Automobile brake assembly abnormality identification method and system based on image recognition

Abstract

The application provides an image recognition-based automobile brake assembly abnormality identification method and system. The method first collects the original image of the hub bearing nut entering the assembly station and records the batch information. The deep learning model is used to extract the visual fingerprint features and compare them with the pre-stored defect features to complete the risk assessment. Then, the images before and after the anchoring operation are collected at the same position before and after the hub bearing nut is tightened. Through image registration and difference operation, the difference image is generated. The machine learning model is used to analyze the microscopic physical change characteristics caused by the anchoring operation, and the quality judgment result is output. The system integrates the early risk assessment result and the quality judgment result to identify whether the assembly state is abnormal. If an abnormality is found, all information is packaged as an abnormal data packet, the encryption hash value is calculated, and it is stored in the distributed block chain. The method realizes the identification and tracing of abnormalities in the brake assembly process, and improves the quality control level of the automobile safety critical components.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a method and system for identifying abnormal assembly of automobile brakes based on image recognition. Background Technology

[0002] In the automotive manufacturing industry, the assembly quality of the braking system directly affects vehicle safety performance, with the correct installation of wheel hub bearing nuts being particularly critical. The most significant problem with existing technologies lies in the difficulty of achieving real-time and accurate identification of minute anomalies during brake assembly, especially those subtle defects that are difficult to detect with the naked eye but could lead to serious safety hazards. Traditional quality inspection methods typically rely on manual visual inspection or simple mechanical measurements, which are insufficient to capture microscopic changes and potential risks during the assembly process.

[0003] Especially on high-speed production lines, inspection time is limited, and manual inspection is easily affected by factors such as fatigue and subjective judgment, leading to a high rate of missed detections. Conventional automated inspection equipment can often only detect obvious geometric deviations or surface defects, and cannot effectively identify hidden problems such as minute deformations and abnormal stress distribution that occur during assembly. Although these hidden problems are not obvious during the assembly stage, they may gradually worsen during vehicle use, eventually leading to braking system failure. Summary of the Invention

[0004] This invention provides a method and system for identifying abnormal assembly of automotive brakes based on image recognition, in order to solve the above-mentioned technical problems.

[0005] In a first aspect, the present invention provides a method for identifying abnormal assembly of automotive brakes based on image recognition, the method comprising the following steps:

[0006] The beneficial effects of this invention are:

[0007] This invention, by integrating deep learning and image differential analysis technologies, achieves accurate identification and full-process traceability of minute anomalies during automotive brake assembly, significantly improving the quality control level of critical automotive safety components. Compared to traditional methods relying on manual inspection or simple mechanical measurement, this invention can automatically capture microscopic physical changes that are difficult to detect with the naked eye, improving detection accuracy to the sub-millimeter level and greatly reducing the missed detection rate and false judgment rate. In particular, through precise registration and differential analysis of images before and after anchoring, this invention can intuitively present the minute deformations and stress distribution anomalies generated during tightening. These hidden defects are often overlooked in traditional inspections but may be the root cause of future safety hazards. The early risk assessment mechanism of this invention, through visual fingerprint feature extraction, can identify potential problems before parts enter the assembly stage, realizing forward quality control and avoiding resource waste caused by unqualified parts flowing into subsequent processes. In addition, this invention innovatively introduces blockchain technology into the quality traceability system, ensuring the authenticity, integrity, and immutability of abnormal data, providing a reliable basis for quality responsibility determination and batch recall. In practical applications, this invention not only improves the automation level and inspection efficiency of production lines, but also forms a self-evolving quality control system through data accumulation and model optimization, continuously enhancing its ability to identify new types of defects. This comprehensive, high-precision, and traceable quality control solution provides strong technical support for the automotive manufacturing industry, effectively ensuring the reliability of braking systems and vehicle driving safety. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating an image recognition-based method for identifying abnormal assembly of automotive brakes in one embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the system structure of the automobile brake assembly anomaly identification system in one embodiment of this application. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0011] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0012] Figure 1 This is a flowchart illustrating an image recognition-based method for identifying assembly anomalies in automotive brake systems, as shown in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps. For example Figure 1 As shown, the image recognition-based method for identifying assembly anomalies in automotive brakes disclosed in this invention specifically includes the following steps:

[0013] S101. Collect the original image of the wheel hub bearing nut entering the assembly station, and identify and record the batch information of the wheel hub bearing nut.

[0014] When the wheel hub bearing nut enters the assembly station, a high-resolution industrial camera takes multi-angle photos of it under standardized lighting conditions, acquiring raw images containing the nut's surface texture, geometry, and material characteristics. Simultaneously, optical character recognition (OCR) technology or a QR code scanner reads the batch identifier on the nut's surface, linking information such as the batch number, production date, and supplier code with the image data. For example, when the batch number "HN2024A001" is detected, the camera captures front and side images of the nut at a resolution of 2048×2048 pixels, ensuring image clarity meets subsequent analysis requirements.

[0015] S102. Use a deep learning model to extract the visual fingerprint features of the wheel hub bearing nut from the original image, and compare the visual fingerprint features with the pre-stored visual defect features to complete the early risk assessment of the wheel hub bearing nut.

[0016] The deep learning model first preprocesses the original image, including Gaussian filtering for noise reduction, histogram equalization, and illumination normalization. Then, it uses the Canny edge detection algorithm to accurately segment the region of interest (ROI) of the nut. The preprocessed image is input into a pre-trained convolutional autoencoder network, which consists of a 5-layer encoder and a 5-layer decoder structure. A non-linear transformation using the ReLU activation function is applied to extract a 512-dimensional latent space feature vector from the bottleneck layer as the initial visual fingerprint. Principal component analysis is then applied to reduce the feature dimension to 128 dimensions, retaining 95% of the variance information. The optimized visual fingerprint features are then compared with defect features stored in a historical risk database using a weighted cosine similarity calculation. When the similarity exceeds a threshold of 0.85, it is marked as a preliminary abnormal risk; otherwise, it is marked as normal.

[0017] S103. Before the wheel hub bearing nut is tightened and anchored, an image before anchoring is captured. After the same wheel hub bearing nut is tightened and anchored, an image after anchoring is captured at the same shooting position.

