An image recognition-based automobile part detection analysis method and system

Abstract

The application discloses an automobile part detection analysis method and system based on image recognition, and relates to the technical field of automobile part detection.The technical key points are as follows: multi-scale fractal feature extraction is performed on each recognized defect area, a defect fractal fingerprint fusing space and time dimension features is generated, fractal evolution entropy values are calculated based on the segmented information entropy of the fractal fingerprint, the disorder degree and initial growth potential of the defect microstructure are quantified, and then the micro growth path and failure critical point of the defect in a preset use cycle are predicted in combination with a process correction coefficient, so that the automobile parts with potential failure risks can be recognized in advance, and the safety and reliability in the use process of the whole vehicle are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of automotive parts inspection technology, and in particular to an image recognition-based method and system for automotive parts inspection and analysis. Background Technology

[0002] As the automotive industry rapidly develops towards intelligence and high reliability, the quality and stability of automotive parts directly determine the safety performance and service life of the entire vehicle. As a core link in ensuring the quality of parts leaving the factory, automated testing technology has gradually replaced traditional manual visual inspection and become the mainstream application solution in the industry.

[0003] Currently, the mainstream image recognition-based automated inspection solutions for automotive parts in the industry typically involve deploying one or more industrial cameras at the inspection station to capture images of the parts' surfaces. After standardized preprocessing such as denoising, enhancement, and binarization, basic features such as the parts' edge contours and grayscale distribution are extracted. The extracted features are then compared with pre-established standard defect templates to identify visible macroscopic defects such as cracks, scratches, and dents. Finally, based on the number, area, and location of the defects, the parts are classified into three levels: qualified, reworkable, or scrapped according to preset judgment criteria. This detection mode is essentially a static snapshot detection, which can only capture the macroscopic surface state of the parts at the moment of detection and cannot deeply quantify the microscopic structural features inside the defects. However, because automotive parts are subjected to complex working conditions such as vibration, temperature changes, and load impact during long-term use, the initial tiny microscopic defects will gradually expand and eventually lead to the failure of the parts. Existing technology cannot assess the initial growth potential of defects, let alone predict their evolution trend and failure threshold throughout their entire life cycle. This results in a large number of parts with potential failure hazards entering the market, posing a serious threat to the safety of the entire vehicle. Furthermore, when batch defects are detected, quality management personnel collect defect samples and organize personnel from production and process departments to conduct manual analysis, systematically checking for potential influencing factors such as processing temperature, pressure, and tool condition, and proposing process adjustment plans based on experience. Then, small-batch trial production is used to verify the adjustment effect. If the defect rate does not meet the standard, the above process is repeated until the defect rate drops to the acceptable range. This process traceability and optimization model, which relies on manual experience, cannot establish a quantitative mapping relationship between defect characteristics and production process parameters. This not only leads to long process problem localization cycles and low optimization efficiency, but also results in a disconnect between inspection data and the production process. Inspection results cannot be automatically fed back to the production system to achieve dynamic process adjustments, failing to fundamentally and continuously improve product quality. Therefore, we propose an image recognition-based method and system for automotive parts inspection and analysis. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an image recognition-based method and system for the inspection and analysis of automotive parts, solving the technical problems of existing technologies such as limited inspection dimensions, inability to predict failures, and difficulty in tracing processes.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: An image recognition-based method for detecting and analyzing automotive parts, comprising the following steps: Multi-view images of automotive parts are acquired and preprocessed to extract three core features: surface texture, geometric dimensions, and edge contours. The three types of core features are compared with the preset defect feature template library to obtain preliminary defect identification results including defect location, defect type and defect coverage area, and then the preliminary quality assessment results are calculated. Multi-scale fractal feature extraction is performed on the identified defective regions to generate a fixed-length digital defect fractal fingerprint that integrates spatial fractal features and temporal fractal features; The fractal evolution entropy value is calculated based on the segmented information entropy of the defect fractal fingerprint, which quantifies the disorder degree and initial growth potential of the defect microstructure. By integrating the defect fractal fingerprints, fractal evolution entropy values, and corresponding production process parameters of all completed inspection parts on a production batch basis, a database linking defect fractal fingerprints and production process parameters is constructed. By using cluster analysis to extract common fractal feature templates of various defects from the association database, a bidirectional nonlinear mapping relationship between the common fractal feature templates and the deviation of production process parameters is established. By combining the production process correction coefficients output by fractal evolution entropy and bidirectional nonlinear mapping relationship, the microscopic growth path and failure critical point of defects within the preset service life can be predicted. The safety level of the defect location is determined based on the preset location safety level mapping relationship. The preliminary quality assessment results are dynamically corrected according to the defect failure threshold and the safety level of the defect location to generate the final quality assessment conclusion.

[0006] Preferably, the step of calculating the preliminary quality assessment result based on the preliminary defect identification result is as follows: Three basic evaluation dimensions were determined: surface defect severity, dimensional deviation, and contour integrity. Individual scores were calculated for each basic evaluation dimension. Surface defect severity was weighted based on the number of defects, the area of ​​a single defect, and the basic importance of the defect's location. Dimensional deviation was scored based on the magnitude of the deviation between the actual size and the standard size. Contour integrity was scored based on the degree of contour offset. The weighting coefficients are determined based on the degree of impact of each basic evaluation dimension on the performance of the components. The individual scores of the three dimensions are multiplied by their corresponding weighting coefficients and then summed to obtain the preliminary quality assessment results.

[0007] Preferably, the defect fractal fingerprint is generated through the following steps: Based on the minimum side length of the circumscribed rectangle of the defect, the defect region is divided into a multi-level continuous scale square grid from a single pixel to a complete defect region; the number of grids containing defect pixels at each scale is counted, and the fractal scale-quantity relationship curve of the defect is plotted; the curve fitting method is used to calculate the local slope of the curve in each scale interval, and the corresponding number of spatial fractal dimension values ​​are obtained. Under the same acquisition conditions, multiple consecutive acquisitions are performed on the same defect area. The change in the corresponding scale fractal dimension between adjacent acquisitions is calculated to obtain the temporal fractal feature value. The spatial fractal feature value and the temporal fractal feature value are quantized into digital codes of a preset number of bits and then concatenated. A check bit is added to form a unique defect fractal fingerprint.

[0008] Preferably, the fractal evolution entropy value is calculated using the following steps: The fixed-length defect fractal fingerprint is divided into multiple continuous equal-length sub-segments, each sub-segment corresponding to a specific fractal scale interval; the information entropy value of the digital code of each sub-segment is calculated to obtain the fractal entropy value of each fractal scale interval. Preset weight coefficients are assigned according to the principle that the weight of the microscale interval is greater than that of the macroscale interval. Each fractal entropy value is multiplied by its corresponding weight coefficient and then summed to obtain the total fractal evolution entropy value, which ranges from 0 to 1. Based on the magnitude of the total fractal evolution entropy value, defects are divided into different growth levels.

[0009] Preferably, the bidirectional nonlinear mapping relationship is established through the following steps: Clustering algorithms are used to perform cluster analysis on the defect fractal fingerprints in the associated database, and clustering parameters that meet the detection requirements are set; common features of all defect fractal fingerprints in each cluster are extracted to form a common fractal feature template for this type of defect; and records of multiple core production process parameters corresponding to each cluster are collected. A forward mapping model from common fractal feature templates to deviations in production process parameters and a reverse mapping model from deviations in production process parameters to common fractal feature templates are trained using artificial neural networks. The bidirectional mapping model is trained and validated using historical production data, and put into use when the model's prediction accuracy reaches a preset threshold.

