A method and system for online detection of rubber hose quality

Abstract

The application discloses a kind of rubber hose quality on-line detection method and system, it is related to rubber product quality detection technical field, including in the continuous traction conveying process of rubber hose, based on axial displacement and circumference sampling angle establishes circumferential development detection coordinate system;Synchronous acquisition rubber hose outer surface visible light image, laser profile data and near infrared spectrum data, and mapping to corresponding circumferential development position, generate rubber hose surface development quality chart;Surface development quality chart is divided into multiple multi-modal detection units, and the multi-modal feature vector of each detection unit is extracted;With each detection unit as node, according to its spatial adjacent relationship, establish circumferential development adjacency graph;Multi-modal feature vector and circumferential development adjacency graph are input into difference topological graph attention model, obtain the quality determination result of each detection unit, and output rubber hose on-line detection result.Can improve rubber hose full circumference defect positioning accuracy and multi-modal quality determination stability.

Description

Technical Field

[0001] This invention relates to the field of rubber product quality testing technology, and in particular to an online method and system for testing the quality of rubber hoses. Background Technology

[0002] In recent years, with the development of rubber hose production equipment towards continuous, automated, and high-speed production, online inspection technologies for the surface quality and dimensional condition of rubber hoses have been gradually applied. Existing solutions typically deploy visual inspection devices, laser measurement devices, or spectral detection devices in the traction and conveying section after the rubber hose has undergone vulcanization and cooling. By acquiring images, contours, or reflectance spectra of the hose's outer surface, these devices identify cracks, scratches, contamination, outer diameter fluctuations, localized bulges, dents, and abnormal material conditions. Compared to manual visual inspection and offline sampling, these methods improve inspection efficiency and quality monitoring capabilities during production, and provide a data foundation for subsequent re-inspection, rejection, or quality traceability.

[0003] However, existing technologies still have shortcomings in handling full-surface quality inspection of rubber hoses under continuous traction. First, the outer surface of the rubber hose is a closed circumferential surface, and the hose continuously moves along the axial direction during inspection. Existing inspection methods are mostly based on local fields of view or single inspection sections, making it difficult to unify the hose's axial displacement, circumferential sampling position, and inspection data into a stable unfolded position. This results in inaccurate localization and segmentation of defects on the entire circumferential surface of the hose. Second, existing inspection schemes typically process visible light appearance information, laser profile information, and spectral material information separately, or only perform simple threshold fusion. It is difficult to combine the local feature differences between adjacent inspection areas and the axial-circumferential spatial continuity for comprehensive judgment. This makes them susceptible to factors such as local reflection, dust adhesion, and random texture fluctuations, thereby reducing the stability of distinguishing appearance defects, geometric deformation defects, and material anomaly defects. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online inspection method and system for rubber hose quality, which solves the problems of inaccurate positioning of full-circumferential surface defects under continuous traction and lack of axial-circumferential spatial relationship fusion judgment in the existing online inspection technology for rubber hose quality.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an online quality inspection method for rubber hoses, comprising, During the continuous traction and conveying process of the rubber hose, a circumferential unfolding detection coordinate system is established based on the axial displacement and circumferential sampling angle of the rubber hose; Based on the circumferential unfolding detection coordinate system, visible light images, laser contour data and near-infrared spectral data of the outer surface of the rubber hose are acquired simultaneously, and the visible light images, laser contour data and near-infrared spectral data are mapped to the corresponding circumferential unfolding positions to generate a surface unfolding quality map of the rubber hose. The surface unfolded quality map of the rubber hose is divided into multiple multimodal detection units, and the multimodal feature vectors of each multimodal detection unit are extracted. Using each multimodal detection unit as a node, a circular unfolded adjacency graph is established based on the spatial adjacency relationship of each multimodal detection unit in the circular unfolded detection coordinate system; The multimodal feature vectors and the circumferentially expanded adjacency graph are input into the differential topology graph attention model. Graph attention fusion is performed based on the local feature differences between adjacent multimodal detection units and the axial-circumferential topological relationship to obtain the quality judgment results of each multimodal detection unit. The online detection results of the rubber hose are then output based on the quality judgment results.

[0007] As a preferred embodiment of the online quality inspection method for rubber hoses according to the present invention, the step of establishing a circumferential unfolding inspection coordinate system based on the axial displacement and circumferential sampling angle of the rubber hose is as follows: During the continuous traction and conveying of the rubber hose, the lateral swing of the rubber hose is restricted by the centering and guiding mechanism, and the central axis of the optical inspection station is used as the axial reference line, the starting section of the rubber hose entering the optical inspection station is used as the axial reference section, and the preset circumferential sampling direction is used as the circumferential zero angle direction. The axial position of each sampling point is determined based on the axial displacement of the rubber hose obtained by the traction encoder, and the circumferential angle of each sampling point is determined based on the angle of each circumferential sampling position relative to the circumferential zero angle direction. The detection points on the outer surface of the rubber hose are represented as circumferential unfolded detection coordinates composed of axial position and circumferential angle.

[0008] As a preferred embodiment of the online quality detection method for rubber hoses described in this invention, the steps for simultaneously acquiring visible light images, laser contour data, and near-infrared spectral data of the outer surface of the rubber hose are as follows: Based on the circumferentially unfolded detection coordinate system, a synchronous trigger signal is generated according to the equidistant axial sampling interval, so that the visible light acquisition unit, the laser contour acquisition unit, and the near-infrared spectral acquisition unit simultaneously acquire visible light image data, laser contour data, and near-infrared spectral data of the outer surface of the rubber hose. Based on the circumferential unfolding detection coordinate system, visible light image data, laser contour data, and near-infrared spectral data under the same sampling sequence number are bound to the corresponding axial position and circumferential angle to form the original optical detection data sequence.

[0009] As a preferred embodiment of the online quality detection method for rubber hoses according to the present invention, the steps for generating the surface unfolded quality map of the rubber hose are as follows: Based on the original optical detection data sequence and the circumferential unfolding detection coordinate system, a two-dimensional circumferential unfolding matrix is ​​established with the axial position as the row direction and the circumferential angle as the column direction. Visible light image data, laser contour data and near-infrared spectral data are mapped to the corresponding circumferential unfolding positions. Visible light texture layer, laser contour layer and near-infrared spectral layer are formed from the mapped visible light image data, laser contour data and near-infrared spectral data respectively. The data of each layer under the same circumferential unfolding position are combined to generate the surface unfolding quality map of the rubber hose.

[0010] As a preferred embodiment of the online quality inspection method for rubber hoses according to the present invention, the steps of dividing the surface unfolded quality map of the rubber hose into multiple multimodal detection units and extracting the multimodal feature vector of each multimodal detection unit are as follows: Based on the axial position and circumferential angle in the surface unfolded quality map of the rubber hose, the surface unfolded quality map of the rubber hose is divided into multiple multimodal detection units, and the unfolded position set corresponding to each multimodal detection unit is determined. Visible light texture features, laser contour features, and near-infrared spectral features are extracted from the visible light texture layer, laser contour layer, and near-infrared spectral layer corresponding to each multimodal detection unit, respectively. The visible light texture features, laser contour features, and near-infrared spectral features of the same multimodal detection unit are stitched together to obtain the multimodal feature vector of each multimodal detection unit.