[0018] During the critical tightening and anchoring operation of the nut, a high-precision camera system with a fixed position is used for image acquisition. Before anchoring, the camera takes a reference image when the nut is not in contact with the tool, recording the initial state of the nut, including surface gloss, texture distribution, and geometric contour. After the anchoring operation is completed, the camera acquires an image of the nut under exactly the same shooting position, angle, and lighting conditions, ensuring that the spatial coordinate deviation between the two shots is controlled within 0.1 mm. A precise mechanical positioning device and laser guidance system ensure high consistency of camera position. For example, when the torque wrench completes a tightening operation of 35 N·m, the camera immediately triggers an image capture, recording the microscopic changes on the nut surface caused by the force.

[0019] S104. Perform image registration and difference operations on the images before and after anchoring to generate a difference image. Use a machine learning model to analyze the microscopic physical changes caused by the anchoring operation in the difference image and output the quality judgment result of the anchoring operation.

[0020] Accurate registration of the images before and after anchoring is a crucial step in detecting microscopic changes. First, the SURF (Accelerated Robust Feature) and ORB algorithms are used to detect local feature points in key structural regions of the two images. A FLANN matcher is then used to establish the correspondence between these feature points. Next, the RANSAC algorithm is used to eliminate mismatches, retaining only accurate correspondences with a confidence level higher than 0.9. Based on these feature points, a homography transformation matrix is ​​calculated to align the post-anchoring image with the pre-anchoring image at the sub-pixel level. After alignment, normalized gray-level difference operations are performed to calculate the gray-level difference at each pixel location, generating an initial difference image containing information on positive and negative changes. Subsequently, Gaussian filtering is applied to smooth noise, and adaptive threshold binarization is used to highlight areas of significant change. The final difference image clearly displays the microscopic physical characteristics caused by the anchoring operation, such as surface deformation, tool contact marks, and material plastic deformation.

[0021] The differential image input is based on an anchoring anomaly recognition model built on the DenseNet architecture. This model is trained using a dataset containing 5000 samples of different anchoring states. The model's deep convolutional layers extract multi-level feature maps representing the morphology of the contact area, the degree of minor plastic deformation of the material, surface gloss, and texture variations. Each feature map contains 64 to 512 different feature channels. Discrete category judgments of anchoring quality are output through fully connected layers and Softmax classification layers, including three levels: "acceptable," "minor anomaly," and "serious anomaly," along with confidence scores. A parallel attention mechanism generates heatmaps to accurately locate the coordinates of anomalous features in the differential image. When the judgment result is unacceptable, the model decodes and outputs a specific anomaly description from the feature map, such as "uneven thread contact" or "over-tightening causing material deformation." All outputs are integrated to form a complete anchoring operation quality judgment report.

[0022] S105. Identify whether there are any assembly abnormalities in the assembly status of the wheel hub bearing nuts by combining the assessment results of the early risk assessment with the quality judgment results.

[0023] The anomaly decision engine, built on weighted fuzzy logic, receives early risk assessment results and anchoring quality judgment results as input variables. Fuzzy membership functions are preset for different input combinations; for example, when the risk assessment is "high risk" and the quality judgment is "severe anomaly," the membership function value is set to 0.9. A rule base containing 25 IF-THEN rules is established, such as "IF early risk is moderate AND anchoring quality is slightly abnormal THEN assembly status is suspicious." The decision engine converts the input values ​​into fuzzy sets through a fuzzification process and applies the Mamdani inference method for fuzzy reasoning. Finally, a comprehensive evaluation score of 0-100 is obtained through centroid defuzzification. The score is compared with preset thresholds: a score below 30 is normal, 30-70 is suspicious and requires re-inspection, and a score above 70 is abnormal and requires processing.

[0024] S106. If there is an assembly abnormality in the assembly status of the wheel hub bearing nut, all image information, batch information, early risk assessment results and quality judgment results of the collected wheel hub bearing nut are packaged into an abnormal data packet, the encrypted hash value of the abnormal data packet is calculated, and the abnormal data packet and the encrypted hash value are stored in the preset distributed blockchain.

[0025] When an assembly anomaly is detected, all relevant data is first timestamped to ensure consistency in time between image information, batch information, evaluation results, and judgment results. The synchronized critical anomaly data is then encoded using the Apache Avro serialization format to generate a compact binary anomaly data packet, typically 2-5MB in size. The Keccak-256 hash function from the SHA-3 algorithm family is applied to encrypt the data packet, generating a 64-bit hexadecimal fixed-length hash value, such as "a1b2c3d4e5f6...", which serves as the unique content fingerprint of the data packet. A smart contract encapsulates the encrypted hash value and the anomaly data packet into a blockchain transaction, which is then verified using an elliptic curve digital signature algorithm. Finally, the data is stored in a pre-defined consortium blockchain or private blockchain. Each block contains a timestamp, the hash value of the previous block, and transaction data, ensuring data immutability and complete traceability, providing a reliable chain of evidence for subsequent investigations of quality issues.

[0026] In one implementation, extracting the visual fingerprint features of the wheel hub bearing nut from the original image using a deep learning model includes the following steps:

[0027] The original image is preprocessed, including image denoising, illumination normalization, and precise segmentation of the region of interest based on contour detection.

[0028] The preprocessed original image is input into a pre-trained convolutional autoencoder network, and the encoder part of the convolutional autoencoder network performs nonlinear compression transformation on the original image.

[0029] After performing nonlinear compression transformation on the original image, the latent spatial feature vector of the original image is extracted from the intermediate bottleneck layer of the convolutional autoencoder network as the initial visual fingerprint of the wheel hub bearing nut.

[0030] Principal component analysis was applied to the initial visual fingerprint for dimensional optimization and feature selection to enhance its ability to express defect-sensitive features, thus forming the visual fingerprint features of the wheel hub bearing nut.

[0031] In this embodiment, preprocessing of the original image is a crucial foundational step to ensure the accuracy of subsequent analysis. First, image denoising is performed using a Gaussian filter to eliminate random noise generated by the camera sensor. The filter kernel size is set to 5×5, the standard deviation σ=1.2, and the filtering formula is as follows: Next, illumination normalization is performed, and a histogram equalization algorithm is used to adjust the image brightness distribution, ensuring that pixel values ​​are evenly distributed within the range of 0-255, thus eliminating the impact of different lighting conditions on image quality. Finally, region of interest segmentation is performed, and the Canny edge detection algorithm is used to identify the nut contour. Dual threshold parameters are set to 50 and 150, and morphological operations are used to fill the contour, accurately extracting the nut region. For example, for a 1024×1024 pixel original image, after preprocessing, a clear 512×512 pixel nut region image is obtained, with background noise reduced by 80% and brightness standard deviation controlled within 15, providing high-quality input data for subsequent feature extraction.