[0010] Preferably, the prediction process for the defect microscopic growth path and failure critical point is as follows: A defect growth dynamics model based on fractal theory is constructed, with fractal evolution entropy as the input and output of the basic growth rate of the defect; the defect fractal fingerprint is input into the forward mapping model to obtain the deviation values ​​of multiple core production process parameters that cause the defect; and the comprehensive production process correction coefficient is calculated by combining the influence of the deviation of each production process parameter on defect growth. The actual growth rate of the defect is obtained by multiplying the basic growth rate by the comprehensive production process correction coefficient; the fractal dimension change and defect area expansion of the defect within the preset service life are simulated based on the actual growth rate to generate the microscopic growth path curve of the defect; when the simulated fractal evolution entropy value reaches the preset failure threshold, the corresponding time point is the failure critical point of the defect.

[0011] Preferably, the analytical method further includes the following steps: The unique identifier of each component, the fractal fingerprint of defects, and the final quality assessment conclusion are written into the electronic identity tag to establish a full lifecycle quality traceability system based on the fractal fingerprint of defects. The establishment method of the full lifecycle quality traceability system is as follows: Each automotive component is assigned a unique identifier. Basic component information, production information, defect fractal fingerprints, final quality assessment results, and predicted failure time are encoded in a preset format and written into the storage area of ​​the electronic identifier. During the transportation, assembly, and maintenance of the component, the electronic identifier information is read by a reader / writer to record the status of each stage and bind the component information to the vehicle information. When a component malfunctions during use, the defect fractal fingerprint of the malfunctioning component is extracted and compared with the production stage defect fractal fingerprint stored in the electronic identifier. When the similarity reaches a preset threshold, it is determined that the malfunction originated from a defect generated during the production stage.

[0012] Preferably, the analytical method further includes the following steps: Based on the analysis results of the associated database, the deviation range of production process parameters corresponding to batch defects is located, and a targeted production process adjustment plan is generated. The adjustment effect is verified by the defect fractal fingerprint data of subsequent production batches. The process of generating and verifying the production process adjustment plan is as follows: The frequency of various defects within a single batch is statistically analyzed. When the frequency of a certain type of defect exceeds a preset threshold, it is determined to be a batch defect. The common fractal feature template of this type of batch defect is input into a forward mapping model to obtain the corresponding predicted deviation value of the production process parameters. The predicted deviation value is compared with the standard production process parameters to determine the specific production process parameters with deviations and the deviation range. Combining the degree of deviation of the production process parameters with the degree of impact on product quality, a production process adjustment plan is generated, which includes adjustment values, adjustment steps, and verification standards. The adjustment plan is implemented in multiple subsequent consecutive production batches. When the average fractal evolution entropy value of the same type of defect in multiple consecutive batches decreases to a preset standard, the production process adjustment is confirmed to be effective and solidified as a new production standard.

[0013] An image recognition-based automotive parts inspection and analysis system, the system comprising: The feature extraction module is used to acquire multi-view synchronous images of automotive parts and perform preprocessing to extract three core features: surface texture, geometric dimensions, and edge contours of the parts. The identification and evaluation module is used to perform multi-feature fusion comparison between the three types of core features and the preset defect feature template library to obtain preliminary defect identification results including defect location, defect type and defect coverage area, and calculate preliminary quality evaluation results. The defect fractal fingerprint generation module is used to extract multi-level continuous scale fractal features from the pixel level to the whole defect region of the identified defect region, and generate a fixed-length binary defect fractal fingerprint that integrates spatial fractal features and temporal fractal features. The fractal evolution entropy calculation module is used to calculate the fractal evolution entropy value based on the segmented information entropy of the defect fractal fingerprint, and to quantify the degree of disorder and initial growth potential of the defect microstructure. The associated database construction module is used to integrate the defect fractal fingerprints, fractal evolution entropy values ​​and corresponding production process parameters of all completed inspection parts in production batches, and to build an associated database of defect fractal fingerprints and production process parameters. The bidirectional mapping establishment module is used to extract common fractal feature templates of various defects in the association database through density clustering analysis, and establish a bidirectional nonlinear mapping relationship between the common fractal feature templates and the deviation of production process parameters. The defect growth prediction module is used to combine the production process correction coefficients output by the fractal evolution entropy value and the bidirectional nonlinear mapping relationship to predict the microscopic growth path and failure critical point of defects within a preset service life. The quality assessment correction module is used to determine the safety level of the location of the defect based on the preset safety level mapping table of critical locations of components, and to dynamically weight and correct the preliminary quality assessment results according to the defect failure threshold and the safety level of the defect location, so as to generate the final quality assessment conclusion. The full lifecycle traceability module is used to write the unique identifier of the component, the defect fractal fingerprint, and the final quality assessment conclusion into the passive electronic identification tag, and establish a full lifecycle quality traceability system based on the defect fractal fingerprint. The process adjustment module is used to locate the deviation range of production process parameters corresponding to batch defects based on the analysis results of the associated database, generate targeted production process adjustment plans, and verify the adjustment effect through defect fractal fingerprint data of subsequent production batches.

[0014] Preferably, the defect fractal fingerprint generation module includes: Multi-scale meshing unit is used to divide the defect region into multi-level continuous-scale square meshes from 1×1 pixels to the complete defect region, based on the minimum side length of the defect's circumscribed rectangle. The fractal curve plotting unit is used to count the number of grids containing at least one defect pixel at each scale and plot the fractal scale-number double logarithmic curve of the defect. The spatial fractal calculation unit is used to fit the local slope of the curve in each scale interval using the least squares method to obtain the corresponding number of spatial fractal dimension values. The time-fractal calculation unit is used to perform multiple consecutive acquisitions of the same defect area under the same acquisition conditions, calculate the absolute difference of the corresponding scale fractal dimension between adjacent acquisitions, and obtain the time-fractal feature value. The fractal fingerprint encoding unit is used to quantize the spatial fractal feature values ​​and the temporal fractal feature values ​​into binary codes of a preset number of bits, then concatenate them and add a check bit to form a unique defect fractal fingerprint. The bidirectional mapping establishment module includes: The density clustering analysis unit is used to perform clustering analysis on the defect fractal fingerprints in the associated database using the density clustering algorithm, and to set clustering parameters that meet the detection requirements. The common template extraction unit is used to extract the mode of each bit encoding of all defect fractal fingerprints in each cluster, forming a common fractal feature template for that type of defect; The production process parameter collection unit is used to collect records of multiple core production process parameters corresponding to each cluster, including processing temperature, processing pressure, processing time, tool speed and coolant flow rate. The bidirectional model training unit is used to train, through a multi-layer neural network, a forward mapping model from common fractal feature templates to production process parameter deviations, and a reverse mapping model from production process parameter deviations to common fractal feature templates. The model validation unit is used to train and validate the bidirectional mapping model using historical production data. It is put into use when the model's prediction accuracy reaches a preset threshold.

[0015] (III) Beneficial Effects 1. By extracting multi-scale fractal features from each identified defect region, a defect fractal fingerprint integrating spatial and temporal dimensions is generated. Based on the segmented information entropy of the fractal fingerprint, the fractal evolution entropy value is calculated to quantify the disorder and initial growth potential of the defect's microstructure. Then, combined with the process correction coefficient, the microscopic growth path and failure threshold of the defect within a preset service life are predicted, identifying components with potential failure risks in advance and effectively improving the safety and reliability of the entire vehicle during use. Furthermore, by integrating the defect fractal fingerprints, fractal evolution entropy values, and corresponding production process parameters of all completed inspection components on a production batch basis, a related database is constructed, and common fractal feature templates of various defects are extracted. A bidirectional nonlinear mapping relationship between defect features and deviations in production process parameters is established. When batch defects occur, the specific process deviation range can be automatically located, and a targeted production process adjustment plan can be generated. The adjustment effect is verified by the defect fractal fingerprint data of subsequent production batches, forming a complete quality control process and achieving continuous improvement in product quality.