[0011] As a preferred embodiment of the online quality inspection method for rubber hoses described in this invention, the steps for establishing a circumferentially expanded adjacency graph based on the spatial adjacency relationship of each multimodal detection unit in the circumferentially expanded detection coordinate system are as follows: Using each multimodal detection unit as a graph node, and taking the axial center position, circumferential central angle, and multimodal feature vector of each multimodal detection unit as node attributes, a graph edge is established based on the axial and circumferential adjacency relationships of each multimodal detection unit in the circumferentially unfolded detection coordinate system. Spatial adjacency weights are calculated based on the axial distance and circumferential arc length between adjacent graph nodes, and a circumferentially unfolded adjacency graph is generated based on graph nodes, graph edges, and spatial adjacency weights.

[0012] As a preferred embodiment of the online quality detection method for rubber hoses of the present invention, the differential topology graph attention model is a pre-trained graph neural network model, which includes a node feature mapping layer, a differential topology attention layer, a neighborhood weighted fusion layer and a quality classification layer connected in sequence. The node feature mapping layer is used to generate initial quality node features for each graph node. The difference topology attention layer is used to generate graph attention weights based on local feature differences and axial-circumferential topological relationships. The neighborhood weighted fusion layer is used to generate a fused quality representation vector. The quality classification layer is used to output the quality judgment results of each multimodal detection unit.

[0013] As a preferred embodiment of the online quality detection method for rubber hoses described in this invention, the steps for obtaining the quality judgment results of each multimodal detection unit are as follows: The multimodal feature vectors of each multimodal detection unit and the circumferentially expanded adjacency graph are input into the differential topology graph attention model, and the multimodal feature vectors are mapped to the initial quality node features of the corresponding graph nodes through the node feature mapping layer. The extended neighborhood set of each graph node is determined based on the circumferentially expanded adjacency graph, and the graph attention weight is calculated by the differential topological attention layer based on the local feature differences between adjacent graph nodes, the axial-circumferential topological relationship, and the spatial adjacency weight. The neighborhood weighted fusion layer uses graph attention weights to weight and fuse the initial quality node features of each graph node in the extended neighborhood set to obtain the fused quality representation vector of each graph node. The quality classification layer then outputs the quality judgment results of each multimodal detection unit based on the fused quality representation vector.

[0014] As a preferred embodiment of the online quality inspection method for rubber hoses described in this invention, the online inspection result of the rubber hose based on the quality judgment result is to group adjacent abnormal detection units into segments to be judged according to the quality judgment result of each multimodal detection unit, and output the qualified, re-inspected or rejected result of the corresponding segment.

[0015] Secondly, the present invention provides an online quality inspection system for rubber hoses, comprising, The coordinate system establishment module is used to establish a circumferential unfolding detection coordinate system based on the axial displacement and circumferential sampling angle of the rubber hose; The optical data acquisition and mapping module is used to simultaneously acquire visible light images, laser contour data and near-infrared spectral data of the outer surface of the rubber hose, and generate a surface unfolding quality map of the rubber hose. The detection unit feature extraction module is used to divide the surface unfolded quality map of the rubber hose into multiple multimodal detection units and extract the multimodal feature vector of each multimodal detection unit. The adjacency graph construction module is used to build a circularly expanded adjacency graph based on the spatial adjacency relationship of each multimodal detection unit as a graph node and determine the spatial adjacency weight. The quality assessment output module is used to input the multimodal feature vector and the circumferentially expanded adjacency graph into the difference topology graph attention model, obtain the quality assessment result, and output the online detection result of the rubber hose.

[0016] The beneficial effects of this invention are as follows: By establishing a circumferential unfolding detection coordinate system, the axial position and circumferential angle during continuous traction are uniformly bound with multi-source detection data, which can improve the accuracy of full-circumferential surface defect localization and segmented representation of rubber hoses; at the same time, visible light, laser profile and near-infrared spectral features are constructed into multimodal detection units, and combined with the circumferential unfolding adjacency graph and difference topology graph attention model for fusion judgment, so that the detection results can simultaneously reflect appearance defects, geometric deformation and material anomalies, reduce misjudgments caused by local reflection, dust or texture fluctuations, and improve the stability and reliability of online quality judgment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the online quality inspection method for rubber hoses in Example 1.

[0019] Figure 2 This is a structural diagram of the online quality inspection system for rubber hoses in Example 1.

[0020] Figure 3 This is a schematic diagram of the mapping between the circumferential unfolding coordinates and multi-source optical data in Example 1.

[0021] Figure 4 This is a schematic diagram of the establishment of the circumferential unfolding detection coordinate system in Example 1. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides an online quality inspection method for rubber hoses, including the following steps: S1. During the continuous traction and conveying process of the rubber hose, a circumferential unfolding detection coordinate system is established based on the axial displacement and circumferential sampling angle of the rubber hose. S1.1: During the continuous traction and conveying of the rubber hose, the lateral swing of the rubber hose is restricted by the centering and guiding mechanism, and the central axis of the optical inspection station is used as the axial reference line, the starting section of the rubber hose entering the optical inspection station is used as the axial reference section, and the preset circumferential sampling direction is used as the circumferential zero angle direction. Specifically, after the rubber hose has undergone vulcanization and cooling, it enters the continuous traction conveying section, where it is guided by a centering and guiding mechanism to the optical inspection station. The central axis of the optical inspection station is used as the axial reference line of the circumferentially unfolded inspection coordinate system, and the initial section of the rubber hose entering the optical inspection station is used as the axial reference section, with the corresponding axial position recorded as x0. Simultaneously, a preset circumferential sampling direction above the optical inspection station is used as the circumferential zero-angle direction, defining the direction of the rubber hose's outer surface around the axis as the circumferential direction. Thus, the inspection position of the rubber hose's outer surface is decomposed into two coordinate quantities: axial position and circumferential angle.

[0026] It should be noted that the centering and guiding mechanism is used to limit the lateral swing of the rubber hose during the traction process, so that the axis of the rubber hose is consistent with the central axis of the optical inspection station. The centering and guiding mechanism is chosen as the positioning basis before coordinate establishment because the rubber hose is prone to radial swing during continuous transportation. If the hose axis is not stabilized first, the detection position obtained by subsequent follow-up circumferential angle expansion will be deviated. The preset circumferential sampling direction is determined based on the installation center line of the first circumferential sampling unit in the optical inspection station. Specifically, the radial direction of the intersection of the optical axis of the first linear array visible light camera located directly above the rubber hose and the outer surface of the hose is taken as the circumferential zero angle direction, and it remains fixed after the equipment is installed and calibrated.

[0027] S1.2: Based on the axial displacement of the rubber hose obtained by the traction encoder, determine the axial position of each sampling point, and based on the angle of each circumferential sampling position relative to the circumferential zero angle direction, determine the circumferential angle of each sampling point, and represent the detection points on the outer surface of the rubber hose as circumferential unfolded detection coordinates composed of axial position and circumferential angle.