[0032] The preprocessed image is input into a pre-trained convolutional autoencoder network for deep feature learning. This network employs a symmetric encoder-decoder architecture, with the encoder containing five convolutional layers, each using a 3×3 kernel with a stride of 2 and ReLU activation. The first layer outputs 64 feature maps, followed by layers with 128, 256, 512, and 1024 feature maps respectively. Each convolutional layer is followed by a batch normalization layer and a max-pooling layer, progressively reducing spatial resolution while increasing feature depth. Through this non-linear compression transformation, the 512×512 input image is progressively compressed into a high-dimensional feature representation of 16×16×1024. The network is pre-trained on a dataset containing 10,000 nut samples, learning abstract representations of nut surface texture, geometry, and material properties, effectively capturing key visual features affecting assembly quality.

[0033] After the convolutional autoencoder completes the nonlinear compression transformation, the latent space feature vector is extracted from the network's intermediate bottleneck layer as the initial visual fingerprint of the nut. The bottleneck layer, located at the last layer of the encoder, contains 16×16×1024 neurons. Global average pooling is used to compress the spatial dimension, resulting in a 1024-dimensional feature vector. This 1024-dimensional vector contains high-level semantic information about the nut image, with each dimension representing a specific visual pattern or feature combination. For example, some dimensions might correspond to the regularity of the threads, while others might reflect surface roughness or material gloss. t-SNE visualization analysis reveals that nuts with similar quality conditions cluster together in the latent space, while nuts with different defect types show clear separation in space, demonstrating the effectiveness and discriminative power of the feature vector.

[0034] Principal component analysis (PCA) was applied to the initial 1024-dimensional visual fingerprint for dimensionality optimization and feature selection to enhance its ability to express defect-sensitive features. First, the covariance matrix of the eigenvectors was calculated. Then, the eigenvalues ​​and eigenvectors are solved, and the eigenvalues ​​are sorted in descending order. The top k principal components with a cumulative variance contribution rate of 95% are selected, typically with k values ​​between 128 and 256. The transformation formula is as follows. ,in The principal component matrix is ​​used. Through PCA transformation, the original 1024-dimensional features are compressed to 128 dimensions while retaining the most important change information.

[0035] In one implementation, comparing visual fingerprint features with pre-stored visual defect features to complete an early risk assessment of wheel hub bearing nuts includes the following steps:

[0036] Access the historical risk database, which pre-stores visual defect features corresponding to nut samples of various defect types;

[0037] Weighted cosine similarity is used to calculate the matching score between visual fingerprint features and visual defect features;

[0038] By combining the matching score and the multi-level matching score thresholds set for different defect types, it is determined whether the visual fingerprint features have a significant statistical association with any defect type.

[0039] If a significant statistical association exists, the early risk assessment result of the preliminary anomaly of the corresponding wheel hub bearing nut mark is based on the associated defect type and the confidence interval of the matching score, and the visual fingerprint features are updated to the historical risk database.

[0040] If there is no significant statistical correlation or the matching score is lower than the preset warning threshold, the early risk assessment result is that the corresponding wheel hub bearing nut marking is initially normal.

[0041] In this implementation, the historical risk database employs a distributed storage architecture, pre-storing visual defect features corresponding to nut samples of 15 common defect types. The database is categorized by defect type, including thread damage, surface cracks, material inhomogeneity, dimensional deviations, and corrosion spots. Each defect type contains 500-2000 samples with 128-dimensional feature vectors. The database uses a B+ tree index structure, supporting fast retrieval and range queries, with an average query time controlled within 10 milliseconds. Access is achieved through the SQL query "SELECT feature_vector, defect_type, confidence_level FROM defect_database WHERE defect_type IN ('thread_damage', 'surface_crack', 'material_uneven')" to retrieve relevant feature data. The database also maintains statistical information for each defect type, including the mean, standard deviation, and distribution parameters of the feature vectors. This efficient database access mechanism provides a complete reference dataset for subsequent similarity calculations.

[0042] The weighted cosine similarity algorithm is used to calculate the matching degree between the visual fingerprint features of the current nut and various defect features in the database. The weighted cosine similarity formula is:

[0043]

[0044] in This is the current nut feature vector. For the feature vector of the defect sample, Let be the weight coefficient of the i-th feature. The weight coefficient is determined based on the feature dimension's ability to distinguish different defect types, and is calculated using the information gain algorithm. The weight of important feature dimensions can reach 2.5, and the weight of secondary dimensions is 0.3. During the calculation process, the similarity is calculated for all samples of each defect type, and then the average value is taken as the overall matching score for that defect type.

[0045] The statistical association between visual fingerprint features and various defect types was determined based on matching scores and a multi-level threshold system. Three levels of matching score thresholds were set for different defect types: high-risk threshold T_h = 0.85, medium-risk threshold T_m = 0.70, and low-risk threshold T_l = 0.55. Statistical association was determined using hypothesis testing: the null hypothesis H_0 was "no association between the current sample and the defect type," and the alternative hypothesis H_1 was "a significant association exists." The t-statistic was calculated. Here, μ_0 represents the historical average similarity of this defect type, σ_0 is the standard deviation, and n is the sample size. When the t-value is greater than the critical value t_0.05 = 1.96, the null hypothesis is rejected, and a significant statistical association is considered to exist. For example, if the matching score between a nut and thread damage is 0.82, exceeding the high-risk threshold of 0.85, and the t-value is 2.34 > 1.96, a significant association is determined. Simultaneously, confidence intervals are considered; when the 95% confidence interval of the matching score is [0.78, 0.86], the reliability of the association is further confirmed.

[0046] When a significant statistical association is detected, a detailed early risk assessment result is assigned to the nut based on the associated defect type and the confidence interval of the matching score. Risk levels are divided into four tiers: extremely high risk (match score > 0.90), high risk (0.80–0.90), medium risk (0.65–0.80), and low risk (0.55–0.65). The assessment result includes defect type, risk level, confidence level, and recommended actions. For example, a thread damage association with a matching score of 0.87 is labeled as "Defect type: thread damage, risk level: high risk, confidence level: 92%, recommendation: prioritize checking thread integrity." Simultaneously, the visual fingerprint features of the current nut are used as new samples to update the historical risk database, and the statistical parameters of the defect type are updated using an incremental learning approach. The update formula is as follows: , .

