Search

CN-122023420-A - Image recognition-based method for detecting surface defects of engine cylinder liner

CN122023420ACN 122023420 ACN122023420 ACN 122023420ACN-122023420-A

Abstract

The invention belongs to the technical field of machine vision and quality detection of parts of an internal combustion engine, and particularly discloses an image recognition-based detection method for surface defects of a cylinder liner of the engine. The method comprises the steps of obtaining a cylinder liner inner wall light intensity image sequence through multi-angle polarization imaging, generating polarization degree and polarization angle images through Stokes vector decomposition, inputting the images as double channels, sending the images into a neural network containing a physical constraint layer for feature extraction and defect initial positioning, applying physical consistency constraint on feature response by combining an optical reflection model, reconstructing three-dimensional morphology of defects based on polarization information, outputting depth, width and distribution parameters, and finally judging fatal defects according to a preset threshold value and generating classification and visual reports. According to the invention, three-dimensional quantitative detection of the surface defects with high precision and high robustness can be realized without a three-dimensional scanning device, and the recognition accuracy of micro cracks and hidden holes is improved.

Inventors

  • ZENG YI

Assignees

  • 重庆秉宪机械制造有限公司

Dates

Publication Date
20260512
Application Date
20260414

Claims (9)

  1. 1. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition is characterized by comprising the following steps of: Carry out multi-angle polarized light irradiation to engine cylinder liner inner wall surface through polarization image device to gather the light intensity image sequence under a plurality of polarization directions in step, include: Configuring an annular polarized light source array, wherein the annular polarized light source array comprises at least four light emitting units with different polarization directions, and the light emitting units are physically distributed according to polarization angle directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees; a linear polaroid is arranged at the front end of each light-emitting unit, and the light-emitting units are uniformly distributed around the axial direction of the cylinder sleeve at preset angle intervals; The rotating device is used for driving the cylinder sleeve of the engine to be in a uniform rotation state, and a trigger instruction is sent to the annular polarized light source array through the synchronous control module, so that the light-emitting units are driven to flash in turn according to a preset time step; the synchronous control module sends synchronous pulse signals to the polarization camera while the light emitting unit flashes, and the polarization camera collects surface reflection images in corresponding polarization states; the polarization camera adopts a global shutter photosensitive element, and sets single-frame exposure time according to the rotation speed of the engine cylinder sleeve, so as to acquire an original light intensity image sequence containing four polarization dimension information; calculating a polarization degree image and a polarization angle image corresponding to each pixel point by using a Stokes vector decomposition algorithm based on the light intensity image sequence; Taking the polarization degree image and the polarization angle image as two-channel input data, and sending the two-channel input data into a physical driving neural network for feature extraction and defect preliminary positioning; embedding a physical constraint layer in the physical driving neural network, and applying physical consistency constraint to the characteristic response of the network intermediate layer by utilizing the physical constraint layer according to the optical model of specular reflection and diffuse reflection; reconstructing three-dimensional morphology information of a defect area based on the feature map subjected to physical constraint, and outputting depth parameters, width parameters and spatial distribution parameters of the defect; And judging the types of the defects according to the three-dimensional morphology information by combining with a preset threshold value, and generating a defect classification result and a visual report.
  2. 2. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition according to claim 1, wherein the calculation of the polarization degree image and the polarization angle image corresponding to each pixel point by using a stokes vector decomposition algorithm specifically comprises: Defining a light intensity value of each pixel point in the light intensity image sequence under a polarization angle of 0 degrees, a light intensity value under a polarization angle of 45 degrees, a light intensity value under a polarization angle of 90 degrees and a light intensity value under a polarization angle of 135 degrees; Summing the light intensity value under the 0-degree polarization angle and the light intensity value under the 90-degree polarization angle to obtain a zeroth Stokes component; Subtracting the light intensity value under the 90-degree polarization angle from the light intensity value under the 0-degree polarization angle to obtain a first Stokes component; subtracting the light intensity value under the 135-degree polarization angle from the light intensity value under the 45-degree polarization angle to obtain a second Stokes component; Determining the mode length of a linear polarization vector by utilizing the first Stokes component and the second Stokes component, and calculating the polarization degree by combining the zeroth Stokes component, wherein the polarization degree represents the polarization characteristic intensity of a local area; And processing the ratio of the second Stokes component to the first Stokes component by using an arctangent function, and calculating to obtain a polarization angle, wherein the polarization angle represents the space orientation of the polarization main direction.
  3. 3. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition according to claim 2, wherein the process of calculating the degree of polarization and the polarization angle specifically comprises the following steps: the calculation formula of the module length is as follows: ; The mode length of the linear polarization vector is represented, Representing a first stokes component of the sample, Representing a second stokes component; The calculation formula of the polarization degree is as follows: ; The degree of polarization is indicated as being, Representing a zeroth stokes component; the calculation formula of the polarization angle is as follows: ; Representing the polarization angle; the polarization degree image reflects the material characteristics and roughness changes of the surface of the engine cylinder liner, and the polarization angle image characterizes the microscopic geometric outline of the surface of the engine cylinder liner.
