Search

CN-122016647-A - Connector terminal semi-finished product appearance detection method and device based on binocular machine vision and AI model

CN122016647ACN 122016647 ACN122016647 ACN 122016647ACN-122016647-A

Abstract

The invention belongs to the technical field of detection equipment and AI (advanced technology attachment), and discloses a connector terminal semi-finished product appearance detection method and device based on binocular machine vision and an AI model. The binocular machine vision detection unit comprises vertical and horizontal direction area array camera modules in an orthogonal layout, and synchronously captures top and side images of a product. The jig comprises a replaceable transparent material embedded panel. And a lightweight deep learning model is built in the industrial personal computer to perform primary defect identification, and a high-performance AI analysis program is built in the cloud server to perform secondary judgment. The invention adopts the edge-cloud cooperative architecture, reduces the requirement on the operation resources of the industrial personal computer, improves the analysis speed and the analysis efficiency, realizes the functions of data convergence analysis, model continuous optimization and intelligent early warning, and is widely suitable for intelligent appearance quality detection of the semi-finished product of the high-speed cable connector.

Inventors

  • LI JINGXING
  • YU CHENQI
  • YANG KENAN

Assignees

  • 东莞市三瑞自动化科技有限公司

Dates

Publication Date
20260512
Application Date
20260212

Claims (10)

