An online detection system for cold-rolled suspension spring forming accuracy based on image recognition

Abstract

The present invention relates to the field of image data processing technology, and specifically to an online detection system for cold-rolled forming accuracy of suspension springs based on image recognition, comprising: an image acquisition and preprocessing module, a spring morphology characterization module, a spring image reconstruction module, a stereo image alignment module, and a dynamic monitoring module. The image acquisition and preprocessing module dynamically acquires multi-angle images of the spring and determines the target area; the spring morphology characterization module performs morphology analysis, establishes a first morphology map, and extracts two-dimensional features; the spring image reconstruction module uses Faster R-CNN to generate a two-dimensional mask, reconstructs a three-dimensional point cloud through multi-angle images, and extracts key nodes; the stereo image alignment module aligns the three-dimensional data with the standard model, calculates the accuracy deviation, and makes a qualified judgment. The dynamic monitoring module can be executed periodically, and a stop signal is triggered when the accuracy meets the standard to realize closed-loop control. In addition, the system adopts a dynamic deviation threshold and hierarchical processing according to the spring type to improve detection efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to an online detection system for cold-rolling forming accuracy of a suspension spring based on image recognition. Background Art

[0002] In the manufacturing process of suspension springs, the cold coil forming process is widely used due to its advantages such as high efficiency and low cost. However, the forming accuracy of the spring directly affects its mechanical properties and service life, so high-precision inspection of the formed spring is crucial. Traditional inspection systems mainly rely on manual measurement or contact measurement equipment, which have problems such as low efficiency, high subjectivity, and easy damage to the product surface. In recent years, with the development of machine vision technology, non-contact inspection systems based on image processing have gradually become a research hotspot, but they still face many technical bottlenecks in practical application.

[0003] Existing machine vision technologies primarily rely on traditional edge detection and feature matching systems, which suffer from significant environmental adaptability deficiencies in practical applications. Dynamic shooting scenarios exhibit significant instability when lighting conditions change: low-contrast scenes can cause edge features to break, while overexposure can lead to loss of valid features. Furthermore, interference from complex light sources such as highlights and reflections can easily generate false edge features, significantly increasing the mismatch rate. Background changes caused by camera movement can cause feature point drift, resulting in tracking failures. Furthermore, non-rigid deformations can undermine the effectiveness of feature descriptors, and scale changes can render fixed-threshold edge detection systems ineffective.

[0004] Therefore, an online detection system for cold-rolling forming accuracy of suspension springs based on image recognition is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide an online detection system for the cold-rolled forming accuracy of suspension springs based on image recognition. Through the integration of multiple technologies and algorithm innovation, it aims to solve the pain point of difficult dynamic posture capture in the traditional cold-rolled forming accuracy detection of suspension springs. The present invention proposes an online detection system for the cold-rolled forming accuracy of suspension springs based on image recognition. An adaptive image enhancement module is used to handle noise interference in complex industrial scenes, and a cascaded convolutional neural network is used to achieve accurate segmentation and reconstruction of spring contour features. A three-dimensional space coordinate mapping model is established with the help of a time series posture estimation algorithm. The detection parameters of springs of different models are autonomously matched through a dynamic weight distribution mechanism, forming a non-contact detection closed loop that integrates optical imaging, intelligent analysis, and real-time feedback. It effectively solves the technical difficulties of online monitoring of suspension spring forming quality and significantly improves detection accuracy and efficiency.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An online detection system for cold-rolling forming accuracy of a suspension spring based on image recognition includes the following steps:

[0008] Dynamically capture multi-angle images of the suspension spring during the cold rolling process to obtain the spring target area;

[0009] Preferably, median filtering and image enhancement processing are performed on the image data within the target area of ​​the spring; gradient calculation, non-maximum suppression, double threshold screening and edge connection technology are used to detect and generate the initial contour of the suspension spring from the pre-processed image data;

[0010] Based on the target area of ​​the spring, multi-scale features are extracted, and contour correction and matrix decomposition processing are performed to establish a first topography image representing the two-dimensional topography of the suspension spring; and the two-dimensional structural features of the spring are further extracted;

[0011] Preferably, the positions of two-dimensional key points of the spring target area image are detected by using the Gaussian difference in scale space, and the directions and feature descriptors of the two-dimensional key points are calculated; feature point extraction or matching processing is performed on the initial contour to obtain a plurality of two-dimensional feature matching points;

[0012] The initial contour is corrected, fitted and skeletonized in combination with the obtained / feature matching points to form a final first topography image, and the two-dimensional structural features are extracted therefrom.

