Intelligent evaluation method for surface quality of scroll disc of new energy automobile air conditioner compressor

Abstract

This invention relates to the field of artificial intelligence technology and discloses an intelligent assessment method for the surface quality of a scroll plate in a new energy vehicle air conditioning compressor. The method includes: acquiring surface images of the scroll plate through multi-angle, multi-spectral imaging; stripping background textures using a physical prior-guided texture decoupling network to generate a defect residual feature map; extracting dual-channel features of macroscopic morphology and microstructure using self-supervised contrastive learning, and determining defect confidence through dynamic weighted fusion; and finally, achieving quantitative grading of defects based on morphological indicators. This invention eliminates the need for manual annotation, effectively suppresses interference from processing textures, and improves the recognition accuracy and production line inspection efficiency of micron-level scratches, pits, and corrosion spots.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an intelligent assessment method for the surface quality of the scroll plate of an air conditioning compressor in a new energy vehicle. Background Technology

[0002] With the continuous rise in the production and sales of new energy vehicles, the intelligent manufacturing and quality control of their core components have become crucial to ensuring the performance and reliability of the entire vehicle. As a vital subsystem affecting driving comfort, the air conditioning system's core actuator—the scroll compressor—requires extremely high manufacturing precision. The scroll plate, as the core structural component of the compression chamber, directly determines sealing performance, friction loss, and overall machine lifespan based on its surface quality. During high-precision machining, defects such as microcracks, scratches, pits, or material peeling may occur on the surface. These defects are often embedded in a complex periodic texture background formed by precision grinding or milling processes, resulting in a high degree of similarity between defective areas and normal textures in grayscale, edge response, and local spectral characteristics.

[0003] Machine vision-based automated surface inspection technology has become a mainstream direction in industrial quality inspection. Traditional methods often employ classic image segmentation algorithms such as threshold segmentation, edge detection, or region growing, relying on manually designed features or fixed criteria for defect identification. However, when faced with complex scenarios involving vortex disk surfaces that combine high-frequency textures, non-uniform illumination reflection, and weak defect signals, these methods are prone to misclassifying normal textures as defects or obscuring real defects in background noise, resulting in high false negative and false positive rates. Even with the introduction of deep learning models such as convolutional neural networks, their performance remains limited by scarce labeled data, insufficient generalization ability, and sensitivity to subtle structural differences, making it difficult to consistently achieve industrial-grade inspection accuracy requirements in actual production lines.

[0004] Existing technologies generally suffer from bottlenecks when dealing with the segmentation of minute defects against highly similar texture backgrounds, including weak model generalization ability, local optima in the optimization objective, and blurred segmentation boundaries. Especially when the defects are irregular in shape, small in scale, and share similar statistical characteristics with the background texture, classical optimization algorithms struggle to construct an effective global energy function to distinguish the foreground from the background. This necessitates a novel intelligent evaluation method for high-speed online inspection of new energy vehicle vortex disks. This method should integrate prior knowledge, possess global optimum search capabilities, and efficiently solve complex segmentation problems to overcome the performance limitations of traditional image segmentation in precision manufacturing surface quality inspection. Summary of the Invention

[0005] This invention provides an intelligent assessment method for the surface quality of the scroll disk of an air conditioning compressor in new energy vehicles. It acquires multi-angle and multi-spectral image data of the scroll disk surface through a high-resolution imaging system, and combines a texture decoupling network based on physical priors to explicitly model and strip complex background textures, generating residual feature maps that retain only potential defect information. Furthermore, it utilizes a self-supervised contrastive learning mechanism under local geometric invariance constraints to construct a defect-sensitive feature space under unlabeled sample conditions, and achieves high-precision identification and quantitative classification of defects such as micron-level scratches, pits, and corrosion spots through a dual-channel discriminator that integrates macroscopic morphological statistical features and microscopic structural anomaly responses.

