Rotating head assembling, positioning and correcting system based on machine vision

Abstract

The invention relates to the technical field of rotary head assembly, and discloses a rotary head assembly positioning and correction system based on machine vision, which comprises a data acquisition and preprocessing module, a multi-modal feature fusion module, a dynamic-static hybrid positioning module and a detection and adjustment module, through cooperative application of multi-modal feature fusion, dynamic-static hybrid positioning and deviation compensation technologies, precise extraction of multiple types of native features such as geometry, texture and light and shadow of the rotary head is concerned, links such as dynamic trajectory prediction, static fine trimming positioning and closed-loop deviation feedback are also included, the positioning result can more accurately meet the precision requirement of rotary head assembly, and the precision of rotary head assembly is improved. And a more scientific basis is provided for efficient production and quality control of the rotary head.

Description

Technical Field

[0001] This invention relates to the field of spinneret assembly technology, and more specifically discloses a spinneret assembly positioning and correction system based on machine vision. Background Technology

[0002] The machining accuracy and surface quality of a workpiece directly depend on the dynamic performance and positioning accuracy of the rotary head. However, due to long-term exposure to high loads, thermal deformation, and mechanical wear, the rotary head is prone to positioning deviations and geometric errors, leading to a series of quality problems such as distorted machining contours and excessive assembly clearances. Therefore, it is necessary to perform precise positioning detection and error correction on the rotary head.

[0003] While existing technologies have achieved preliminary positioning and dynamic tracking based on machine vision, possessing a certain level of automation and quality assurance capabilities, they largely rely on single visual features and lack multi-feature fusion and adaptive weighting mechanisms. Furthermore, when faced with changes in the rotating head material, surface reflection, or dirt interference, their closed-loop feedback and parameter self-adjustment capabilities are insufficient, which restricts their versatility and accuracy stability under complex working conditions. Summary of the Invention

[0004] The main technical problem solved by this invention is to provide a machine vision-based rotating head assembly positioning and correction system, which can solve the problems raised in the background art.

[0005] To address the aforementioned technical problems, according to one aspect of the present invention, more specifically, a machine vision-based spin head assembly positioning and correction system, comprising: a data acquisition and preprocessing module, a multimodal feature fusion module, a dynamic-static hybrid positioning module, and a detection and adjustment module. The data acquisition and preprocessing module acquires spin head-related data and preprocesses the data, while simultaneously receiving feedback information to adjust the data processing method. The multimodal feature fusion module extracts multiple features of the spin head from the processed data, evaluates the effectiveness of each feature, and dynamically allocates feature weights based on the evaluation results to achieve feature fusion. The dynamic-static hybrid positioning module predicts the position trajectory of the spin head in motion and can perform static positioning correction for spin heads entering a designated area. The detection and adjustment module detects deviations after spin head assembly, generates compensation information based on the deviation data, and adjusts parameters according to the compensation information.

[0006] Furthermore, the data acquisition and preprocessing module includes: a data acquisition module, a data preprocessing module, and an output and feedback module;

[0007] Data acquisition module: includes a high-speed tracking camera, a high-resolution positioning camera, and three light sources: coaxial light, ring light, and diffused light. The high-speed tracking camera acquires dynamic motion images of the rotating head, and the high-resolution positioning camera acquires static high-definition images of the rotating head. During image acquisition, the conveyor speed, station number, and rotating head material information are recorded simultaneously.

[0008] Data preprocessing module: Gaussian filtering, Retinex algorithm, adaptive threshold segmentation, and morphological opening operation are used to perform timestamp alignment and resolution normalization on dynamic and static images;

[0009] Output and feedback module: Outputs preprocessed dynamic trajectory images and static positioning images to the multimodal feature fusion module, while receiving positioning deviation data from the detection and adjustment module and adjusting the preprocessing parameters.

[0010] Furthermore, the multimodal feature fusion module includes: a feature extraction module, a confidence evaluation module, and a dynamic weight fusion module;

[0011] Feature extraction module: Canny edge detection + Hough circle transform is used to extract geometric features from static images, local binary mode is used to extract texture features from static images, gray-level gradient analysis is used to extract light and shadow features from static images, and optical flow method is used to extract motion features from dynamic images.