[0016] 2. By writing the unique identifier of each component, the fractal fingerprint of defects, and the final quality assessment conclusion into the electronic identity tag, relevant information can be obtained through reading and writing devices at each stage of transportation, assembly, and use and maintenance. This enables full-process quality tracking of components from production to final disposal, providing accurate basis for product failure analysis, recall management, and recycling. Secondly, a multi-feature fusion comparison method is used for preliminary defect identification. The method combines three core features—surface texture, geometric size, and edge contour—for comprehensive judgment. Cross-verification of multi-view images eliminates the influence of interference factors such as shooting light and image noise, reducing the occurrence of misjudgment and omission. Attached Figure Description

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0018] Figure 1 This is an architecture diagram of an embodiment of the present invention; Figure 2 This is an overall flowchart of an embodiment of the present invention; Figure 3 This is a diagram illustrating the defect growth path and failure prediction in an embodiment of the present invention. Detailed Implementation

[0019] This application provides an image recognition-based method and system for detecting and analyzing automotive parts, which solves the technical problems of existing technologies, such as limited detection dimensions, inability to predict failures, and difficulty in tracing processes.

[0020] The technical solution of this invention will be described below using a practical batch testing scenario of automotive engine connecting rods. This embodiment is only used to explain this invention and is not intended to limit the scope of protection of this invention. The overall workflow is as follows: Figure 2 As shown; on the engine connecting rod production line in the factory, the inspection station is set after the machining process and before the assembly process. The connecting rod to be inspected is transported to the inspection station by a conveyor belt and clamped and positioned by a pneumatic positioning fixture. Initially, we used a simple stop block positioning, but found that the repeatability error exceeded 0.1mm, causing the dimensional deviation of the same part in different inspections to exceed the allowable range. Therefore, we switched to a precision positioning fixture with V-blocks and pneumatic ejectors. The pneumatic ejectors push the connecting rod against the V-blocks to achieve precise positioning, ensuring that the part position is consistent in each inspection and avoiding the impact of shooting position deviations on subsequent inspection results. The specific solution is as follows: Example 1: I. Image Processing: Traditional inspection stations only deploy three industrial cameras to capture images of the front, side, and top surfaces. However, in actual operation, we found that the detection rate of micro-cracks and porosity defects on the back and bottom surfaces of the connecting rod remained high, especially the axial cracks on the inner wall of the large-end hole, which were the most severely missed. To address this, we tried increasing the camera tilt angle, but this resulted in increased image distortion and decreased dimensional measurement accuracy, failing to meet the requirements. Ultimately, we decided to deploy one industrial area array camera in each of the six directions, corresponding to the front, back, left, right, top, and bottom surfaces of the connecting rod, forming a hexahedral surround acquisition layout. Furthermore, each camera was equipped with a supplementary light. The light source uses diffused illumination. The reason for using diffused illumination instead of direct or strip light sources is that direct light sources produce strong reflections, and the gray value of the reflected area is close to 255, making it impossible to extract effective features. Strip light sources provide uneven illumination, with the brightness of the edge area being 30% lower than that of the center. Both of these can lead to reflections or dark areas being misjudged as defects. Therefore, diffused illumination is used to avoid specular reflections on the metal surface of the connecting rod. Then, the camera synchronously acquires data through a hardware trigger signal issued by the PLC controller, ensuring that six images correspond to the status of the component at the same moment, avoiding mismatches of defects from different perspectives due to slight displacement of the component.

[0021] The camera calibration adopts the Zhang Zhengyou calibration method, which is a camera calibration method based on a planar checkerboard grid. It has the advantages of high accuracy, simple operation, and suitability for industrial field applications. Specifically, a 12×12 checkerboard calibration board is used, with each square having a side length of 10mm. Each camera acquires 20 images of the calibration board at different angles and distances. The camera intrinsic parameter matrix, distortion coefficients, and extrinsic parameter matrix are calculated by solving the homography matrix. The camera intrinsic parameter matrix includes focal length and principal point coordinates, the distortion coefficients include radial and tangential distortion, and the extrinsic parameter matrix includes the rotation and translation vectors of the camera relative to the world coordinate system. The calibration error is controlled within 0.5 pixels. Images without extrinsic parameter calibration cannot be accurately stitched together, and the defect positions cannot be corresponding in three-dimensional space. Therefore, an extrinsic parameter calibration step is added to establish the coordinate transformation relationship between the cameras, laying the foundation for subsequent three-dimensional defect matching. This method is not limited to the existing one; any method with the same functionality can be used.

[0022] The raw color images captured by the camera are first sent to an industrial computer for preprocessing. The specific steps are as follows: The first step was noise reduction. We tested three methods: mean filtering, median filtering, and Gaussian filtering. Mean filtering replaced the center pixel with the average value of neighboring pixels, which severely blurred the defect edges and caused the loss of features of tiny cracks. Median filtering replaced the center pixel with the median value of neighboring pixels, which had the best suppression effect on salt-and-pepper noise and strong edge preservation ability. Therefore, we used a 3×3 window median filter to eliminate salt-and-pepper noise caused by electromagnetic interference from workshop motors and frequency converters. Then, we used a 5×5 window Gaussian filter to smooth the image and remove Gaussian noise generated by the camera sensor itself. The σ value was set to 1.5, which achieved the best balance between noise suppression and detail preservation. If σ is too small, it will not effectively remove noise, and if σ is too large, it will blur the defect details. The second step is contrast enhancement. Contrast-limited adaptive histogram equalization is used, which divides the image into 8×8 sub-blocks and performs equalization processing on each block. The cropping factor is set to 0.02, which enhances the contrast between defective and normal areas while avoiding excessive amplification of noise. The third step is geometric correction. Since the camera cannot be mounted perfectly vertically, the captured images will have angular distortion. We pre-calibrate the intrinsic and extrinsic parameters of each camera using a calibration board, correct the shooting angle deviation through affine transformation, and then use the distortion coefficient obtained by Zhang Zhengyou's calibration method to correct the radial and tangential distortion of the lens. The fourth step is to convert the corrected color image into a grayscale image and perform binarization using the Otsu adaptive thresholding algorithm. This algorithm can automatically calculate the optimal segmentation threshold based on the grayscale distribution of the image, avoiding the problem of poor adaptability of fixed thresholds under different lighting conditions.

[0023] After image processing, three core features are extracted from the standardized grayscale image. First, surface texture features are calculated using the gray-level co-occurrence matrix (GLCM) method. The GLCM describes the joint grayscale distribution probability of two pixels in an image at a certain distance and direction, effectively reflecting the roughness, directionality, and uniformity of the texture. Since LBP is sensitive to changes in illumination, HOG is more suitable for large-scale target detection, while the GLCM is more sensitive to minute texture changes on metal surfaces and is more suitable for detecting small defects such as cracks and scratches. Therefore, a grayscale level of 16 and a step size of 1 are set, and texture parameters in four directions (0°, 45°, 90°, and 135°) are calculated. The average value is then used to obtain the texture roughness, texture direction, and texture parameters. The system employs three feature values: uniformity, uniformity, and geometric dimensions. First, it uses the Canny edge detection algorithm to extract the connecting rod's edge contour. The Canny algorithm achieves accurate edge extraction through non-maximum suppression and dual-threshold detection, with high and low thresholds set to 0.7 and 0.3 respectively. This parameter combination effectively extracts true edges while suppressing false edges. Then, it uses the least squares method to fit the contour curve and calculates six parameters: length, width, large end hole diameter, small end hole diameter, rod thickness, and coaxiality. Third, it calculates three parameters—contour integrity, contour curvature, and contour offset—by comparing the deviation between the actual contour and the standard contour. All feature parameters are linearly normalized and mapped to the 0-1 interval to facilitate subsequent feature comparison and eliminate the influence of different feature units.