  1. 1. A connector terminal semi-manufactured product appearance detection device based on binocular machine vision and AI model, characterized by comprising: The rack comprises a desktop base panel, an internal bracket and a shell; The binocular machine vision detection unit is arranged on the rack and comprises a vertical direction area array camera module and a horizontal direction area array camera module, wherein the vertical direction area array camera module and the horizontal direction area array camera module are in orthogonal or oblique layout, optical focusing points of the two groups of camera modules are fixed target positions on the jig, an effective detection area of the vertical direction area array camera module covers the upper end face of the jig, and an effective detection area of the horizontal direction area array camera module covers the front end face of the jig; the fixture is arranged on the base panel and comprises a fixed support made of transparent materials and a detachable embedded panel, wherein an embedded groove matched with the appearance structure of a detected product is formed in the embedded panel and is used for embedding the detected product in the embedded groove and positioning the detected product during detection, so that the upper end face of the detected product is flush with the upper end face of the embedded panel, and the front end face of the detected product is flush with the front end face of the embedded panel; The control and display unit comprises an industrial personal computer, an input keyboard and a display, wherein an AI model is arranged in the industrial personal computer and is used for carrying out preliminary defect identification on the acquired image; The cloud server is in communication connection with the industrial personal computer through a wired or wireless network, and is internally provided with a high-performance AI analysis program and used for receiving detection data and a primary analysis result uploaded by the industrial personal computer and carrying out secondary judgment.
  2. 2. The detection device according to claim 1, wherein the vertical direction area camera module comprises a vertical industrial camera and a vertical light source adjusting module, a lens focusing point of the vertical industrial camera is aligned to the center of the top end face of the embedded object groove, the vertical light source adjusting module is a high-angle annular light source composed of multi-ring multi-angle LED lamp beads, and the irradiation angle is adjustable by 30-60 degrees.
  3. 3. The detection device according to claim 1, wherein the horizontal direction area array camera module comprises a horizontal industrial camera and a horizontal light source adjusting module, wherein a lens focusing point of the horizontal industrial camera is aligned to the center of the lateral end face of the embedded object groove, the horizontal light source adjusting module is a high-angle annular light source composed of multi-ring multi-angle LED lamp beads, and the irradiation angle is adjustable by 30-60 degrees.
  4. 4. The detection device of claim 1, wherein the fixing support further comprises a left support leg and a right support leg which are symmetrically arranged, the bottom end face of each support leg is detachably connected with the base panel through a screw hole, the embedded panel is provided with an elastic cantilever type buckle which vertically and downwards extends, the vertical face of each support leg is provided with a clamping groove matched with the buckle, after the elastic cantilever type buckle of the embedded panel is embedded into the clamping groove of each support leg, the embedded panel is detachably connected with the support leg, the top end face of each support leg is provided with a cylindrical locating pin, and locating holes are correspondingly formed in the embedded panel, and auxiliary guiding and locating are achieved in the process that the embedded panel is detachably connected with the support leg.
  5. 5. The detection device of claim 1, wherein the industrial personal computer is internally provided with an AI model and is a lightweight deep learning model, the industrial personal computer comprises a multi-scale feature extraction module, a rapid classifier, a multi-scale welding spot area focusing module and an adaptive soft tag strategy module, the multi-scale welding spot area focusing module is used for performing multi-scale downsampling, central area enhancement and local texture sharpening on welding spot images, the adaptive soft tag strategy module is used for generating multimodal flexible tags and modeling class transition relations according to welding spot fusion degrees, the industrial personal computer is internally provided with an AI model reasoning acceleration module, a CPU neural network acceleration mode of SIMD instruction vectorization, memory layout rearrangement and operator fusion is adopted, or a OpenVINO + TensorRT mixed acceleration mode is adopted, and the high-performance AI analysis program internally arranged in the cloud server comprises a precise analysis model based on a Transformer and a time sequence data analysis module.
  6. 6. A connector terminal semi-finished product appearance detection method based on binocular machine vision and an AI model, adopting the detection equipment as claimed in any one of claims 1-5, characterized by comprising the following steps: s1, placing a product, namely placing a connector terminal semi-finished product to be detected in an embedded groove of a jig, and sending a trigger signal to an industrial personal computer after a pressure sensor in the embedded groove detects that the product is in place; S2, acquiring local images, namely synchronously controlling the vertical direction area array camera module and the horizontal direction area array camera module to acquire images after the industrial personal computer receives the trigger signal, and respectively acquiring an upper end face image and a front end face image of a product; s3, carrying out local preliminary analysis, namely inputting an embedded lightweight deep learning model to carry out preliminary defect identification after preprocessing an acquired image by the industrial personal computer, and generating a local preliminary analysis result, wherein the preliminary analysis result comprises a defect candidate region, a preliminary classification confidence coefficient and a feature vector; S4, uploading data, namely uploading original image data, a preliminary analysis result and product information to a cloud server through a wired or wireless network by the industrial personal computer; s5, cloud secondary judgment, namely, after the cloud server receives the uploaded data, performing secondary judgment by utilizing a built-in high-performance AI analysis program, wherein the secondary judgment comprises fine defect classification based on a transducer model, historical trend analysis based on time sequence data and batch comparison analysis with similar products; S6, result feedback, namely returning the secondary judgment result and the confidence score to the industrial personal computer by the cloud server; And S7, outputting a result, namely integrating a local primary analysis result and a cloud secondary judgment result by the industrial personal computer, generating a final detection report, binding and storing the detection result and the product bar code information, displaying a detection image, a defect mark and a judgment result on a display, switching the system to an off-line detection mode when the network connection between the industrial personal computer and the cloud server is interrupted, completing detection by the industrial personal computer only depending on a local lightweight deep learning model, temporarily storing data in a local memory, and automatically uploading temporary storage data to the cloud server for supplementary analysis after network recovery.
  7. 7. The detection method according to claim 6, wherein the specific process of the local preliminary analysis in step S3 includes feature extraction of the input image by a multi-scale feature extraction module, and then generating a preliminary judgment result by a fast classifier, and the lightweight deep learning model adopts MobileNetV or ShuffleNet architecture.
  8. 8. The detection method according to claim 6, wherein the specific process of cloud secondary judgment in step S5 includes: S5.1, fine defect classification, namely inputting received image data into a fine analysis model based on a transducer by a cloud server, and carrying out fine classification on a defect candidate region which is primarily identified, wherein the fine analysis model adopts a Swin transducer or DeiT architecture; s5.2, historical trend analysis, namely inquiring historical detection data of the product according to the bar code information of the product by the cloud server, performing time sequence analysis, and judging whether the defect is sporadic or has a worsening trend; And S5.3, carrying out batch comparison analysis, namely carrying out statistical comparison on the detection result of the current product and batch detection data of the similar products by the cloud server, and identifying whether the batch quality problem exists.
  9. 9. The detection method according to claim 6, wherein the specific rule for generating the final detection report in step S7 is to directly use the local primary analysis result as the final detection result when the local primary analysis result is consistent with the cloud secondary judgment result, and to use the result as the final detection result when the local primary analysis result is inconsistent with the cloud secondary judgment result, and to use the result with higher confidence as the result if the local primary analysis result is inconsistent with the cloud secondary judgment result, and to mark that the local primary analysis result is required to be manually checked if the confidence is lower than the set threshold.
  10. 10. The method according to claim 6, wherein the step S4 of uploading the data uses a hierarchical uploading strategy, namely only uploading product information and analysis result summaries for qualified products with the preliminary analysis confidence of more than or equal to 90%, and uploading complete original image data and detailed analysis results for products with the preliminary analysis confidence of less than 90% or identified as defects.