[0012] Confidence assessment module: Employs a lightweight MobileNetCNN model to output the confidence values ​​of each feature in the range of 0-1;

[0013] Dynamic weight fusion module: Dynamically allocates fusion weights based on confidence values, and generates a comprehensive feature vector of the spiral head through weighted summation.

[0014] Furthermore, the dynamic-static hybrid positioning module includes: a dynamic trajectory prediction module and a static correction positioning module;

[0015] Dynamic trajectory prediction module: Based on dynamic images from a high-speed camera, the trajectory is tracked using optical flow, combined with the conveyor speed of the assembly line encoder, and a long short-term memory network is used to predict the coordinates of the rotating head entering the static positioning area;

[0016] Static correction positioning module: It captures high-definition images with a high-resolution camera, calculates the positional deviation by combining comprehensive feature vectors, and corrects the positioning results.

[0017] Furthermore, the detection and adjustment module includes: an assembly deviation detection module, a deviation compensation module, and a parameter adjustment module;

[0018] Assembly Deviation Detection Module: Uses a laser coaxiality detector to detect coaxiality deviation after assembly, collects assembly gripping position data, and calculates X / Y direction position deviation and axis angle deviation;

[0019] Deviation Compensation Module: A self-supervised deviation compensation model is constructed using an incremental learning algorithm. The comprehensive feature vector output by the multimodal feature fusion module is associated with the deviation data obtained by the assembly deviation detection module to generate feature weight adjustment coefficients and dynamic trajectory prediction compensation coefficients.

[0020] Parameter adjustment module: Sends the compensation coefficients generated by the deviation compensation module to the data acquisition and preprocessing module, the multimodal feature fusion module, and the dynamic-static hybrid positioning module respectively, and adjusts the filtering intensity and light source intensity of the data acquisition and preprocessing module, the feature weight allocation ratio of the multimodal feature fusion module, and the long short-term memory network prediction parameters of the dynamic-static hybrid positioning module.

[0021] Furthermore, in the confidence evaluation module of the multimodal feature fusion module, the input of the MobileNetCNN model is geometric feature edge continuity data, texture feature contrast data, light and shadow feature gradient integrity data, and motion feature trajectory smoothness data.

[0022] Furthermore, the input features of the Long Short-Term Memory Network include optical flow vectors of high-speed tracking camera images and real-time delivery speed values ​​of the pipeline encoder.

[0023] The beneficial effects of the machine vision-based spinner assembly positioning and correction system of this invention are as follows: Through the synergistic application of multimodal feature fusion, dynamic-static hybrid positioning, and deviation compensation technology, it not only focuses on the accurate extraction of various original features such as spinner geometry, texture, and lighting, but also incorporates dynamic trajectory prediction, static fine-tuning positioning, and closed-loop deviation feedback, making the positioning results more accurately meet the assembly precision requirements and providing a more scientific basis for efficient production and quality control of spinners. In addition, the combination of dynamic trajectory prediction and static correction achieves an effective balance between high-speed production lines and high-precision positioning, breaking the limitations of traditional static positioning on production line cycle time or the dependence of dynamic synchronization on hardware costs, thus improving positioning efficiency and equipment economy simultaneously. At the same time, through the adaptive weight allocation of multimodal features and deviation compensation technology, it can flexibly cope with changes in different materials, models, and surface defects of spinners. Furthermore, based on real-time feature confidence and assembly feedback, it can quickly provide parameter self-adjustment suggestions under production line switching or abnormal working conditions, greatly improving the versatility of spinner visual positioning and its ability to cope with complex production situations. Attached Figure Description

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0025] Figure 1 This is a schematic diagram of the system module architecture. Detailed Implementation

[0026] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0027] According to one aspect of the invention, such as Figure 1 As shown, a machine vision-based spinner assembly positioning and correction system is provided, including: a data acquisition and preprocessing module, which acquires spinner-related data and preprocesses the data, while receiving feedback information to adjust the data processing method. This module includes:

[0028] Data acquisition module: includes a high-speed tracking camera, a high-resolution positioning camera, and three light sources: coaxial light, ring light, and diffused light. The high-speed tracking camera acquires dynamic motion images of the rotating head, and the high-resolution positioning camera acquires static high-definition images of the rotating head. During image acquisition, the conveyor speed, station number, and rotating head material information are recorded simultaneously.