[0024] The extracted feature vectors are compared with a pre-set defect feature template library using multi-feature fusion. The template library stores feature templates for six common defects: cracks, scratches, dents, deformations, pores, and burrs. Each template contains corresponding surface texture, geometric dimensions, and edge contour features. Multi-feature fusion comparison is used to judge the similarity of the three types of features. The cosine similarity between the feature vector to be detected and each template is calculated. The cosine similarity measures the similarity between two vectors in a direction, with a value ranging from 0 to 1. The larger the value, the higher the similarity. When the similarity exceeds 0.85, it is initially determined to be a defect of the corresponding type. At the same time, the deviation of the geometric dimension parameters from the standard dimensions is compared. When the deviation exceeds 0.1 mm, it is determined to be a dimensional defect.

[0025] To reduce false positives, we cross-validate the recognition results from six perspectives. If a defect is detected only in one perspective and not in the other five, it is considered a false positive due to shooting interference and is excluded. If a defect is detected in two or more adjacent perspectives, and the two-dimensional pixel coordinates of different perspectives can be converted into three-dimensional spatial coordinates using a binocular stereo vision algorithm, it is confirmed as a genuine defect. The binocular stereo vision algorithm calculates the disparity using the extrinsic matrix of two adjacent cameras and obtains the three-dimensional coordinates of the defect based on the principle of triangulation. It judges based solely on the number of perspectives. Dust or reflections may appear simultaneously in two adjacent perspectives, so a three-dimensional spatial coordinate matching step is added, further reducing the false positive rate. The final preliminary defect recognition result includes the precise pixel coordinates of each defect, the defect type, and the area of ​​the defect's bounding rectangle.

[0026] II. Preliminary Quality Assessment: Preliminary quality assessment results are calculated based on the initial defect identification results. The assessment system includes three basic dimensions: surface defect severity, dimensional deviation, and contour integrity. Each dimension has a maximum score of 100 points. The surface defect severity score is calculated by weighting the number of defects, the area of ​​a single defect, and the location of the defect. The number of defects has a weight of 0.3, the area of ​​a single defect has a weight of 0.4, and the location importance has a weight of 0.3.

[0027] By performing finite element stress analysis on the connecting rod, the importance level of different locations was determined: The inner walls of the connecting rod's large end bore, small end bore, the transition fillet between the rod body and the large end, and the transition fillet between the rod body and the small end are critical stress-bearing locations, bearing the largest alternating tensile and compressive loads. The importance weighting coefficient for these defect locations is 1.0. The side of the rod body, the large end face, and the small end face are important locations, bearing smaller loads, with a weighting coefficient of 0.6. Unmachined surfaces and chamfered areas are general locations, bearing almost no load, with a weighting coefficient of 0.3. The dimensional deviation score is calculated based on the absolute deviation between the actual size and the standard size, with 1 point deducted for every 0.01mm deviation, up to a maximum deduction of 0.01mm. The contour integrity score is calculated based on the average contour offset, with 2 points deducted for every 0.01mm offset, up to a maximum deduction of 0.01mm. Finally, based on the impact of each dimension on the connecting rod's performance, the weights for surface defect severity (0.5), dimensional deviation (0.3), and contour integrity (0.2) are determined using the analytic hierarchy process. The individual scores for each of the three dimensions are multiplied by their corresponding weight coefficients and then summed to obtain a preliminary quality assessment result ranging from 0 to 100 points. When multiple defects exist, the surface defect severity score is the weighted sum of all defect scores, with the weight proportional to the defect area; larger defects have a greater impact on quality.

[0028] III. Defect Fractal Fingerprint Generation: This is a key improvement we made. A problem we encountered in production was that two crack defects with identical external dimensions exhibited drastically different propagation rates in subsequent fatigue tests. Some reached the critical size after 1 million cycles, while others showed no significant change after 10 million cycles. Traditional detection methods can only measure the macroscopic size and location of defects and cannot distinguish these differences in microstructure, leading to some potentially high-risk components entering the market. To address this, we adopted fractal theory to describe the microstructural characteristics of irregular shapes. Defects such as cracks and pores in materials have self-similarity, meaning that the local structure is similar to the overall structure. The fractal dimension is closely related to the growth potential of the defect. The higher the fractal dimension, the more micro-cracks and pores are inside the defect, and the faster the growth rate. Therefore, we decided to introduce fractal geometry methods to quantitatively characterize defects.

[0029] First, based on the minimum side length L of the circumscribed rectangle of the defect, the defect region is divided into 16 consecutive square grids of different scales. The formula for calculating the side length of the grid at the i-th scale is: s_i = L × 2^((i-1) / 15), where i is the scale number, ranging from 1 to 16; s_i is the side length of the grid at the i-th scale; L is the minimum side length of the circumscribed rectangle of the defect; the minimum grid is 1×1 pixel, the maximum grid is the entire defect region, and the ratio of the side lengths of adjacent grids is 2^(1 / 15) to ensure uniform scale distribution. Initially, we tried equal-interval scale division, that is, increasing the grid side length of each scale by a fixed number of pixels. We found that for small defects, large-scale grids are meaningless, and for large defects, the number of small-scale grids is too large, resulting in excessive computation.

[0030] The number of grids N_i containing at least one defect pixel at each scale was counted. A fractal scale-number double logarithmic curve was plotted with the logarithm of scale log2(s_i) as the x-axis and the logarithm of grid number log2(N_i) as the y-axis. The local slope of the double logarithmic curve in 16 scale intervals was fitted by the least squares method to obtain 16 spatial fractal dimension values ​​D_i. The fractal dimension values ​​range from 1 to 2. The larger the value, the more complex the defect structure and the more micro-cracks and pores inside.

[0031] During testing, we found that spatial fractal features alone were insufficient to completely distinguish defects with different growth potentials. Some defects had the same spatial fractal dimension, but exhibited different microscopic dynamic changes during continuous imaging. We hypothesized that this was due to the subtle opening and closing of microcracks within the defect under the minute influences of vibration and illumination. This change was related to the internal stress state and growth potential of the defect; the greater the internal stress, the more pronounced the opening and closing of microcracks. Therefore, we decided to incorporate temporal fractal features. Under the same acquisition conditions, three consecutive image acquisitions were performed on the same defect region at 100-millisecond intervals. The time interval ensures that the defect has enough time to undergo minor changes, but does not cause the defect position to shift due to excessively long intervals. The 16 spatial fractal dimension values ​​D_i1, D_i2, and D_i3 obtained from three acquisitions are calculated respectively. Then, the absolute differences in scale fractal dimensions between two adjacent acquisitions are calculated as ΔD_i1=|D_i2-D_i1| and ΔD_i2=|D_i3-D_i2|. The average value is taken to obtain 15 time fractal feature values ​​T_i=(ΔD_i1+ΔD_i2) / 2. The time fractal features can reflect the microscopic dynamic changes and internal stress state of the defect surface.

[0032] Next, quantization encoding is performed using a 4-bit encoding scheme. The range of each fractal feature value is divided into 16 equal intervals, with each interval corresponding to a 4-bit binary number. While ensuring accuracy, the length is controlled to 128 bits. The 16 spatial fractal feature values ​​and 15 temporal fractal feature values ​​are quantized into 4-bit binary codes respectively, and then concatenated in sequence to generate a 124-bit basic code. Finally, a 4-bit CRC check bit is added at the end. The CRC check can detect single-bit and multi-bit errors during transmission and storage, ultimately generating a 128-bit unique defect fractal fingerprint. Each defect's fractal fingerprint is unique and can serve as a digital identity identifier for the defect.

