Automobile shock absorber rod multi-type defect detection method and system based on machine vision

Abstract

The application discloses a machine vision-based automobile shock absorber rod multi-type defect detection method and system, relates to the related technical field of shock absorber rod defect detection, and comprises the following steps: placing an automobile shock absorber rod on a detection equipment pipeline to perform automobile shock absorber rod cleaning treatment; conveying to an AOI detection station, activating a multi-modal acquisition unit to perform multi-modal perception of the automobile shock absorber rod; performing positioning analysis on the automobile shock absorber rod in the collection field of view; performing hierarchical perception processing according to a multi-modal image set and geometric pose information; performing defect identification scoring according to the hierarchical perception processing result, and utilizing the defect identification scoring to perform shunt conveying management of the automobile shock absorber rod. The technical problems that the prior art cannot comprehensively cope with multi-type defects on the surface of the automobile shock absorber rod and is insufficient in complex working condition adaptability, resulting in low defect detection precision and efficiency are solved, and the technical effects of multi-type defect comprehensive coverage detection and improved detection precision and efficiency are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of shock rod defect detection, and specifically relates to a shock rod multi-type defect detection method and system based on machine vision. BACKGROUND

[0002] As a core component of the automobile suspension system, the performance of the shock rod directly affects the safety, stability and comfort of the vehicle. In the manufacturing process of the shock rod, due to factors such as material, process or environment, the surface or structure may produce various types of defects, such as cracks, scratches, rust or geometric size deviation, which not only reduce the service life of the shock rod, but also may cause abnormal vibration or failure of the vehicle during driving, and further cause safety hazards. However, traditional shock rod defect detection cannot meet the needs of large-scale production and cannot fully cover the diversity of surface defects. Defect detection based on machine vision can non-contact capture the subtle features of the part surface, realize fast and objective defect identification, but still faces many challenges, such as the shock rod surface may have complex geometric shape and light reflection characteristics, which makes the image acquisition easy to be disturbed by light, shadow or noise; the defect types are various, from macroscopic size deviation to microscopic crack, which requires multi-level perception strategy, and the existing method is difficult to balance multiple defect detection requirements, thereby affecting the defect detection efficiency and accuracy.

[0003] Therefore, in the related art at present, there is a technical problem that the multi-type defects on the surface of the automobile shock rod cannot be comprehensively dealt with and the adaptability to complex working conditions is insufficient, resulting in low defect detection precision and efficiency. SUMMARY

[0004] The present application provides a shock rod multi-type defect detection method and system based on machine vision, which solves the technical problem that the multi-type defects on the surface of the automobile shock rod cannot be comprehensively dealt with and the adaptability to complex working conditions is insufficient in the prior art, resulting in low defect detection precision and efficiency, and achieves the technical effect of comprehensive multi-type defect detection and improved detection precision and efficiency.

[0005] The application provides a machine vision-based multi-type defect detection method for automobile shock rods, which comprises the following steps: placing an automobile shock rod on a detection equipment pipeline, transmitting the automobile shock rod into a roller through a feeding mechanism, and starting an air knife to perform cleaning treatment on the automobile shock rod; conveying the automobile shock rod after the cleaning treatment to an AOI detection station, activating a multi-modal acquisition unit to perform multi-modal sensing of the automobile shock rod, and establishing a multi-modal image set; performing positioning analysis on the automobile shock rod in the acquisition field of view, and outputting geometric pose information; performing hierarchical sensing processing according to the multi-modal image set and the geometric pose information, wherein the hierarchical sensing processing comprises coarse screening sensing layer processing and attention sensing layer processing; performing defect identification scoring according to the hierarchical sensing processing result, and using the defect identification scoring to perform shunt conveying management of the automobile shock rod.

[0006] The machine vision-based multi-type defect detection method for automobile shock rods further performs the following processing: acquiring position importance data of the automobile shock rod, identifying attention scales of the multi-modal image set according to the position importance data and the geometric pose information, and establishing a multi-scale pyramid representation; activating the coarse screening sensing layer, using low-resolution units in the multi-scale pyramid representation to perform rapid identification of single-frame visible light images and near-infrared images of a lightweight neural network, and establishing a candidate defect set; performing multi-modal fusion identification of the candidate defect set by the attention sensing layer to establish the hierarchical sensing processing result.