[0047] When the matching scores of the visual fingerprint features with all defect types are all below the preset warning threshold of 0.55, or there is no significant statistical correlation, the corresponding nut mark is considered a preliminary normal early risk assessment result. The normal state judgment employs a multiple verification mechanism: first, it checks whether the matching scores of all 15 defect types are below their respective low-risk thresholds; then, it verifies whether the highest matching score is below the global warning threshold of 0.55. In the statistical correlation test, the t-statistics for all defect types are less than the critical value of 1.96, and the p-value is greater than 0.05, thus accepting the null hypothesis. The normal state mark includes detailed information: "Risk level: Normal, Highest matching score: 0.42 (surface roughness), Confidence level: 96%, Recommendation: Continue assembly according to standard procedures." To ensure the reliability of the judgment, Mahalanobis distance is also calculated. This is used to measure the distance between the current sample and the center of the normal sample distribution. When D_M < 3.0, the normal state is further confirmed.

[0048] In one embodiment, performing image registration and difference operations on the pre-anchoring image and the post-anchoring image to generate a difference image includes the following steps:

[0049] Local feature points were detected in key structural regions of the pre-anchoring and post-anchoring images using accelerated robust features and ORB algorithms, respectively.

[0050] FLANN matching is performed based on the descriptors of local feature points to establish the initial correspondence between local feature points in the pre-anchoring and post-anchoring images. Then, erroneous matching point pairs are eliminated by random sampling consensus algorithm to obtain the filtered precise corresponding feature points.

[0051] Based on the precise correspondence of feature points, the homography transformation model is used to align the image after anchoring with the image before anchoring at the sub-pixel level.

[0052] The normalized gray values ​​are subtracted from the corresponding pixel positions in the aligned pre-anchoring image and the post-anchoring image to obtain an initial difference image containing positive and negative difference information.

[0053] The initial difference image is subjected to Gaussian filtering to smooth the noise, and an adaptive threshold binarization method is applied to highlight the microscopic physical changes caused by the anchoring operation, thus obtaining the difference image.

[0054] In this embodiment, SURF (Speed-Up Robust Features) and ORB (Oriented FAST and Rotated BRIEF) algorithms are applied to detect local feature points in key structural regions of the images before and after anchoring, respectively, to provide basic anchor points for subsequent image registration. The SURF algorithm first constructs a scale space and detects interest points using the determinant of the Hessian matrix, calculated as follows: ,when The time is determined as a feature point. The ORB algorithm combines FAST corner detection and the BRIEF descriptor, using the Harris corner response function. Stable feature points are selected, where M is the structure tensor matrix and k is typically set to 0.04. In practical applications, the SURF algorithm detects approximately 150 feature points in the threaded region of the nut, while the ORB algorithm detects approximately 200 feature points in the surface texture region. The combined use of the two algorithms ensures that a sufficient number of feature points can be stably detected under different lighting and angle variations, laying a solid foundation for accurate registration.

[0055] Based on the 128-dimensional SURF descriptor and 32-dimensional ORB descriptor of the detected local feature points, the FLANN (Fast Nearest Neighbor Search) matching algorithm is used to establish the initial correspondence between feature points in the images before and after anchoring. FLANN uses a KD-tree or LSH hash table for fast nearest neighbor search, with the matching criterion being the minimum Euclidean distance. The distance calculation formula is as follows: To improve matching quality, the Lowe's ratio test is used. A match is considered valid when the ratio of the nearest neighbor distance to the second nearest neighbor distance is less than 0.7. Initial matching typically generates 300-500 matching point pairs, but includes a large number of false matches. Subsequently, the RANSAC (Random Sample Consensus) algorithm is applied to remove erroneous matches. During the iteration process, four matching points are randomly selected to calculate the homography matrix, and the number of inliers is counted. Points with a reprojection error of less than 3 pixels are considered inliers.

[0056] Based on the selected precisely corresponding feature points, a homography transformation model is used to achieve sub-pixel-level precise alignment between the image after anchoring and the image before anchoring. The homography matrix H is a 3×3 homogeneous transformation matrix, and the transformation relationship is as follows:

[0057] ,in .

[0058] The eight unknown parameters are solved using the least squares method, and the objective function is: To achieve sub-pixel accuracy, bilinear interpolation is used for pixel value resampling. The interpolation formula is as follows:

[0059]

[0060] Where α and β are the decimal parts. The registration accuracy after transformation is usually within 0.1 pixels, effectively eliminating the small displacements, rotations and scale changes between the two shots, creating conditions for accurately detecting the real physical changes caused by the anchoring operation.

[0061] After precise alignment, the images before and after anchoring are subjected to a subtraction operation of normalized grayscale values ​​to generate an initial difference image containing positive and negative difference information. First, grayscale normalization is performed on both images to ensure consistent pixel value ranges. The normalization formula is as follows: Then perform a difference operation at the same pixel position. This yields difference values ​​in the range [-255, 255]. For ease of display and subsequent processing, these difference values ​​are mapped to the range [0, 255] using the following mapping formula: Positive difference values ​​(bright areas) indicate that the area becomes brighter after anchoring, usually corresponding to material compression or tool contact marks; negative difference values ​​(dark areas) indicate that the area becomes darker, which may reflect material deformation or changes in shadow.

[0062] The initial difference image is smoothed using Gaussian filtering to eliminate noise interference, and then adaptive threshold binarization is applied to highlight significant microscopic physical changes. The Gaussian filtering uses a 5×5 convolution kernel with a standard deviation σ=1.5, and the filter weights are calculated according to a two-dimensional Gaussian distribution. This effectively suppresses high-frequency noise while preserving edge information. Adaptive threshold binarization uses local statistical properties to determine the threshold, calculating the mean and standard deviation of each pixel's neighborhood. The threshold formula is as follows: , where μ is the local mean, σ is the local standard deviation, and k is an adjustment parameter, usually set to 0.2. It is set to 255 (white) when the pixel value is greater than the local threshold, and 0 (black) otherwise.

[0063] In one implementation, analyzing the microscopic physical changes in the difference image caused by the anchoring operation using a machine learning model and outputting a quality judgment result for the anchoring operation includes the following steps:

[0064] The differential image is input into the anchoring anomaly recognition model built on a densely connected network. The anchoring anomaly recognition model is pre-trained by supervised learning using differential image samples containing different anchoring states.