[0013] Preferably, the step of detecting the suspension spring using a two-stage object detector Faster R-CNN specifically includes:

[0014] performing pixel-level segmentation of the suspension spring in the initial formed image information using the two-stage object detector to generate an accurate two-dimensional mask;

[0015] The precise two-dimensional mask is used to guide the spring stereoscopic reconstruction algorithm to generate the initial three-dimensional point cloud;

[0016] Preferably, the spring three-dimensional morphology reconstruction algorithm specifically includes the following steps:

[0017] Based on the accurate two-dimensional mask generated by the two-stage object detector, feature points of the suspension spring in the multi-angle images are matched to establish correspondence between the images;

[0018] Sparse point cloud reconstruction: calculate the camera pose through the motion recovery SfM algorithm, and generate the initial sparse 3D point cloud by combining the matching feature points;

[0019] Dense point cloud reconstruction uses the multi-view stereo vision MVS algorithm to optimize sparse point clouds and generate high-precision real-time 3D point clouds; point cloud post-processing performs outlier filtering, smoothing, and normal vector estimation on the real-time 3D point clouds.

[0020] Preferably, the spring target area located in the multi-angle real-time image is used to generate a real-time three-dimensional point cloud of the suspension spring using a spring stereoscopic morphology reconstruction algorithm; and based on the real-time three-dimensional point cloud, key nodes representing its three-dimensional geometric shape are extracted;

[0021] Preferably, the real-time three-dimensional point cloud or its key nodes are imported into a virtual space coordinate system and aligned with a preset standard CAD model of the suspension spring through rigid transformation; based on the aligned three-dimensional point cloud or key nodes and the standard CAD model of the suspension spring, the cold-rolled forming precision deviation of the suspension spring is calculated and qualified, where the precision deviation includes at least one of length deviation, diameter deviation, number of turns deviation, turn spacing deviation, helical curvature deviation, and verticality deviation;

[0022] Preferably, the steps are continuously and periodically performed to obtain time series accuracy deviation data of the suspension spring during the cold rolling process. The online detection system proposed in the present invention can automatically call the preset three-dimensional reconstruction parameters, key node extraction strategy or accuracy deviation calculation model for the specific type to perform accuracy detection based on the classified spring type or model.

[0023] Preferably, the specific steps of deviation threshold analysis and qualification determination include:

[0024] The deviation threshold is not a fixed value, but can be dynamically adjusted or selected based on at least one of the following:

[0025] Pre-set tolerance standards corresponding to the specific type or model of suspension spring identified by the two-stage object detector or subsequent classifier;

[0026] Dynamic thresholds generated based on historical test data and statistical process control (SPC) principles to adapt to normal fluctuations in the production process or detect process drift;

[0027] The combined threshold is set by considering the mutual influence and correlation between multiple precision deviation parameters. Once the accuracy is determined to be qualified, a stop signal is triggered to control the cold roll forming equipment to terminate the forming process;

[0028] Preferably, an online detection system for cold-rolled suspension spring forming accuracy based on image recognition comprises:

[0029] Image acquisition and target positioning module: Responsible for dynamically capturing multi-angle spring images and quickly locking the spring in the spring target area in the image, providing raw visual data and precise input range for subsequent analysis.

[0030] Spring morphology characterization module: Extracts the multi-scale features of the spring based on the target area image of the spring and establishes the first morphology map. At the same time, it extracts its two-dimensional structural parameters for preliminary evaluation or auxiliary three-dimensional analysis.

[0031] Spring image reconstruction module: Use the spring target area in the multi-angle image to generate real-time 3D point cloud data of the spring, and extract key geometric nodes from it to realize the conversion from 2D to 3D information.

[0032] Stereo Image Alignment Module: Based on the spring contour features extracted by the two-stage object detector, the module matches the preset spring model database to determine the type of the current suspension spring; based on the identified spring type, the module automatically associates the preset detection parameters corresponding to the model, including the 3D reconstruction accuracy threshold, key node definition rules, and tolerance standards; and calculates the cold-rolled forming accuracy deviation of the suspension spring, where the accuracy deviation includes at least one of the following: length deviation, diameter deviation, number of turns deviation, turn spacing deviation, helix curvature deviation, and perpendicularity deviation.

[0033] The dynamic monitoring module continuously and periodically executes the inspection process to obtain real-time accuracy deviation data during the forming process. It also receives real-time feedback from the acceptance judgment and result output module. It then optimizes the adaptability of inspection parameters using statistical process control (SPC) algorithms. It is configured to compare the calculated accuracy deviation with preset tolerances. If both accuracy deviations are within their corresponding tolerances, the suspension spring is deemed to have passed the current forming accuracy test; otherwise, it is deemed unqualified. After the determination is complete, the test results are output. Once the spring is deemed to have passed the accuracy test, a signal is triggered to control the cold coil forming equipment to terminate the forming process.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. Median filtering and image enhancement are performed on the collected multi-angle images to effectively suppress noise and improve image clarity and contrast, thereby enhancing the recognizability of the spring outline and details. Furthermore, multi-scale feature extraction is used to comprehensively capture the spring's structural information at different scales, including overall shape, local curvature, and subtle surface features. This hierarchical feature fusion strategy enhances the system's sensitivity to spring defects at different scales. Through feature complementation, it effectively avoids false and missed detections at a single scale, improving the reliability of suspension spring precision testing.