[0007] The machine vision-based multi-type defect detection method for automobile shock rods further performs the following processing: performing multi-modal feature fusion in each multi-scale pyramid representation, wherein the multi-modal features include texture features, depth profile features, thermal texture features and spectral response features; establishing a joint representation according to the multi-modal feature fusion, inputting the joint representation and the candidate defect set into a model set composed of multiple heterogeneous models to perform reasoning, and establishing a defect identification result, wherein the defect identification result is provided with a confidence score identifier; outputting the defect identification result with the confidence score identifier as the hierarchical sensing processing result.

[0008] The machine vision-based multi-type defect detection method for automobile shock rods further performs the following processing: determining whether the confidence score meets a calibration score threshold; if the confidence score cannot meet the calibration score threshold, performing corresponding automobile shock rod abnormal positioning, performing additional attention acquisition according to the abnormal positioning result, and outputting the defect identification result after compensation using the additional attention acquisition result as the hierarchical sensing processing result.

[0009] The machine vision-based multi-type defect detection method for automobile shock rods further performs the following processing: the plurality of heterogeneous models include a convolutional neural network model and a transformer model of a self-attention mechanism, the convolutional neural network model is used to extract local texture and edge features, and the transformer model of the self-attention mechanism is used to extract cross-scale global dependency relationships; and joint inference is performed after cross fusion of the convolutional neural network model and the transformer model of the self-attention mechanism.

[0010] The machine vision-based multi-type defect detection method for automobile shock rods further performs the following processing: a double-main-stream channel is established, including a qualified main-stream channel and an unqualified main-stream channel; if the defect identification score meets the qualified threshold, the qualified main-stream channel is used for transmission processing; and if the defect identification score does not meet the qualified threshold, the automobile shock rod is uploaded to the unqualified main-stream channel, and defect sub-stream transmission management is performed according to the main defect type.

[0011] The machine vision-based multi-type defect detection method for automobile shock rods further performs the following processing: the defect identification score is uploaded to the system synchronously, data synchronization binding is performed after reading the unique code of the automobile shock rod, and traceability management is performed according to the data synchronization binding result.

[0012] The application also provides a machine vision-based multi-type defect detection system for automobile shock rods, which includes: a cleaning processing module, which is used to place the automobile shock rod in a detection equipment pipeline, transmit the automobile shock rod into a roller through a feeding mechanism, and start a wind knife to perform cleaning processing of the automobile shock rod; a multi-modal perception module, which is used to deliver the automobile shock rod after cleaning to an AOI detection station, activate a multi-modal acquisition unit to perform multi-modal perception of the automobile shock rod, and establish a multi-modal image set; a positioning analysis module, which is used to perform positioning analysis of the automobile shock rod in a collection field of view and output geometric pose information; a hierarchical perception processing module, which is used to perform hierarchical perception processing according to the multi-modal image set and the geometric pose information, the hierarchical perception processing including coarse screening perception layer processing and attention perception layer processing; and a shunt conveying management module, which is used to perform defect identification scoring according to the hierarchical perception processing result, and perform shunt conveying management of the automobile shock rod by using the defect identification score.

[0013] The method and system for detecting multiple types of defects of automobile shock rods based on machine vision provided in the application can place the automobile shock rod in a detection equipment pipeline to perform cleaning treatment on the automobile shock rod, convey the automobile shock rod to an AOI detection station, activate a multi-modal acquisition unit to perform multi-modal sensing of the automobile shock rod, perform positioning analysis on the automobile shock rod in the collection field of view, perform hierarchical perception processing according to the multi-modal image set and geometric pose information, perform defect identification scoring according to the hierarchical perception processing result, and use the defect identification score to perform shunt conveying management of the automobile shock rod. The method and system solve the technical problems of the prior art that cannot comprehensively cope with multiple types of defects on the surface of the automobile shock rod and have insufficient adaptability to complex working conditions, resulting in low defect detection precision and efficiency, and achieve the technical effects of comprehensive coverage of multiple types of defects and improvement of detection precision and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A method for detecting multiple types of defects of automobile shock rods based on machine vision is provided.

[0015] Figure 2 A system structure diagram for detecting multiple types of defects of automobile shock rods based on machine vision is provided.

[0016] The following table explains the reference numerals: cleaning treatment module 10, multi-modal sensing module 20, positioning analysis module 30, hierarchical perception processing module 40, and shunt conveying management module 50. DETAILED DESCRIPTION

[0017] To further illustrate the technical means and effects adopted by the application to achieve the predetermined object, the specific embodiments, structures, features, and effects of the application are described in detail below in combination with the drawings and preferred embodiments.

[0018] The application provides a method for detecting multiple types of defects of automobile shock rods based on machine vision, as shown in Figure 1 The method includes the following steps.