[0029] Three types of light sources are paired with the camera for targeted acquisition. Coaxial light is used to highlight the light and shadow gradient on the micro-arc surface of the top of the spinner to assist in attitude calibration. Ring light is used to enhance the outline clarity of the outer edge and the central shaft hole to optimize geometric feature extraction. Diffuse light is used to weaken surface reflection interference to fully preserve the injection molding texture details. The spinner material information recorded simultaneously can be matched with the preset light source-camera parameter combination. For example, for spinners with highly reflective materials, the intensity of the corresponding light source is reduced and the camera exposure parameters are adapted to avoid overexposure or underexposure of the acquired images, providing high-quality raw data for subsequent preprocessing.

[0030] The high-speed tracking camera continuously captures the movement trajectory of the rotating head, and combined with the synchronously recorded conveyor speed of the production line, provides a continuous sequence of motion images for dynamic trajectory prediction. The high-resolution positioning camera focuses on acquiring static high-definition images when the rotating head enters the designated area, ensuring the clear presentation of details such as geometry and texture. In addition, the recording of the workstation number can realize the association between the image and the specific production workstation, which is convenient for subsequent deviation tracking and parameter adjustment.

[0031] In addition, all collected data is stored in a structured format, associated with information such as the material of the rotating head, the speed of the production line, and the station number, so that the subsequent preprocessing module can flexibly call up the data according to different working conditions, and at the same time provide data support for the parameter adjustment of the output and feedback modules.

[0032] Data preprocessing module: Gaussian filtering, Retinex algorithm, adaptive threshold segmentation, and morphological opening operation are used to perform timestamp alignment and resolution normalization on dynamic and static images;

[0033] In the noise suppression stage, a two-dimensional Gaussian filtering algorithm is used to achieve a balance between noise removal and feature preservation. The core formula is as follows:

[0034]

[0035] In the formula For Gaussian kernel weights, The x and y values ​​are the relative coordinates of pixels within the kernel, which are the parameters controlling the filter intensity.

[0036] Meanwhile, during the timestamp alignment process, the conveyor speed and workstation number recorded synchronously during the data acquisition phase are combined to accurately match the time dimension of dynamic motion images and static high-definition images, eliminating image timing deviations caused by the rotation head movement, ensuring that images of the same rotation head in different acquisition phases can be correlated, and resolution normalization avoids feature scale mismatch problems caused by image scale differences during subsequent feature extraction by unifying the image pixel size, thus providing a unified image foundation for multimodal feature fusion.

[0037] Furthermore, to address common defects on the surface of the spinner, a differentiated preprocessing strategy is adopted: the Retinex algorithm is used to suppress surface reflection interference and restore the true color and outline of the spinner; the stain area is separated by adaptive threshold segmentation to avoid the interference of stains on feature extraction; and morphological opening operation is used to remove burr defects and ensure the integrity of edge features. The three algorithms work together to effectively solve the problem that a single preprocessing algorithm cannot suppress multiple defects at the same time.

[0038] During the final preprocessing process, the image processing effect is monitored in real time, such as edge continuity and texture contrast. If a certain indicator fails to meet the standard, the parameters of the corresponding preprocessing algorithm are adjusted.

[0039] Output and feedback module: Outputs preprocessed dynamic trajectory images and static positioning images to the multimodal feature fusion module, and simultaneously receives positioning deviation data from the detection and adjustment module to adjust preprocessing parameters;

[0040] Specifically, during output, in addition to transmitting the preprocessed image data, image quality indicators (such as edge sharpness, texture integrity, and consistency of light and shadow gradients) are also included. This facilitates the multimodal feature fusion module in quickly determining the initial feasibility of feature extraction and reducing invalid feature processing steps. Furthermore, these quality indicators can be linked to the confidence assessment results of the multimodal feature fusion module. The confidence normalization weight allocation formula is as follows:

[0041]

[0042] In the formula, For the first The fusion weights of class features To score the confidence level of the features, The total number of features, based on feature confidence. Determine whether insufficient preprocessing is causing low confidence levels for a certain type of feature, and then make adjustments accordingly;

[0043] Meanwhile, when receiving positioning deviation data from the detection and adjustment module, the system first analyzes whether the source of the deviation is related to preprocessing. For example, if the deviation originates from texture feature distortion, the Gaussian filter value of the corresponding image is reduced to enhance texture detail preservation, or the preprocessing parameters corresponding to the diffuse light intensity are adjusted. If the deviation originates from edge feature blurring, the image preprocessing intensity corresponding to the ring light is increased, or the parameters of the morphological opening operation are optimized. By dynamically adjusting the preprocessing parameters, a closed-loop optimization of "preprocessing-fusion positioning-deviation feedback-parameter adjustment" is formed to ensure that the preprocessing effect always adapts to the subsequent positioning requirements.