[0006] This invention provides an intelligent assessment method for the surface quality of the scroll plate of an air conditioning compressor in a new energy vehicle, comprising: By using an industrial-grade high-resolution linear array camera in conjunction with a ring polarization light source and an oblique incident structured light illumination device, synchronous image acquisition is performed on the surface of the vortex disk from at least three different azimuth angles to obtain a multimodal raw image sequence containing information on surface reflection intensity, phase gradient, and polarization difference. Pixel-level registration and radiometric consistency correction are performed on the multimodal original image sequence to generate a fused image dataset in a unified coordinate system; The fused image dataset is input into a pre-trained texture decoupling neural network model. This texture decoupling neural network model adopts an encoder-decoder architecture. The encoder consists of five levels of convolutional residual blocks. The output of each level is connected to the corresponding decoder level. A physical prior constraint module based on the manufacturing process parameters of the vortex disk is introduced in the bottleneck layer to suppress periodic texture interference caused by machining marks and material grain orientation, thereby outputting a residual feature map representing the real defect area. Based on the residual feature map, the local curvature change rate, gray-level co-occurrence matrix energy value and edge direction histogram distribution are extracted as macroscopic morphological statistical features. At the same time, the energy concentration and directional entropy of high-frequency wavelet coefficients are calculated on the residual feature map region by region through the sliding window mechanism to form microstructural anomaly response features. The macroscopic morphological statistical features and microstructural anomaly response features are respectively input into two independent fully connected discriminant branches, and finally the confidence score of each region belonging to the defect category is output. The confidence scores of the two discrimination branches are weighted and fused. The weight coefficients are dynamically adjusted according to the overall noise level of the current batch of images. The fusion result is thresholded and segmented to generate a binary defect mask. The defects are classified according to the mask area, perimeter ratio and morphological compactness index, and the final quality assessment report is output.

[0007] Preferably, pixel-level registration and radiometric consistency correction are performed on the multimodal original image sequence to generate a fused image dataset in a unified coordinate system, including: A non-rigid registration algorithm based on maximizing mutual information is used to perform pixel-level alignment of the three-view images. The grid spacing of the registration control points is set to eight pixels, and cubic smooth interpolation is used as the interpolation method. The original images of each channel are subjected to pixel-by-pixel gain compensation and offset correction using the standard whiteboard reflection response curve to generate a radiometrically consistent image. The corrected images are stitched together into a nine-channel fused image dataset according to viewpoint and information type. The channel order is as follows: viewpoint 1 bright field, viewpoint 1 phase gradient x component, viewpoint 1 phase gradient y component, viewpoint 1 polarization difference, viewpoint 2 bright field, viewpoint 2 phase gradient x component, viewpoint 2 phase gradient y component, viewpoint 2 polarization difference, and viewpoint 3 bright field.

[0008] Preferably, inputting the fused image dataset into a pre-trained texture decoupling neural network model includes: The decoder upsamples step by step through transposed convolution, and then performs channel concatenation with the feature maps of the corresponding encoder level before fusing them through two 3x3 convolutional layers. The bottleneck layer receives parameters such as spindle speed, feed rate, and tool path curvature radius, maps them into a two-dimensional frequency domain template, and applies a band-stop filter operation to the feature tensor in the Fourier transform domain to filter out response components in a specific frequency range related to the machining texture. This frequency range is calculated and determined by machining parameters such as feed rate, tool diameter, and spindle speed.

[0009] Preferably, the extraction of macroscopic morphological statistical features based on the residual feature map includes: Gaussian smoothing is applied to the residual feature map; Calculate the absolute value of the largest eigenvalue of the second-order partial derivative matrix at each pixel location, and use it as the local curvature change rate; Calculate the energy value of the gray-level co-occurrence matrix and take the mean value; After detecting edge points using the Canny operator, the frequency percentage of edge normal vector angles falling within multiple equal-width intervals is statistically analyzed to form an edge direction histogram distribution.

[0010] Preferably, the energy concentration and directional entropy of high-frequency wavelet coefficients are calculated region by region on the residual feature map using a sliding window mechanism to form microstructural anomaly response features, including: Traverse the residual feature map; Perform three-level wavelet decomposition on each window image block and extract the wavelet coefficients of the three high-frequency sub-bands in the third level (corresponding to the horizontal-vertical, vertical-horizontal, and diagonal directions, respectively); The energy concentration of high-frequency wavelet coefficients is calculated, and this index is obtained by statistically analyzing the energy proportion of all wavelet coefficients. The directional entropy is calculated by statistically analyzing the proportion of energy in the three directional sub-bands (horizontal, vertical, and diagonal) to the total high-frequency energy.