[0033] The fractal evolution entropy value is calculated based on the fractal fingerprint of defects. Different segments of the fractal fingerprint correspond to different scale intervals. The segments at the microscale contain more information about the internal structure of the defect and have a greater impact on defect growth. Therefore, by calculating the information entropy of each segment and summing them by weight, a quantitative index that can comprehensively reflect the degree of disorder and growth potential of the defect microstructure can be obtained.

[0034] The 128-bit defect fractal fingerprint was divided into 16 consecutive 8-bit segments, each corresponding to a fractal scale interval. The fractal entropy value of each segment was calculated using the information entropy formula: H_i = -Σ(p_j × log2(p_j)), where H_i is the fractal entropy value of the i-th fractal scale interval; p_j is the probability of the j-th binary combination occurring in the segment, with j ranging from 0 to 255, since there are 256 combinations of 8 bits. Information entropy is an indicator used in information theory to measure the uncertainty of random variables; a larger entropy value indicates higher uncertainty and a higher degree of disorder in the microstructure of the defect. This formula comes from Shannon's information entropy theory, and we apply it to quantify the degree of disorder in fractal fingerprints because the degree of disorder in fractal fingerprints is directly related to the number of microcracks and pores inside the defect.

[0035] Assign a weight coefficient of 0.08 to the first 8 microscale intervals and a weight coefficient of 0.045 to the last 8 macroscale intervals. The sum of the weights is 8 × 0.08 + 8 × 0.045 = 1.0. Multiply the 16 fractal entropy values ​​by their corresponding weight coefficients and sum them to obtain the total fractal evolution entropy value, which ranges from 0 to 1: H_total = Σ(w_i × H_i), where w_i is the weight coefficient of the i-th scale interval, which is 0.08 for the first 8 and 0.045 for the last 8.

[0036] Based on the magnitude of the total fractal evolution entropy value, defects are divided into three growth levels: Entropy values ​​less than 0.3 indicate slow-growing defects. These defects have a dense microstructure and low internal stress, and will hardly expand under normal use conditions. Entropy values ​​between 0.3 and 0.7 indicate stable-growing defects. These defects expand slowly with increasing use time. Entropy values ​​greater than 0.7 indicate fast-growing defects. These defects have a loose microstructure, high internal stress, and rapid expansion, posing a high risk of failure.

[0037] IV. Building a Related Database: Information on all completed connecting rods is stored in a MySQL database, organized by production batch, to construct a database linking defect fractal fingerprints with production process parameters. Each record contains three parts: basic component information, inspection information, and production process information. The basic component information includes the connecting rod number, model, production batch, and production date. The inspection information includes the defect fractal fingerprint, fractal evolution entropy value, preliminary quality assessment result, final quality assessment result, and predicted failure time. The production process information includes processing temperature, processing pressure, processing time, tool speed, coolant flow rate, equipment number, and operator. The database is sharded and partitioned by production month to avoid excessive data volume in a single table, which could reduce query efficiency. A hash index is also created for the fractal fingerprint, mapping the 128-bit fractal fingerprint to a 32-bit hash value, improving data query and analysis efficiency, with the query time for a single record controlled within 10ms.

[0038] After accumulating more than 1,000 defect data in the database, a bidirectional nonlinear mapping relationship between common fractal feature templates and deviations in production process parameters is established; an artificial neural network model is adopted, whose powerful nonlinear fitting ability can handle this complex relationship well.

[0039] First, the DBSCAN density clustering algorithm is used to perform cluster analysis on the defect fractal fingerprints in the associated database. DBSCAN is a density-based clustering algorithm that can automatically discover clusters of arbitrary shapes without pre-specifying the number of clusters, making it suitable for processing high-dimensional data such as fractal fingerprints. The neighborhood radius is set to 0.1 and the minimum number of samples is set to 5. The neighborhood radius is used to measure the distance between two samples, and the minimum number of samples is used to define the core point. The mode of each bit encoding of all defect fractal fingerprints in each cluster is extracted to form the common fractal feature template of this type of defect. The mode can reflect the most representative feature in the cluster. All production process parameter records corresponding to each cluster are collected, and the average process parameter deviation corresponding to each type of defect is calculated, that is, the difference between the average process parameter of this type of defect and the standard process parameter.

[0040] Two three-layer backpropagation (BP) neural networks were constructed as the forward mapping model and the backward mapping model, respectively. The forward mapping model has an input layer of 128 neurons corresponding to a 128-bit fractal fingerprint, and a hidden layer of 64 neurons using the ReLU activation function. The ReLU activation function can solve the gradient vanishing problem and improve the model training speed. The output layer has 5 neurons corresponding to the deviation values ​​of five core production process parameters, namely processing temperature, processing pressure, processing time, tool speed, and coolant flow rate. The backward mapping model has an input layer of 5 neurons corresponding to the deviation values ​​of the five process parameters, a hidden layer of 64 neurons, and an output layer of 128 neurons corresponding to a 128-bit fractal fingerprint.

[0041] The model was trained and validated using historical data from the first 10 production batches, with 80% of the data used as the training set and 20% as the test set. The Adam optimizer was used for training, combining the advantages of momentum and adaptive learning rate, resulting in fast and stable convergence. The initial learning rate was 0.001, decreasing by 10% every 100 iterations. The batch size was 32, and the mean squared error loss function was used to measure the difference between the model's predictions and the actual values. An early stopping strategy was employed to avoid overfitting; training stopped when the validation set loss did not decrease for 50 consecutive iterations. The model was completed and put into use when its prediction accuracy on the test set reached over 95%. A reverse mapping model was used to verify the reasonableness of the process deviations output by the forward mapping model. The process deviations output by the forward mapping were input into the reverse model to obtain the predicted fractal fingerprint. When the similarity between the predicted fractal fingerprint and the actual fractal fingerprint was less than 90%, the forward mapping was recalculated to improve prediction accuracy.

[0042] We predict the microscopic growth path and failure critical point of defects by combining fractal evolution entropy and bidirectional nonlinear mapping. Since there is a power-law relationship between fractal evolution entropy and defect growth rate, we construct a defect growth kinetic model based on fractal theory: v = k × H^α, where v is the basic defect growth rate in mm / 10^6 cycles; H is the fractal evolution entropy; k is a material constant related to the mechanical properties of the material; and α is an exponent reflecting the influence of fractal evolution entropy on the growth rate. This formula originates from the Paris formula for fatigue crack propagation, which describes the power-law relationship between crack propagation rate and stress intensity factor range. We improve it to reflect the relationship between fractal evolution entropy and growth rate, as fractal evolution entropy can comprehensively reflect the microstructure and stress state of defects. For the 40Cr alloy steel used in the connecting rod, we calibrated k = 0.001 and α = 2.5 through fatigue tests. During the calibration process, we tested the defect propagation rate under different stress levels and obtained these two parameters through least squares fitting.

[0043] To achieve rapid adaptation to different materials, a universal parameter calibration process is established: First, standard samples of the material are prepared, and defects with different fractal evolution entropy values ​​are pre-formed. Then, fatigue tests are conducted, and the defect growth rate under different entropy values ​​is recorded. Finally, the values ​​of k and α are obtained by fitting using the least squares method. Taking 6061 aluminum alloy as an example, k=0.0015 and α=2.2 are obtained through this process. Verification shows that the prediction error under these parameters is within 18%, which meets the requirements of industrial applications.