[0044] The multimodal feature fusion module extracts multiple features of the spinner from the processed data, evaluates the effectiveness of each feature, and dynamically assigns weights to each feature based on the evaluation results to achieve feature fusion. This module includes:

[0045] Feature extraction module: Canny edge detection + Hough circle transform is used to extract geometric features from static images, local binary mode is used to extract texture features from static images, gray-level gradient analysis is used to extract light and shadow features from static images, and optical flow method is used to extract motion features from dynamic images.

[0046] Specifically, Canny edge detection accurately identifies the contour boundary between the outer edge of the rotating head and the central shaft hole through dual threshold processing. Then, combined with Hough circle transform, the edge pixels are fitted into regular circular features, thereby obtaining core geometric parameters such as the diameter of the rotating head and the coordinates of the center of the shaft hole, providing a rigid benchmark for positioning. The local binary mode generates texture feature encoding by comparing the grayscale difference between the pixels on the rotating head surface and the neighboring pixels, completely preserving the unique texture distribution formed during the injection molding process. When the geometric features are distorted due to burrs or stains, this texture feature can be used as a posture reference.

[0047] Meanwhile, grayscale gradient analysis captures the gradient distribution pattern of the ring-shaped light and shadow by calculating the grayscale change rate of the micro-arc surface at the top of the rotating head under coaxial light illumination. This gradient information is strongly correlated with the perpendicularity of the rotating head axis and can be used to calibrate the attitude deviation during the positioning process. Optical flow method generates motion feature vectors by tracking the motion trajectory of the rotating head pixels in dynamic images. This vector can reflect the speed and direction of the rotating head on the assembly line, providing motion state data support for the dynamic trajectory prediction module.

[0048] Confidence assessment module: Employs a lightweight MobileNetCNN model to output the confidence values ​​of each feature in the range of 0-1;

[0049] Specifically, the model takes the geometric feature edge continuity data, texture feature contrast data, light and shadow feature gradient integrity data, and motion feature trajectory smoothness data output by the feature extraction module as input, and analyzes the quality of each feature through a lightweight convolutional layer: for example, judging whether there are gaps in the geometric feature edges, whether the texture feature contrast meets the discrimination requirements, whether the light and shadow feature gradient is continuous without abrupt changes, and whether the motion feature trajectory has no abnormal fluctuations.

[0050] Then, the confidence values ​​corresponding to each feature are output through the fully connected layer: if a certain feature meets the localization requirements (e.g., continuous geometric feature edges without obvious distortion), the confidence value is close to 1; if the feature is affected by defects (e.g., texture features are blurred due to stains), the confidence value is close to 0. This evaluation result is directly used as the core basis for dynamic weight allocation to ensure that the weight is tilted towards high-quality features.

[0051] Dynamic weight fusion module: dynamically allocates fusion weights based on confidence values, and generates a comprehensive feature vector of the vortex head through weighted summation;

[0052] The fusion weight allocation is based on the confidence assessment results: when the confidence of the geometric feature is the highest, it is assigned a higher weight, making it the main reference for localization; if the confidence of the geometric feature decreases due to the deviation of the ellipticity of the shaft hole, the weight of the texture feature or the lighting feature is increased, and the shortcomings of the single feature are made up for by the complementarity of multiple features.

[0053] During the weighted summation process, the core parameters of each feature (such as the center coordinates of geometric features, the encoding vector of texture features, the gradient value of light and shadow features, and the velocity vector of motion features) are multiplied by their corresponding weights and then summed to generate a comprehensive feature vector containing the position, attitude, and motion state of the spinning head. This vector can be directly input into the static correction positioning module of the dynamic-static hybrid positioning module to provide a comprehensive and reliable feature basis for position deviation calculation.

