the same analysis node, which facilitates time series analysis.

[0106] (3) The processing capacity of a single AI analysis node (A10 GPU) is ≥120 items / minute, and the total processing capacity of a single cluster (10 nodes) is ≥1200 items / minute.

[0107] IV. Operational Test Results of the Prototype in this Embodiment Changeover time ≤ 15 seconds (replacing the insert panel + adjusting the camera angle); defect detection rate at bending points 98.5%; bending angle measurement accuracy ±0.5°; average cloud analysis latency 120ms. This demonstrates the versatility and flexibility of the edge-cloud collaborative detection architecture of this invention.

[0108] Example 3 This embodiment, based on the aforementioned embodiments, further focuses on the inspection of solderless connector terminal semi-finished products, emphasizing the detection of defects such as parallel pin parallelism, injection molding snap-fit ​​positioning accuracy, metal semi-enclosed protective shell deformation, and cable shielding layer position. The difference lies in: I. Differences in Equipment Configuration 1. Insert panel: The insert groove is 12mm deep, suitable for the cable part of products without solder joints; the inner wall is equipped with a 0.5mm thick silicone buffer layer to prevent the pins from being deformed by collision.

[0109] 2. Light source parameters: The vertical light source illumination angle is adjusted to 30°-45° to highlight the top outline of the parallel pin; the horizontal light source illumination angle is adjusted to 40°-55° to highlight the defects of the injection molded buckle and the metal protective shell.

[0110] 3. Local model on industrial control computer: Using YOLOv5s target detection network, the model size is 7MB and the inference time is 18ms. Detected targets include: pins (category 1), clips (category 2), protective shells (category 3), and shielding layers (category 4).

[0111] 4. High-performance cloud-based model: Adopting the YOLOv5x architecture, with 86M parameters and mAP@0.5 reaching 97.2%. An added pin parallelism measurement algorithm: Extracting the pin edge contour, fitting a straight line to calculate the included angle, and determining whether the parallelism is qualified (threshold ±0.05mm corresponds to an angle deviation of ±0.3°).

[0112] II. Edge-Cloud Collaborative Detection Process Steps S1-S2 (product placement, local image acquisition): Same as in Example 1.

[0113] Step S3: Local Preliminary Analysis: The industrial computer runs the YOLOv5s model to perform target detection, outputting the bounding box coordinates, class confidence scores, and pin endpoint coordinates for each target. A preliminary estimate of pin parallelism is calculated based on the pin endpoint coordinates. Inference time is 18ms, and the preliminary analysis results are temporarily saved.

[0114] Step S4, Data Upload: A tiered upload strategy is adopted. (1) All detections upload target detection results (bounding box, category, confidence, pin endpoint coordinates), with a data volume of approximately 2KB.

[0115] (2) For products whose parallelism estimate is in the boundary area (±0.05mm±0.02mm), the original image is uploaded for detailed cloud analysis.

[0116] Step S5, Secondary judgment in the cloud: S5.1 Fine Object Detection: The cloud-based YOLOv5x model performs more accurate object detection on images, improving bounding box localization accuracy (IoU improvement of approximately 5%).

[0117] S5.2 Precise Measurement of Pin Parallelism: Based on cloud-based detection results, a sub-pixel edge detection algorithm (Canny + Zernike moments) is used to extract the pin edges. The pin centerline is then fitted using the least squares method to calculate the angle between pins, achieving a measurement accuracy of 0.01mm. Specifically, the sub-pixel edge detection algorithm first extracts the coarse outline of the pin edges using the Canny operator, then uses the Zernike moment operator to perform sub-pixel-level fitting on the coarse outline, calculating the sub-pixel coordinates of the edge points with a fitting accuracy of 0.01 pixels. Finally, the pin centerline is fitted using the least squares method to calculate the parallelism deviation.

[0118] S5.3 Batch Statistical Analysis: Statistically analyze the pin parallelism distribution of products in the same batch to identify whether there is a systematic offset (such as gradual change in parallelism caused by mold wear).

[0119] Steps S6-S7 (Result Feedback and Output): The cloud returns detailed inspection results and accurate parallelism measurement values, and the industrial control computer generates the final report. The inspection report includes: inspection results for each target, pin parallelism measurement values, comparison conclusions with specifications, and batch statistical prompts.