[0036] 2. Accurate positioning and 3D reconstruction optimization are achieved with the help of Faster R-CNN. The invention emphasizes the use of a two-stage target detector, Faster R-CNN, to detect suspension springs, and further uses it for pixel-level segmentation to generate accurate two-dimensional masks to further guide the spring's three-dimensional morphology reconstruction algorithm. This means that the reconstruction process can better focus on the spring itself, reduce errors caused by changes in the spring's posture, thereby generating more accurate key nodes, effectively eliminating the spring's translation and rotation in space, and more accurately assessing its true morphology deviation, providing a highly reliable data foundation for subsequent precision calculations, further improving the system's detection accuracy and efficiency.

[0037] 3. Real-time dynamic online detection and closed-loop control are achieved. The system can continuously and periodically execute the detection process to obtain the time series accuracy deviation data of the suspension spring during the cold roll forming process. This online monitoring method can promptly detect anomalies in the forming process and avoid the production of defective products. More importantly, the system is equipped with a cold roll forming dynamic monitoring and stop control module. When it is detected that the spring accuracy meets the standard, the system can trigger a stop signal to control the cold roll forming equipment to terminate the forming. At the same time, the system can automatically call the preset three-dimensional reconstruction parameters, key node extraction strategy or accuracy deviation calculation model for this specific type based on the classification results. This shows that the detection system has the ability to support the accuracy measurement of various types of springs, has higher flexibility and applicability, and can meet different production needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of an online detection system for cold-rolled suspension springs based on image recognition, proposed in an embodiment of the present invention;

[0039] Figure 2 A diagram of the system anomaly detection adjustment steps for the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of the detection system proposed in the embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] The present invention provides an online detection system for cold-rolled suspension springs based on image recognition and designs a set of online detection systems for cold-rolled suspension springs. As an embodiment of the present invention, referring to Figure 1 , which is a flow chart of an online detection system for cold-rolling forming accuracy of a suspension spring based on image recognition proposed in an embodiment of the present invention.

[0043] As a key functional component in the automotive suspension system, the molding quality of the spring directly affects the safety and comfort of the entire vehicle. Traditional detection systems suffer from low detection accuracy, long detection time, and the inability to reflect the actual deformation of the spring after being subjected to force online, making it difficult to meet the high standards required for modern automotive intelligent manufacturing. To address these shortcomings, the present invention proposes an online detection system for spring cold-rolling forming accuracy based on high-resolution visual imaging and deep learning. This system combines multi-scale feature extraction and spring three-dimensional morphology reconstruction algorithms to achieve high-precision online real-time measurement of suspension springs, greatly improving detection accuracy and efficiency. At the same time, it has strong adaptability and can meet the online measurement needs of springs of different specifications.

[0044] An automotive parts manufacturer has successfully deployed and operated an online cold coil forming accuracy detection system based on image recognition on its suspension spring production line. To illustrate the system and its effectiveness, the present invention will be described in detail with reference to the accompanying drawings and the following two examples.

[0045] Example 1

[0046] A production plant has applied an online detection system for cold-rolled suspension springs based on image recognition on its production line. Figure 1 Referring to the flowchart, the specific implementation steps are as follows:

[0047] To capture images of the spring in its initial state, three industrial cameras with a resolution of 1920x1080 pixels were used to simultaneously capture the initial formation of the suspension spring from 0°, 60°, and 120°. The spring being tested is numbered "SPG-20230728-001."

[0048] Furthermore, grayscale processing is performed to enhance the contrast:

[0049] Specifically, for high-grayscale areas (the main body of the spring), a grayscale threshold is set and linear stretching is used to enhance contrast. For pixels exceeding the threshold, the grayscale value is proportionally increased, capped at its maximum value. For low-grayscale areas (the background), local mean replacement is used to smooth noise. A median filter is also applied to remove salt-and-pepper noise and optimize the clarity of the spring's outline.

[0050] During the image preprocessing stage, the input color image is first converted to grayscale to generate a single-channel grayscale image. Furthermore, histogram stretching is performed on high-grayscale regions to enhance subject contrast, while local mean integration is performed on low-grayscale regions to suppress background interference. This segmented processing effectively reduces noise while enhancing the saliency of spring features.

[0051] A 3x3 median filter is applied to remove noise. For each pixel, the median grayscale value of the pixels in the 3x3 neighborhood is taken to eliminate noise and make the spring contour smoother.

[0052] Furthermore, the initial contour of the suspension spring is detected and generated from the preprocessed image data using gradient calculation, non-maximum suppression, double threshold screening and edge connection techniques;

[0053] The Canny algorithm was applied to the identified spring region image for edge detection, convolving the image with the convolution kernel. The gradient magnitude and direction in the x and y directions were calculated, and non-maximum suppression was used to refine the edges, ensuring each edge was only one pixel wide. A double thresholding algorithm was then applied: a high threshold of 100 and a low threshold of 50 were used for filtering and edge connection, resulting in three clear images of the initial spring contours.