[0044] The fractal fingerprint of the defect is input into the forward mapping model to obtain the deviation values ​​of five production process parameters that cause the defect. The comprehensive production process correction coefficient is calculated based on the degree of influence of each process parameter deviation on defect growth: C = 1 + Σ(w_i × |Δp_i|), where C is the comprehensive production process correction coefficient; w_i is the weight coefficient of the i-th process parameter, determined through orthogonal experiments to be 0.3 for machining temperature, 0.25 for tool speed, 0.2 for machining pressure, 0.15 for machining time, and 0.1 for coolant flow rate; Δp_i is the relative deviation value of the i-th process parameter, i.e., the difference between the actual process parameter and the standard process parameter divided by the standard process parameter; the correction coefficient is limited to a range of 0.5 to 2.0 to avoid extreme values; the larger the deviation of the process parameter, the stronger its promoting effect on defect growth, and the larger the correction coefficient; different process parameters have different degrees of influence, therefore different weights are assigned.

[0045] The actual growth rate of the defect, v_actual = v × C, is obtained by multiplying the basic growth rate by the comprehensive production process correction factor. Based on the actual growth rate, the fractal dimension change and defect area expansion of the defect are simulated for each year within a 10-year service life, generating a microscopic growth path curve for the defect. When the simulated fractal evolution entropy value reaches the preset failure threshold of 0.9, the corresponding time point is the failure critical point of the defect. This threshold is determined through material fracture testing. Figure 3 As shown.

[0046] When multiple defects exist, a fault tree analysis method is used to comprehensively consider the mutual influence between defects: First, a fault tree for connecting rod failure is established, with the top event being connecting rod fracture, intermediate events being failure of defects at various locations, and the bottom event being a single defect; then, the failure probability of each defect is calculated. When the distance between two defects is less than 3 times the defect size, the stress concentration superposition effect is considered, and their joint failure probability is multiplied by an amplification factor of 1.3; finally, the earliest failure time among all defects is taken as the overall failure time of the component. For example, there are two adjacent cracks near the connecting rod big end hole. When they exist alone, their failure times are 5 years and 6 years, respectively. Since their distance is only 2 times the defect size, the joint failure time is shortened to 3.5 years.

[0047] The preliminary quality assessment results are dynamically revised based on the predicted failure threshold of the defect and the safety level of its location. A safety level mapping table for key positions of the link is pre-established, dividing different positions of the link into three safety levels: Safety level 1 is a critical stress position, where defect failure will directly lead to link breakage and cause serious safety accidents; Safety level 2 is an important position, where defect failure will affect the performance of the link but will not immediately lead to breakage; Safety level 3 is a general position, where defect failure has a relatively small impact on the performance of the link.

[0048] Calculate the dynamic correction factor based on the safety level of the defect location and the predicted failure time: For safety level 1 defects, the correction factor is 0.5 if the failure time is less than 3 years, 0.7 if the failure time is 3 years ≤ failure time < 6 years, and 0.9 if the failure time is ≥ 6 years. For safety level 2 defects, the correction factor is 0.7 if the failure time is less than 3 years, 0.85 if the failure time is 3 years ≤ failure time < 6 years, and 0.95 if the failure time is ≥ 6 years. The correction factor for safety level 3 defects is uniformly 1.0. When multiple defects exist, the minimum value among all defect correction factors is taken as the final correction factor, because failure of any defect in a critical position will lead to the failure of the entire component. The preliminary quality assessment result is multiplied by the dynamic correction factor to obtain the final quality assessment result. Based on the final quality assessment result, the connecting rod is divided into three levels: a final score ≥ 80 is qualified and allowed to enter the assembly process; a score 60 ≤ final score < 80 is reworked and requires corresponding repair; and a final score < 60 is scrapped and sent directly to the scrap area.

[0049] V. Full Lifecycle Quality Traceability System: Establish a full lifecycle quality traceability system based on defect fractal fingerprints. Defect fractal fingerprints are unique digital identifiers for each defect. Once written into an electronic tag, they can be read at any time throughout the entire lifecycle of the component, enabling full-process quality tracking from production to scrap.

[0050] Each link is assigned a unique 16-bit hexadecimal identifier. Basic component information, testing information, and production information are encoded according to a preset format and written into the user data area of ​​the passive UHF RFID tag. We tested active tags, low-frequency tags, and high-frequency tags: active tags require batteries and have a short lifespan, generally 3-5 years, which cannot meet the component's lifespan requirement of more than 10 years; low-frequency tags have a short read / write distance, <10cm, which is not suitable for rapid reading in industrial environments; high-frequency tags have poor anti-metal capabilities, and the read / write distance on metal surfaces is significantly shortened. Ultimately, we chose a passive UHF anti-metal tag with a working frequency of 915MHz, a storage capacity of 512 bytes, no battery required, a lifespan of over 20 years, and the ability to withstand the high temperature, high humidity, and vibration environments of industrial environments. The data format is: 16 bytes of identifier, 8 bytes of model, 8 bytes of production batch, 16 bytes of fractal fingerprint, 2 bytes of final quality assessment result, 4 bytes of predicted failure time, 4 bytes of production date, 2 bytes of equipment number, and the remaining bytes are reserved for future expansion. The tag is affixed to the non-machined surface of the link to avoid affecting the assembly and use of the component.

[0051] During transportation, fixed RFID readers at the warehouse entrance read tag information to record transportation time, route, and vehicle information. During assembly, handheld RFID readers read tag information to bind connecting rod information to the engine VIN code, establishing a correspondence between parts and the complete vehicle. During use and maintenance, RFID readers connected to the vehicle's OBD interface read tag information to update the connecting rod's usage time and maintenance records. When a connecting rod malfunctions during use, the faulty connecting rod is removed, and its defect image is re-acquired in the laboratory. The defect fractal fingerprint is extracted and compared with the cosine similarity of the production stage defect fractal fingerprint stored in the electronic tag. When the similarity reaches 90% or higher, it is determined that the fault originated from a defect generated during the production stage, providing accurate evidence for quality problem tracing and product recall.

[0052] VI. Production Process Adjustment: Adjustments to the production process are made based on the analysis results of the associated database. This is mainly achieved by analyzing the common fractal characteristics of batch defects, inferring the deviations in process parameters that lead to defects, adjusting the process parameters, and verifying the adjustment effect through the test data of subsequent batches to reduce the occurrence of defects.

[0053] The frequency of various defects within a single batch is statistically analyzed. When the frequency of a certain type of defect exceeds 5%, it is determined to be a batch defect. The common fractal feature template of this type of batch defect is input into the forward mapping model to obtain the corresponding predicted deviation value of the production process parameters. The predicted deviation value is compared with the standard production process parameters to determine the specific process parameters with deviation and the deviation range. Combining the degree of deviation of the process parameters and the degree of impact on product quality, a production process adjustment plan is generated, which includes precise adjustment values, adjustment steps, and verification standards. For example, if the model predicts that the processing temperature is 5°C higher, the adjustment plan is to reduce the processing temperature by 5°C and monitor the frequency of this type of defect in subsequent production.

[0054] The adjustment plan is implemented in the subsequent three consecutive production batches. Here, consecutive production batches refer to three batches produced consecutively after the process parameters are adjusted, to avoid the impact of the production of other product models on the verification results. The fractal fingerprint of the connecting rods produced after the adjustment is collected in real time, and the average fractal evolution entropy value of the same type of defect is calculated. When the average fractal evolution entropy value of the same type of defect in three consecutive batches decreases by more than 30%, the process adjustment is confirmed to be effective and the adjusted process parameters are solidified as the new production standard. If the expected effect is not achieved, the above process is repeated to re-analyze and adjust the process parameters.