[0054] Furthermore, the weight allocation process will indirectly link with the data acquisition and preprocessing modules: if a certain type of feature has a persistently low confidence level, it will provide a reverse prompt to the preprocessing module that there may be a parameter adaptation problem (for example, the low confidence level of light and shadow features may be due to improper coaxial light preprocessing intensity), thereby assisting the output and feedback modules in adjusting the preprocessing parameters and indirectly improving the quality of subsequent feature extraction.

[0055] A dynamic-static hybrid positioning module predicts the position trajectory of the rotating head in motion and performs static positioning correction for the rotating head entering a designated area. This module includes:

[0056] Dynamic trajectory prediction module: Based on dynamic images from a high-speed camera, the trajectory is tracked using optical flow, combined with the conveyor speed of the assembly line encoder, and a long short-term memory network is used to predict the coordinates of the rotating head entering the static positioning area;

[0057] Specifically, the optical flow method tracks key pixels (such as edge feature points and texture feature points) on the surface of the rotating head in dynamic images, calculates the motion displacement and direction of the pixels in real time, and generates the initial motion trajectory of the rotating head. At the same time, it combines the conveying speed fed back in real time by the pipeline encoder to calibrate the initial trajectory, eliminate the trajectory deviation caused by the asynchrony between the camera frame rate and the pipeline speed, and ensure that the motion trajectory is consistent with the actual conveying state.

[0058] The Long Short-Term Memory (LSTM) network receives historical motion trajectory data generated by optical flow and encoder speed change data. By learning the motion pattern of the rotating head at different conveying speeds (such as the acceleration phase when the production line starts and the constant speed phase when it is running stably), a trajectory prediction model is established. When the rotating head is about to perform static positioning, the model predicts the coordinates of the rotating head when it reaches the center of the static positioning area based on the real-time collected motion data, providing an accurate pre-positioning reference for subsequent static correction.

[0059] Meanwhile, the prediction process also incorporates motion features output by the multimodal feature fusion module (such as velocity parameters in the motion feature vector). If the motion features indicate that the spinning head has an attitude shift (such as slight tilt), attitude calibration is performed when predicting coordinates to further reduce the deviation range of subsequent static corrections and avoid increasing the difficulty of correction due to attitude issues.

[0060] Static correction positioning module: Captures high-definition images with a high-resolution camera, calculates positional deviations by combining comprehensive feature vectors, and corrects the positioning results;

[0061] Specifically, when the spinner is transported to the vicinity of the predicted coordinates, the high-resolution camera focuses on the spinner area to capture a static high-definition image. After noise reduction and enhancement processing by the data preprocessing module, the image can clearly present the geometric features (such as the center of the shaft hole and the outer circle contour), texture features (such as the surface injection molding texture), and light and shadow features (such as the top ring light and shadow). Subsequently, the system calls the comprehensive feature vector generated by the multimodal feature fusion module to extract the core positioning parameters (such as the theoretical coordinates of the center of the shaft hole and the theoretical angle of the spinner axis) from the vector.

[0062] Simultaneously, the parameters in the comprehensive feature vector are compared with the actual feature parameters extracted from the static high-definition image (such as the actual coordinates of the center of the shaft hole and the actual angle of the axis). The positional deviation of the rotating head in the X and Y directions and the angle deviation of the axis are calculated. Based on the deviation value, a correction command is generated to drive the positioning actuator (such as a pneumatic fine-tuning device) to fine-tune the position and attitude of the rotating head, so that the actual position of the rotating head is consistent with the reference position required for assembly, and finally high-precision positioning is achieved.

[0063] In addition, the statically corrected positioning results will be synchronously fed back to the dynamic trajectory prediction module to optimize the prediction parameters of the LSTM model and reduce the prediction error of the subsequent spin head. At the same time, it will also be transmitted to the detection and adjustment module as a reference for subsequent deviation compensation, forming a collaborative optimization mechanism between modules.

[0064] The detection and adjustment module detects deviations after the rotating head is assembled, generates compensation information based on the deviation data, and adjusts parameters according to the compensation information. This module includes:

[0065] Assembly Deviation Detection Module: Uses a laser coaxiality detector to detect coaxiality deviation after assembly, collects assembly gripping position data, and calculates X / Y direction position deviation and axis angle deviation;

[0066] The laser coaxiality detector precisely aligns with the assembly joint between the rotating head and the main body during testing. It scans the relative position of the axes of the two through a laser beam, capturing the coaxiality deviation value in real time, thus avoiding the problems of low efficiency and large errors caused by manual testing.