[0028] Specifically, during the continuous traction and conveying of the rubber hose, the axial displacement of the rubber hose is obtained by a meter-counting wheel-type incremental encoder set at the front end of the optical inspection station; at the k-th sampling, the cumulative number of pulses output by the encoder since the start of the inspection is read to determine the axial position corresponding to the k-th sampling. Through this process, the detection points on the surface of the rubber hose in continuous motion are converted into defined axial positions, avoiding axial positioning errors caused by fluctuations in traction speed when the position is calculated solely based on sampling time. After determining the axial position, the preset circumferential zero-angle direction in the optical inspection station is used as the angular reference, and the installation angle of the m-th circumferential sampling position relative to the circumferential zero-angle direction is recorded as... , where represents the center circumferential sampling angle of the m-th circumferential sampling position; if the m-th circumferential sampling position includes multiple lateral sampling points, then the circumferential sampling angle corresponding to the u-th lateral sampling point is determined based on the pre-calibrated center sampling point number and angular resolution of that circumferential sampling position. ; After obtaining the axial position and circumferential sampling angle Then, the detection point on the outer surface of the rubber hose corresponding to the u-th lateral sampling point within the m-th circumferential sampling position at the k-th sampling time is represented as the circumferential unfolded detection coordinate. ,in, This is used to characterize the position of the detection point along the length of the rubber hose. This is used to characterize the position of the detection point on the circumference of the rubber hose. When the calculated circumferential sampling angle is less than 0 or greater than or equal to 2π, the circumferential sampling angle is normalized to the range of 0 to 2π to ensure that the detection points located on both sides of the zero angle of the circumference are still positioned according to the same circumferential period. Thus, a circumferential unfolding detection coordinate system is established based on the axial displacement of the rubber hose and the circumferential sampling angle, so that the subsequently acquired optical detection data of the hose surface can be unfolded into a two-dimensional detection map according to the same coordinate reference.

[0029] Determine the axial position corresponding to the kth sampling. The expression is:

[0030] In the formula, This indicates the axial reference position corresponding to the starting section of the test. This represents the cumulative number of pulses output by the encoder from the start of detection to the kth sampling time. This represents the axial displacement of the rubber hose corresponding to a single encoder pulse. This indicates the initial pulse count value of the encoder at the start of the detection. Determine the circumferential sampling angle corresponding to the u-th lateral sampling point. The expression is:

[0031] In the formula, This represents the circumferential sampling angle corresponding to the u-th horizontal sampling point within the m-th circumferential sampling position. This represents the angle of the m-th circumferential sampling position relative to the center circumference in the zero-angle direction of the circumference. This indicates the index of the reference lateral sampling point within the m-th circumferential sampling position, corresponding to the angle with the central circumference. This represents the circumferential angular resolution between adjacent lateral sampling points within the m-th circumferential sampling position.

[0032] It should be noted that the meter-counting wheel incremental encoder includes a meter-counting wheel that rolls against the outer surface of the rubber hose, an incremental rotary encoder coaxially connected to the meter-counting wheel, and a pulse counting unit. When the rubber hose is continuously traction and conveyed, the rubber hose drives the meter-counting wheel to rotate synchronously. The incremental rotary encoder outputs pulse signals as the meter-counting wheel rotates. The pulse counting unit accumulates the number of pulses and converts the accumulated number of pulses into the actual axial displacement of the rubber hose based on the axial displacement corresponding to each pulse.

[0033] S2. Based on the circumferential unfolding detection coordinate system, the visible light image, laser contour data and near-infrared spectral data of the outer surface of the rubber hose are acquired simultaneously, and the visible light image, laser contour data and near-infrared spectral data are mapped to the corresponding circumferential unfolding position to generate the unfolding quality map of the rubber hose surface. S2.1: Based on the circumferential unfolding detection coordinate system, a synchronous trigger signal is generated according to the equidistant axial sampling interval, so that the visible light acquisition unit, laser contour acquisition unit and near-infrared spectrum acquisition unit can synchronously acquire visible light image data, laser contour data and near-infrared spectrum data of the outer surface of the rubber hose; Specifically, after establishing the circumferential unfolding detection coordinate system, the upper limit of the equidistant axial sampling interval is set according to the minimum defect length that needs to be identified in the length direction of the rubber hose to be inspected, and the axial displacement corresponding to a single encoder pulse is used as the basis for the measurement. Select the number of encoder pulses between two adjacent synchronous acquisitions. This makes the actual equidistant axial sampling interval For example, when this embodiment requires the identification of cracks, scratches, local bulges, or local dents with a length of not less than 2 mm, the axial distance between two adjacent synchronous acquisitions is set to not more than 1 mm, and the selection of parameters that meet the requirements is... positive integers This ensures that any defect of 2mm or more is covered by at least two consecutive synchronous acquisitions when the rubber hose passes through the optical inspection station; during the continuous traction and conveying of the rubber hose, the acquisition controller accumulates the number of pulses output by the meter wheel incremental encoder, and for each accumulated pulse... Each pulse generates a synchronization trigger signal, and the same synchronization trigger signal is simultaneously sent to each linear array visible light camera, each laser line profile sensor and each near-infrared diffuse reflection spectrum acquisition head, so that each optical acquisition unit completes a synchronous acquisition under the same sampling sequence number. During the k-th synchronization trigger, according to Determine the axial position corresponding to this synchronous acquisition, and record the sampling sequence number k and sampling time. Upon arrival of the same synchronization trigger signal, each linear array visible light camera acquires visible light stripe images of the outer surface of the rubber hose within its corresponding circumferential sampling area, thus obtaining visible light image data. Each laser line profile sensor collects the profile height and outer diameter profile points within its corresponding circumferential sampling area to obtain laser profile data. Each near-infrared diffuse reflectance spectral acquisition head collects the near-infrared reflectance spectrum within its corresponding circumferential sampling area, thus obtaining near-infrared spectral data. Thus, visible light image data, laser profile data, and near-infrared spectral data under the same sampling number k correspond to the same axial position. .

[0034] S2.2: Based on the circumferential unfolded detection coordinate system, the visible light image data, laser contour data and near-infrared spectral data under the same sampling sequence number are bound to the corresponding axial position and circumferential angle to form the original optical detection data sequence.

[0035] Specifically, based on the circumferential sampling angle, the three types of optical data under the same sampling number k are respectively bound to the corresponding circumferential unfolding detection coordinates; among them, visible light image data The horizontal sampling points in the middle are bound to laser contour data Contour sampling points in the data are bound to Near-infrared spectral data Spectral sampling points in the data are bound to This ensures that visible light image data, laser profile data, and near-infrared spectral data at the same axial position and corresponding circumferential angle have a unified position index; the visible light image data, laser profile data, and near-infrared spectral data with completed coordinate binding are stored sequentially according to the sampling sequence number to form a sequence of original optical detection data. Each set of original optical detection data includes the sampling sequence number, axial position, circumferential angle, visible light image data, laser profile data, near-infrared spectral data, and rubber hose production batch number.