[0054] Furthermore, the positions of two-dimensional key points of the spring target area image are detected by using the Gaussian difference in scale space, and the directions and feature descriptors of the two-dimensional key points are calculated; feature point extraction or matching processing is performed on the initial contour to obtain a plurality of two-dimensional feature matching points;

[0055] Apply Gaussian difference detection to find the local extreme points of Gaussian difference response at different scales and spatial positions to determine the location of key points. The calculation method is as follows:

[0056] ;

[0057] in, are the pixel coordinates of the image, is the scale parameter indicating the degree of Gaussian blur, is a scale factor that defines the proportional relationship between the blur levels of adjacent scales. The setting starts from 1.6, and the scale factor k is set to , detect 3 octaves, each octave contains 4 scale layers.

[0058] Furthermore, for each detected key point, the gradient direction distribution of its surrounding neighborhood is calculated, the main direction of the key point is determined and a feature descriptor with rotation invariance is generated. The corner detection algorithm is used to extract feature points on the initial contour of the spring, and the initial contour is feature matched with the preset spring two-dimensional model contour to obtain multiple corresponding two-dimensional feature matching points.

[0059] Specifically, we assume that an average of 150 to 200 key points are detected per image. We then calculate the key point descriptors and match them with the 180 feature points in the pre-stored 2D spring standard model outline. Ultimately, we successfully match an average of 120 to 160 feature point pairs per image.

[0060] With the feature matching points as strong constraints, the initial spring contour is corrected, fitted, and skeletonized to generate an accurate two-dimensional shape representation - the "first topography image".

[0061] By dynamically capturing suspension springs from multiple angles and accurately identifying target spring areas, combined with median filtering and image enhancement processing, the system can comprehensively capture spring status information, significantly improve image quality, suppress noise interference, and provide a high-quality geometric basis for subsequent contour extraction.

[0062] Extract key 2D structural parameters from the first topography image to assist in initial estimation and verification of 3D reconstruction and quantify the geometric properties of the spring, including:

[0063] (1) Length: Calculate the length of the skeleton, or the distance between key points on the contour.

[0064] (2) Diameter: Analyze the local width of the profile, or measure the length of the section perpendicular to the skeleton.

[0065] (3) Spring pitch: Analyze the number of bending times of the skeleton or the number of periodic fluctuations of the profile in the axial direction to estimate the effective number of turns of the spring in the two-dimensional projection.

[0066] (4) Number of turns: Analyze the number of bends of the skeleton or the number of periodic fluctuations of the profile in the axial direction to estimate the effective number of turns of the spring in the two-dimensional projection.

[0067] (5) Perpendicularity: The maximum allowable angle or linear offset between the center axis and the theoretical installation reference plane, which is used to measure the axial alignment accuracy of the spring after assembly.

[0068] (6) Curvature: Calculate the local curvature of each point on the contour or skeleton.

[0069] By combining gradient calculation and non-maximum suppression techniques to generate an initial contour, the stability of suspension spring edge detection is improved. Feature point matching and contour correction create a more accurate secondary topography map, laying the foundation for 2D structural feature extraction and improving detection reliability.

[0070] The detection branch of Faster R-CNN follows the RPN and performs more precise classification and bounding box regression for each candidate region. The Mask R-CNN branch is integrated for pixel-level segmentation. This adds a small fully convolutional network to Faster R-CNN to predict the segmentation mask. This is trained and optimized using a cross-entropy loss function. This generates a precise two-dimensional mask, where the pixel value indicates whether the corresponding image pixel belongs to the suspension spring.

[0071] Through pixel-level segmentation prediction, a mask is generated to accurately locate the 2D area of ​​the suspension spring, avoiding false detections. The 2D mask also matches the original image size, ensuring accurate spatial alignment during 3D reconstruction and improving model consistency.

[0072] The key steps of SfM include feature matching, by matching feature points in different images;

[0073] Extremal geometric constraints use the fundamental matrix F or the essential matrix E to constrain matching points: or ;

[0074] in and are the matching image points. The fundamental matrix F contains the relative rotation and translation information of the two cameras. In normalized image coordinates, the essential matrix E contains the relative rotation and translation information.

[0075] The MVS algorithm is applied to multiple images with known camera poses. Through pixel-level matching and depth estimation, corresponding pixels are found in images of different perspectives, and their three-dimensional coordinates are calculated to generate a dense point cloud.

[0076] The MVS algorithm uses the sum of the square differences of corresponding pixels between two image blocks as the breadth consistency cost function, and a smaller value indicates a higher similarity.

[0077] Multi-view feature matching and MVS algorithms generate high-precision sparse point clouds, providing a reliable data foundation for 3D reconstruction; outlier filtering and normal vector estimation are used to optimize point cloud quality, enhance the ability to restore details of the 3D model, and ensure the credibility of precision detection.

[0078] Furthermore, a pixel backprojection step is performed to convert the pixels in the 2D image back to the 3D camera coordinate system. For each reference image, the depth value of each pixel is estimated. For a pixel in the reference image, assuming a depth, the pixel can be backprojected into 3D space to obtain a 3D point P. Then, P is projected into other images with known camera poses to obtain the corresponding pixel. The photometric consistency cost between the reference image patch and the target image patch is calculated. By searching within the depth range, the depth value that minimizes the cost function is found.