[0067] When calculating the X / Y position deviation, the rotating head positioning coordinates (i.e., the target position in the positioning stage) output by the multimodal feature fusion module are combined with the actual grasping position after assembly to clarify whether the deviation is due to positioning error or assembly operation error. When calculating the axis angle deviation, the rotating head axis attitude data extracted in the positioning stage are referenced and compared with the actual axis angle after assembly to accurately locate the link that caused the deviation and provide a basis for subsequent targeted compensation.

[0068] In addition, all detected deviation data will be associated with and stored with the corresponding spinner station number, material information, positioning parameters and other data to form a complete deviation traceability file, which is convenient for subsequent analysis of deviation patterns (such as the types of deviations that are likely to occur with spinners of specific materials) and provides traceable data support for production line quality control.

[0069] Deviation Compensation Module: A self-supervised deviation compensation model is constructed using an incremental learning algorithm. The comprehensive feature vector output by the multimodal feature fusion module is associated with the deviation data obtained by the assembly deviation detection module to generate feature weight adjustment coefficients and dynamic trajectory prediction compensation coefficients.

[0070] Specifically, the incremental learning algorithm uses a comprehensive feature vector (such as the fusion information of spin geometry, texture, lighting and motion features) as input features, and uses the coaxiality deviation, X / Y position deviation and axis angle deviation obtained by the assembly deviation detection module as supervision labels to gradually train a self-supervised deviation compensation model. The model will explore the intrinsic relationship between features and deviations. For example, when poor continuity of geometric feature edges leads to axis angle deviation, it will identify the correlation pattern of "low confidence of geometric features → angle deviation".

[0071] Parameter adjustment module: Sends the compensation coefficients generated by the deviation compensation module to the data acquisition and preprocessing module, the multimodal feature fusion module, and the dynamic-static hybrid localization module respectively, and adjusts the filtering intensity and light source intensity of the data acquisition and preprocessing module, the feature weight allocation ratio of the multimodal feature fusion module, and the long short-term memory network prediction parameters of the dynamic-static hybrid localization module.

[0072] Specifically, when sending compensation coefficients to the data acquisition and preprocessing module, the filtering strength and light source intensity are adjusted accordingly: for example, when the deviation data shows that the texture features have lost details due to excessive filtering, the filtering strength of the corresponding image is reduced; when the reflection interference deviation is caused by the light source being too strong, the intensity of the corresponding light source (such as coaxial light) is reduced to ensure that the preprocessed data is more suitable for subsequent feature extraction.

[0073] When the feature weight adjustment coefficient is sent to the multimodal feature fusion module, it will guide the module to update the fusion ratio of each feature in real time: for example, if the compensation coefficient requires an increase in the weight of texture features, the module will increase the proportion of texture features during the dynamic weight fusion process to make up for the positioning deviation caused by geometric feature distortion.

[0074] When sending dynamic trajectory prediction compensation coefficients to the dynamic-static hybrid positioning module, they are used to optimize the prediction parameters of the long short-term memory network. For example, based on the compensation coefficient of the trajectory deviation in the Y direction, the weight of the pipeline velocity feature in the network is adjusted to reduce the predicted coordinate deviation of the subsequent rotating head entering the static positioning area. Finally, through the coordinated adjustment of the parameters of each module, a closed loop of "deviation feedback - parameter optimization - positioning accuracy improvement" is formed to ensure that the system maintains a high-precision positioning state for a long time.

[0075] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also within the protection scope of the present invention.

Claims

1. A machine vision-based rotating head assembly positioning and correction system, characterized in that, include: The system comprises a data acquisition and preprocessing module, a multimodal feature fusion module, a dynamic-static hybrid positioning module, and a detection and adjustment module. The data acquisition and preprocessing module acquires and preprocesses data related to the spinning head, while simultaneously receiving feedback to adjust the data processing method. The multimodal feature fusion module extracts multiple features of the spinning head from the processed data, evaluates the effectiveness of each feature, and dynamically allocates feature weights based on the evaluation results to achieve feature fusion. The dynamic-static hybrid positioning module predicts the position trajectory of the spinning head in motion and can perform static positioning correction for spinning heads entering designated areas. The detection and adjustment module detects deviations after the spinning head is assembled, generates compensation information based on the deviation data, and adjusts parameters according to the compensation information.