[0036] It should be noted that, , and These represent the positions of the visible light lateral sampling point, the laser profile sampling point, and the near-infrared spectral sampling point within the m-th circumferential sampling position along the circumferential direction of the rubber hose.

[0037] S2.3: Based on the original optical detection data sequence and the circumferential unfolding detection coordinate system, a two-dimensional circumferential unfolding matrix is ​​established with the axial position as the row direction and the circumferential angle as the column direction, and the visible light image data, laser contour data and near-infrared spectral data are mapped to the corresponding circumferential unfolding positions; Specifically, after forming the original optical detection data sequence, a two-dimensional circular unfolding matrix is ​​established using the circular unfolding detection coordinate system as the mapping reference, with the axial position as the row direction and the circular angle as the column direction. The axial position is determined by the axial position corresponding to each synchronous sampling number, and the circular angle is discretized according to the circular sampling angle based on a preset circular angle resolution, so that each unfolding position in the two-dimensional circular unfolding matrix corresponds to a detection position on the outer surface of the rubber hose. Based on the coordinate binding relationship, the visible light image data, laser profile data and near-infrared spectral data in the original optical detection data sequence are written into the corresponding circular unfolding positions in the two-dimensional circular unfolding matrix, so that the three types of optical data have a unified axial position and circular angle index in the same matrix structure.

[0038] S2.4: Form visible light texture layer, laser contour layer and near-infrared spectral layer respectively for the mapped visible light image data, laser contour data and near-infrared spectral data, and combine the data of each layer under the same circumferential unfolding position to generate the surface unfolding quality map of the rubber hose.

[0039] Specifically, brightness correction is performed on the visible light image data mapped to the two-dimensional circular unfolding matrix. Dark field calibration results are used to eliminate camera background response, and uniform white board calibration results are used to correct brightness differences between different circular sampling positions and different pixels, resulting in corrected visible light grayscale values. The corrected visible light image data is then written into the visible light texture layer. When adjacent circular sampling positions have overlapping sampling values ​​at the same unfolding position, the corrected grayscale value corresponding to the sampling position whose center angle is closest to that unfolding position is selected as the visible light texture value for that unfolding position. The surface height and radius values ​​from the laser contour data are written into the laser contour layer. When a certain circular... When the unfolded position is located between two adjacent laser contour sampling points, the linear interpolation result of the two adjacent contour sampling points is used as the contour value of that unfolded position. The near-infrared spectral data is converted into absorbance features and written into the near-infrared spectral layer. The absorbance features are calculated based on the reference reflectance intensity obtained by the same near-infrared diffuse reflectance spectral acquisition head from the standard white board before detection. Finally, the visible light texture layer, laser contour layer and near-infrared spectral layer under the same unfolded position are combined to generate the unfolded quality map of the rubber hose surface, so that each unfolded position corresponds to multimodal quality data containing visible light texture information, laser contour information and near-infrared spectral information.

[0040] Corrected visible light gray value The expression is:

[0041] In the formula, This represents the original grayscale value obtained from the u-th horizontal sampling point at the m-th circular sampling position during the k-th synchronous acquisition. This represents the dark field grayscale value corresponding to the u-th horizontal sampling point at the m-th circular sampling position under dark field calibration, used to characterize the camera's background response. This represents the flat-field grayscale value corresponding to the u-th lateral sampling point at the m-th circumferential sampling position under uniform whiteboard calibration, used to characterize the response of this sampling point under uniform lighting conditions. This indicates an extremely small positive number that prevents the denominator from being zero; It should be noted that, It is obtained by acquiring the grayscale response of the u-th horizontal sampling point at the m-th circular sampling position in dark field conditions. The grayscale response of the uth lateral sampling point at the mth circumferential sampling position is obtained under uniform whiteboard illumination and stored in the acquisition controller for online calibration before formal testing.

[0042] Absorbance characteristics The expression is:

[0043] In the formula, This indicates that during the k-th synchronous acquisition, at the circumferential angle... The near-infrared reflectance intensity of the rubber hose surface was collected at wavelength λ. This represents the reference reflectance intensity at wavelength λ obtained by acquiring a standard white board using the same near-infrared diffuse reflectance spectroscopy acquisition head before testing. This represents the dark field reflection intensity at wavelength λ obtained by the same near-infrared diffuse reflectance spectroscopy acquisition head under light-shielding conditions. This represents the smallest positive number that prevents the denominator from being zero.

[0044] S3. Divide the surface unfolded quality map of the rubber hose into multiple multimodal detection units, and extract the multimodal feature vector of each multimodal detection unit; S3.1: Based on the axial position and circumferential angle in the unfolded quality map of the rubber hose surface, the unfolded quality map of the rubber hose surface is divided into multiple multimodal detection units, and the unfolded position set corresponding to each multimodal detection unit is determined; Specifically, using the axial position and circumferential angle in the unfolded quality map of the rubber hose surface as the dividing benchmark, the unfolded quality map of the rubber hose surface is divided into multiple multimodal detection units, and the axial length and circumferential angle width of the multimodal detection units are determined. Starting from the detection start section and the circumferential zero angle direction, the unfolded quality map of the rubber hose surface is continuously divided according to the axial length and circumferential angle width, so that each multimodal detection unit corresponds to a local unfolded area on the outer surface of the rubber hose. For the remaining area at the axial end that is less than a complete axial length, the remaining area is merged into the adjacent previous multimodal detection unit. The circumferential direction is divided continuously according to the period, so that the multimodal detection units on both sides of the circumferential zero angle remain adjacent. The axial center position, circumferential central angle, and the set of unfolded positions covered by the i-th multimodal detection unit are recorded.

[0045] Axial length of multimodal detection unit The expression is:

[0046] In the formula, This indicates the minimum axial defect length that needs to be identified in product inspection requirements. Indicates the equidistant axial sampling interval; Circumferential angle width of the multimodal detection unit The expression is:

[0047] In the formula, This indicates the minimum circumferential defect arc length width that needs to be identified in product inspection requirements. It is used to characterize the minimum actual width of the defect after it is expanded along the circumferential direction of the outer surface of the rubber hose. This indicates the standard radius of the rubber hose of this specification. This represents the circumferential angular resolution of a two-dimensional circular expansion matrix.