[0019] In step S100, the automobile shock rod is placed in a detection equipment pipeline, the automobile shock rod is transmitted into a roller through a feeding mechanism, and a wind knife is started to perform cleaning treatment on the automobile shock rod.

[0020] Preferably, the detection equipment pipeline is an automated detection conveying line composed of a motor-driven conveyor belt or chain, and the automobile shock absorber is placed in the detection equipment pipeline. The automobile shock absorber is transmitted into the roller through the feeding mechanism, wherein the feeding mechanism can be a mechanical arm, a vibrating disc or a special pushing device, which is used to accurately grasp and place the shock absorber to be detected from the material frame to the starting position of the detection pipeline, realizing full-automatic production. The roller can be a roller conveying line or a cleaning cabin, and the shock absorber is sent into it. The roller slowly rotates forward, that is, by allowing the shock absorber to rotate itself while advancing, it can ensure that its entire outer surface can be exposed in turn under the field of view of the cleaning and subsequent camera, thereby providing smooth and bump-free transmission, preventing the parts from being damaged by knocking during the transmission process; then start the air knife to perform cleaning treatment of the automobile shock absorber, wherein the air knife is an industrial air blowing dust removal equipment, which is usually a long nozzle that can produce a uniform, high-speed and knife-like high-pressure air curtain. When the rotating shock absorber passes above or beside the air knife, the air curtain blows over the entire surface of the shock absorber, thereby removing dust, metal debris, oil stains, fibers and other small particulate matters attached to the surface of the shock absorber during production, handling and feeding. Specifically, any small dust under the high-resolution lens of the camera can be misjudged as a pit, scratch or rust spot, resulting in false detection. The air knife cleaning treatment will not cause secondary scratches on the surface of the shock absorber, so as to ensure the accuracy and reliability of visual detection.

[0021] Step S200, conveying the automobile shock absorber after cleaning to the AOI detection station, activating the multi-modal acquisition unit to perform multi-modal perception of the automobile shock absorber, and establishing a multi-modal image set.

[0022] Preferably, the AOI detection station refers to an automatic optical inspection station, that is, a station that uses optical cameras and image processing to automatically detect product defects. It is usually a closed illumination box with specific angle, white or blue LED light sources inside to eliminate external light interference and optimally highlight the microscopic features of the shock rod surface. After cleaning, the automobile shock rod is transported to the AOI detection station for detection, and the multi-modal acquisition unit is activated to perform multi-modal perception of the automobile shock rod. The multi-modal acquisition unit is composed of multiple different types of sensors, which detect the shock rod from different dimensions. The sensors may include high-resolution visible light cameras for capturing visible defects such as scratches, rust, stains, and paint flaws; near-infrared cameras sensitive to certain material or coating structures, moisture, or specific contaminants, identifying defects invisible to the naked eye; 3D profilers / laser scanners for accurately measuring the three-dimensional shape and depth of the surface to detect geometric defects such as pits, bumps, deformations, and size deviations. When the shock rod is rotating or stationary in the detection station, all sensors are triggered synchronously or sequentially for acquisition. The visible light camera takes a set of color photos, the 3D scanner generates a depth map containing height information for each point, and the near-infrared camera generates a grayscale image. Finally, for the same shock rod, a multi-modal image set is formed, containing surface appearance information, three-dimensional geometric information, and material / internal structure auxiliary information, thereby improving defect detection accuracy.

[0023] Step S300, performing positioning analysis of the automobile shock rod within the acquisition field of view and outputting geometric pose information.

[0024] Preferably, the acquisition field of view refers to a fixed area within which the sensors on the AOI detection station can detect. The positioning analysis of the automobile shock rod in this area involves analyzing the captured image to obtain the pixel-level position of the shock rod in the camera frame through edge detection or template matching, identifying and segmenting the outline of the automobile shock rod from the background, and then accurately identifying the current position of the shock rod and outputting geometric pose information, including the geometric position and geometric attitude of the shock rod. The geometric position refers to the precise coordinates of the shock rod in two-dimensional or three-dimensional space, and the geometric attitude refers to the rotation angle and inclination degree of the shock rod, usually represented by a rotation matrix containing yaw, pitch, and roll elements, thereby ensuring the accuracy of defect detection.

[0025] Step S400, performing hierarchical perception processing based on the multi-modal image set and the geometric pose information, which includes coarse screening perception layer processing and attention perception layer processing.