Description

Connector terminal semi-finished product appearance detection method and device based on binocular machine vision and AI model Technical Field The invention belongs to the technical field of detection equipment and AI (advanced technology), and particularly relates to a connector terminal semi-finished product appearance detection method and device based on binocular machine vision and an AI model, which are particularly suitable for multi-dimensional appearance quality detection and intelligent quality management of a high-speed cable connector semi-finished product. Background With the rapid development of artificial intelligence technology, the application of low-voltage high-speed high-current cable connectors in the manufacture of equipment such as AI servers, intelligent computation centers, high-performance computers and the like is becoming wider and wider. For example, 224Gbps cable connectors for high-speed transmission are widely used for high-speed signal transmission between chips and boards, such as interconnection systems of frames of the english da GB300 NVLink. The plug-in cable terminal semi-finished product for manufacturing the connector generally has a relatively complex structure and a plurality of quality detection items, and has various models and specifications. Taking 224Gbps high-speed cable connector as an example, a semi-finished product adopts a double-shaft cable design, and comprises an insulating cable (containing a wire harness with a metal core), a metal semi-package protective shell (containing a positioning hole), an injection molding buckle, a parallel terminal pin and a cable shielding layer. The parallel pins are connected with the core wire bundle through laser welding (soldering points), and the semi-finished product is in the form of finished terminal crimping, preliminary encapsulation, but not assembled into a complete connector assembly. When quality detection is carried out on the semi-finished products, more detection items are involved, and the quality detection method mainly comprises a welding spot connection state, whether parallel terminal pins are parallel, whether a metal semi-package protection shell is deformed, whether injection molding buckle positioning is accurate, whether parallel terminal pins and a core wire bundle contact a cable shielding layer and the like. The semi-finished product types generally comprise products with welding spots, non-welding semi-finished products, linear semi-finished products, bending semi-finished products and the like. At present, the existing detection equipment mainly has the following problems: (1) The detection dimension is single, namely the existing equipment can only collect images with a single visual angle, and can not synchronously acquire information of a plurality of end faces of a product, so that the detection is incomplete. (2) The fixture has poor suitability, the fixture structure of the existing equipment is fixed, only products with one specification can be detected, and the fixture can not adapt to the detection requirements of connector terminal semi-finished products with various types and structures. (3) AI models are difficult to deploy, high-performance deep learning models generally require expensive hardware resources such as GPUs, while industrial field devices are often limited by cost and space, complex models are difficult to deploy, and detection accuracy is limited. (4) The existing equipment is mostly operated in a single machine, detection data are stored in a scattered mode, cross-equipment and cross-batch statistical analysis cannot be performed, and the quality problem of the system is difficult to find. (5) The model updating is difficult, AI model updating of single machine equipment needs to be operated one by one, the maintenance cost is high, and continuous optimization of the model is difficult to realize. For example, CN107478164a discloses a machine vision-based connector terminal detection device, but the machine vision detection system can only collect images of the upper end face of the connector terminal in the vertical direction, cannot synchronously collect images of the front end face of the terminal, and the fixture is fixed in structure, and can only detect one product, and the whole structure is complex, the detection end face is few, the speed is slow, the efficiency is low, and the adaptability is poor. The machine vision method for detecting the quality of the miniature wiring terminal disclosed by CN107895362A adopts a traditional image processing algorithm (such as Canny operator and Blob analysis) to detect the quality, does not adopt a deep learning technology, has limited identification capability on complex defects, can only detect images with a single visual angle, and cannot realize multi-dimensional synchronous detection. Disclosure of Invention The invention aims to provide a connector terminal semi-finished product appearance detection method and