[0065] The deep convolutional layer of the anchoring anomaly identification model extracts hierarchical feature maps from the difference image to represent the morphology of the tool contact area, the degree of minor plastic deformation of the material, and the changes in surface gloss and texture.

[0066] Based on hierarchical feature maps, the model outputs discrete category judgments of anchorage quality through fully connected layers and Softmax classification layers of the anchorage anomaly recognition model. At the same time, through parallel regression branches or attention mechanisms in the anchorage anomaly recognition model, the model outputs confidence scores for anchorage quality judgments and heatmaps indicating the location of abnormal micro-features in the difference image.

[0067] If the discrete category judgment result is a non-qualified state, then decode from the hierarchical feature map and output a textual description of the microscopic physical change features associated with the non-qualified state;

[0068] The text description and all outputs of the anchoring anomaly identification model are integrated into a quality judgment result for the anchoring operation.

[0069] In this embodiment, the differential image input is based on an anchoring anomaly detection model built on the DenseNet architecture. This model employs a densely connected block structure to achieve feature reuse and gradient flow optimization. The model contains four dense blocks, and the layers within each block are directly connected. The connection method is as follows: ,in This indicates the concatenation of the outputs of all preceding layers. The composite function of layer l includes batch normalization, ReLU activation, and 3×3 convolution operations. The model is pre-trained using a dataset containing 8000 differential image samples of different anchoring states, covering 12 state categories including normal anchoring, over-tightening, under-tightening, and eccentric contact. The training process employs the cross-entropy loss function.

[0070] The deep convolutional layers of the anchoring anomaly identification model employ a multi-scale feature extraction strategy, extracting hierarchical feature maps from the difference image layer by layer to characterize the morphology of the tool contact area, material plastic deformation, surface gloss, and texture changes. The first convolutional layer extracts 64 basic 7×7 feature maps, capturing edge and basic texture information; the second layer outputs 128 5×5 feature maps to identify the geometry of the tool contact; the third layer generates 256 3×3 feature maps to detect subtle deformation patterns in the material; and the fourth layer produces 512 feature maps to analyze the variation patterns of surface gloss. The calculation formula for each layer's feature map is as follows: ,in It is the ReLU activation function. For convolution kernel weights, This represents a convolution operation. Through dense connections, deep feature maps can access all shallow information, forming rich feature representations. For example, when detecting over-tightening of threads, shallow feature maps show the sharpening of contact edges, while deep feature maps capture the changes in texture density caused by material compression.

[0071] Based on the extracted hierarchical feature maps, discrete category judgments of anchorage quality are output through fully connected layers and softmax classification layers. Simultaneously, confidence scores and anomaly location heatmaps are generated using parallel regression branches and attention mechanisms. The fully connected layer maps the 960-dimensional feature vector to predicted scores for 12 categories, and the softmax function calculates the probability of each category. ,in Let be the original score for the i-th class. The final class classification is the class with the highest probability. The parallel regression branch outputs a confidence score, which is then processed by the sigmoid function. The regression values ​​are mapped to the [0,1] interval. The attention mechanism uses the Grad-CAM algorithm to generate a heatmap, and the calculation formula is as follows: ,in The weights of the k-th feature map are... This is the corresponding feature map.

[0072] When the discrete category judgment result is a non-qualified state, a detailed textual description of the microscopic physical change features associated with the non-qualified state is decoded from the hierarchical feature map and output. The decoding process employs a feature-text mapping network based on an attention mechanism, which is pre-trained to establish a correspondence between the feature map and the descriptive text. First, the feature map is compressed into a feature vector through global average pooling. Then, the text description is generated by inputting it into the LSTM decoder. The hidden state update formula of the LSTM is: The probability distribution of the output words is as follows For example, for the "overtightened" state, the decoding output is "obvious indentations appear in the thread contact area, the material undergoes plastic deformation, the surface gloss decreases by 15%, the texture density increases, and it is recommended to check the torque setting value".

[0073] The text description and all outputs of the anchoring anomaly identification model are integrated into a complete anchoring operation quality judgment report, forming a standardized quality assessment document. The integration process organizes information according to a predefined template structure, including six parts: basic information, classification results, confidence assessment, anomaly location, feature description, and handling recommendations. Basic information records the detection time, image resolution, and processing time; classification results display the probability distribution of 12 categories and the final judgment; confidence assessment provides a quantitative indicator of the judgment's reliability; anomaly location accurately marks the problem area using heatmap coordinates; feature descriptions use natural language to detail the detected physical changes; and handling recommendations provide operational guidance based on the anomaly type and severity. For example, the complete report format is: "Detection time: 2024-03-15 14:32:05, Classification result: Overtightening (probability 0.89), Confidence: 89%, Anomaly location: (128, 156) radius 25 pixel area, Feature description: Thread root indentation depth 0.02mm exceeds the standard, Handling recommendation: Adjust the torque to the standard value of 30±2 N·m and reassemble."

[0074] In one embodiment, identifying whether there is an assembly abnormality in the assembly state of the wheel hub bearing nut by combining the assessment results of the early risk assessment with the quality judgment results includes the following steps:

[0075] An anomaly decision engine is constructed based on weighted fuzzy logic. The inputs of the anomaly decision engine are the results of early risk assessment and quality judgment.

[0076] Pre-set fuzzy membership functions and IF-THEN rule bases for different combinations of early risk assessment results and quality judgment results;

[0077] By combining fuzzy membership functions and the IF-THEN rule base, and through the fuzzification, fuzzy reasoning and defuzzification processes of the anomaly decision engine, the comprehensive evaluation score of the wheel hub bearing nut assembly status is calculated.

[0078] The comprehensive evaluation score is compared with the preset multi-level thresholds, and the assembly status of the wheel hub bearing nut is judged based on the comparison results to determine whether there is an assembly abnormality.

[0079] In this embodiment, an anomaly decision engine is constructed based on weighted fuzzy logic theory. This engine receives early risk assessment results and quality judgment results as dual input variables for comprehensive decision analysis. The decision engine adopts the Mamdani fuzzy inference system architecture, which includes four core components: a fuzzification interface, a rule base, an inference engine, and a defuzzification interface. The input range for the early risk assessment results is [0,1], where 0 represents no risk and 1 represents extremely high risk; the input range for the quality judgment results is also [0,1], where 0 represents fully qualified and 1 represents severely abnormal. The engine uses a weighted mechanism to handle the difference in importance between the two inputs, and the weight allocation formula is as follows: The decision engine outputs a comprehensive evaluation score, ranging from [0, 100]. A higher score indicates a greater likelihood of assembly anomalies. Through the uncertainty handling capabilities of fuzzy logic, the engine effectively integrates two different types of evaluation information, overcoming the limitations of traditional binary logic in handling boundary conditions, and providing more flexible and accurate decision support for complex assembly quality judgments.