[0026] The step S400 further comprises a step S410 of acquiring position importance data of the automobile shock rod, performing attention scale identification on the multi-modal image set according to the position importance data and the geometric pose information, and establishing a multi-scale pyramid representation; a step S420 of activating a coarse screening perception layer, performing fast identification on single-frame visible light images and near-infrared images by a lightweight neural network using low-resolution units in the multi-scale pyramid representation, and establishing a candidate defect set; and a step S430 of performing multi-modal fusion identification on the candidate defect set by an attention perception layer to establish a hierarchical perception processing result.

[0027] Preferably, the hierarchical perception processing of the multi-modal image set and the geometric pose information comprises coarse screening perception layer processing and attention perception layer processing. The coarse screening perception layer processing refers to fast scanning of the preprocessed images by a lightweight neural network with small calculation amount and extremely fast speed to identify all regions where defects may exist. The attention perception layer processing refers to identifying the coarse screening perception layer processing result by a convolutional neural network (CNN) + Transformer fusion model to accurately diagnose defects of the shock rod. Then, position importance data of the automobile shock rod is acquired. The position importance data refers to prior knowledge of the automobile shock rod, which is used to define key regions such as threaded connections, welding seams, and force rods of the shock rod, and non-key regions such as non-load-bearing smooth surfaces. The geometric pose information includes real-time position and attitude of the shock rod. Then, attention scale identification is performed on the multi-modal image set according to the position importance data and the geometric pose information, that is, intelligent decision is made to perform high-resolution identification on the key regions and low-resolution scanning on the non-key regions, thereby generating multiple sub-images with gradually reduced resolutions for the same original image, forming an image set like a pyramid, that is, a multi-scale pyramid representation. The tower top representing a low-resolution unit is small in picture and fast in processing speed, and is used for fast browsing of the global; the tower bottom representing a high-resolution unit is large in picture and rich in details, and is used for fine analysis, thereby realizing on-demand allocation of computing resources.

[0028] Preferably, the coarse screening perception layer is activated, and low-resolution units in the multi-scale pyramid representation, that is, tower top images or middle and upper layer images, are used to process single-frame visible light images and near-infrared images by a lightweight neural network, to quickly identify and determine regions where defects may exist, and form a candidate defect set, thereby ensuring processing speed and efficiency. Then, multi-modal fusion identification is performed on the candidate defect set by the attention perception layer, including deep analysis and inference by a fusion model of a larger and more precise convolutional neural network model combined with a self-attention mechanism, fusion of complex multi-modal features, and accurate diagnosis of whether it is a defect, a defect type, and generation of a defect recognition confidence, thereby realizing high-speed, high-precision, and high-robustness automatic detection of multiple types of defects of the automobile shock rod in a complex industrial environment.

[0029] Further, step S430 further comprises step S431, performing multi-modal feature fusion in each multi-scale pyramid representation, the multi-modal features including texture features, depth profile features, thermal texture features, and spectral response features; step S432, establishing a joint representation according to the multi-modal feature fusion, inputting the joint representation and the candidate defect set into a model set composed of multiple heterogeneous models to perform reasoning, and establishing a defect recognition result, wherein the defect recognition result is provided with a confidence score identifier; and step S433, outputting the defect recognition result with the confidence score identifier as a hierarchical perception processing result.

[0030] Preferably, multi-modal feature fusion is performed in each multi-scale pyramid representation, that is, candidate defect regions are understood from different dimensional perceptions, and the multi-modal features include texture features, depth profile features, thermal texture features, and spectral response features. The texture features are used to describe the roughness, frosting, scratches, cracks, rust, and other texture changes of the surface, the depth profile features are used to describe the concave-convex undulations of the surface, the thermal texture features are used to describe the surface temperature distribution patterns caused by different materials and different structures, and the spectral response features are used to describe the response of the shock absorber under illumination of different wavebands. By fusing different dimensional features, a joint representation of the defect region is established, and a unified information-rich digital description is obtained.

[0031] Preferably, the joint representation and the candidate defect set are input into a model set composed of multiple heterogeneous models to perform reasoning, wherein the heterogeneous models are fusion models of a convolutional neural network and a self-attention mechanism model. The convolutional neural network is used to extract local detail features such as edges, corners, and tiny texture patterns from images to accurately determine the direction of scratches and the bifurcation of cracks. The self-attention mechanism model is used to grasp the global context and long-distance dependency relationship, understand the overall morphology of the entire defect region, and understand the correlation between the defect and the surrounding normal region. The multiple heterogeneous models jointly analyze the joint representation, and obtain a defect recognition result through cross-validation fusion, such as that the region is a xx type of defect. At the same time, a corresponding confidence score identifier is calculated for the defect recognition result, for example, 95% confident that it is a scratch, and 60% confident that it is a pit. Finally, the defect recognition result with the confidence score identifier is output as a hierarchical perception processing result, containing a structured defect diagnosis report, to ensure high-precision and quantifiable intelligent recognition of defects of the shock absorber.