[0055] Example 2: like Figure 1As shown, a three-tier architecture of edge acquisition, edge computing, and cloud backup is adopted, suitable for 24-hour continuous operation of the production line. Initially, a cloud computing architecture was considered, but it was found that the network in the industrial site was unstable and had high latency, which could not meet the production line's cycle time requirement of 30 pieces per minute. Therefore, an edge computing architecture was adopted instead, where all image processing and model inference are completed locally at the edge, and only the final results and backup data are uploaded to the cloud. Details are as follows: It mainly consists of an image acquisition device, an edge computing device, a database server, an RFID reader / writer, a human-machine interface device, and a cloud backup server. The image acquisition device includes six industrial cameras, six ring LED light sources, a PLC controller, a pneumatic positioning fixture, and a conveyor belt, responsible for synchronous image acquisition and component positioning and transmission. The edge computing device uses an industrial control computer; all image processing, feature extraction, defect identification, fractal calculation, and model inference are completed at the edge, with a single link detection time not exceeding 2 seconds, meeting production line cycle time requirements. The database server uses a Dell PowerEdge server running MySQL 8.0, storing all inspection data, model parameters, and inspection reports. The RFID reader / writer includes fixed readers, handheld readers, and vehicle-mounted readers. The human-machine interface device includes a touchscreen display, keyboard, and mouse, installed next to the inspection station for operators to view inspection results, set system parameters, and export inspection reports. The cloud backup server uses an Alibaba Cloud server, automatically backing up local database data to the cloud at 2 AM daily to prevent data loss.

[0056] The various components interact with each other via industrial Ethernet and communicate with the production MES system via the OPC UA protocol. OPC UA is an industrial communication protocol that enables standardized data exchange between devices from different manufacturers. It can automatically obtain production process parameters from the MES system and send inspection results and process adjustment suggestions to the MES system. During operation, the image acquisition device sends the acquired images to the edge computing device for processing. The processing results are stored in the local database and displayed on the touch screen. When a batch of defects is detected, the system automatically generates a process adjustment plan and sends it to the MES system. Every night, the system automatically backs up the day's data to the cloud. Every Sunday at 2:00 AM, the system automatically performs a collaborative incremental update on all models using newly added samples. During the update, a transfer learning algorithm is used to update only the parameters of the last two layers of the model, retaining the feature extraction capabilities of the previous layers to avoid catastrophic forgetting. The update process does not affect the normal operation of the system and maintains the accuracy of the model. It has a built-in complete anomaly handling mechanism, covering various anomaly scenarios commonly encountered in industrial production: When a single frame of data is missing during the detection process and the missing percentage is ≤30%, a Kriging interpolation algorithm based on causal completion is used to complete the data in combination with adjacent time-series data. Kriging interpolation is an interpolation method based on spatial correlation. We construct a covariance function by introducing the causal relationship between process parameters and defect features, which improves the completion accuracy by more than 40% compared with ordinary interpolation algorithms; when the missing percentage is >30%, valid data from two adjacent frames at the same location are retrieved for mean completion, and a data anomaly log is recorded simultaneously; when the detection... When a sensor detects no signal or a signal exceeding its range for more than 3 seconds, the system automatically switches to a pre-deployed backup detection channel at the same location. Simultaneously, it generates an equipment fault warning for the central control system via the Modbus-TCP protocol. Modbus-TCP is a commonly used communication protocol in industrial settings for equipment control and data acquisition, and it does not interrupt the current detection process. When the model's output results deviate from the measured values ​​by more than 15% for 3 consecutive frames, the model inference is paused, and the most recent 100 sets of valid samples from the same scenario are retrieved to complete the temporary fine-tuning and calibration of the model. After calibration, inference is resumed, and the model deviation log is recorded.

[0057] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for detecting and analyzing automotive parts based on image recognition, characterized in that, The analytical method involves the following steps: Multi-view images of automotive parts are acquired and preprocessed to extract three core features: surface texture, geometric dimensions, and edge contours. The three types of core features are compared with the preset defect feature template library to obtain preliminary defect identification results including defect location, defect type and defect coverage area, and then the preliminary quality assessment results are calculated. Multi-scale fractal feature extraction is performed on the identified defective regions to generate a fixed-length digital defect fractal fingerprint that integrates spatial fractal features and temporal fractal features; The fractal evolution entropy value is calculated based on the segmented information entropy of the defect fractal fingerprint, which quantifies the disorder degree and initial growth potential of the defect microstructure. By integrating the defect fractal fingerprints, fractal evolution entropy values, and corresponding production process parameters of all completed inspection parts on a production batch basis, a database linking defect fractal fingerprints and production process parameters is constructed. By using cluster analysis to extract common fractal feature templates of various defects from the association database, a bidirectional nonlinear mapping relationship between the common fractal feature templates and the deviations of production process parameters is established. By combining the production process correction coefficients output by fractal evolution entropy and bidirectional nonlinear mapping relationship, the microscopic growth path and failure critical point of defects within the preset service life can be predicted. The safety level of the defect location is determined based on the preset location safety level mapping relationship. The preliminary quality assessment results are dynamically corrected according to the defect failure threshold and the safety level of the defect location to generate the final quality assessment conclusion.

2. The method for detecting and analyzing automotive parts based on image recognition according to claim 1, characterized in that, The steps for calculating the preliminary quality assessment results based on the preliminary defect identification results are as follows: Three basic evaluation dimensions were determined: surface defect severity, dimensional deviation, and contour integrity. Individual scores were calculated for each basic evaluation dimension. Surface defect severity was weighted based on the number of defects, the area of ​​a single defect, and the basic importance of the defect's location. Dimensional deviation was scored based on the magnitude of the deviation between the actual size and the standard size. Contour integrity was scored based on the degree of contour offset. The weighting coefficients are determined based on the degree of impact of each basic evaluation dimension on the performance of the components. The individual scores of the three dimensions are multiplied by their corresponding weighting coefficients and then summed to obtain the preliminary quality assessment results.

3. The method for detecting and analyzing automotive parts based on image recognition according to claim 1, characterized in that, The defect fractal fingerprint is generated through the following steps: Based on the minimum side length of the circumscribed rectangle of the defect, the defect region is divided into a multi-level continuous scale square grid from a single pixel to a complete defect region; the number of grids containing defect pixels at each scale is counted, and the fractal scale-quantity relationship curve of the defect is plotted; the curve fitting method is used to calculate the local slope of the curve in each scale interval, and the corresponding number of spatial fractal dimension values ​​are obtained. Under the same acquisition conditions, multiple consecutive acquisitions are performed on the same defect area, and the change in the corresponding scale fractal dimension between adjacent acquisitions is calculated to obtain the time fractal characteristic value. The spatial fractal feature values ​​and the temporal fractal feature values ​​are quantized into digital codes of a preset number of bits and then concatenated. A check bit is added to form a unique defect fractal fingerprint.

4. The method for detecting and analyzing automotive parts based on image recognition according to claim 3, characterized in that, The fractal evolution entropy value is calculated using the following steps: The fixed-length defect fractal fingerprint is divided into multiple continuous equal-length sub-segments, each sub-segment corresponding to a specific fractal scale interval; the information entropy value of the digital code of each sub-segment is calculated to obtain the fractal entropy value of each fractal scale interval. Preset weight coefficients are assigned according to the principle that the weight of the microscale interval is greater than that of the macroscale interval. Each fractal entropy value is multiplied by its corresponding weight coefficient and then summed to obtain the total fractal evolution entropy value, which ranges from 0 to 1. Based on the magnitude of the total fractal evolution entropy value, defects are divided into different growth levels.