[0048] S3.2: Extract visible light texture features, laser contour features, and near-infrared spectral features from the visible light texture layer, laser contour layer, and near-infrared spectral layer corresponding to each multimodal detection unit, respectively; For the i-th multimodal detection unit, visible light texture features are extracted from the visible light texture layer. Specifically, the corrected grayscale values ​​within the unfolded position set are read, and the grayscale mean and standard deviation are calculated. Then, the adjacent units along the upper axis, the adjacent units along the lower axis, the adjacent units on the left circumference, and the adjacent units on the right circumference of the i-th multimodal detection unit are used as local reference windows. When the i-th multimodal detection unit is located at the axial start end or axial end end, the local reference window... It consists only of actual axially adjacent units and circumferentially adjacent units; when the i-th multimodal detection unit is located on both sides of the circumferential zero angle, its left and right circumferentially adjacent units are determined according to the circumferential periodicity; then the texture anomaly intensity is calculated; thus forming the visible light texture features of the i-th multimodal detection unit, including gray-scale mean, gray-scale standard deviation and texture anomaly intensity; For the same i-th multimodal detection unit, laser contour features are extracted from the laser contour layer. Specifically, the laser contour radius values ​​within the unfolded position set are read, the average radius, maximum radius, and minimum radius within the multimodal detection unit are calculated, and the local equivalent outer diameter deviation index, radius fluctuation index, and local height abrupt change index are calculated according to the formula. Thus, the laser contour features of the i-th multimodal detection unit are formed, including the local equivalent outer diameter deviation index, radius fluctuation index, and local height abrupt change index. For the same i-th multimodal detection unit, near-infrared spectral features are extracted from the near-infrared spectral layer; Specifically, the near-infrared absorbance at each characteristic wavelength within the unfolded position set is read, and the average absorbance of the i-th multimodal detection unit at wavelength λ is calculated; the characteristic wavelength set is then... The average absorbance is used to form the absorbance feature vector. Then calculate the near-infrared spectral residual; thus forming the near-infrared spectral characteristics of the i-th multimodal detection unit, including the absorbance feature vector and the near-infrared spectral residual.

[0049] It should be noted that the set of characteristic wavelengths is determined by the difference in near-infrared absorbance between qualified rubber hose samples and defective samples of the same specification during the offline calibration stage. Wavelengths that can characterize the vulcanization state of the rubber surface, the uniformity of mixing, or the difference in material composition are selected as characteristic wavelengths.

[0050] Calculate texture anomaly intensity The expression is:

[0051] In the formula, This represents the average visible light gray value of the i-th multimodal detection unit. Represents a local reference window The average visible light gray level within the range, Represents a local reference window The visible light gray standard deviation within; Calculate the local equivalent outer diameter deviation index The expression is:

[0052] In the formula, This represents the average value of the laser profile radius within the i-th multimodal detection unit. This indicates the standard outer diameter of the rubber hose of this specification. This indicates the upper limit of the allowable outer diameter deviation for this specification of rubber hose; Calculate the radius fluctuation index The expression is:

[0053] In the formula, and Let these represent the maximum and minimum radius values ​​within the i-th multimodal detection unit, respectively. This indicates the upper limit of the permissible local radius fluctuation for this specification of rubber hose; Calculate the local height abrupt change index The expression is:

[0054] In the formula, This represents the median value of the laser profile radius within the local reference window. This indicates the upper limit of the permissible local bulge or indentation height for this specification of rubber hose; calculate the average absorbance. The expression is:

[0055] In the formula, It is an expanded set of positions. This indicates the absorbance of the near-infrared spectral layer written in step S3; Calculate near-infrared spectral residuals The expression is:

[0056] In the formula, This represents the average absorbance feature vector of qualified rubber hose samples of the same specification. This represents the covariance matrix of the absorbance eigenvectors of qualified rubber hose samples of the same specification. This represents a regularization term used to prevent the covariance matrix from becoming non-invertible. I is a very small positive number, and I represents the identity matrix.

[0057] It should be noted that the upper limit of outer diameter deviation, the upper limit of local radius fluctuation, and the upper limit of local protrusion or depression height are determined before testing based on the product drawings, enterprise inspection specifications, or applicable product standards of the rubber hose to be tested. If no corresponding upper limit is given directly, the laser profile data of qualified rubber hose samples of the same specification are used for statistical analysis. The normal fluctuation upper limit of outer diameter deviation, local radius difference, and local height deviation in qualified samples is taken as the corresponding upper limit, and remains unchanged during the online testing of rubber hoses of the same specification. covariance matrix The near-infrared absorbance feature vectors of qualified rubber hose samples of the same specification were statistically obtained during the offline calibration phase. Specifically, near-infrared absorbance feature vectors of multiple qualified rubber hose samples of the same specification, confirmed by manual re-inspection, were collected. First, the average value of each absorbance feature vector was calculated. Then, the covariance matrix was calculated based on the deviation of each absorbance feature vector from the average value. This serves as a fixed reference parameter for online testing of rubber hoses of this specification.

[0058] S3.3: The visible light texture features, laser contour features and near-infrared spectral features of the same multimodal detection unit are spliced ​​together to obtain the multimodal feature vector of each multimodal detection unit.

[0059] Specifically, after obtaining the visible light texture features, laser contour features, and near-infrared spectral features of the i-th multimodal detection unit, the three types of features are merged into the multimodal feature vector of the multimodal detection unit in a fixed order: visible light texture features first, laser contour features in the middle, and near-infrared spectral features last. The multimodal feature vector is used to simultaneously characterize the surface texture state, contour deformation state, and material spectral state of the local area of ​​the rubber hose corresponding to the multimodal detection unit.

[0060] S4. Using each multimodal detection unit as a node, establish a circular unfolded adjacency graph based on the spatial adjacency relationship of each multimodal detection unit in the circular unfolded detection coordinate system; S4.1: Using each multimodal detection unit as a graph node, and taking the axial center position, circumferential central angle and multimodal feature vector of each multimodal detection unit as node attributes, establish graph edges based on the axial and circumferential adjacency relationships of each multimodal detection unit in the circumferentially unfolded detection coordinate system; Specifically, after obtaining each multimodal detection unit and its multimodal feature vector, each multimodal detection unit is treated as a graph node in the circumferentially expanded adjacency graph, and the axial center position, circumferential center angle, and multimodal feature vector of the i-th multimodal detection unit are used as the node attributes of the corresponding graph node. Based on the continuous division of the surface development quality map of the rubber hose, the axial sequence number a corresponding to the i-th multimodal detection unit is determined. i and the circumference number c i The circumferential indices are set according to the circumferential periodicity, ensuring that the multimodal detection unit corresponding to the largest circumferential indices is adjacent to the multimodal detection unit corresponding to the smallest circumferential indices in the circumferential direction; for any two graph nodes, if they satisfy... And c i =c j If they are axially adjacent nodes, then they are determined to be adjacent nodes; if they satisfy a i =a j And c i With c j If two nodes are adjacent in the circumferential direction, they are determined to be circumferentially adjacent nodes. When two graph nodes are determined to be axially adjacent nodes or circumferentially adjacent nodes, a graph edge is established between the two graph nodes. Thus, a graph edge set is formed based on the spatial adjacency relationship of the multimodal detection unit in the circumferential unfolded detection coordinate system.

[0061] S4.2: Calculate the spatial adjacency weight based on the axial distance and circumferential arc length between adjacent graph nodes, and generate a circumferentially unfolded adjacency graph based on graph nodes, graph edges and spatial adjacency weights.