[0032] Further, step S432 further comprises step S4321, wherein the multiple heterogeneous models include a convolutional neural network model and a transformer model of a self-attention mechanism. The convolutional neural network model is used to extract local texture and edge features, and the transformer model of the self-attention mechanism is used to extract cross-scale global dependency relationships; and step S4322, performing joint reasoning according to cross-fusion of the convolutional neural network model and the transformer model of the self-attention mechanism.

[0033] Preferably, the plurality of heterogeneous models include a convolutional neural network model and a transformer model of self-attention mechanism, i.e., a collaborative architecture of convolutional neural network CNN and Transformer, wherein the convolutional neural network model slides locally on the image through its convolution kernel, focusing on capturing subtle correlations between pixels to identify local, detailed patterns such as edges of scratches, directions of cracks, contours of rust spots, and further extract local texture and edge features; the transformer of self-attention mechanism simultaneously focuses on all parts of the image and calculates the correlation weights between them, for example, whether a local concave is related to a convex in another area, or whether a tiny texture anomaly is repeated in the global range, and further identifies the overall morphology of the defect and its relationship with the surrounding environment to determine the cross-scale global dependency.

[0034] Preferably, the convolutional neural network model and the transformer model of self-attention mechanism are cross-fused, specifically, the fine local feature maps extracted by the convolutional neural network model are spliced or weighted with the feature maps rich in global context information extracted by the Transformer; or the attention map calculated by the transformer model of self-attention mechanism is fed back to the convolutional neural network model to guide its convolution kernel to focus on the key area for secondary fine feature extraction. Then joint inference is performed, i.e., the local features provided by the convolutional neural network model and the context dependency provided by the Transformer are comprehensively considered, and then a more reliable and accurate defect recognition result is output, thereby greatly reducing the risk of false detection and missed detection.

[0035] Further, step S433 further includes step S4331 of judging whether the confidence score meets a calibrated score threshold; and step S4332 of performing corresponding automobile shock rod abnormal positioning if the confidence score fails to meet the calibrated score threshold, and performing additional attention collection according to the abnormal positioning result, and using the additional attention collection result to compensate the defect recognition result before outputting as the hierarchical perception processing result.

[0036] Preferably, the calibration score threshold is a preset pass line according to historical data, for example, set to 90%, for determining whether the confidence is reliable, and the confidence score is compared with the calibration score threshold to determine whether the judgment is very reliable if the confidence score is higher than the calibration score threshold; if the confidence score is lower than the calibration score threshold, it is considered that the judgment is doubtful; if the confidence score cannot meet the calibration score threshold, the corresponding automobile shock rod abnormal positioning is performed, that is, the precise position of the abnormality on the actual shock rod is calculated by using the geometric pose information and the coordinates of the defect in the image to obtain the abnormal positioning result; then additional attention collection is performed based on the abnormal positioning result, specifically, the camera parameters are adjusted, for example, the light angle is changed to highlight the texture, the focal length is adjusted to obtain clearer images or specific optical filters are enabled, or a high-power optical microscope, a 3D profiler and the like are called to perform super-high-precision scanning on the candidate defect area to obtain super-clear and rich image data of the doubtful area as the additional attention collection result; and the additional attention collection result is used for defect recognition result compensation, including inputting the additional attention collection result into the heterogeneous model for reasoning to obtain a more accurate compensation result with higher confidence and outputting the compensation result as the hierarchical perception processing result of the defect area, thereby greatly reducing the misjudgment rate and improving the reliability of the shock rod defect detection.

[0037] Step S500, according to the hierarchical perception processing result, defect identification scoring is performed, and the automobile shock rod is managed by using the defect identification scoring.

[0038] Preferably, the hierarchical perception processing result is converted into defect identification scoring based on a preset scoring rule, different weights can be given according to different types and severity of defects and weighted calculation of the comprehensive score, for example, defect identification scoring = (scratch length x 0.3) + (pit depth x 0.5) + (rust area x 0.2), and a defect grade mechanism is established, wherein the higher the score, the worse the quality, for example, 0-60 points are qualified products and 61-100 points are unqualified products; then the defect identification scoring is used for the shunt conveying management of the automobile shock rod, the shock rod with defect identification scoring meeting the qualified threshold is conveyed through the qualified main flow channel, and the shock rod with defect identification scoring not meeting the qualified threshold is conveyed through the unqualified main flow channel, so as to realize multi-grade management such as qualified, repair and scrap.