[0080] Detailed fuzzy membership functions and a complete IF-THEN rule base are pre-defined for different combinations of early risk assessment and quality judgment results. Early risk assessment uses triangular membership functions to define five linguistic variables:

[0081] Extremely low risk

[0082] Low risk

[0083] Medium risk

[0084] High risk

[0085] Extremely high risk

[0086] The quality assessment results adopt the same five-level classification structure. The rule base contains 25 IF-THEN rules, covering all possible input combinations, such as "IF Early risk is high risk AND Quality assessment is moderate abnormality THEN Assembly status is severe abnormality". Each rule is configured with a corresponding weight coefficient, with key rules having a weight of up to 1.0 and secondary rules having a weight of 0.6.

[0087] Combining a pre-defined fuzzy membership function and rule base, the anomaly decision engine calculates a comprehensive evaluation score for the assembly status of the wheel hub bearing nut through three stages: fuzzification, fuzzy inference, and defuzzification. The fuzzification stage converts the precise input values ​​into fuzzy sets and calculates the membership values ​​of each linguistic variable. The fuzzy inference stage uses a minimum-maximum composition operation to calculate and output a fuzzy set for each activation rule, using the formula: ,in Let be the activation strength of the i-th rule. This corresponds to the output membership function. Defuzzing uses the centroid method to calculate the final precise output value, as shown in the formula. ,in These are the discrete points of the output universe of discourse.

[0088] The calculated comprehensive evaluation score is compared with a preset multi-level threshold system to accurately determine whether there are any abnormalities in the assembly status of the wheel hub bearing nut. The multi-level threshold system includes four key nodes: normal threshold... Suspicious threshold Abnormal threshold Severe abnormal threshold The judgment logic is: when When it is determined to be in a normal state, no special treatment is required; when If the condition is deemed suspicious, a manual re-examination is recommended; when The initial assessment indicated a minor abnormality, requiring reassembly; when The condition was initially determined to be a moderate abnormality, requiring replacement of the part; when If an error is detected, it is considered a serious anomaly, and the system must be shut down for inspection. Threshold settings are determined based on ROC curve analysis and cost-benefit assessment to ensure a false positive rate of less than 3%. For example, if the comprehensive evaluation score of 72.3 exceeds the anomaly threshold of 65, the system classifies it as a moderate anomaly and automatically triggers a parts replacement process.

[0089] In one implementation, the process of packaging all image information, batch information, early risk assessment results, and quality judgment results of the collected wheel hub bearing nuts into an abnormal data packet, calculating the encrypted hash value of the abnormal data packet, and storing the abnormal data packet and the encrypted hash value into a preset distributed blockchain includes the following steps:

[0090] The collected image information, batch information, early risk assessment results, and quality judgment results of the wheel hub bearing nuts are time-stamped and synchronized, and integrated into key anomaly data.

[0091] The integrated critical anomaly data is encoded according to the Avro serialization format to form a binary anomaly data packet;

[0092] An anomalous data packet is processed using a cryptographic hash function based on the SHA-3 algorithm family to obtain a fixed-length cryptographic hash value. This cryptographic hash value is the unique content fingerprint of the anomalous data packet on the blockchain.

[0093] The encrypted hash value and the abnormal data packet are encapsulated and stored in a pre-defined distributed blockchain.

[0094] In this implementation, all relevant data of the collected wheel hub bearing nuts undergo precise timestamp synchronization to ensure data consistency and integrity. Timestamp synchronization uses the Coordinated Universal Time (UTC) standard with millisecond-level accuracy. All image information includes original images, images before and after anchoring, and differential images. Each image is accompanied by a capture timestamp, camera parameters, and image quality indicators. Batch information includes nut number, production date, supplier code, and material specifications. Early risk assessment results include risk level, confidence score, and defect type prediction. Quality judgment results include classification results, anomaly descriptions, and handling recommendations. The integration process is organized according to a predefined JSON data structure, forming a unified key anomaly data format. The integrated key anomaly data is efficiently encoded using the Apache Avro serialization format to generate compact binary anomaly data packets. Avro employs a schema evolution mechanism, first defining a data schema, including field names, data types, and constraints. For example, image data is defined as bytes, timestamps as long, and text descriptions as string. The serialization process converts the JSON-formatted anomaly data into binary format, using variable-length encoding technology to compress integers and strings, typically achieving a compression ratio of 60-70%. The encoding algorithm removes duplicate data, and the LZ4 compression algorithm is applied to image data to further reduce its size. The serialized binary data packet has cross-platform compatibility and high read / write performance, supports fast data transmission and storage operations, and provides an optimized data format for distributed storage in blockchain.

[0095] The Keccak-256 hash function, based on the SHA-3 algorithm family, is applied to encrypt serialized abnormal data packets, generating a fixed-length unique content fingerprint. The SHA-3 algorithm employs a sponge construction structure, comprising two phases: absorption and extrusion, with a state array size of 1600 bits. The absorption phase processes the input data in blocks of 576 bits each, using a permutation function... Perform 24 rounds of transformation, the transformation formula is as follows ,in This is the current state. The input message block is used for the extrusion phase, which extracts a 256-bit hash value from the final state, outputting 64 hexadecimal characters. The hash calculation process exhibits an avalanche effect; small changes in the input data can lead to significant differences in the output hash value, ensuring the reliability of data integrity verification. The generated hash value serves as a unique identifier for the data packet on the blockchain, supporting fast data retrieval and integrity verification.

[0096] The calculated encrypted hash value and the abnormal data packet are encapsulated and then stored in a pre-defined distributed blockchain network to achieve immutable data storage. The encapsulation process constructs a blockchain transaction structure, comprising four parts: a transaction header, a hash value, a data packet, and a digital signature. The transaction header records the timestamp, version number, and data size; the hash value is stored as a content index in the transaction's key fields; the abnormal data packet is Base64 encoded and stored in the transaction's data payload; the digital signature is generated using the Elliptic Curve Digital Signature Algorithm (ECDSA). The blockchain network adopts a consortium blockchain architecture with five verification nodes, employing the PBFT consensus algorithm to ensure data consistency. The storage process first broadcasts the encapsulated transaction to the network. Each node verifies the transaction's validity and signature correctness, and upon reaching consensus, packages the transaction into a new block. Each block contains a block header and a transaction list. The block header includes the hash of the previous block, the Merkle root, and a timestamp, forming an immutable chain structure. After storage, the abnormal data obtains a unique blockchain address and transaction ID, supporting rapid retrieval and verification via hash value, providing reliable technical assurance for quality traceability and auditing.