[0062] Specifically, the periodic circumferential angle difference, axial distance, and circumferential arc length distance between the i-th and j-th graph nodes with established graph edges are calculated. The spatial adjacency weights between adjacent graph nodes are calculated based on the axial distance and circumferential arc length distance. For two graph nodes without spatial adjacency, no graph edges are established, and no spatial adjacency weights are calculated. After establishing graph edges and calculating spatial adjacency weights for each spatially adjacent node, a circular unfolded adjacency graph G=(V,E,W) is generated, where V represents the set of graph nodes composed of each multimodal detection unit, E represents the set of graph edges established based on spatial adjacency, and W represents the set of spatial adjacency weights corresponding to each graph edge. Thus, using each multimodal detection unit as a node, a circular unfolded adjacency graph is established based on the spatial adjacency of each multimodal detection unit in the circular unfolded detection coordinate system, and the spatial adjacency weights between adjacent nodes are determined.

[0063] Calculate the difference in circumferential angles during the calculation period The expression is:

[0064] In the formula, and Let represent the circumferential center angles of the i-th and j-th multimodal detection units, respectively; Calculate axial distance and the distance of the circumferential arc The expression is:

[0065] In the formula, r0 represents the standard radius of the rubber hose of this specification; Calculate the spatial adjacency weight w ij The expression is:

[0066] In the formula, L u Indicates the axial length of the multimodal detection unit. This represents the circumferential angle width of the multimodal detection unit. This represents the arc length and width of the multimodal detection unit in the circumferential direction.

[0067] It should be noted that, , , and All measurements are expressed in radians, and radians are treated as dimensionless quantities in arc length calculations. and Both represent the actual arc length along the circumference.

[0068] S5. Input the multimodal feature vector and the circumferentially expanded adjacency graph into the difference topology graph attention model. Perform graph attention fusion based on the local feature differences between adjacent multimodal detection units and the axial-circumferential topological relationship to obtain the quality judgment results of each multimodal detection unit. Output the online detection results of the rubber hose based on the quality judgment results.

[0069] Before inputting the difference topology graph attention model, the multimodal feature vectors of each multimodal detection unit are normalized to convert visible light texture features, laser contour features, and near-infrared spectral features into dimensionless features. In the difference topology graph attention model, the output dimension of each feature transformation matrix is ​​pre-set to the same hidden feature dimension, so that node features, local difference features, and positional difference features can be spliced, weighted aggregated, and residuals added.

[0070] S5.1: Input the multimodal feature vectors of each multimodal detection unit and the circumferentially expanded adjacency graph into the differential topology graph attention model, and map each multimodal feature vector to the initial quality node features of the corresponding graph node through the node feature mapping layer; Specifically, after obtaining the circumferentially expanded adjacency graph, the multimodal feature vectors of each multimodal detection unit are used as input features for graph nodes, and the graph edge set and spatial adjacency weight set are used as graph topology inputs, which are then input into a pre-trained differential topology graph attention model. The differential topology graph attention model first performs linear mapping on the multimodal feature vectors of each multimodal detection unit through a node feature mapping layer, transforming visible light texture features, laser contour features, and near-infrared spectral features to the same hidden feature dimension, thus obtaining the initial quality node features of the corresponding graph nodes. This processing ensures that multimodal detection features from different sources and with different dimensions have a unified expression form before entering the graph attention calculation, providing an input foundation for subsequent local difference analysis and neighborhood fusion based on graph topology relationships.

[0071] Initial mass node characteristics The expression is:

[0072] In the formula, M f F represents the node feature transformation matrix obtained through pre-training. i Let represent the multimodal feature vector of the i-th multimodal detection unit.

[0073] S5.2: Determine the extended neighborhood set of each graph node based on the circumferentially expanded adjacency graph, and calculate the graph attention weights through the differential topology attention layer based on the local feature differences between adjacent graph nodes, the axial-circumferential topological relationship, and the spatial adjacency weights. Specifically, for any graph node, the adjacent graph nodes connected to it are determined based on the established graph edge relationships in the circumferentially expanded adjacency graph, and the graph node itself is added to the range of the adjacent graph nodes, forming an extended neighborhood set containing itself and its adjacent nodes. The spatial adjacency weights are topologically normalized to obtain normalized spatial adjacency weights, which characterize the adjacency strength between the current graph node and each extended neighborhood node. The differential topology attention layer subtracts the initial quality node features of the current graph node from the initial quality node features of the adjacent graph nodes along the same feature dimension to obtain local feature differences, which characterize abrupt changes in appearance texture, geometric contour, and near-infrared spectral response between adjacent multimodal detection units. Simultaneously, based on the axial center position difference and circumferential center angle difference between the current graph node and each extended neighborhood node, An axial-circumferential position difference vector is constructed to characterize the axial-circumferential topological relationship. This axial-circumferential topological relationship represents the relative spatial relationship between adjacent multimodal detection units along the length and circumference of the rubber hose. The axial-circumferential position difference vector serves as the quantized input to the axial-circumferential topological relationship in the graph attention weight calculation. The initial quality node features of the current graph node, the initial quality node features of the extended neighbor nodes, local feature differences, the axial-circumferential position difference vector, and the normalized spatial adjacency weights are input into the attention scoring sublayer to obtain the attention scores of the current graph node for each extended neighbor node. The attention scores within the same extended neighbor set are normalized so that the sum of the graph attention weights corresponding to each extended neighbor node is one, resulting in the graph attention weights used for subsequent neighborhood weighted fusion.

[0074] Normalized spatial adjacency weight The expression is:

[0075] In the formula, d i d represents the adjacency weight of the i-th graph node within its extended neighborhood. j Let q represent the adjacency weight of the j-th graph node in its corresponding extended neighborhood. Any extended neighbor node of a graph node, w represents the extended neighborhood set of the i-th graph node; when j=i, w ii This represents the adjacency weight of the self-connected space, and is usually set to 1. Axial-circular position difference vector P ij The expression is:

[0076] In the formula, Let x represent the signed circumferential angle difference between the j-th multimodal detection unit and the ith multimodal detection unit. i and xj These represent the axial center positions of the i-th and j-th multimodal detection units, respectively. Attention score of current graph node to each extended neighbor node The expression is:

[0077] In the formula, This represents the attention parameter vector corresponding to the h-th attention head. This represents the transpose of the attention parameter vector. This represents the node feature transformation matrix corresponding to the h-th attention head. This shows the local difference feature transformation matrix corresponding to the h-th attention head. This represents the position difference transformation matrix corresponding to the h-th attention head. and Let represent the initial quality node features of the i-th and j-th graph nodes, respectively. This represents the difference in local features between the i-th graph node and the j-th graph node. Let represent the positional difference vector between the i-th graph node and the j-th graph node. This represents a linear unit activation function with leakage correction, used to generate non-linear attention scores; Graph attention weights The expression is:

[0078] In the formula, This represents the attention score between the i-th graph node and the q-th extended neighbor node.

[0079] S5.3: Through the neighborhood weighted fusion layer, the initial quality node features of each graph node in the extended neighborhood set are weighted and fused using graph attention weights to obtain the fused quality representation vector of each graph node. The quality classification layer outputs the quality judgment result of each multimodal detection unit based on the fused quality representation vector.