[0011] Preferably, the macroscopic morphological statistical features and microscopic structural anomaly response features are respectively input into two independent fully connected discriminant branches, including: The macro-discrimination branch receives an eighteen-dimensional feature vector and outputs a one-dimensional defect confidence score. The micro-discrimination branch receives two-dimensional feature vectors and outputs a one-dimensional defect confidence score. The network weights of the two branches are optimized through a self-supervised contrastive learning mechanism. The self-supervised contrastive loss function is used to bring the distance between different enhanced views of the same defect area closer and push the distance between the defect and the normal area further apart.

[0012] Preferably, the confidence scores of the two discriminant branches are weighted and fused, including: Calculate the standard deviation and mean of the current batch of fused image datasets; The fusion strategy is dynamically selected based on the ratio of the standard deviation to the mean: when the ratio is in a low range, a weighted fusion strategy dominated by macro features is adopted; when the ratio is in a high range, a weighted fusion strategy dominated by micro features is adopted, and finally a comprehensive confidence score is obtained (the confidence scores of both macro and micro discriminant branches are taken into account during fusion).

[0013] Preferably, the fusion result is used to generate a binary defect mask after threshold segmentation, including: An adaptive threshold segmentation method is used to binarize the fused confidence map; The binarized mask is subjected to opening and closing operations in sequence, and the structuring element is a circular kernel with a radius of two pixels.

[0014] Preferably, defects are classified into grades based on mask area, perimeter ratio, and morphological compactness indicators, including: Calculate the area, perimeter, and morphological compactness of each connected region; A Level 1 defect is defined as a defect area that is small in size and has high morphological compactness. Secondary defects are defined as defects with a moderate area or moderate morphological compactness. Level 3 defects are defined as defects with a large area or low morphological compactness.

[0015] Preferably, the ratio of high to low thresholds of the Canny operator is set to 2:1.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a texture decoupling mechanism that integrates multimodal optical imaging and physical manufacturing priors, the inherent processing texture of the vortex disk surface and the real defect signal are separated, solving the problem of high misjudgment rate caused by background texture interference in traditional image segmentation algorithms; 2. A self-supervised contrastive learning framework that does not require manual annotation is introduced, which reduces the dependence on large-scale labeled datasets and improves the generalization ability of the model in production environment. 3. By adopting a dual-channel discrimination strategy of macroscopic statistics and microscopic structure, the global morphological characteristics and local detail anomalies of defects are taken into account, and high-sensitivity identification and accurate quantitative classification of micron-level surface defects are achieved. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the texture decoupling network based on physical prior guidance in this invention; Figure 3 This is a logical flowchart of the multimodal image acquisition and preprocessing stage in this invention; Figure 4 This is a logical flowchart of the dual-channel discrimination mechanism of macroscopic morphological statistical features and microscopic structural anomaly response in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between residual feature map generation and defect-sensitive feature space construction in this invention; Figure 6 This is a logical flowchart of the defect quantification and grading and quality assessment report output stage in this invention. Detailed Implementation

[0018] refer to Figures 1 to 6 This invention provides an intelligent assessment method for the surface quality of the scroll disk of an air conditioning compressor in new energy vehicles. Its core lies in achieving high-precision identification and quantitative classification of micron-level defects on the scroll disk surface through multimodal optical imaging, physical prior-guided texture decoupling, self-supervised feature learning, and a dual-channel discrimination mechanism. The following will detail the specific implementation of each step in the method flow.

[0019] The method first performs step S1: using an industrial-grade high-resolution linear array camera in conjunction with a ring polarization light source and an oblique incident structured light illumination device, synchronous image acquisition is performed on the surface of the vortex disk from at least three different azimuth angles to obtain a multimodal raw image sequence containing information on surface reflection intensity, phase gradient, and polarization difference.