[0120] III. Dynamic load balancing between edge and cloud This embodiment adopts a dynamic load balancing strategy based on network conditions and detection timeliness requirements: (1) High bandwidth and low latency scenario (bandwidth ≥ 100Mbps, latency ≤ 20ms): The industrial control computer uploads the original image, and the cloud server performs complete target detection and parallelism measurement. The industrial control computer is only responsible for image acquisition and result display. In this mode, the industrial control computer has the lowest resource consumption (CPU < 20%).

[0121] (2) Low bandwidth scenario (bandwidth < 50Mbps): The industrial control computer performs complete local analysis (YOLOv5s + simple parallelism estimation), and only uploads the analysis results and key feature data (approximately 5KB). Batch statistical analysis is performed in the cloud, without processing individual products. In this mode, the network transmission time is ≤ 100ms.

[0122] (3) Balanced Scenario (Default): The industrial control computer performs YOLOv5s target detection and preliminary parallelism estimation, while the cloud performs fine measurement and batch analysis. The overall detection time in this mode is about 200ms.

[0123] The load balancing strategy is automatically decided by the cloud server based on real-time network detection results and is updated every 30 seconds.

[0124] IV. Prototype Operation Test Results in this Embodiment The detection accuracy is ≥99.5% (target detection mAP@0.5 97.2%); the pin parallelism measurement accuracy is 0.01mm; the detection speed is ≥50fps (local mode) or ≥20fps (cloud fine mode); it effectively solves the problems of different detection items and high detection difficulty of solderless products, and adapts to the production and detection needs of solderless products.

[0125] Example 4 This embodiment is based on the foregoing embodiments and is further applied to a hybrid production line that simultaneously produces semi-finished high-speed cable connector terminals of various models, including straight, bent, soldered, and solderless types. This production line needs to quickly switch between testing models to meet the production needs of multiple varieties in small batches. Its difference from the foregoing embodiments lies in: I. Main Equipment Parameters Vertical area array camera module: adopts a 25-megapixel industrial area array camera with a 12-36mm zoom lens, an aperture of F2.8, and a frame rate of 25fps; the vertical light source adjustment module is an adjustable LED ring light source with an illumination angle that can be adjusted between 20° and 60°, and a brightness adjustable range of 0-1500cd / ㎡.

[0126] Horizontal area array camera module: adopts a 25-megapixel industrial area array camera with a 12-36mm zoom lens, an aperture of F2.4, and a frame rate of 25fps; the horizontal light source adjustment module is an adjustable LED ring light source with an illumination angle that can be adjusted between 20° and 60°, and a brightness adjustable range of 0-1500cd / ㎡.

[0127] Fixture: Quick-change transparent PC material insert panel, equipped with 10 different insert slots, which can be quickly changed by clips, with a change time of less than 30 seconds. The trigger threshold of the pressure sensor in the insert slot can be adjusted to 4-7N according to the product type.

[0128] Industrial PC: Intel Core i9-12900K processor, 64GB DDR4 memory, 1TB SSD storage, equipped with OpenVINO+TensorRT hybrid acceleration toolkit, network bandwidth 100Mbps.

[0129] Cloud server: Huawei Cloud ECS server, configured with 64 cores, 128GB memory, 2TB SSD storage, 1000Mbps bandwidth, built-in deep learning classification model based on EfficientNet-B0, supports multi-model parallel inference, synchronizes model updates with industrial control computers, and supports parallel processing by multiple industrial control computers.

[0130] AI Model: A deep learning classification model based on an improved EfficientNet-B0. The multi-scale solder joint region focusing module has four scales (1×1, 0.75×0.75, 0.5×0.5, 0.25×0.25). The Gaussian smoothing coefficient of the adaptive soft label can be automatically adjusted according to the model. The model training uses 20,000 terminal image samples of various models, covering 10 different types of terminal semi-finished products.

[0131] II. Prototype Operation Test Results in this Embodiment Detection speed: The initial detection time of the industrial control computer is less than 0.2 seconds, the secondary judgment time of the cloud server is less than 0.05 seconds, and the overall single-product detection time is less than 0.25 seconds; compared with the detection method without cloud server collaboration, the detection speed is improved by 30%.

[0132] Inspection accuracy: The inspection accuracy of all models is above 99.2%, with a size inspection accuracy of ±0.01mm. Compared with the 80% accuracy of manual inspection, the inspection accuracy is improved by 19.2%; compared with the inspection method without cloud server collaboration, the inspection accuracy is improved by 0.2%.