[0080] Specifically, after obtaining the graph attention weights corresponding to each extended neighborhood node, the neighborhood weighted fusion layer performs feature transformation on the initial quality node features of each graph node in the extended neighborhood set, and performs weighted summation according to the corresponding graph attention weights to obtain the neighborhood fusion feature of the current graph node. At the same time, the initial quality node features of the current graph node itself are included as residual retention terms in the neighborhood fusion feature to obtain the fusion quality representation vector of the current graph node, so as to avoid the weakening of micro-cracks, local bulges or local material anomalies in the neighborhood fusion process. For the case of using multi-head attention, multiple attention heads independently complete attention scoring, weight normalization and neighborhood fusion, and then the fusion results obtained by multiple attention heads are averaged to form the final fusion quality representation vector. The quality classification layer classifies the fusion quality representation vector and outputs the probability of the corresponding multimodal detection unit belonging to the quality category of qualified, appearance defect, geometric deformation defect and material anomaly defect, and takes the category with the highest probability as the quality judgment result of the multimodal detection unit.

[0081] Fusion quality representation vector The expression is:

[0082] In the formula, K represents the total number of attention heads. This represents the neighborhood feature transformation matrix corresponding to the h-th attention head. Represents a nonlinear activation function; Quality category probability The expression is:

[0083] In the formula, U represents the pre-trained classification layer weight matrix, b represents the pre-trained classification layer bias term, and Softmax represents the normalization exponential function, which is used to convert the classification layer output into probabilities of different quality categories.

[0084] It should be noted that the fused quality characterization vector includes the appearance, contour and spectral quality information of the current multimodal detection unit itself, as well as the topological context information provided by its axially adjacent and circumferentially adjacent multimodal detection units. This is used to enhance the judgment response of the real continuous distribution area of ​​defects and reduce misjudgments caused by isolated reflections, local dust or random texture fluctuations. In the differential topology graph attention model, the node feature mapping matrix, the feature transformation matrix corresponding to each attention head, the attention parameter vector, the classification layer weight matrix, and the classification layer bias term are all trainable parameters of the model and are obtained through offline training before online detection. The training data for the difference topology graph attention model includes historical qualified hose samples consistent with the specifications of the rubber hose to be inspected, as well as defective hose samples confirmed by manual re-inspection, cross-sectional re-inspection, or standard quality inspection. The defective hose samples include at least appearance defect samples, geometric deformation defect samples, and material anomaly defect samples. For each training sample, the aforementioned steps are followed to generate a surface unfolded quality map of the rubber hose, divide it into multimodal detection units, extract the multimodal feature vectors of each multimodal detection unit, and establish a circumferential unfolded adjacency graph containing spatial adjacency weights. Simultaneously, based on the confirmed axial position, circumferential position, defect range, and defect type of the defect in the defect sample, the corresponding multimodal detection unit is labeled as an appearance defect, geometric deformation defect, or material anomaly defect. Multimodal detection units located within the inspection-confirmed qualified area are labeled as qualified. This forms a training sample set with multimodal feature vectors and the circumferential unfolded adjacency graph as input and quality category labels of each multimodal detection unit as supervised output. During training, The training sample set is input into the differential topology graph attention model. The model outputs the probability of each multimodal detection unit belonging to the quality category of qualified, appearance defect, geometric deformation defect, and material abnormality defect. The classification loss is calculated based on the difference between the quality category probability and the corresponding quality category label. When the number of samples in each quality category is unbalanced, the class weights are set according to the number of samples in each category, and the weighted classification loss is calculated. Then, the trainable parameters of the node feature mapping layer, differential topology attention layer, neighborhood weighted fusion layer, and quality classification layer in the model are updated through backpropagation and gradient optimization until the classification loss on the verification samples is stable or the preset training rounds are reached, resulting in the trained differential topology graph attention model. In the online detection stage, the multimodal feature vector and circumferential unfolded adjacency graph corresponding to the current rubber hose are input into the trained differential topology graph attention model. Only forward inference is performed, and the quality category probability and quality judgment result of each multimodal detection unit are output. No online update of model parameters is performed.

[0085] S5.4: Based on the quality judgment results of each multimodal detection unit, merge adjacent abnormal detection units into segments to be judged, and output the qualified, re-inspected, or rejected results of the corresponding segments.

[0086] Specifically, the axial position, circumferential angle, and quality category of each multimodal detection unit are read; multimodal detection units with quality categories of appearance defects, geometric deformation defects, or material anomalies are identified as anomaly detection units, and based on the circumferential expansion of the adjacency graph, spatially adjacent anomaly detection units with the same quality category or belonging to the same processing level are grouped into the same segment to be judged; for each segment to be judged, the axial range of the segment is determined according to the axial position of each anomaly detection unit in the segment to be judged, and the circumferential range of the segment is determined according to the circumferential angle of each anomaly detection unit in the segment to be judged; if the segment to be judged contains geometric deformation defects or material anomalies, the segment to be judged is output as a rejection segment; if the segment to be judged only contains appearance defects, the segment to be judged is output as a re-inspection segment; rubber hose areas not included in the segment to be judged are output as qualified segments. After obtaining the qualified segments, re-inspected segments, and rejected segments, the axial range, circumferential range, quality category, and online inspection results of each segment are output as the online inspection results of the rubber hose. For re-inspected or rejected segments, the corresponding action time is determined based on the distance between the axial position of the segment and the marking or rejection station, combined with the current traction speed. At the corresponding action time, the re-inspected segment is marked or the rejected segment is separated.

[0087] It should be noted that the segmented axial range is determined by the minimum and maximum axial positions of the anomaly detection unit, and the segmented circumferential range is determined according to the circumferential periodic relationship in the circumferential unfolding detection coordinate system.

[0088] This embodiment also provides an online quality inspection system for rubber hoses, including: The coordinate system establishment module is used to establish a circumferential unfolding detection coordinate system based on the axial displacement and circumferential sampling angle of the rubber hose; The optical data acquisition and mapping module is used to simultaneously acquire visible light images, laser contour data and near-infrared spectral data of the outer surface of the rubber hose, and generate a surface unfolding quality map of the rubber hose. The detection unit feature extraction module is used to divide the surface unfolded quality map of the rubber hose into multiple multimodal detection units and extract the multimodal feature vector of each multimodal detection unit. The adjacency graph construction module is used to build a circularly expanded adjacency graph based on the spatial adjacency relationship of each multimodal detection unit as a graph node and determine the spatial adjacency weight. The quality assessment output module is used to input the multimodal feature vector and the circumferentially expanded adjacency graph into the difference topology graph attention model, obtain the quality assessment result, and output the online detection result of the rubber hose.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for online quality inspection of rubber hoses, characterized in that, include: During the continuous traction and conveying process of the rubber hose, a circumferential unfolding detection coordinate system is established based on the axial displacement and circumferential sampling angle of the rubber hose; Based on the circumferential unfolding detection coordinate system, visible light images, laser contour data and near-infrared spectral data of the outer surface of the rubber hose are acquired simultaneously, and the visible light images, laser contour data and near-infrared spectral data are mapped to the corresponding circumferential unfolding positions to generate a surface unfolding quality map of the rubber hose. The surface unfolded quality map of the rubber hose is divided into multiple multimodal detection units, and the multimodal feature vectors of each multimodal detection unit are extracted. Using each multimodal detection unit as a node, a circular unfolded adjacency graph is established based on the spatial adjacency relationship of each multimodal detection unit in the circular unfolded detection coordinate system; The multimodal feature vectors and the circumferentially expanded adjacency graph are input into the differential topology graph attention model. Graph attention fusion is performed based on the local feature differences between adjacent multimodal detection units and the axial-circumferential topological relationship to obtain the quality judgment results of each multimodal detection unit. The online detection results of the rubber hose are then output based on the quality judgment results.