  4. 4. The method for detecting surface defects of cylinder liners of engine based on image recognition according to claim 3, wherein the physical driving neural network adopts an encoder and decoder architecture, and specifically comprises: the encoder is composed of a plurality of stages of convolution modules, the first stage of convolution module performs feature extraction on the double-channel input data by utilizing convolution check, and a nonlinear activation layer is connected behind each stage of convolution module; The encoder is internally provided with a maximum pooling layer for downsampling operation, and the spatial resolution of a feature map is reduced and the number of channels is increased after pooling once, so that local microscopic defect details and global macroscopic geometric distribution are captured; the decoder recovers spatial resolution by a transposed convolution or bilinear interpolation upsampling operation; And establishing jump connection between the coding layer of the coder and the corresponding layer of the decoder, fusing multi-scale characteristic information of the coding stage into a corresponding characteristic diagram of the decoding stage, and outputting a probability diagram representing the confidence that each pixel belongs to a defect area.
  5. 5. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition according to claim 4, wherein a physical constraint layer is embedded in the physical driving neural network, and physical consistency constraint is applied to characteristic response of a network intermediate layer by using the physical constraint layer according to an optical model of specular reflection and diffuse reflection, and the method specifically comprises the following steps: a light reflection physical model for distinguishing specular reflection components and diffuse reflection components is built in the physical constraint layer; obtaining geometrical parameters of the current pixel point through the physical constraint layer, wherein the geometrical parameters comprise geometrical relations among a surface normal direction, an incident light direction and an observation direction; Calculating the theoretical polarization degree of each pixel point under the current normal prediction according to the principle that the polarization state of reflected light accords with the Fresnel reflection law; Comparing the theoretical polarization degree with the polarization degree characteristics actually extracted by the physical driving neural network; When the polarization degree characteristic violates an optical reflection rule, the physical constraint layer generates a physical consistency error item, and the physical consistency error item is fed back to the weight updating process of the physical driving neural network as regularization constraint.
  6. 6. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition according to claim 5, wherein the reconstructing of the three-dimensional morphological information of the defective region specifically comprises: converting the polarization angle image into a gradient vector field of the surface by utilizing the mapping relation between the polarization angle and the normal direction of the surface, wherein the gradient vector field comprises a horizontal direction gradient and a vertical direction gradient; performing divergent operation on the gradient vector field to obtain a Laplacian response; performing discrete cosine transform on the Laplace response, and processing the Laplace response by using an integral filter in a frequency domain; performing inverse discrete cosine transform on the data processed in the frequency domain, and returning to the space domain to obtain a height map representing microscopic undulations of the surface; And determining the edge and bottom morphology of the defect by combining local curvature analysis, and calculating the maximum depth value, average depth value, surface opening width and depth-to-width ratio parameters of the cross section of the defect.
  7. 7. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition according to claim 6, wherein the judging of the types of the defects specifically comprises: setting a preset threshold system comprising a depth threshold, a width threshold, an aspect ratio threshold and an area threshold, comparing the depth of the defect obtained by reconstruction with a preset distance, and identifying potential risk points; When the depth-to-width ratio of the defect is in a preset range and the depth exceeds a preset critical value, judging the defect as a fatal defect and triggering an alarm mechanism; dividing the defects into scratches, holes, cracks or compound defects by utilizing a multi-class classifier, and outputting position coordinates, geometric dimensions and risk grades aiming at each class of defects; the visual report is presented in a pseudo-color depth graph mode, areas with different depth values correspond to different color schemes, and the pseudo-color depth graph and the original polarization degree image are displayed in a superposition fusion mode.
  8. 8. The method for detecting the surface defects of the cylinder liner of the engine based on image recognition according to claim 7, wherein the training process of the physical driving neural network specifically comprises the following steps: Introducing a polarized image sample marked with real depth information, wherein the polarized image sample covers a defect form sample and a honing line sample; optimizing the physical driving neural network by adopting a joint loss function, wherein the joint loss function comprises a pixel-level reconstruction error term and a physical consistency regular term generated by the physical constraint layer; Performing sample enhancement by using a generating countermeasure network, and generating virtual defect samples with different geometric dimensions, depth-to-width ratios and illumination conditions based on physical parameters extracted from a physical driving network; and after the virtual defect sample is subjected to rationality verification through the physical constraint layer, adding the virtual defect sample into a training set to participate in iterative optimization, and establishing a nonlinear mapping relation between polarization characteristics and three-dimensional morphology.
  9. 9. The image recognition-based engine cylinder liner surface defect detection method according to claim 8, wherein when a lubricating oil film exists on the surface of the engine cylinder liner, the detection process further comprises: Configuring narrowband light sources with different wavelengths in the polarization imaging device, and independently carrying out Stokes vector decomposition aiming at each wavelength channel to obtain a multispectral polarization characteristic set; Calculating the polarization degree ratio under different wavelength channels, and determining a polarization deviation correction factor according to the polarization degree ratio, wherein the polarization deviation correction factor is used for compensating polarization distortion caused by oil film surface reflection; introducing a multilayer physical model containing a film interference effect into the physical constraint layer, and establishing a mapping relation among the thickness of an oil film, the incident angle of light and the reflective polarization state; Calculating the equivalent oil film thickness of each pixel point according to the mapping relation, subtracting oil film reflection components from the total reflection signals, and extracting substrate defect characteristics; And solving the surface height field by using redundancy constraint generated by multispectral polarization information through a variation energy functional minimization method.