5. The method for detecting and analyzing automotive parts based on image recognition according to claim 1, characterized in that, The bidirectional nonlinear mapping relationship is established through the following steps: Clustering algorithms are used to perform cluster analysis on the defect fractal fingerprints in the associated database, and clustering parameters that meet the detection requirements are set; common features of all defect fractal fingerprints in each cluster are extracted to form a common fractal feature template for this type of defect; Collect records of multiple core production process parameters corresponding to each cluster; A forward mapping model from common fractal feature templates to deviations in production process parameters and a reverse mapping model from deviations in production process parameters to common fractal feature templates are trained using artificial neural networks. The bidirectional mapping model is trained and validated using historical production data, and put into use when the model's prediction accuracy reaches a preset threshold.

6. The method for detecting and analyzing automotive parts based on image recognition according to claim 5, characterized in that, The prediction process for the defect microscopic growth path and failure critical point is as follows: A defect growth dynamics model based on fractal theory is constructed, with fractal evolution entropy as the input and output of the basic growth rate of the defect; the defect fractal fingerprint is input into the forward mapping model to obtain the deviation values ​​of multiple core production process parameters that cause the defect; and the comprehensive production process correction coefficient is calculated by combining the influence of the deviation of each production process parameter on defect growth. The actual growth rate of the defect is obtained by multiplying the basic growth rate by the comprehensive production process correction coefficient; the fractal dimension change and defect area expansion of the defect within the preset service life are simulated based on the actual growth rate to generate the microscopic growth path curve of the defect; when the simulated fractal evolution entropy value reaches the preset failure threshold, the corresponding time point is the failure critical point of the defect.

7. The method for detecting and analyzing automotive parts based on image recognition according to claim 1, characterized in that, The analytical method also includes the following steps: The unique identifier of each component, the fractal fingerprint of defects, and the final quality assessment conclusion are written into the electronic identity tag to establish a full lifecycle quality traceability system based on the fractal fingerprint of defects. The establishment method of the full lifecycle quality traceability system is as follows: Each automotive component is assigned a unique identifier. Basic component information, production information, defect fractal fingerprints, final quality assessment results, and predicted failure time are encoded in a preset format and written into the storage area of ​​the electronic identifier. During the transportation, assembly, and maintenance of the component, the electronic identifier information is read by a reader / writer to record the status of each stage and bind the component information to the vehicle information. When a component malfunctions during use, the defect fractal fingerprint of the malfunctioning component is extracted and compared with the production stage defect fractal fingerprint stored in the electronic identifier. When the similarity reaches a preset threshold, it is determined that the malfunction originated from a defect generated during the production stage.

8. The method for detecting and analyzing automotive parts based on image recognition according to claim 5, characterized in that, The analytical method also includes the following steps: Based on the analysis results of the associated database, the deviation range of production process parameters corresponding to batch defects is located, and a targeted production process adjustment plan is generated. The adjustment effect is verified by the defect fractal fingerprint data of subsequent production batches. The process of generating and verifying the production process adjustment plan is as follows: The frequency of various defects within a single batch is statistically analyzed. When the frequency of a certain type of defect exceeds a preset threshold, it is determined to be a batch defect. The common fractal feature template of this type of batch defect is input into a forward mapping model to obtain the corresponding predicted deviation value of the production process parameters. The predicted deviation value is compared with the standard production process parameters to determine the specific production process parameters with deviations and the deviation range. Combining the degree of deviation of the production process parameters with the degree of impact on product quality, a production process adjustment plan is generated, which includes adjustment values, adjustment steps, and verification standards. The adjustment plan is implemented in multiple subsequent consecutive production batches. When the average fractal evolution entropy value of the same type of defect in multiple consecutive batches decreases to a preset standard, the production process adjustment is confirmed to be effective and solidified as a new production standard.

9. An image recognition-based automotive parts inspection and analysis system, characterized in that, The analysis system includes: The feature extraction module is used to acquire multi-view synchronous images of automotive parts and perform preprocessing to extract three core features: surface texture, geometric dimensions, and edge contours of the parts. The identification and evaluation module is used to perform multi-feature fusion comparison between the three types of core features and the preset defect feature template library to obtain preliminary defect identification results including defect location, defect type and defect coverage area, and calculate preliminary quality evaluation results. The defect fractal fingerprint generation module is used to extract multi-level continuous scale fractal features from the pixel level to the whole defect region of the identified defect region, and generate a fixed-length binary defect fractal fingerprint that integrates spatial fractal features and temporal fractal features. The fractal evolution entropy calculation module is used to calculate the fractal evolution entropy value based on the segmented information entropy of the defect fractal fingerprint, and to quantify the degree of disorder and initial growth potential of the defect microstructure. The associated database construction module is used to integrate the defect fractal fingerprints, fractal evolution entropy values ​​and corresponding production process parameters of all completed inspection parts in production batches, and to build an associated database of defect fractal fingerprints and production process parameters. The bidirectional mapping establishment module is used to extract common fractal feature templates of various defects in the association database through density clustering analysis, and establish a bidirectional nonlinear mapping relationship between the common fractal feature templates and the deviation of production process parameters. The defect growth prediction module is used to combine the production process correction coefficients output by the fractal evolution entropy value and the bidirectional nonlinear mapping relationship to predict the microscopic growth path and failure critical point of defects within a preset service life. The quality assessment correction module is used to determine the safety level of the location of the defect based on the preset safety level mapping table of critical locations of components, and to dynamically weight and correct the preliminary quality assessment results according to the defect failure threshold and the safety level of the location of the defect, so as to generate the final quality assessment conclusion. The full lifecycle traceability module is used to write the unique identifier of the component, the defect fractal fingerprint, and the final quality assessment conclusion into the passive electronic identification tag, and establish a full lifecycle quality traceability system based on the defect fractal fingerprint. The process adjustment module is used to locate the deviation range of production process parameters corresponding to batch defects based on the analysis results of the associated database, generate targeted production process adjustment plans, and verify the adjustment effect through defect fractal fingerprint data of subsequent production batches.

10. The image recognition-based automotive parts inspection and analysis system according to claim 9, characterized in that, The defect fractal fingerprint generation module includes: Multi-scale meshing unit is used to divide the defect region into multi-level continuous-scale square meshes from 1×1 pixels to the complete defect region, based on the minimum side length of the defect's circumscribed rectangle. The fractal curve plotting unit is used to count the number of grids containing at least one defect pixel at each scale and plot the fractal scale-number double logarithmic curve of the defect. The spatial fractal calculation unit is used to fit the local slope of the curve in each scale interval using the least squares method to obtain the corresponding number of spatial fractal dimension values. The time-fractal calculation unit is used to perform multiple consecutive acquisitions of the same defect area under the same acquisition conditions, calculate the absolute difference of the corresponding scale fractal dimension between adjacent acquisitions, and obtain the time-fractal feature value. The fractal fingerprint encoding unit is used to quantize the spatial fractal feature values ​​and the temporal fractal feature values ​​into binary codes of a preset number of bits, then concatenate them and add a check bit to form a unique defect fractal fingerprint. The bidirectional mapping establishment module includes: The density clustering analysis unit is used to perform clustering analysis on the defect fractal fingerprints in the associated database using the density clustering algorithm, and to set clustering parameters that meet the detection requirements. The common template extraction unit is used to extract the mode of each bit encoding of all defect fractal fingerprints in each cluster, forming a common fractal feature template for that type of defect; The production process parameter collection unit is used to collect records of multiple core production process parameters corresponding to each cluster, including processing temperature, processing pressure, processing time, tool speed and coolant flow rate. The bidirectional model training unit is used to train, through a multi-layer neural network, a forward mapping model from common fractal feature templates to production process parameter deviations, and a reverse mapping model from production process parameter deviations to common fractal feature templates. The model validation unit is used to train and validate the bidirectional mapping model using historical production data. It is put into use when the model's prediction accuracy reaches a preset threshold.

Images