[0133] Adaptability: A single production line is compatible with the testing of 10 different types of terminal semi-finished products without the need to change equipment. Only the fixture needs to be changed and the corresponding AI model needs to be loaded. The production line switchover time is less than 5 minutes.

[0134] Computing resources: The CPU utilization rate of the industrial control computer was reduced from 90% to 55%, and the memory utilization rate was reduced from 75% to 40%, which significantly reduced the computing pressure on the industrial control computer and extended its service life.

[0135] Data Management: The cloud server can automatically save all test data, generate multi-model defect data statistical reports and defect trend analysis reports, provide data support for production line quality analysis, realize cross-production line quality data comparison and analysis, and provide data basis for product design optimization.

[0136] The inspection equipment provided in the above embodiments of the present invention combines binocular machine vision, replaceable fixtures, improved AI models, and cloud server collaboration to achieve multi-face, multi-model, high-precision, high-speed, and low-computational-pressure appearance inspection of connector terminal semi-finished products. The inspection method based on this equipment employs a multi-scale solder joint area focusing module and an adaptive soft-label strategy deep learning classification model, combined with CPU acceleration technology and cloud server collaboration to achieve high-speed and high-precision inspection. This invention solves the problems of existing inspection equipment, such as limited inspection faces, slow speed, low efficiency, poor adaptability, and high computational resource requirements, and can be widely applied to the appearance quality inspection of various high-speed cable connector semi-finished products.

[0137] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A device for inspecting the appearance of semi-finished connector terminals based on binocular machine vision and AI models, characterized in that, include: A rack, the rack comprising a desktop base panel, an internal support, and a housing; A binocular machine vision inspection unit is mounted on the frame and includes a vertical area array camera module and a horizontal area array camera module. The vertical area array camera module and the horizontal area array camera module are arranged orthogonally or obliquely. The optical focus points of the two camera modules are fixed target positions on the fixture. The effective detection area of ​​the vertical area array camera module covers the upper end face of the fixture, and the effective detection area of ​​the horizontal area array camera module covers the front end face of the fixture. The fixture, mounted on the base panel, includes a transparent fixing bracket and a detachable insert panel. The insert panel has an insert groove that matches the shape and structure of the product being tested. During testing, the product is inserted into the insert groove and positioned so that the upper surface and the front surface of the product are flush with the upper surface and the front surface of the insert panel. The fixing bracket and the insert panel are detachably connected by a positioning pin and a snap-fit ​​structure. A pressure sensor is also installed in the insert groove. When the connector terminal semi-finished product is fed into the insert groove, the sensor sends a trigger signal to the industrial control computer to start the test. The control and display unit includes an industrial computer, an input keyboard, and a display. The industrial computer has a built-in AI model for preliminary defect identification of the acquired images. The cloud server communicates with the industrial control computer via a wired or wireless network. The cloud server has a built-in high-performance AI analysis program, which is used to receive the detection data and preliminary analysis results uploaded by the industrial control computer and make secondary judgments.

2. The detection device according to claim 1, characterized in that, The vertical area array camera module includes a vertical industrial camera and a vertical light source adjustment module. The lens of the vertical industrial camera is focused on the center of the top end face of the insert slot. The vertical light source adjustment module is a high-angle ring light source composed of multi-ring, multi-angle LED beads, with an adjustable illumination angle of 30°-60°.

3. The detection device according to claim 1, characterized in that, The horizontal area array camera module includes a horizontal industrial camera and a horizontal light source adjustment module. The lens of the horizontal industrial camera is focused on the center of the side end face of the insert slot. The horizontal light source adjustment module is a high-angle ring light source composed of multi-ring, multi-angle LED beads, with an adjustable illumination angle of 30°-60°.

4. The detection device according to claim 1, characterized in that, The fixed bracket also includes two symmetrically arranged left and right legs, the bottom end face of each leg being detachably connected to the base panel through screw holes; the insert panel is provided with a vertically extending elastic cantilever buckle, and the vertical surface of the leg is provided with a groove matching the buckle; after the elastic cantilever buckle of the insert panel is inserted into the groove of the leg, a detachable connection is formed between the insert panel and the leg; a cylindrical positioning pin is provided on the top end face of the leg, and a corresponding positioning hole is provided on the insert panel, which assists in guiding and positioning during the process of forming a detachable connection between the insert panel and the leg.