[0079] The algorithm backprojects pixels into points in three-dimensional space, projecting the three-dimensional point P onto pixels on the image plane of another camera. For each pixel in a reference image, the algorithm searches within a certain depth range. For each hypothesized depth, the pixel is backprojected into three-dimensional space and reprojected onto the other known camera views. The photometric consistency cost is then calculated between the image patch surrounding the pixel in the reference image and the image patch projected onto the corresponding area in the other view. The depth value that minimizes this cost, that is, the one with the highest similarity, is selected as the estimated depth for that pixel.

[0080] Once the depth value of each pixel is estimated, the camera's intrinsic and extrinsic parameters can be used to back-project each pixel into three-dimensional space to obtain the three-dimensional point coordinates corresponding to the pixel.

[0081] Finally, by rotating and translating the extrinsic parameters of the reference camera, the three-dimensional points in the camera coordinate system can be converted to coordinates in the world coordinate system; depth estimation and backprojection are performed on the pixels of all reference images, and coordinate transformation is performed to obtain a dense three-dimensional point cloud of the scene.

[0082] The first topography image is verified and corrected through projection, reducing the cumulative error between multi-view images. The 3D point cloud is then rigidly transformed in the virtual space coordinate system to align with the standard model, directly quantifying the cold roll forming deviation, simplifying the determination process and improving efficiency.

[0083] Furthermore, a spring stereoscopic reconstruction algorithm is used to generate a real-time 3D point cloud of the spring based on multi-angle images and an accurate 2D mask. The Faster R-CNN two-stage object detector is again applied to extract key nodes from the 3D point cloud that characterize the spring's 3D geometry. These key nodes represent the spring's primary geometric features and are used in subsequent error calculations.

[0084] Furthermore, an iterative closest point algorithm is used to achieve precise alignment between the 3D point cloud and the standard model. The algorithm iteratively finds the closest point pairs between the two point clouds and calculates the rotation and translation transformations that minimize the distance between these point pairs. The objective function is set to minimize the sum of the squared distances between the point pairs.

[0085] It also automatically identifies spring type or model, loads corresponding test parameters, standard models, and tolerances, and assists in dynamic threshold adjustment. It continuously and periodically performs tests, acquiring real-time accuracy deviation data and triggering a stop signal to control the forming equipment when accuracy meets the standard. The calculated accuracy deviation is compared with the standard tolerance for acceptance, and the test results are summarized and output in a report.

[0086] Specifically, the spring accuracy deviation calculation includes:

[0087] Length deviation: It can calculate the difference between the distance between the actual spring key points and the distance between the corresponding points of the standard model;

[0088] Diameter deviation: Fit the helix of the actual spring and calculate the difference between its average diameter and the diameter of the standard model;

[0089] Coil Deviation: Analyze the structural features of the 3D point cloud to identify the actual number of spring coils and compare it with the number of coils of the standard model;

[0090] Coil spacing deviation: Calculate the difference between the average axial distance between adjacent coils of the actual spring and the corresponding distance of the standard model;

[0091] Perpendicularity deviation: Analyze the angle between the spring axis and the reference plane, and calculate its deviation from the vertical angle of the standard model;

[0092] Helix Curvature Deviation: Fits the helix of the actual spring and calculates the difference in curvature from the helix of the standard model.

[0093] Table 1: 3D accuracy calculation and deviation analysis

[0094]

[0095] Furthermore, in order to establish a unified error measurement standard, the deviation threshold between the two topography images should be calculated. Specifically, the topography deviation is measured by the corresponding distance between the two groups of points, and the deviation index is measured by the root mean square error.

[0096] Furthermore, the shape comparison and deviation calculation in the virtual space are carried out, and the deviation threshold standard is set by the maximum deviation. If the threshold standard is exceeded, it is judged that the molding accuracy is unqualified, otherwise it is qualified.

[0097] Specifically, the aligned measurement point cloud and the standard model point cloud are used to calculate the corresponding point distance and the shape deviation is: ; Calculate the maximum deviation as ;

[0098] Accurate calculation of spring curvature is the key to accurately assess internal stress and predict fatigue life, ensuring the safety and reliability of spring design. If the geometric mutation points of the spring can be identified based on this, the weak links where stress is most concentrated can be located, guiding the optimization of the design and preventing premature failure. The curvature detection abnormality condition is set as when the curvature exceeds the standard by 5% and the curvature gradient of the adjacent points is greater than 5%. When the curvature deviation exceeds the process tolerance, it is marked as a mutation point, and when the curvature deviation exceeds the process tolerance, it is marked as a molding defect area and fed back to the dynamic monitoring module for early warning and process parameter adjustment.