[0020] The line scan camera boasts a pixel resolution of 5120×5120, employing a global shutter CMOS sensor to ensure no motion blur even in high-speed conveyor belt environments. The ring-polarized light source consists of 16 independently adjustable LED groups, each equipped with a linear polarizer, with polarization directions evenly distributed at 22.5-degree intervals. The oblique-incident structured light illumination device uses digital micromirrors to project sinusoidal fringe patterns at a 45-degree angle with a fringe period of 0.5 millimeters.

[0021] Three acquisition angles were located at 0°, 120°, and 240° azimuths in an orthogonal coordinate system above the vortex disk. Each angle triggered an image exposure synchronously to ensure temporal consistency among the three sets of images. The acquired raw image sequence contained three types of information: the first type was a conventional bright-field image, recording the surface reflection intensity; the second type was a structured light phase map, which obtained surface height gradient information after demodulation using a four-step phase-shifting method; and the third type was a cross-polarized image pair, which obtained the difference in surface scattering characteristics by switching the polarization directions of the light source and the camera.

[0022] All image data is stored in 16-bit grayscale format, along with timestamps and metadata for light source configuration parameters.

[0023] Then, step S2 is performed: pixel-level registration and radiometric consistency correction are applied to the original multimodal image sequence to generate a fused image dataset in a unified coordinate system. In this step, pixel-level registration employs a non-rigid registration algorithm based on maximizing mutual information. First, key feature points are extracted from the three-view images to construct an initial coarse registration transformation matrix; then, a grid of control points is laid out across the entire image domain with a grid spacing of 8 pixels, and the displacement vector of each control point is iteratively optimized by maximizing the local mutual information objective function.

[0024] The interpolation method employs cubic smooth interpolation to ensure the continuity of image edges after registration. Radiometric consistency correction relies on a standard whiteboard calibration process: before each detection task begins, the system automatically acquires the response image of the standard whiteboard under the same light source configuration, establishing gain and offset correction curves for each channel. Specifically, for each pixel location, its corrected grayscale value is calculated by combining the original grayscale value with the offset and gain coefficient, both of which are obtained by fitting the whiteboard image.

[0025] After the above processing is completed, the three-view images are mapped to the same Cartesian coordinate system and stitched together into a nine-channel fused image dataset. The channel order is as follows: View 1 bright field, View 1 phase gradient x component, View 1 phase gradient y component, View 1 polarization difference, View 2 bright field, View 2 phase gradient x component, View 2 phase gradient y component, View 2 polarization difference, and View 3 bright field.

[0026] The aforementioned channel selection rules are based on the following: In the image acquisition stage, all three perspectives fully acquire the full modal information of bright field, phase gradient, and polarization difference. The core purpose is to ensure the registration accuracy of multi-view images and the imaging coverage of the entire surface of the vortex disk without blind spots. In the preprocessing channel stitching stage, based on the geometric layout of the three perspectives (0°, 120°, and 240° orthogonal orientation distribution), the phase gradient components of perspective one (0°) and perspective two (120°) have covered the height gradient information in the x and y directions in the plane. The phase gradient component of perspective three (240°) is highly linearly correlated with the phase gradient information of the first two perspectives, with extremely high redundancy. At the same time, the polarization difference channels of perspective one and perspective two have fully covered the full-angle detection requirements of surface scattering characteristics. The polarization difference channel of perspective three has no new effective information, with significant redundancy. Retaining only the bright-field channel of viewpoint three is to supplement the imaging information of the blind area formed by the light blockage of the side wall of the vortex disk profile due to the first two viewpoints. This not only ensures the imaging coverage of the entire surface of the vortex disk, but also greatly reduces the computational load of subsequent model inference, adapting to the cycle time requirements of high-speed inspection on the production line.

[0027] Next, step S3 is executed: the fused image dataset is input into a pre-trained texture decoupling neural network model. This texture decoupling neural network model adopts an encoder-decoder architecture. The encoder consists of five levels of convolutional residual blocks. The output of each level is connected to the corresponding decoder level. Furthermore, a physical prior constraint module based on the manufacturing process parameters of the vortex disk is introduced in the bottleneck layer to suppress periodic texture interference caused by machining marks and material grain orientation, thereby outputting a residual feature map representing the real defect area.