Description

Image recognition-based method for detecting surface defects of engine cylinder liner Technical Field The invention belongs to the technical field of machine vision and quality detection of parts of an internal combustion engine, and particularly relates to an image recognition-based detection method for surface defects of an engine cylinder liner. Background With the continuous progress of precision manufacturing technology, the surface quality of an engine cylinder sleeve serving as a core power component directly influences the sealing performance, friction loss and service life of an engine. In the field of industrial vision detection, real-time monitoring and defect identification are performed on the inner wall of a cylinder sleeve by utilizing an automation technology, so that the method has become a key link for guaranteeing the efficiency of a production line and the consistency of products. By combining optical imaging and a computer vision algorithm, various surface flaws such as scratches, holes, cracks and the like generated in the cylinder liner during processing or service can be identified, so that the overall reliability of the internal combustion engine equipment is improved. The defect detection scheme based on image processing is widely applied to the quality control of the cylinder sleeve by virtue of the advantages of non-contact, high precision and the like. The technology generally utilizes an industrial camera to collect characteristic information of the inner surface of the cylinder liner, and combines deep learning or a traditional characteristic extraction algorithm to segment, position and classify the collected image. Aiming at the high requirements of complex honing reticulate patterns and assembly precision of the inner wall of the cylinder sleeve, the system needs to have strong anti-interference capability so as to ensure that the real geometric shape deviation can be accurately extracted and distinguished from complex texture background. In the prior art, when the defects on the surface of the cylinder liner are treated, specular reflection interference is often faced, and a lubricating oil film remained on the surface or strong light spots generated by a metal texture can cause local saturation of an image, so that defect characteristic masking is caused or false alarm false detection is generated. The conventional two-dimensional imaging mode lacks the perception capability of depth information, and is difficult to accurately judge the depth-to-width ratio of scratches or holes through the change of pixel intensity, so that the recognition accuracy of fatal defects such as deep cracks and the like is limited. The conventional convolutional neural network lacks physical and optical constraint in the feature extraction process, cannot effectively strip feature coupling of honing lines and tiny flaws, has obvious robustness deficiency when dealing with nonlinear light and shadow changes and tiny three-dimensional morphology reconstruction, and is difficult to meet the requirement of the high-precision tip manufacturing field on defect quantitative analysis. Disclosure of Invention The invention aims to provide an image recognition-based detection method for detecting surface defects of an engine cylinder liner, which solves the problems in the background art. In order to achieve the above object, the present invention provides a method for detecting surface defects of a cylinder liner of an engine based on image recognition, comprising the steps of: Performing multi-angle polarized light irradiation on the inner wall surface of the engine cylinder liner by using a polarized imaging device, and synchronously acquiring light intensity image sequences in a plurality of polarization directions; calculating a polarization degree image and a polarization angle image corresponding to each pixel point by using a Stokes vector decomposition algorithm based on the light intensity image sequence; Taking the polarization degree image and the polarization angle image as two-channel input data, and sending the two-channel input data into a physical driving neural network for feature extraction and defect preliminary positioning; embedding a physical constraint layer in the physical driving neural network, and applying physical consistency constraint to the characteristic response of the network intermediate layer by utilizing the physical constraint layer according to the optical model of specular reflection and diffuse reflection; reconstructing three-dimensional morphology information of a defect area based on the feature map subjected to physical constraint, and outputting depth parameters, width parameters and spatial distribution parameters of the defect; And judging the types of the defects according to the three-dimensional morphology information by combining with a preset threshold value, and generating a defect classification result and a visual report. Preferably, the multi-ang