5. The detection device according to claim 1, characterized in that, The industrial control computer has a built-in AI model, which is a lightweight deep learning model. It includes a multi-scale feature extraction module, a fast classifier, a multi-scale solder joint region focusing module, and an adaptive soft labeling strategy module. The multi-scale solder joint region focusing module is used to perform multi-scale downsampling, central region enhancement, and local texture sharpening on the solder joint image. The adaptive soft labeling strategy module is used to generate multi-peak flexible labels based on the degree of solder joint melting and model the category transition relationship. The industrial control computer also has a built-in AI model inference acceleration module, which uses a CPU neural network acceleration method with SIMD instruction vectorization, memory layout rearrangement, and operator fusion, or a hybrid acceleration method of OpenVINO+TensorRT. The high-performance AI analysis program built into the cloud server includes a Transformer-based fine analysis model and a time series data analysis module.

6. A method for appearance inspection of connector terminal semi-finished products based on binocular machine vision and AI model, using the inspection equipment described in any one of claims 1-5, characterized in that, Includes the following steps: S1. Product Placement: Place the semi-finished connector terminal to be tested into the insert groove of the fixture. After the pressure sensor in the insert groove detects that the product is in place, it sends a trigger signal to the industrial control computer. S2. Local Image Acquisition: After receiving the trigger signal, the industrial control computer synchronously controls the vertical area array camera module and the horizontal area array camera module to acquire images, respectively obtaining the top surface image and the front surface image of the product; S3. Local Preliminary Analysis: After the industrial control computer preprocesses the acquired images, it inputs the built-in lightweight deep learning model to perform preliminary defect identification and generate local preliminary analysis results. The preliminary analysis results include defect candidate regions, preliminary classification confidence, and feature vectors. S4. Data Upload: The industrial control computer uploads the original image data, preliminary analysis results, and product information to the cloud server via wired or wireless network; S5. Secondary cloud-based assessment: After receiving the uploaded data, the cloud server uses its built-in high-performance AI analysis program to perform a secondary assessment. The secondary assessment includes: fine-grained defect classification based on the Transformer model, historical trend analysis based on time-series data, and batch comparative analysis with similar products; the secondary assessment results are generated, and the detection data is saved to the cloud database. S6. Result Feedback: The cloud server returns the secondary judgment result and confidence score to the industrial control computer; S7. Result Output: The industrial control computer integrates the local preliminary analysis results and the cloud secondary judgment results to generate a final inspection report. The inspection results are bound and stored with the product barcode information, and the inspection images, defect marks and judgment results are displayed on the monitor. When the network connection between the industrial control computer and the cloud server is interrupted, the system switches to offline inspection mode. The industrial control computer only relies on the local lightweight deep learning model to complete the inspection and temporarily stores the data in the local storage. After the network is restored, the temporarily stored data is automatically uploaded to the cloud server for supplementary analysis.

7. The detection method according to claim 6, characterized in that, The specific process of the local preliminary analysis in step S3 includes: extracting features from the input image through a multi-scale feature extraction module, and then generating preliminary judgment results through a fast classifier; the lightweight deep learning model adopts the MobileNetV3 or ShuffleNet architecture.

8. The detection method according to claim 6, characterized in that, The specific process of the cloud-based secondary judgment in step S5 includes: S5.1 Fine Defect Classification: The cloud server inputs the received image data into a fine analysis model based on Transformer to perform fine classification on the initially identified defect candidate regions. The fine analysis model adopts the Swing Transformer or DeiT architecture. S5.2 Historical Trend Analysis: The cloud server queries the historical testing data of the product based on the product barcode information, performs time-series analysis, and determines whether the defect is sporadic or has a worsening trend. S5.3 Batch Comparison Analysis: The cloud server statistically compares the test results of the current product with the batch test data of similar products to identify whether there are batch quality problems.

9. The detection method according to claim 6, characterized in that, The specific rules for generating the final test report in step S7 are as follows: when the local preliminary analysis result is consistent with the cloud secondary judgment result, the result is directly used as the final test result; when the two results are inconsistent, the result with higher confidence is used; if the confidence of both is lower than the set threshold, it is marked as requiring manual review.

10. The detection method according to claim 6, characterized in that, The data upload in step S4 adopts a tiered upload strategy: for qualified products with a preliminary analysis confidence level ≥90%, only product information and analysis result summary are uploaded; for products with a preliminary analysis confidence level <90% or identified as defective, complete original image data and detailed analysis results are uploaded.

Images