CN-122016653-A - Online visual inspection and defect positioning method for flexible hose composite material production line

CN122016653ACN 122016653 ACN122016653 ACN 122016653ACN-122016653-A

Abstract

The invention relates to the technical field of image processing, in particular to an online visual detection and defect positioning method of a flexible hose composite material production line, which comprises the following steps that firstly, a linear array hyperspectral camera and a production line synchronous encoder are arranged on the flexible hose composite material production line along the conveying direction, and a geometric space coordinate system of the flexible hose composite material is established; step two, constructing a cylindrical unfolding hyperspectral double-branch depth self-encoder model, merging adjacent spectrum anomaly super-pixel areas to obtain a spectrum anomaly candidate area set, and step three, mapping the spectrum anomaly candidate area to the actual surface of the flexible hose composite material to generate an initial geometric candidate defect area set, and extracting position parameters of a stable defect area as a defect positioning result. The invention can effectively identify the deep material difference which is difficult to capture by the traditional two-dimensional vision, and improves the positioning accuracy and the boundary stability of the defect region through geometric space mapping and morphological processing.

Inventors

Wei chuang
WEI ZHIXIANG
MAO XIANG
XIE JINGFU
YANG FU
GONG YONG

Assignees

湖南天卓管业有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The on-line visual detection and defect positioning method for the flexible hose composite material production line is characterized by comprising the following steps of: Arranging a linear array hyperspectral camera and a production line synchronous encoder on a flexible hose composite material production line along a conveying direction, acquiring multichannel spectral linear array images in an axial direction by the linear array hyperspectral camera under the triggering control of the production line synchronous encoder, and performing cylindrical unfolding resampling by an industrial control calculation unit according to an axial position pulse signal provided by the production line synchronous encoder, pre-calibrated outer diameter information of the flexible hose composite material, an installation geometric relationship between an optical axis of the linear array hyperspectral camera and a central axis of the flexible hose composite material and an axial displacement interval between adjacent triggering moments of the linear array hyperspectral camera, combining the spectral linear arrays into a cylindrical unfolding hyperspectral image sequence, arranging the cylindrical unfolding hyperspectral image sequence into a spectral space data cube, and establishing a geometric space coordinate system of the flexible hose composite material; step two, constructing a cylindrical unfolding hyperspectral double-branch depth self-encoder model, performing spectrum normalization processing and super-pixel region division on a cylindrical unfolding hyperspectral image, inputting an average spectrum sequence of a super-pixel region into the cylindrical unfolding hyperspectral double-branch depth self-encoder model to obtain a reconstructed spectrum sequence, calculating a reconstruction difference index, determining an online reconstruction difference segmentation threshold according to the frequency distribution of the reconstruction difference index, marking the super-pixel region with the reconstruction difference index not lower than the online reconstruction difference segmentation threshold as a spectrum anomaly super-pixel region, and merging adjacent spectrum anomaly super-pixel regions to obtain a spectrum anomaly candidate region set; Mapping the spectrum abnormal candidate region to the actual surface of the flexible hose composite material to generate an initial geometric candidate defect region set, performing image segmentation and morphological processing on the initial geometric candidate defect region to obtain a stable defect region mask, and extracting position parameters of the stable defect region as a defect positioning result.
2. The method according to claim 1, wherein in the step one, the industrial control computing unit combines the spectral linear arrays around the flexible hose composite material for a whole circumference into a cylindrical unfolding hyperspectral image sequence which is planar in the axial direction and the circumferential unfolding direction, the cylindrical unfolding hyperspectral image sequence is arranged into a spectral space data cube in the industrial control computing unit in the spectral channel direction, the axial direction and the circumferential unfolding direction, and a geometric space coordinate system of the flexible hose composite material is established by the axial position pulse accumulation result of the production line synchronous encoder and the line and column index of the cylindrical unfolding hyperspectral image, so that the axial position, the circumferential position and the actual position of each pixel in the spectral space data cube on the flexible hose composite material have a one-to-one correspondence.
3. The method of claim 1, wherein the super-pixel region division in the second step adopts a region growing mode, specifically comprising using a plurality of interval pixels as seed pixels, merging pixels adjacent to the seed pixels and having reflection intensity differences within a preset similarity range on all spectrum channels into the same super-pixel region, and continuing to expand the adjacent pixels until the spectrum differences of the adjacent pixels exceed the preset similarity range or the super-pixel region reaches a preset area upper limit, so as to form a super-pixel region division result covering the whole cylindrical unfolded hyperspectral image.
4. The method of claim 1, wherein the cylindrical unfolding hyperspectral double-branch depth self-encoder model in the second step comprises two encoding and decoding paths, namely a single-point spectrum branch and an axial neighborhood spectrum branch, wherein the single-point spectrum branch takes an average spectrum sequence of a single super-pixel area as an input, sequentially reduces the dimension of a spectrum channel layer by layer through a plurality of layers of fully-connected encoding layers to form a compressed spectrum feature vector, restores the compressed spectrum feature vector to a reconstructed spectrum sequence through a plurality of layers of fully-connected decoding layers symmetrical to the encoding layer structure, takes the average spectrum sequences of three super-pixel areas adjacent in the axial direction in a cylindrical unfolding coordinate system as the input, sequentially splices the three average spectrum sequences into an axial spectrum neighborhood block in an axial direction, performs sliding convolution on the axial spectrum neighborhood block in the axial direction through a one-dimensional convolution encoding layer to extract an axial change feature, compresses a convolution result to form an axial compressed spectrum feature vector through a fully-connected layer, and restores the axial compressed spectrum feature vector to an axial reconstructed spectrum feature sequence corresponding to three positions through a one-dimensional deconvolution decoding layer and a fully-connected decoding layer combination.