[0039] Further, step S500 further includes step S510 of establishing a double main flow channel, the double main flow channel including a qualified main flow channel and an unqualified main flow channel; step S520, if the defect identification scoring meets the qualified threshold, the automobile shock rod is conveyed through the qualified main flow channel; step S530, if the defect identification scoring does not meet the qualified threshold, the automobile shock rod is uploaded to the unqualified main flow channel, and then defect branch flow transmission management is performed according to the main defect type.

[0040] Preferably, a double main flow channel is established, including a qualified main flow channel and an unqualified main flow channel, and according to the calculated defect identification score, the sorting device on the assembly line is automatically controlled to manage the diversion of the shock absorber, specifically, if the defect identification score meets the qualified threshold, the sorting mechanism is activated to send the shock absorber into the qualified main flow channel, and finally flows to the packaging or assembly workshop production line; if the defect identification score does not meet the qualified threshold, the shock absorber is sent to the unqualified main flow channel for more precise secondary sorting, that is, the main defect type is identified and the defect branch flow is managed according to the main defect type, for the shock absorber with repairable defects, it is transported to the repair branch line for simple repair; for the shock absorber with non-repairable defects, such as structural cracks, it is transported to the waste box; for the pending shock absorber, it is transported to the re-inspection area for manual re-inspection and judgment.

[0041] Further, step S500 further includes step S540 of synchronously uploading the defect identification score to the system, and after reading the unique code of the automobile shock absorber, performing data synchronization binding, and performing traceability management according to the data synchronization binding result.

[0042] Preferably, the defect identification score is synchronously uploaded to the system, the unique code of each automobile shock absorber is obtained through a code reader, such as a DPM code, a barcode or an RFID electronic tag engraved on the shock absorber, and then the unique code is synchronously bound with the defect identification score of this defect detection, as well as detailed defect type, position, size and other data, to obtain a data synchronization binding result; finally, according to the data synchronization binding result, traceability management is performed, which may include forward tracking from parts to data and reverse traceability from problem to root cause, specifically, if a certain coded shock absorber has a problem, all detection records and defect images during production are immediately retrieved to quickly determine whether it is a production problem, an assembly problem or a transportation damage; if a batch of shock absorbers generally have a certain defect, all part codes with the same problem are quickly located by querying the database, so as to initiate accurate recall and avoid expanding losses, and by analyzing the association between defect data and production time, shift, equipment parameters, the root cause of the quality problem is accurately located, and then the production process parameters are optimized to reduce the generation of shock absorber defects from the source, thereby realizing the upgrade of quality management from passive inspection to active prevention.

[0043] In the foregoing, the machine vision-based multi-type defect detection method for automobile shock absorbers according to the embodiments of the present application is described in detail. Next, the machine vision-based multi-type defect detection system for automobile shock absorbers according to the embodiments of the present application will be described with reference to the accompanying drawings. Figure 1 In the foregoing, the machine vision-based multi-type defect detection method for automobile shock absorbers according to the embodiments of the present application is described in detail. Next, the machine vision-based multi-type defect detection system for automobile shock absorbers according to the embodiments of the present application will be described with reference to the accompanying drawings. Figure 2 In the foregoing, the machine vision-based multi-type defect detection method for automobile shock absorbers according to the embodiments of the present application is described in detail. Next, the machine vision-based multi-type defect detection system for automobile shock absorbers according to the embodiments of the present application will be described with reference to the accompanying drawings.

[0044] The machine vision-based automobile shock absorber rod multi-type defect detection system according to the embodiment of the present application is used to solve the technical problem that the prior art cannot comprehensively cope with the multi-type defects on the surface of the automobile shock absorber rod and has insufficient adaptability to complex working conditions, resulting in low defect detection precision and efficiency, and achieves the technical effects of comprehensive multi-type defect detection and improved detection precision and efficiency. Figure 2 As shown in FIG. 1, the machine vision-based automobile shock absorber rod multi-type defect detection system comprises a cleaning processing module 10, a multi-modal perception module 20, a positioning analysis module 30, a hierarchical perception processing module 40, and a shunt conveying management module 50.