[0028] The encoder part of the texture decoupling neural network has a first-stage convolutional kernel size of 7×7, a stride of 2, and 64 output channels. The second to fifth stages all use 3×3 convolutional kernels with a stride of 2, and the number of output channels are 128, 256, 512, and 1024 respectively. Each convolutional residual block contains two convolutional layers and one skip connection, and the activation function is a modified linear unit.

[0029] The decoder uses transposed convolution for upsampling. After each upsampling stage, the feature maps from the corresponding encoder level are concatenated, and then fused using two 3×3 convolutional layers. The bottleneck layer is located between the fifth stage of the encoder and the first stage of the decoder, where a physical prior constraint module is introduced. This module receives real-time feedback from the CNC machining center regarding spindle speed, feed rate, and toolpath curvature radius parameters.

[0030] The spindle speed ranges from 3000 to 8000 rpm, the feed rate from 50 to 200 mm per minute, and the toolpath curvature radius from 5 to 20 mm. These parameters are mapped to a two-dimensional frequency domain template: first, the spatial frequency of the theoretically machined texture is calculated based on the spindle speed and feed rate; then, the texture direction distribution is determined by combining this with the toolpath curvature radius. The resulting frequency domain template applies a band-stop filter to the feature tensor output by the bottleneck layer in the Fourier transform domain, filtering out specific frequency range response components related to the machined texture. This frequency range is calculated and determined by machining parameters such as feed rate, tool diameter, and spindle speed.

[0031] After this operation, the periodic components in the feature tensor related to the known processed texture are significantly suppressed. The decoder finally outputs a single-channel residual feature map, where the pixel values ​​represent the probability of the presence of non-periodic anomalous structures at the location, i.e., potential defect regions.

[0032] Then, step S4 is executed: Based on the residual feature map, the local curvature change rate, gray-level co-occurrence matrix energy value and edge direction histogram distribution are extracted as macroscopic morphological statistical features; at the same time, the energy concentration and directional entropy of high-frequency wavelet coefficients are calculated on the residual feature map region by region through the sliding window mechanism to form microstructural anomaly response features.

[0033] For macroscopic feature extraction, the residual feature map is first Gaussian smoothed with a Gaussian kernel standard deviation of 1.5 pixels to suppress high-frequency noise. Then, the absolute value of the largest eigenvalue of the second-order partial derivative matrix at each pixel location is calculated and used as the local curvature change rate, reflecting the maximum degree of curvature in the neighborhood. The calculation window for the gray-level co-occurrence matrix energy value is 16 pixels × 16 pixels. Within the window, the sum of the gray-level values ​​of all pixel pairs is calculated. The direction set includes four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Finally, the average of the energy values ​​in the four directions is taken as the feature of the window.

[0034] Edge orientation histogram distribution is achieved by detecting edge points using the Canny operator, with a Canny high-low threshold ratio of 2:1. The edge point normal vector angle is quantized into 16 equal-width intervals (each interval being 22.5 degrees). The proportion of edge points in each interval to the total number of edge points is calculated to form a 16-dimensional histogram vector. For micro-feature extraction, a sliding window mechanism is used to traverse the residual feature map, with a window size of 8 pixels × 8 pixels and a stride of 4 pixels.

[0035] A three-level wavelet decomposition is performed on the image patch within each window, using the 4th-order Daubechies wavelet (db4) as the wavelet basis. This wavelet basis has a 4th-order vanishing moment, exhibiting excellent capture capability for high-frequency components of subtle surface anomalies, which matches the detection requirements of this scheme for micron-level defects. The implementation of the three-level wavelet decomposition is as follows: the input window image patch is decomposed level by level, with each level generating one low-frequency approximate sub-band and three high-frequency detail sub-bands (horizontal, vertical, and diagonal). After completing the three-level decomposition, a total of three high-frequency sub-bands are obtained. The three high-frequency sub-bands output from the third level decomposition are used for subsequent feature calculations.