5. The method according to claim 1 or 4, wherein in the second step, the industrial control computing unit inputs the average spectrum sequence of each super-pixel region in the training sample set into a single-point spectrum branch and an axial neighborhood spectrum branch respectively to obtain a single-point reconstructed spectrum sequence and an axial reconstructed spectrum sequence, averages the absolute values of channel differences between the input average spectrum sequence and the single-point reconstructed spectrum sequence and the absolute values of channel differences between the reconstructed spectrum sequences at corresponding positions in the input average spectrum sequence and the axial reconstructed spectrum sequence, calculates a single-point reconstructed difference index and an axial reconstructed difference index, uses the sum of the two reconstructed difference indexes as a total reconstructed difference index of the training sample, and adjusts network parameters of the cylindrical unfolding hyperspectral double-branch depth self-encoder model in a gradient descent mode to gradually reduce the total reconstructed difference index of the training sample until the variation amplitude is in a stable range in a preset training round.
6. The method according to claim 1, wherein in the second step, the industrial control computing unit generates an online cylindrical expansion hyperspectral image for each spectrum space data cube in a cylindrical expansion mode when the production line runs formally, generates an online superpixel area set by adopting the same spectrum normalization preprocessing and superpixel area construction mode as the training stage, computes an average spectrum sequence for each superpixel area in the online superpixel area set, inputs the average spectrum sequence into a trained cylindrical expansion hyperspectral double-branch depth self-encoder model, obtains a corresponding single-point reconstruction spectrum sequence and an axial reconstruction spectrum sequence, computes a single-point reconstruction difference index and an axial reconstruction difference index of each superpixel area, adds the single-point reconstruction difference index and the axial reconstruction difference index to form an online total reconstruction difference index, and fills all online total reconstruction difference indexes into the online reconstruction difference index map according to axial and circumferential positions of the superpixel areas in a cylindrical expansion coordinate system.
7. The method according to claim 6, wherein the industrial control computing unit reads all online total reconstruction difference indexes in the online reconstruction difference index map, counts frequency distribution of the online total reconstruction difference indexes, selects a continuous online total reconstruction difference index section with highest frequency as a normal material reconstruction difference section, selects an online reconstruction difference segmentation threshold in a section of online total reconstruction difference index section outside an upper boundary of the normal material reconstruction difference section, marks a super-pixel region with the online total reconstruction difference index not lower than the online reconstruction difference segmentation threshold as a spectrum anomaly super-pixel region, obtains a spectrum anomaly super-pixel region set, and merges adjacent spectrum anomaly super-pixel regions in a cylindrical expansion coordinate system according to communication conditions of the spectrum anomaly super-pixel regions in an axial direction and a circumferential expansion direction, so as to obtain a spectrum anomaly candidate region set.
8. The method according to claim 1, wherein in the second step, the industrial control computing unit calls a cylindrical expansion hyperspectral double-branch depth self-encoder training module in the equipment installation and debugging stage, a plurality of spectrum space data cubes are collected from the flexible hose composite material sample which is confirmed to be qualified by manpower, a cylindrical expansion hyperspectral image sequence for training is generated in the same cylindrical expansion mode as in the first step, and dark current correction and reflection intensity normalization processing are performed on each spectral channel in the industrial control computing unit according to the spectral channel sequence for each cylindrical expansion hyperspectral image of the cylindrical expansion hyperspectral image sequence for training, so that a spectrum normalized cylindrical expansion hyperspectral image is obtained.
9. The method according to claim 1, wherein in the third step, the industrial control computing unit maps the axial position index and the circumferential position index of each spectral anomaly candidate region in the spectral anomaly candidate region set to the actual surface of the flexible hose composite material through the flexible hose composite material geometric space coordinate system established in the first step, generates a corresponding initial geometric candidate defect region set, selects a brightness channel and a preset critical spectral channel from the original cylindrical unfolded hyperspectral image for each initial geometric candidate defect region to form a two-dimensional gray scale map, and the industrial control computing unit performs fixed threshold segmentation on the two-dimensional gray scale map to obtain the initial binary defect mask.
10. The method according to claim 9, wherein the industrial control computing unit sequentially performs morphological expansion operation and morphological corrosion operation on the initial binary defect mask to eliminate isolated fine noise points and to make boundaries of real defect areas coherent, performs morphological opening operation and smooth boundary processing on the expanded and corroded binary defect mask to obtain stable defect area masks, corresponds the stable defect area masks to a flexible hose composite geometrical space coordinate system, extracts a start position and an end position of each stable defect area along an axial direction, a start position and an end position along a circumferential direction and corresponding actuator trigger positions in a conveying direction of a production line, and sends the position parameters and profile information of the stable defect areas in the cylindrical expansion coordinate system to the production line actuator as defect positioning results for online marking, rejecting or grading of defect sections of the flexible hose composite.