[0045] The cleaning processing module 10 is configured to place the automobile shock absorber rod in a detection equipment flow line, transmit the automobile shock absorber rod into a roller through a feeding mechanism, and start a wind knife to perform cleaning processing on the automobile shock absorber rod. The multi-modal perception module 20 is configured to convey the automobile shock absorber rod after the cleaning processing to an AOI detection station, activate a multi-modal acquisition unit to perform multi-modal perception on the automobile shock absorber rod, and establish a multi-modal image set. The positioning analysis module 30 is configured to perform positioning analysis on the automobile shock absorber rod in a collection field of view and output geometric pose information. The hierarchical perception processing module 40 is configured to perform hierarchical perception processing according to the multi-modal image set and the geometric pose information, wherein the hierarchical perception processing comprises coarse screening perception layer processing and attention perception layer processing. The shunt conveying management module 50 is configured to perform defect identification scoring according to the hierarchical perception processing result and perform shunt conveying management of the automobile shock absorber rod by using the defect identification scoring.

[0046] Next, the specific configuration of the hierarchical perception processing module 40 is described in detail. The hierarchical perception processing module 40 further comprises the following: obtaining position importance data of the automobile shock absorber rod, performing attention scale identification of the multi-modal image set according to the position importance data and the geometric pose information, and establishing a multi-scale pyramid representation; activating a coarse screening perception layer, performing rapid identification of single-frame visible light images and near-infrared images by using a lightweight neural network in a low-resolution unit of the multi-scale pyramid representation, and establishing a candidate defect set; performing multi-modal fusion identification of the candidate defect set by using an attention perception layer to establish a hierarchical perception processing result.

[0047] Next, the specific configuration of the hierarchical perception processing module 40 is described in detail. The hierarchical perception processing module 40 further comprises the following: performing multi-modal feature fusion in each multi-scale pyramid representation, wherein the multi-modal features comprise texture features, depth profile features, thermal texture features, and spectral response features; establishing a joint representation according to the multi-modal feature fusion, inputting the joint representation and the candidate defect set into a model set composed of a plurality of heterogeneous models to perform reasoning, and establishing a defect identification result, wherein the defect identification result is provided with a confidence score identifier; and outputting the defect identification result with the confidence score identifier as the hierarchical perception processing result.

[0048] Next, the specific configuration of the hierarchical perception processing module 40 is described in detail. The hierarchical perception processing module 40 further comprises: determining whether the confidence score meets the calibrated score threshold; if the confidence score cannot meet the calibrated score threshold, performing corresponding automobile shock rod abnormal positioning, collecting additional attention according to the abnormal positioning result, and outputting the defect identification result compensation result as the hierarchical perception processing result.

[0049] Next, the specific configuration of the hierarchical perception processing module 40 is described in detail. The hierarchical perception processing module 40 further comprises: the plurality of heterogeneous models comprise a convolutional neural network model and a transformer model of a self-attention mechanism, the convolutional neural network model is used to extract local texture and edge features, and the transformer model of the self-attention mechanism is used to extract cross-scale global dependency; after cross fusion of the convolutional neural network model and the transformer model of the self-attention mechanism, joint inference is performed.

[0050] Next, the specific configuration of the hierarchical perception processing module 40 is described in detail. The hierarchical perception processing module 40 further comprises: establishing a double-mainstream channel, the double-mainstream channel comprising a qualified mainstream channel and an unqualified mainstream channel; if the defect identification score meets the qualified threshold, the transmission process is performed through the qualified mainstream channel; if the defect identification score does not meet the qualified threshold, the automobile shock rod is uploaded to the unqualified mainstream channel, and the defect branch transmission management is performed according to the main defect type.

[0051] Next, the specific configuration of the hierarchical perception processing module 40 is described in detail. The hierarchical perception processing module 40 further comprises: synchronously uploading the defect identification score to the system, and performing data synchronization binding after reading the unique code of the automobile shock rod, and performing traceability management according to the data synchronization binding result.

[0052] The machine vision-based automobile shock rod multi-type defect detection system provided in the embodiments of the present application can perform the machine vision-based automobile shock rod multi-type defect detection method provided in any of the embodiments of the present application, and has the corresponding functional modules and beneficial effects of the execution method.