[0036] The coefficients of the three high-frequency sub-bands in the third layer (corresponding to the horizontal-vertical, vertical-horizontal, and diagonal directions, respectively) are extracted, and the energy concentration of the high-frequency wavelet coefficients is calculated. This index is obtained by statistically analyzing the energy proportion of all wavelet coefficients. At the same time, the directional entropy is calculated, which is obtained by statistically analyzing the proportion of energy of the horizontal, vertical, and diagonal sub-bands to the total high-frequency energy. Each window outputs a two-dimensional microscopic feature vector.

[0037] Then, step S5 is executed: the macroscopic morphological statistical features and microscopic structural anomaly response features are respectively input into two independent fully connected discriminant branches. Each branch contains three nonlinear transformation layers, and the activation function is a modified linear unit. Finally, the confidence score of each region belonging to the defect category is output.

[0038] The macroscopic discriminant branch receives the macroscopic feature vector from each 16x16 window, with a dimension of 18 (one dimension for local curvature, one dimension for gray-level co-occurrence matrix energy, and 16 dimensions for edge orientation histogram). This macroscopic feature vector is first mapped to 128 dimensions through the first fully connected layer, then compressed to 32 dimensions through the second layer, and the third layer outputs a one-dimensional scalar representing the confidence that a defect exists in the central region of the window.

[0039] The micro-discrimination branch receives the micro-feature vector of each 8×8 window, with a dimension of 2. This micro-feature vector is expanded to 64 dimensions by the first fully connected layer, to 16 dimensions by the second layer, and the third layer also outputs a one-dimensional scalar confidence score.

[0040] The network weights of the two branches are optimized during the training phase through a self-supervised contrastive learning mechanism: positive sample pairs (different enhanced views of the same defective region) and negative sample pairs (defective region and normal texture region) are constructed, and a self-supervised contrastive loss function is used to bring positive samples closer and push negative samples further apart, without the need for manual labeling. After training, the weights of the two branches are fixed and used only for feature discrimination during the inference phase.

[0041] Finally, step S6 is executed: the confidence scores of the two discrimination branches are weighted and fused. The weight coefficients are dynamically adjusted according to the overall noise level of the current batch of images. After thresholding, the fusion result generates a binary defect mask. The defects are classified according to the mask area, perimeter ratio and morphological compactness index, and the final quality assessment report is output.

[0042] The weighted fusion process is as follows: First, the standard deviation and mean of the current batch of fused image datasets are calculated; then, a fusion strategy is dynamically selected based on the ratio of the standard deviation to the mean: when the ratio is low, a weighted fusion strategy dominated by macroscopic features is adopted; when the ratio is high, a weighted fusion strategy dominated by microscopic features is adopted, ultimately obtaining a comprehensive confidence score (considering the confidence scores of both macroscopic and microscopic discriminant branches during fusion). The fused confidence map is binarized using an adaptive threshold segmentation method to generate a preliminary defect mask. To eliminate isolated noise points, a combination of opening and closing morphological operations is performed on the preliminary defect mask, with the structuring element being a circular kernel with a radius of two pixels, opened first and then closed.

[0043] Subsequently, three quantitative indicators are calculated for each connected region: area, perimeter, and morphological compactness. A Level 1 defect is defined as a small defect area with high morphological compactness; a Level 2 defect is defined as a medium defect area or medium morphological compactness; and a Level 3 defect is defined as a large defect area or low morphological compactness. The final quality assessment report includes the total number of defects, the number of defects at each level, the maximum defect area and its location coordinates, and indicates whether the defect exceeds the company's quality standard threshold (e.g., a Level 3 defect count greater than 0 is considered unqualified).

[0044] At the system level, this invention provides an intelligent assessment system for the surface quality of the scroll plate of an air conditioning compressor in a new energy vehicle. This system includes a multimodal image acquisition unit, an image preprocessing unit, a texture decoupling processing unit, a feature extraction unit, a dual-channel discrimination unit, and a quality assessment output unit.