2. The online quality inspection method for rubber hoses as described in claim 1, characterized in that: The steps for establishing a circumferential unfolding detection coordinate system based on the axial displacement and circumferential sampling angle of the rubber hose are as follows: During the continuous traction and conveying of the rubber hose, the lateral swing of the rubber hose is restricted by the centering and guiding mechanism, and the central axis of the optical inspection station is used as the axial reference line, the starting section of the rubber hose entering the optical inspection station is used as the axial reference section, and the preset circumferential sampling direction is used as the circumferential zero angle direction. The axial position of each sampling point is determined based on the axial displacement of the rubber hose obtained by the traction encoder, and the circumferential angle of each sampling point is determined based on the angle of each circumferential sampling position relative to the circumferential zero angle direction. The detection points on the outer surface of the rubber hose are represented as circumferential unfolded detection coordinates composed of axial position and circumferential angle.

3. The online quality inspection method for rubber hoses as described in claim 2, characterized in that: The steps for synchronously acquiring visible light images, laser contour data, and near-infrared spectral data of the outer surface of the rubber hose are as follows: Based on the circumferentially unfolded detection coordinate system, a synchronous trigger signal is generated according to the equidistant axial sampling interval, so that the visible light acquisition unit, laser contour acquisition unit, and near-infrared spectral acquisition unit synchronously acquire visible light image data, laser contour data, and near-infrared spectral data of the outer surface of the rubber hose. Based on the circumferential unfolding detection coordinate system, visible light image data, laser contour data, and near-infrared spectral data under the same sampling sequence number are bound to the corresponding axial position and circumferential angle to form the original optical detection data sequence.

4. The online quality inspection method for rubber hoses as described in claim 3, characterized in that: The steps for generating the surface development quality diagram of the rubber hose are as follows: Based on the original optical detection data sequence and the circumferential unfolding detection coordinate system, a two-dimensional circumferential unfolding matrix is ​​established with the axial position as the row direction and the circumferential angle as the column direction. Visible light image data, laser contour data and near-infrared spectral data are mapped to the corresponding circumferential unfolding positions. Visible light texture layer, laser contour layer and near-infrared spectral layer are formed from the mapped visible light image data, laser contour data and near-infrared spectral data respectively. The data of each layer under the same circumferential unfolding position are combined to generate the surface unfolding quality map of the rubber hose.

5. The online quality inspection method for rubber hoses as described in claim 4, characterized in that: The steps for dividing the surface unfolded quality map of the rubber hose into multiple multimodal detection units and extracting the multimodal feature vector of each multimodal detection unit are as follows: Based on the axial position and circumferential angle in the surface unfolded quality map of the rubber hose, the surface unfolded quality map of the rubber hose is divided into multiple multimodal detection units, and the unfolded position set corresponding to each multimodal detection unit is determined. Visible light texture features, laser contour features, and near-infrared spectral features are extracted from the visible light texture layer, laser contour layer, and near-infrared spectral layer corresponding to each multimodal detection unit, respectively. The visible light texture features, laser contour features, and near-infrared spectral features of the same multimodal detection unit are stitched together to obtain the multimodal feature vector of each multimodal detection unit.

6. The online quality inspection method for rubber hoses as described in claim 5, characterized in that: The steps for establishing a circular unfolded adjacency graph based on the spatial adjacency relationship of each multimodal detection unit in the circular unfolded detection coordinate system are as follows: Using each multimodal detection unit as a graph node, and taking the axial center position, circumferential central angle, and multimodal feature vector of each multimodal detection unit as node attributes, a graph edge is established based on the axial and circumferential adjacency relationships of each multimodal detection unit in the circumferentially unfolded detection coordinate system. Spatial adjacency weights are calculated based on the axial distance and circumferential arc length between adjacent graph nodes, and a circumferentially unfolded adjacency graph is generated based on graph nodes, graph edges, and spatial adjacency weights.

7. The online quality inspection method for rubber hoses as described in claim 6, characterized in that: The differential topology graph attention model is a pre-trained graph neural network model, which includes a node feature mapping layer, a differential topology attention layer, a neighborhood weighted fusion layer, and a quality classification layer connected in sequence. The node feature mapping layer is used to generate initial quality node features for each graph node. The difference topology attention layer is used to generate graph attention weights based on local feature differences and axial-circumferential topological relationships. The neighborhood weighted fusion layer is used to generate a fused quality representation vector. The quality classification layer is used to output the quality judgment results of each multimodal detection unit.

8. The online quality inspection method for rubber hoses as described in claim 7, characterized in that: The steps to obtain the quality judgment results of each multimodal detection unit are as follows: The multimodal feature vectors of each multimodal detection unit and the circumferentially expanded adjacency graph are input into the differential topology graph attention model, and the multimodal feature vectors are mapped to the initial quality node features of the corresponding graph nodes through the node feature mapping layer. The extended neighborhood set of each graph node is determined based on the circumferentially expanded adjacency graph, and the graph attention weight is calculated by the differential topological attention layer based on the local feature differences between adjacent graph nodes, the axial-circumferential topological relationship, and the spatial adjacency weight. The neighborhood weighted fusion layer uses graph attention weights to weight and fuse the initial quality node features of each graph node in the extended neighborhood set to obtain the fused quality representation vector of each graph node. The quality classification layer then outputs the quality judgment results of each multimodal detection unit based on the fused quality representation vector.

9. The online quality inspection method for rubber hoses as described in claim 8, characterized in that: The online detection result of the rubber hose based on the quality judgment result is to group adjacent abnormal detection units into segments to be judged according to the quality judgment result of each multimodal detection unit, and output the qualified, re-inspection or rejection result of the corresponding segment.

10. An online quality inspection system for rubber hoses, based on the online quality inspection method for rubber hoses according to any one of claims 1 to 9, characterized in that, include: The coordinate system establishment module is used to establish a circumferential unfolding detection coordinate system based on the axial displacement and circumferential sampling angle of the rubber hose; The optical data acquisition and mapping module is used to simultaneously acquire visible light images, laser contour data and near-infrared spectral data of the outer surface of the rubber hose, and generate a surface unfolding quality map of the rubber hose. The detection unit feature extraction module is used to divide the surface unfolded quality map of the rubber hose into multiple multimodal detection units and extract the multimodal feature vector of each multimodal detection unit. The adjacency graph construction module is used to build a circularly expanded adjacency graph based on the spatial adjacency relationship of each multimodal detection unit as a graph node and determine the spatial adjacency weight. The quality assessment output module is used to input the multimodal feature vector and the circumferentially expanded adjacency graph into the difference topology graph attention model, obtain the quality assessment result, and output the online detection result of the rubber hose.

Images