[0097] In one embodiment, the method includes the following steps:

[0098] Receive traceback requests from authorized users that contain one or more query identifiers;

[0099] By interacting with the distributed blockchain through the application programming interface and utilizing the query identifier in the traceability request, matching candidate transaction records are retrieved from the on-chain index of the distributed blockchain.

[0100] For each retrieved candidate transaction record, verify the integrity of the candidate transaction record block hash chain and the validity of the transaction signature;

[0101] Extract the encrypted abnormal data packets and their corresponding encrypted hash values ​​from the verified candidate transaction records.

[0102] In this implementation, the traceability system receives query requests from authorized users. These requests contain one or more query identifiers for locating specific abnormal records. User authentication employs a JWT (JSON Web Token)-based authentication mechanism, ensuring that only users with the appropriate permissions can access sensitive quality data. Query identifiers support various formats, including nut batch numbers, production time ranges, exception type codes, SHA-3 hash values, or blockchain transaction IDs. Requests use a RESTful API interface with the standard format POST / api / trace. The request body contains query parameters and user credentials. For example, a typical traceability request is "batch_id: HN2024A001, time_range: 2024-03-15 to 2024-03-16, defect_type: thread_damage". The system performs parameter validation and format checks on received requests to ensure the validity and completeness of the query identifiers. Request processing employs an asynchronous queue mechanism, supporting concurrent processing of multiple query requests, with an average response time controlled within 200 milliseconds.

[0103] The API interacts with the distributed blockchain, using a query identifier in the traceability request to retrieve matching candidate transaction records from the blockchain's on-chain index. The blockchain index employs a multi-level index structure, including a main index, time index, batch index, and hash index, supporting efficient multi-dimensional queries. The main index is sorted by transaction ID, the time index is constructed as a B+ tree based on block timestamps, the batch index uses a hash table to map batch numbers to transaction records, and the hash index directly maps SHA-3 values ​​to specific transactions. The retrieval algorithm first selects the optimal index path based on the query type. API calls communicate with blockchain nodes using the gRPC protocol, with the query format "SELECT transaction_id, block_hash, timestamp FROM blockchain_index WHERE batch_id = 'HN2024A001' AND timestamp BETWEEN '2024-03-15' AND '2024-03-16'". The retrieval process supports fuzzy matching and range queries, and uses a Bloom filter to pre-screen candidate records, reducing unnecessary disk access. The search results are arranged in chronological order, making it easy for users to track the process and trend of anomalies along a timeline.

[0104] Each retrieved candidate transaction record undergoes rigorous block hash chain integrity verification and transaction digital signature validity checks to ensure data authenticity and tamper-proofing. Block hash chain verification begins with the genesis block and verifies the correctness of the hash value calculation for each block sequentially. Recursive verification ensures the complete chain from the genesis block to the target block remains intact. Transaction signature verification employs the Elliptic Curve Digital Signature Algorithm (ECDSA), a two-step process: first, calculating the hash value of the transaction data, and then verifying the signature using the public key. The verification process also includes timestamp checks to ensure transaction times are within a reasonable range, preventing replay attacks. For a query containing 15 candidate records, the verification process checks the hash chains of 15 transaction signatures and related blocks. The verification results include detailed status information, such as "Block hash verification: Passed, Signature verification: Passed, Timestamp verification: Passed," providing users with a clear indication of data credibility.

[0105] The process involves extracting encrypted anomaly data packets and their corresponding encrypted hash values ​​from verified candidate transaction records to complete data recovery and parsing. The extraction process begins by retrieving Base64-encoded anomaly data packets from the data payload field of the blockchain transaction, followed by decoding to obtain the original Avro serialized binary data. The integrity of the extracted data is then verified by recalculating the SHA-3 hash value and comparing it with the stored hash value. After successful integrity verification, an Avro deserializer is used to restore the binary data into a structured anomaly record, containing complete information such as image data, batch information, risk assessment results, and quality judgment results. Finally, a complete anomaly record report is generated, including the original data, verification status, and extraction time, providing comprehensive and reliable data support for quality analysis and issue tracing.

[0106] The present invention also discloses an image recognition-based automotive brake assembly anomaly identification system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the image recognition-based automotive brake assembly anomaly identification method as described in any of the above embodiments.

[0107] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0108] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of internal storage units and external storage devices of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0109] The present invention also discloses a computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the image recognition-based automotive brake assembly anomaly identification method described in any of the above embodiments.

[0110] The computer program can be stored in a machine-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The machine-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the machine-readable medium includes, but is not limited to, the above-mentioned components.

[0111] The image recognition-based automotive brake assembly anomaly identification method described in the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0112] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0113] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for identifying assembly anomalies in automotive brakes based on image recognition, characterized in that, Includes the following steps: Acquire the original images of the wheel hub bearing nuts entering the assembly station, and identify and record the batch information of the wheel hub bearing nuts; Visual fingerprint features of wheel hub bearing nuts are extracted from the original images using a deep learning model, and then compared with pre-stored visual defect features to complete an early risk assessment of wheel hub bearing nuts. Before the wheel hub bearing nut is tightened and anchored, an image before anchoring is captured; after the same wheel hub bearing nut is tightened and anchored, an image after anchoring is captured at the same shooting position. Image registration and difference operations are performed on the images before and after anchoring to generate a difference image. A machine learning model is used to analyze the microscopic physical changes caused by the anchoring operation in the difference image and output the quality judgment result of the anchoring operation. An anomaly decision engine is constructed based on weighted fuzzy logic. The inputs of the anomaly decision engine are the results of early risk assessment and quality judgment. Pre-set fuzzy membership functions and IF-THEN rule bases for different combinations of early risk assessment results and quality judgment results; By combining fuzzy membership functions and the IF-THEN rule base, and through the fuzzification, fuzzy reasoning and defuzzification processes of the anomaly decision engine, the comprehensive evaluation score of the wheel hub bearing nut assembly status is calculated. The comprehensive evaluation score is compared with the preset multi-level thresholds, and the assembly status of the wheel hub bearing nut is judged based on the comparison results to determine whether there is an assembly abnormality. If there is an assembly abnormality in the assembly status of the wheel hub bearing nut, all image information, batch information, early risk assessment results and quality judgment results of the collected wheel hub bearing nut will be packaged into an abnormal data packet, the encrypted hash value of the abnormal data packet will be calculated, and the abnormal data packet and the encrypted hash value will be stored in the preset distributed blockchain.