[0053] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting multiple types of defects in a vehicle shock absorber rod based on machine vision, characterized in that, The method includes: The car shock absorber bar is placed on the testing equipment production line. The car shock absorber bar is transferred to the roller through the feeding mechanism. The air knife is then activated to perform the car shock absorber bar cleaning process. The cleaned car shock absorber bar is transported to the AOI inspection station, where the multimodal acquisition unit is activated to perform multimodal perception of the car shock absorber bar and establish a multimodal image set. The positioning analysis of the vehicle shock absorber bar within the field of view is performed, and the geometric pose information is output. The hierarchical perception processing is performed based on the multimodal image set and the geometric pose information, and the hierarchical perception processing includes coarse-screening perception layer processing and attention perception layer processing. Based on the results of hierarchical perception processing, defect identification scores are generated, and the defect identification scores are used for the diversion and transportation management of automotive shock absorber bars. The hierarchical perception processing based on the multimodal image set and the geometric pose information includes: The location importance data of the car shock absorber is obtained. Based on the location importance data and the geometric pose information, the attention scale recognition of the multimodal image set is performed to establish a multi-scale pyramid representation. The location importance data refers to the prior knowledge of the car shock absorber, which is used to define the key areas of the shock absorber threaded connection, weld seam, load-bearing rod body, and non-critical areas of the non-load-bearing smooth surface. Activate the coarse-screening perception layer and use the low-resolution units in the multi-scale pyramid representation to quickly identify single-frame visible light and near-infrared images using a lightweight neural network, thereby establishing a candidate defect set. The candidate defect set is subjected to multimodal fusion identification at the attention perception layer to establish a hierarchical perception processing result; The multimodal fusion identification of the candidate defect set by the attention perception layer includes: Multimodal feature fusion is performed in each multi-scale pyramid representation. The multimodal features include texture features, depth profile features, thermal texture features, and spectral response features. A joint representation is established based on the fusion of multimodal features. The joint representation and the candidate defect set are input into a model set composed of multiple heterogeneous models to perform inference and establish a defect identification result. The defect identification result is set with a confidence score label. The defect identification results, labeled with confidence scores, are output as the results of hierarchical perception processing.

2. The machine vision-based multi-type defect detection method for automotive shock absorber bars as described in claim 1, characterized in that, The defect identification result, labeled with a confidence score, is output as the hierarchical perception processing result, including: Determine whether the confidence score meets the calibration score threshold; If the confidence score does not meet the calibration score threshold, the corresponding abnormal positioning of the car shock absorber is performed. Additional attention is collected based on the abnormal positioning result. After the defect identification result is compensated using the additional attention collection result, it is output as the hierarchical perception processing result.

3. The machine vision-based multi-type defect detection method for automotive shock absorber bars as described in claim 1, characterized in that, The joint representation and the candidate defect set are input into a model set composed of multiple heterogeneous models to perform inference, including: The multiple heterogeneous models include a convolutional neural network model and a transformer model with a self-attention mechanism. The convolutional neural network model is used to extract local texture and edge features, and the transformer model with a self-attention mechanism is used to extract cross-scale global dependencies. Joint inference is performed by cross-fusing convolutional neural network models and self-attention mechanism transformer models.

4. The machine vision-based multi-type defect detection method for automotive shock absorber bars as described in claim 1, characterized in that, The use of the aforementioned defect identification scoring for the diversion and delivery management of automotive shock absorber bars includes: Establish a dual-path mainstream channel, which includes a qualified mainstream channel and an unqualified mainstream channel; If the defect identification score meets the passing threshold, it will be transmitted through the qualified mainstream channel. If the defect identification score does not meet the pass threshold, the car shock absorber rod will be uploaded to the non-compliant main channel, and then defect branch transmission management will be carried out according to the main defect type.

5. The machine vision-based multi-type defect detection method for automotive shock absorber bars as described in claim 1, characterized in that, The defect identification score is uploaded to the system synchronously, and after reading the unique code of the car shock absorber, the data is synchronized and bound, and the source management is carried out based on the data synchronization and binding results.

6. A machine vision-based multi-type defect detection system for automotive shock absorber bars, characterized in that, The system is used to implement the machine vision-based multi-type defect detection method for automotive shock absorber bars according to any one of claims 1 to 5, and the system includes: The cleaning module is used to place the car shock absorber rods on the testing equipment production line, and the car shock absorber rods are transferred to the rollers through the feeding mechanism. The air knife is then activated to perform the cleaning process on the car shock absorber rods. The multimodal perception module is used to transport the cleaned car shock absorber to the AOI inspection station, activate the multimodal acquisition unit to perform multimodal perception of the car shock absorber, and establish a multimodal image set; The positioning analysis module is used to perform positioning analysis on the vehicle shock absorber within the field of view and output geometric pose information. The hierarchical perception processing module is used to perform hierarchical perception processing based on the multimodal image set and the geometric pose information. The hierarchical perception processing includes coarse-screening perception layer processing and attention perception layer processing. The diversion and delivery management module is used to score defects based on the results of hierarchical perception processing, and to manage the diversion and delivery of automotive shock absorbers using the defect score.

Images