[0099] Table 2 Morphology comparison and qualification judgment

[0100]

[0101] Based on this, the spring "SPG-20230728-001" is judged to be qualified. The entire process from image acquisition to result output takes about 3.5 seconds. The online system detects and outputs results in real time. The system collects the spring status in real time and automatically executes the above steps; the test results are fed back immediately, including graphic deviation display and qualified judgment information; data reports are provided to production line managers to support traceability and quality control. When the system continues to be abnormal, the parameters will be adjusted and imported, and reference Figure 2 This is a diagram of the steps for adjusting abnormality detection in the system of the present invention;

[0102] Results screen: Displays the 3D reconstructed spring model and a heat map showing deviations from the standard model. Lists the measured values, standard values, and deviations for each geometric parameter. Clearly indicates a "pass" or "fail" decision. Failures are highlighted.

[0103] Data Reporting: Automatically generates a test report containing the above information and stores it in a database for quality traceability. This system can adapt to the production needs of different spring types, enabling flexible testing through parameter adaptation and expanding the system's application scenarios.

[0104] Furthermore, excessive cold-rolling of suspension springs can lead to serious consequences such as dimensional deviations, abnormal stiffness, and a sudden decrease in fatigue life, potentially even causing safety accidents. Therefore, the system incorporates full-process quality monitoring to ensure forming accuracy. This continuous monitoring of spring deviation time series data prevents over-intervention and enables timely correction of deviations. Adjustment results are fed back to the equipment control system, forming a closed "test-verify-termination" loop to continuously optimize the production process.

[0105] Table 3 "SPG-DYNAMIC-001" molding process time series accuracy data record

[0106]

[0107] Note: At T0+2.3s, although individual parameters may be on the verge of passing, the system may be designed to require all key parameters to stably meet the passing criteria for two consecutive inspection cycles, or for the comprehensive topography deviation to meet the standard, before a final pass is determined. This is to avoid misjudgments caused by critical fluctuations. Here, the system simply states that a pass is determined when all parameters pass.

[0108] Specifically, when the inspection system determines at time T0+2.5s that all key geometric parameters of spring "SPG-DYNAMIC-001" are within the preset deviation thresholds, i.e., if the comprehensive judgment is qualified, the system immediately performs the following operations:

[0109] Generate qualified judgment signal: the internal status of the detection system is updated to "molding qualified".

[0110] Send a stop command: Send a stop signal to the PLC of the cold roll forming equipment through the preset communication interface.

[0111] Equipment response: After receiving the stop signal, the PLC of the cold coil forming equipment immediately executes its preset program, safely stops the material feeding and winding action, and completes the current spring forming process. Finally, the results are output and data is recorded:

[0112] The real-time interface displays the final qualified spring's 3D model, deviation heat map, parameter list, and qualification judgment.

[0113] Data Reporting: Automatically generate an inspection report containing time series data and final qualified status and store it in the database. This report not only includes the accuracy of the final product but also records its production process, providing richer data for process analysis and optimization.

[0114] Reduced manual intervention: The fully automated adjustment process reduces dependence on operators, improving production efficiency while reducing the risk of human error.

[0115] By monitoring the rate of change of multi-dimensional parameters in real time and dynamically adjusting system thresholds, we effectively address process fluctuations during production and avoid misjudgments caused by fixed thresholds. Based on the device status, we automatically adjust the sampling density and feature extraction parameters of the spring 3D topography reconstruction algorithm, significantly reducing vibration interference and ensuring the accuracy of point cloud data. A real-time calibration mechanism verifies the effectiveness of adjustments, preventing new errors introduced by parameter changes and ensuring the consistency and reliability of test results.

[0116] This online cold-rolled suspension spring precision inspection system uses an intelligent inspection process to precisely control key quality parameters. It combines multiple indicators, including spring length deviation, elastic modulus, and surface defect score, and dynamically adjusts the segmented thresholds for each indicator in real time to form a threshold combination, streamlining the judgment process. The weights are set as follows: length 1.2, diameter 0.8, pitch 0.5, perpendicularity 1.0, and curvature 0.7. The overall score is calculated as: 0.35 × 1.2 + 0.12 × 0.8 + 0.09 × 0.5 + 0.25 × 1.0 + 2 × 0.7 = 2.231 points. The dynamic acceptance threshold is set at 2.5, and the suspension spring is deemed qualified. A veto on key parameters can also be set. If the curvature deviation rises to 4%, the acceptance threshold may be tightened to 2.0, resulting in a failure.

[0117] Example 2

[0118] A certain automobile brand has a specially designed variable diameter and pitch suspension spring (smaller coil diameters at both ends, larger coil diameters in the middle, and uneven pitch to achieve specific nonlinear stiffness characteristics). This spring requires extremely high molding precision. The production line deployed the online inspection system described in this invention.

[0119] The differences from the first embodiment are: potential improvement in the number of images and resolution, increase in the number of feature points, improvement in the density of the three-dimensional point cloud, and more complex geometric analysis algorithms.

[0120] Specifically, to address the more complex contours and potential self-occlusion of variable diameter and variable pitch springs, a more diverse and comprehensive camera angle combination (0°, 45°, 90°, 135°, 180°, and 225°) was employed. This wraparound camera layout, with the potential for enhanced focus on specific complex areas, ensures that all key geometric transitions on the spring, particularly those with varying diameters and pitches, are clearly and unobstructedly captured in at least a few images. This provides a more complete and redundant data source for subsequent high-precision 3D reconstruction. This contrasts with the relatively uniform viewing angle requirements of conventional springs.