2. The machine vision-based rotating head assembly positioning and correction system according to claim 1, characterized in that: The data acquisition and preprocessing module includes: a data acquisition module, a data preprocessing module, and an output and feedback module; Data acquisition module: includes a high-speed tracking camera, a high-resolution positioning camera, and three light sources: coaxial light, ring light, and diffused light. The high-speed tracking camera acquires dynamic motion images of the rotating head, and the high-resolution positioning camera acquires static high-definition images of the rotating head. During image acquisition, the conveyor speed, station number, and rotating head material information are recorded simultaneously. Data preprocessing module: Gaussian filtering, Retinex algorithm, adaptive threshold segmentation, and morphological opening operation are used to align timestamps and normalize resolution for dynamic and static images; Output and feedback module: Outputs preprocessed dynamic trajectory images and static positioning images to the multimodal feature fusion module, while receiving positioning deviation data from the detection and adjustment module and adjusting the preprocessing parameters.

3. The machine vision-based rotating head assembly positioning and correction system according to claim 1, characterized in that: The multimodal feature fusion module includes: a feature extraction module, a confidence evaluation module, and a dynamic weight fusion module; Feature extraction module: Canny edge detection + Hough circle transform is used to extract geometric features from static images, local binary mode is used to extract texture features from static images, gray-level gradient analysis is used to extract light and shadow features from static images, and optical flow method is used to extract motion features from dynamic images. Confidence assessment module: Employs a lightweight MobileNetCNN model to output the confidence values ​​of each feature in the range of 0-1; Dynamic weight fusion module: Dynamically allocates fusion weights based on confidence values, and generates a comprehensive feature vector of the spiral head through weighted summation.

4. The machine vision-based rotating head assembly positioning and correction system according to claim 1, characterized in that: The dynamic-static hybrid positioning module includes: a dynamic trajectory prediction module and a static correction positioning module; Dynamic trajectory prediction module: Based on dynamic images from a high-speed camera, the trajectory is tracked using optical flow, combined with the conveyor speed of the assembly line encoder, and a long short-term memory network is used to predict the coordinates of the rotating head entering the static positioning area; Static correction positioning module: It captures high-definition images with a high-resolution camera, calculates the positional deviation by combining comprehensive feature vectors, and corrects the positioning results.

5. The machine vision-based rotating head assembly positioning and correction system according to claim 1, characterized in that: The detection and adjustment module includes: an assembly deviation detection module, a deviation compensation module, and a parameter adjustment module; Assembly Deviation Detection Module: Uses a laser coaxiality detector to detect coaxiality deviation after assembly, collects assembly gripping position data, and calculates X / Y direction position deviation and axis angle deviation; Deviation Compensation Module: A self-supervised deviation compensation model is constructed using an incremental learning algorithm. The comprehensive feature vector output by the multimodal feature fusion module is associated with the deviation data obtained by the assembly deviation detection module to generate feature weight adjustment coefficients and dynamic trajectory prediction compensation coefficients. Parameter adjustment module: Sends the compensation coefficients generated by the deviation compensation module to the data acquisition and preprocessing module, the multimodal feature fusion module, and the dynamic-static hybrid localization module respectively, and adjusts the filtering intensity and light source intensity of the data acquisition and preprocessing module, the feature weight allocation ratio of the multimodal feature fusion module, and the long short-term memory network prediction parameters of the dynamic-static hybrid localization module.

6. The machine vision-based rotating head assembly positioning and correction system according to claim 3, characterized in that: In the confidence evaluation module of the multimodal feature fusion module, the input of the MobileNetCNN model is geometric feature edge continuity data, texture feature contrast data, light and shadow feature gradient integrity data, and motion feature trajectory smoothness data.

7. The machine vision-based rotating head assembly positioning and correction system according to claim 4, characterized in that: The input features of the long short-term memory network include the optical flow vector of the high-speed tracking camera image and the real-time delivery speed value of the pipeline encoder.

Images