[0045] The multimodal image acquisition unit consists of three industrial-grade high-resolution linear array cameras, a ring polarization light source array, and an oblique incidence structured light projector, which are installed on a rotating detection platform to ensure full coverage illumination and imaging of the outer edge and inner cavity of the vortex disk.

[0046] The image preprocessing unit is deployed on an embedded GPU module, running a non-rigid registration algorithm based on maximizing mutual information and a radiometric correction procedure, with a processing latency of less than 200 milliseconds.

[0047] The texture decoupling processing unit is a dedicated neural network inference engine that integrates a physical prior constraint module. It supports real-time reception of process parameter streams from CNC machining centers and dynamic updates to the frequency domain filtering template. The feature extraction unit implements parallel sliding window computation and utilizes the SIMD instruction set to accelerate the solution of second-order partial derivative matrices and wavelet transforms.

[0048] The dual-channel discrimination unit comprises two independent fully connected neural network hardware accelerators, processing macroscopic and microscopic feature flows respectively. The quality assessment output unit is responsible for morphological post-processing, defect quantification, and report generation, and uploads the results to the production management system via an industrial Ethernet interface. The entire system has a single-piece inspection cycle of 2.7 seconds, meeting the production line cycle requirement of 1300 pieces per hour.

[0049] The above implementation method fully realizes the automated evaluation process from multimodal image acquisition to defect quantification and grading. Through physical prior-guided texture decoupling, the processed texture and the real defect are separated; through a dual-channel discrimination mechanism constructed by self-supervised contrastive learning, high-sensitivity defect identification is achieved under unlabeled conditions; and through dynamic weight fusion and multi-index quantification and grading, the robustness and engineering applicability of the evaluation results are ensured.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently evaluating the surface quality of the scroll plate of an air conditioning compressor in a new energy vehicle, characterized in that, include: By using an industrial-grade high-resolution linear array camera in conjunction with a ring polarization light source and an oblique incident structured light illumination device, synchronous image acquisition is performed on the surface of the vortex disk from at least three different azimuth angles to obtain a multimodal raw image sequence containing information on surface reflection intensity, phase gradient, and polarization difference. Pixel-level registration and radiometric consistency correction are performed on the multimodal original image sequence to generate a fused image dataset in a unified coordinate system; The fused image dataset is input into a pre-trained texture decoupling neural network model. This texture decoupling neural network model adopts an encoder-decoder architecture. The encoder consists of five levels of convolutional residual blocks. The output of each level is connected to the corresponding decoder level. A physical prior constraint module based on the manufacturing process parameters of the vortex disk is introduced in the bottleneck layer to suppress periodic texture interference caused by machining marks and material grain orientation, thereby outputting a residual feature map representing the real defect area. Based on the residual feature map, the local curvature change rate, gray-level co-occurrence matrix energy value and edge direction histogram distribution are extracted as macroscopic morphological statistical features. At the same time, the energy concentration and directional entropy of high-frequency wavelet coefficients are calculated on the residual feature map region by region through the sliding window mechanism to form microstructural anomaly response features. The macroscopic morphological statistical features and microstructural anomaly response features are respectively input into two independent fully connected discriminant branches, and finally the confidence score of each region belonging to the defect category is output. The confidence scores of the two discrimination branches are weighted and fused. The weight coefficients are dynamically adjusted according to the overall noise level of the current batch of images. The fusion result is thresholded and segmented to generate a binary defect mask. The defects are classified according to the mask area, perimeter ratio and morphological compactness index, and the final quality assessment report is output.

2. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 1, characterized in that, Pixel-level registration and radiometric consistency correction are performed on the multimodal original image sequence to generate a fused image dataset in a unified coordinate system, including: A non-rigid registration algorithm based on maximizing mutual information is used to perform pixel-level alignment of the three-view images. The grid spacing of the registration control points is set to eight pixels, and cubic smooth interpolation is used as the interpolation method. The original images of each channel are subjected to pixel-by-pixel gain compensation and offset correction using the standard whiteboard reflection response curve to generate a radiometrically consistent image. The corrected images are stitched together into a nine-channel fused image dataset according to viewpoint and information type. The channel order is as follows: viewpoint 1 bright field, viewpoint 1 phase gradient x component, viewpoint 1 phase gradient y component, viewpoint 1 polarization difference, viewpoint 2 bright field, viewpoint 2 phase gradient x component, viewpoint 2 phase gradient y component, viewpoint 2 polarization difference, and viewpoint 3 bright field.

3. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 2, characterized in that, The fused image dataset is input into a pre-trained texture decoupling neural network model, including: The decoder upsamples step by step through transposed convolution, and then performs channel concatenation with the feature maps of the corresponding encoder level before fusing them through two 3x3 convolutional layers. The bottleneck layer receives parameters such as spindle speed, feed rate, and tool path curvature radius, maps them into a two-dimensional frequency domain template, and applies a band-stop filter operation to the feature tensor in the Fourier transform domain to filter out response components in a specific frequency range related to the machining texture. This frequency range is calculated and determined by machining parameters such as feed rate, tool diameter, and spindle speed.

4. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 3, characterized in that, Macroscopic morphological statistical features are extracted based on the residual feature map, including: Gaussian smoothing is applied to the residual feature map; Calculate the absolute value of the largest eigenvalue of the second-order partial derivative matrix at each pixel location, and use it as the local curvature change rate; Calculate the energy value of the gray-level co-occurrence matrix and take the mean value; After detecting edge points using the Canny operator, the frequency percentage of edge normal vector angles falling within multiple equal-width intervals is statistically analyzed to form an edge direction histogram distribution.

5. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 4, characterized in that, The energy concentration and directional entropy of high-frequency wavelet coefficients are calculated region by region on the residual feature map using a sliding window mechanism to form microstructural anomaly response features, including: Traverse the residual feature map; Perform three-level wavelet decomposition on each window image block and extract the wavelet coefficients of the three high-frequency sub-bands in the third level; The energy concentration of high-frequency wavelet coefficients is calculated, and this index is obtained by statistically analyzing the energy proportion of all wavelet coefficients. The directional entropy is calculated by statistically analyzing the proportion of energy in the three directional sub-bands (horizontal, vertical, and diagonal) to the total high-frequency energy.

6. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 5, characterized in that, The macroscopic morphological statistical features and microscopic structural anomaly response features are respectively input into two independent fully connected discriminant branches, including: The macro-discrimination branch receives an eighteen-dimensional feature vector and outputs a one-dimensional defect confidence score. The micro-discrimination branch receives two-dimensional feature vectors and outputs a one-dimensional defect confidence score. The network weights of the two branches are optimized through a self-supervised contrastive learning mechanism. The self-supervised contrastive loss function is used to bring the distance between different enhanced views of the same defect area closer and push the distance between the defect and the normal area further apart.

7. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 6, characterized in that, The confidence scores of the two discriminant branches are weighted and fused, including: Calculate the standard deviation and mean of the current batch of fused image datasets; The fusion strategy is dynamically selected based on the ratio of the standard deviation to the mean: when the ratio is in a low range, a weighted fusion strategy dominated by macro features is adopted; when the ratio is in a high range, a weighted fusion strategy dominated by micro features is adopted, and finally a comprehensive confidence score is obtained.

8. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 7, characterized in that, The fusion result, after thresholding, generates a binary defect mask, including: An adaptive threshold segmentation method is used to binarize the fused confidence map; The binarized mask is subjected to opening and closing operations in sequence, and the structuring element is a circular kernel with a radius of two pixels.

9. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 8, characterized in that, Defects are classified into different levels based on mask area, perimeter ratio, and morphological compactness, including: Calculate the area, perimeter, and morphological compactness of each connected region; A Level 1 defect is defined as a defect area that is small in size and has high morphological compactness. Secondary defects are defined as defects with a moderate area or moderate morphological compactness. Level 3 defects are defined as defects with a large area or low morphological compactness.

10. The intelligent evaluation method for the surface quality of the scroll plate of a new energy vehicle air conditioning compressor according to claim 9, characterized in that, The ratio of high to low thresholds for the Canny operator is set to 2:1.

Images