2. The method for identifying abnormal assembly of automotive brakes based on image recognition according to claim 1, characterized in that, The process of extracting the visual fingerprint features of the wheel hub bearing nut from the original image using a deep learning model includes the following steps: The original image is preprocessed, including image denoising, illumination normalization, and precise segmentation of the region of interest based on contour detection. The preprocessed original image is input into a pre-trained convolutional autoencoder network, and the encoder part of the convolutional autoencoder network performs nonlinear compression transformation on the original image. After performing nonlinear compression transformation on the original image, the latent spatial feature vector of the original image is extracted from the intermediate bottleneck layer of the convolutional autoencoder network as the initial visual fingerprint of the wheel hub bearing nut. Principal component analysis was applied to the initial visual fingerprint for dimensional optimization and feature selection to enhance its ability to express defect-sensitive features, thus forming the visual fingerprint features of the wheel hub bearing nut.

3. The method for identifying abnormal assembly of automotive brakes based on image recognition according to claim 2, characterized in that, The process of comparing visual fingerprint features with pre-stored visual defect features to complete the early risk assessment of wheel hub bearing nuts includes the following steps: Access the historical risk database, which pre-stores visual defect features corresponding to nut samples of various defect types; Weighted cosine similarity is used to calculate the matching score between visual fingerprint features and visual defect features; By combining the matching score and the multi-level matching score thresholds set for different defect types, it is determined whether the visual fingerprint features have a significant statistical association with any defect type. If a significant statistical association exists, the early risk assessment result of the preliminary anomaly of the corresponding wheel hub bearing nut mark is based on the associated defect type and the confidence interval of the matching score, and the visual fingerprint features are updated to the historical risk database. If there is no significant statistical correlation or the matching score is lower than the preset warning threshold, the early risk assessment result is that the corresponding wheel hub bearing nut marking is initially normal.

4. The method for identifying abnormal assembly of automotive brakes based on image recognition according to claim 1, characterized in that, The process of performing image registration and difference operations on the pre-anchoring image and the post-anchoring image to generate a difference image includes the following steps: Local feature points were detected in key structural regions of the pre-anchoring and post-anchoring images using accelerated robust features and ORB algorithms, respectively. FLANN matching is performed based on the descriptors of local feature points to establish the initial correspondence between local feature points in the pre-anchoring and post-anchoring images. Then, erroneous matching point pairs are eliminated by random sampling consensus algorithm to obtain the filtered precise corresponding feature points. Based on the precise correspondence of feature points, the homography transformation model is used to align the image after anchoring with the image before anchoring at the sub-pixel level. The normalized gray values ​​are subtracted from the corresponding pixel positions in the aligned pre-anchoring image and the post-anchoring image to obtain an initial difference image containing positive and negative difference information. The initial difference image is subjected to Gaussian filtering to smooth the noise, and an adaptive threshold binarization method is applied to highlight the microscopic physical changes caused by the anchoring operation, thus obtaining the difference image.

5. The method for identifying abnormal assembly of automotive brakes based on image recognition according to claim 4, characterized in that, The process of using a machine learning model to analyze the microscopic physical changes in the difference image caused by the anchoring operation and outputting a quality judgment result for the anchoring operation includes the following steps: The differential image is input into the anchoring anomaly recognition model built on a densely connected network. The anchoring anomaly recognition model is pre-trained by supervised learning using differential image samples containing different anchoring states. The deep convolutional layer of the anchoring anomaly identification model extracts hierarchical feature maps from the difference image to represent the morphology of the tool contact area, the degree of minor plastic deformation of the material, and the changes in surface gloss and texture. Based on hierarchical feature maps, the model outputs discrete category judgments of anchorage quality through fully connected layers and Softmax classification layers of the anchorage anomaly recognition model. At the same time, through parallel regression branches or attention mechanisms in the anchorage anomaly recognition model, the model outputs confidence scores for anchorage quality judgments and heatmaps indicating the location of abnormal micro-features in the difference image. If the discrete category judgment result is a non-qualified state, then decode from the hierarchical feature map and output a textual description of the microscopic physical change features associated with the non-qualified state; The text description and all outputs of the anchoring anomaly identification model are integrated into a quality judgment result for the anchoring operation.

6. The method for identifying abnormal assembly of automotive brakes based on image recognition according to claim 1, characterized in that, The process of packaging all image information, batch information, early risk assessment results, and quality judgment results of the collected wheel hub bearing nuts into an abnormal data packet, calculating the encrypted hash value of the abnormal data packet, and storing the abnormal data packet and the encrypted hash value into a preset distributed blockchain includes the following steps: The collected image information, batch information, early risk assessment results, and quality judgment results of the wheel hub bearing nuts are time-stamped and synchronized, and integrated into key anomaly data. The integrated critical anomaly data is encoded according to the Avro serialization format to form a binary anomaly data packet; An anomalous data packet is processed using a cryptographic hash function based on the SHA-3 algorithm family to obtain a fixed-length cryptographic hash value. This cryptographic hash value is the unique content fingerprint of the anomalous data packet on the blockchain. The encrypted hash value and the abnormal data packet are encapsulated and stored in a pre-defined distributed blockchain.

7. The method for identifying abnormal assembly of automotive brakes based on image recognition according to claim 6, characterized in that, The method includes the following steps: Receive traceback requests from authorized users that contain one or more query identifiers; By interacting with the distributed blockchain through the application programming interface and utilizing the query identifier in the traceability request, matching candidate transaction records are retrieved from the on-chain index of the distributed blockchain. For each retrieved candidate transaction record, verify the integrity of the candidate transaction record block hash chain and the validity of the transaction signature; Extract the encrypted abnormal data packets and their corresponding encrypted hash values ​​from the verified candidate transaction records.

8. A vehicle brake assembly anomaly detection system based on image recognition, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the image recognition-based method for identifying abnormal assembly of automobile brakes as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the image recognition-based method for identifying abnormal automotive brake assembly according to any one of claims 1 to 7.

Images