Description

Online visual inspection and defect positioning method for flexible hose composite material production line Technical Field The invention belongs to the technical field of image processing, and particularly relates to an online visual detection and defect positioning method for a flexible hose composite material production line. Background Flexible hose composites are widely used in the fields of automobiles, engineering machinery, petrochemical transportation, industrial equipment, and the like. Along with the continuous improvement of the requirements of the working environment on the pressure grade, the corrosion resistance and the flexibility, the manufacturing process of the flexible hose composite material continuously evolves, and the states of the inner reinforcing layer, the outer rubber layer and the bonding interface layer have decisive influence on the product performance. However, the surface and internal defects of the flexible hose composite material often have the characteristics of small size, various types, discrete distribution, irregular shape and the like, and the stability is difficult to ensure due to the fact that the production line speed is high, so that the dependence on an automatic visual detection technology has become an industry trend. In the prior art, the mainstream method for detecting the flexible hose composite material generally adopts a detection mode including a two-dimensional camera, a line laser profilometer, an infrared camera or a traditional visible light spectrometer. These detection methods can identify defects including bubbles, breakage, adhesion of foreign matter, welding unevenness, or thickness abnormality of the overcoat layer based on brightness differences, surface texture changes, or simple spectral reflectance characteristics. However, the light reflection characteristics of the flexible hose composite surface vary greatly with color, surface roughness, illumination direction, and process variations, so that the accuracy of detection based on two-dimensional images is significantly limited. For example, in a visible light image, local texture fluctuations of normal materials are similar to slight defects in gradation or color, making it difficult for the system to achieve stable segmentation judgment. In addition, the two-dimensional vision system cannot acquire spectral dimension information of the material and cannot reveal deep features of material difference, so that defect recognition effects on uneven transparent coating, impurity mixing, local aging and the like which are difficult to embody through color or brightness contrast are weak. Disclosure of Invention The invention mainly aims to provide an on-line visual detection and defect positioning method for a flexible hose composite material production line. In order to solve the problems, the technical scheme of the invention is realized as follows: the on-line visual detection and defect positioning method for the flexible hose composite material production line comprises the following steps: Arranging a linear array hyperspectral camera and a production line synchronous encoder on a flexible hose composite material production line along a conveying direction, acquiring multichannel spectral linear array images in an axial direction by the linear array hyperspectral camera under the triggering control of the production line synchronous encoder, and performing cylindrical unfolding resampling by an industrial control calculation unit according to an axial position pulse signal provided by the production line synchronous encoder, pre-calibrated outer diameter information of the flexible hose composite material, an installation geometric relationship between an optical axis of the linear array hyperspectral camera and a central axis of the flexible hose composite material and an axial displacement interval between adjacent triggering moments of the linear array hyperspectral camera, combining the spectral linear arrays into a cylindrical unfolding hyperspectral image sequence, arranging the cylindrical unfolding hyperspectral image sequence into a spectral space data cube, and establishing a geometric space coordinate system of the flexible hose composite material; step two, constructing a cylindrical unfolding hyperspectral double-branch depth self-encoder model, performing spectrum normalization processing and super-pixel region division on a cylindrical unfolding hyperspectral image, inputting an average spectrum sequence of a super-pixel region into the cylindrical unfolding hyperspectral double-branch depth self-encoder model to obtain a reconstructed spectrum sequence, calculating a reconstruction difference index, determining an online reconstruction difference segmentation threshold according to the frequency distribution of the reconstruction difference index, marking the super-pixel region with the reconstruction difference index not lower than the online reconstruction difference se