[0121] Furthermore, preprocessing parameters were fine-tuned to account for the potentially more dramatic light and shadow variations in complex springs, or stray highlights caused by complex surfaces. In this embodiment, the grayscale threshold for contrast enhancement was adjusted to 170, and the stretch factor α was increased to 1.6 to better distinguish the spring from the background. The local mean replacement neighborhood was expanded to 7x7 to more effectively smooth background noise, especially when dealing with complex backgrounds or slight oil reflections. The median filter window was also increased to 5x5 to more effectively remove salt and pepper noise and small highlights caused by complex surfaces, ensuring contour continuity and smoothness. These adjustments reflect greater adaptability to complex image features.

[0122] Furthermore, to account for the multi-scale features resulting from the large apparent diameter variations and uneven pitch density of variable-diameter spring wire, Gaussian segmentation detection was expanded to four octaves, each containing five scale layers. This increases the scale range and density of feature detection, better capturing features ranging from coarse to fine, particularly in areas of varying diameter and pitch. Accordingly, the pre-stored standard 2D model outline contains approximately 250 feature points specifically distributed across the variable diameter and pitch features to cover the more complex geometric details. Ultimately, the number of successfully matched feature point pairs also increased to approximately 240. This demonstrates a stronger ability to capture and match complex geometric features.

[0123] Specifically, the characteristics of variable diameter and pitch springs require more complex and localized parameter definitions. Two-dimensional parameter extraction adds projected minimum outer diameter (ends 1 and 2), projected end 1 pitch, and projected middle pitch. Three-dimensional precision calculations go a step further, including maximum outer diameter (middle), minimum outer diameter (ends 1 and 2), end 1 average pitch (first two turns), middle average pitch (middle three turns), and critical section curvature radius. This global and localized parameter refinement is key to meeting the functional requirements of complex springs and enables a more comprehensive assessment of their forming quality.

[0124] Due to the refinement of parameters, deviation thresholds are also set for different locations and types of parameters. Acceptance is also determined based on these more refined parameters. In this example, the "average pitch deviation of end 1" exceeded the tolerance, resulting in overall failure, even though other macro parameters may have passed. This demonstrates the focused quality control of key functional areas of complex springs.

[0125] Table 4 Morphology comparison and qualification judgment (SPG-VARIO-20230915-077)

[0126]

[0127] The test results show that all nine error parameters are within the allowable tolerances and meet the standards. Only the end pitch deviation, +0.51mm, exceeds the ±0.5mm threshold and is considered unqualified, requiring special attention. Therefore, with the exception of the end pitch, all other indicators are qualified, and overall quality is under control. However, adjustments to the end pitch process are required.

[0128] The in-line inspection system for this automotive brand's variable diameter and variable pitch suspension spring production line successfully overcomes the challenges of inspecting complex spring geometry through an upgraded multi-dimensional data acquisition and intelligent analysis architecture. The system utilizes a high-resolution, simultaneous imaging solution with six viewing angles (0°-225° in 45° increments), achieving over 200% feature point coverage in key areas such as the diameter change zone and pitch transition. Field measurements show a 3D point cloud density three times that of conventional springs. Combined with an improved NURBS surface fitting algorithm, the system achieves contour reconstruction error within ±0.08mm in the variable diameter zone and ±0.12mm in the nonlinear pitch zone, fully meeting the stringent tolerance requirements of this specialized spring.

[0129] Compared to the uniform spring inspection in Example 1, this system effectively addresses the self-occlusion detection challenge of variable diameter and pitch springs through innovative adaptive view fusion technology and an intelligent segmented inspection algorithm, enabling accurate capture of complex geometric features. Furthermore, the integrated nonlinear stiffness verification function upgrades traditional single-dimensional inspection to a comprehensive quality assessment encompassing functional parameters, significantly enhancing process control capabilities for specialized spring products.

[0130] Reference Figure 3 , which is a structural diagram of the detection system proposed in the embodiment of the present invention.

[0131] A system for online inspection of cold-rolled suspension springs based on image recognition has been developed. The system's detection accuracy has been fully verified using two examples presented in this paper. This system innovatively integrates multi-scale feature extraction and Faster R-CNN target detection technology. By dynamically capturing multi-angle spring images, extracting multi-scale features, performing three-dimensional reconstruction, and analyzing virtual space transformations, it enables simultaneous measurement of key spring parameters including length, diameter, number of coils, pitch, and verticality. This system can adapt to the continuous operation requirements of industrial production lines and achieves over 10 times the efficiency of traditional manual inspection, demonstrating the system's significant effectiveness in improving product quality and reducing labor costs.

[0132] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An online detection system for cold-rolled suspension spring forming accuracy based on image recognition, characterized in that: include: Image acquisition and preprocessing module: dynamically acquires multi-angle images of the suspension spring to obtain the spring target area; Spring morphology characterization module: Extracts multi-scale features of the target spring area, including the overall shape, local curvature, and subtle surface features of the suspension spring. It then performs contour correction and matrix decomposition on the multi-scale features to obtain the first morphology map and extract the two-dimensional structural features of the spring. Extracting and matching feature points of the initial contour to obtain a plurality of two-dimensional feature matching points; correcting, fitting, and skeletonizing the initial contour based on the obtained two-dimensional feature matching points to form a second topography image, and extracting two-dimensional structural features from the second topography image; Spring image reconstruction module: Segment the spring target area, use the two-stage object detector Faster R-CNN to locate the spring position, frame the spring projection shape to generate a 2D mask; use the spring stereo morphology reconstruction algorithm to generate a real-time 3D point cloud and extract key nodes from it; Stereo image alignment module: After rigid transformation alignment, it is compared with the standard CAD model of the suspension spring, the suspension spring forming accuracy deviation is calculated, and the qualified judgment is made, and the test results are summarized; Dynamic monitoring module: obtains time series data of suspension spring accuracy deviation and terminates molding when it is determined to be qualified; The dynamic threshold adjustment of the dynamic monitoring module is as follows: Calculate the change rate of each parameter in real time based on time series data, and trigger an early warning when the change rate exceeds the process tolerance; The equipment operating status is incorporated into the spring three-dimensional morphology reconstruction algorithm as a boundary condition to establish a mapping relationship. When the parameter change rate continues to exceed the tolerance range, the judgment threshold and sampling frequency of the subsequent detection cycle are automatically optimized, and parameter adjustments are performed according to the degree of excess.

2. The online detection system for cold-rolling precision of suspension springs based on image recognition according to claim 1 is characterized in that: The image acquisition and preprocessing module includes: Perform dynamic acquisition of multiple angles during the cold rolling process of the suspension spring to obtain spring image information, and then determine the spring target area based on the spring image information; Performing median filtering and image enhancement processing on image data in the spring target area to obtain a first topography image; The initial contour of the suspension spring is detected and generated from the first topography image by utilizing gradient calculation, non-maximum suppression, double threshold screening and edge connection technology.

3. The online detection system for cold-rolling precision of suspension springs based on image recognition according to claim 1 is characterized in that: The spring image reconstruction module includes: Segmenting the suspension spring in the first topography image using the two-stage object detector Faster R-CNN to generate a two-dimensional mask; The detection branch of Faster R-CNN follows the RPN and performs more precise classification and bounding box regression on each candidate region. The integrated Mask R-CNN branch performs pixel-level segmentation. This adds a small fully convolutional network to Faster R-CNN to predict the segmentation mask. This network is trained and optimized using a cross-entropy loss function. This generates an accurate two-dimensional mask, where the pixel value indicates whether the corresponding image pixel belongs to the suspension spring. Based on the two-dimensional mask, feature point matching is performed on the suspension springs in the multi-angle images to establish a correspondence between the spring images; and a sparse point cloud is generated by combining the feature matching points; The multi-view stereo vision MVS algorithm is used to optimize the sparse point cloud, generate real-time 3D point cloud and further enhance the features.

4. The online detection system for cold-rolling precision of suspension springs based on image recognition according to claim 1 is characterized in that: The stereo image alignment module includes: Projecting the key nodes back into the two-dimensional coordinate system of the corresponding multi-angle image; Verifying, correcting and enhancing the first topography image using the projected two-dimensional key node positions and their neighborhood image information in the corresponding image; Through rigid transformation, the key nodes of the real-time point cloud are aligned with the standard CAD model, the cold roll forming deviation is calculated and the qualified judgment result is output.

5. The online detection system for cold-rolling precision of suspension springs based on image recognition according to claim 4 is characterized in that: The suspension springs are classified into the following types: The system obtains a spring model according to the classification, and automatically calls preset three-dimensional reconstruction parameters, key node extraction strategy and deviation calculation model for the model to perform accuracy detection.

6. The online detection system for cold-rolling precision of suspension springs based on image recognition according to claim 1 is characterized in that: The dynamic monitoring module includes: Real-time collection of multi-parameter time series data during the cold coil forming process. When all key parameters reach the preset thresholds for the first time and all parameters remain up to standard for N consecutive test cycles, the product is considered qualified. When the final qualification is determined, a graded stop command is sent to the PLC, the material feed is cut off first, and the machine is completely shut down after the winding kinetic energy decays.

7. The online detection system for cold-rolling precision of suspension springs based on image recognition according to claim 1 is characterized in that: The stereo image alignment module also includes curvature analysis: Based on the real-time three-dimensional point cloud data, the curvature distribution of the central axis of the spring is calculated by a differential geometry algorithm to identify the local bending mutation point of the spring; The extreme points of the spring curvature are compared and analyzed with the theoretical curvature of the standard CAD model. When the curvature deviation exceeds the process tolerance, it is marked as a forming defect area and fed back to the dynamic monitoring module for early warning and process parameter adjustment.

Images