Automatic detection and treatment system for surface defects of aluminum metal decoration strip

Abstract

The invention relates to the technical field of metal surface detection, and discloses an automatic detection and treatment system for surface defects of an aluminum metal decoration strip. A defect data acquisition module of the system acquires processing demand data based on surface defects; the decision generation module inquires a preset strategy space and generates a first processing decision scheme through matching; the quality safety monitoring module monitors the change trend of the characteristic value of the defect degree in real time, and triggers a safety control instruction in case of sudden change; a feature depth optimization module performs energy distribution spectrum analysis on the defect type feature vectors, screens high-confidence feature vectors and updates a defect feature library; the decision dynamic compensation module predicts a defect evolution trend based on the surface state offset characteristic matrix, and optimizes the initial decision scheme to generate an optimization scheme; and the execution control module analyzes the optimization scheme and drives an execution mechanism to position and process defects. The system realizes integration of defect detection, decision generation, dynamic optimization and execution control, and improves adaptability and reliability of defect processing.

Description

Technical Field

[0001] This invention relates to the field of metal surface inspection technology, specifically to an automatic detection and processing system for surface defects in aluminum metal trim strips. Background Technology

[0002] Aluminum metal trim strips are widely used in many fields such as automobile manufacturing, architectural decoration, and electronic equipment due to their excellent mechanical properties, corrosion resistance, and aesthetic appearance. The surface quality of these trim strips directly affects the overall performance and market competitiveness of the products; therefore, the detection and treatment of surface defects are crucial.

[0003] The methods for detecting and handling surface defects in aluminum metal trim strips still have many shortcomings. Traditional manual inspection methods rely on the experience and visual observation of operators, which is not only labor-intensive and inefficient, but also prone to missed or misjudged defects due to factors such as fatigue and subjective judgment differences, making it difficult to meet the accuracy and efficiency requirements of large-scale industrial production.

[0004] With the development of automation technology, some enterprises have introduced machine vision-based inspection equipment. However, existing equipment has significant limitations in the defect handling process. Most systems can only identify and label defects, lacking the ability to make automated decision-making for different defect types. Manual intervention is still required to formulate handling solutions, leading to a disconnect between the inspection and handling processes and affecting production continuity. Furthermore, the defect feature databases of existing systems are outdated and struggle to adapt to new defect types arising from changes in raw material batches and adjustments in process parameters during the production of aluminum metal trim, easily resulting in feature matching errors.

[0005] The lack of a real-time monitoring mechanism for changes in defect severity during defect handling means that a sudden shift in defect characteristics cannot be addressed promptly, potentially escalating minor defects into serious quality issues. Furthermore, traditional handling decisions are fixed once generated, failing to adapt to dynamic changes in the surface condition of the trim strips. This makes them ill-suited for unforeseen defect evolution during production, hindering the stability and reliability of the handling results. These issues collectively restrict the improvement of aluminum trim strip production quality and efficiency, necessitating an automated system that integrates defect detection, decision generation, dynamic optimization, and execution control. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic detection and processing system for surface defects in aluminum metal trim strips, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an automatic detection and processing system for surface defects in aluminum metal trim strips, the system comprising:

[0008] The defect data acquisition module acquires defect processing requirement data based on defects on the surface of aluminum metal decorative strips;

[0009] The defect handling decision generation module is used to query a preset defect handling strategy space based on the defect handling requirement data and match and generate a first defect handling decision scheme.

[0010] The quality and safety monitoring module is used to monitor the change trend of the defect degree characteristic value of the aluminum metal trim surface in real time, and trigger a safety control command when a sudden change in defect characteristics is detected.

[0011] The defect feature depth optimization module is used to perform feature energy distribution spectrum analysis on the defect type feature vector, filter high-confidence defect feature vectors and update the defect feature library.

[0012] The decision dynamic compensation module is used to predict the future frame defect evolution trend based on the surface state offset feature matrix, and to dynamically compensate and optimize the first defect handling decision scheme according to the prediction result, so as to generate an optimized defect handling decision scheme.

[0013] The execution control module is used to analyze the optimized defect handling decision scheme and drive the execution mechanism to locate and process the surface defects of the aluminum metal trim.

[0014] Preferably, the defect data acquisition module further includes:

[0015] The real-time surface image acquisition submodule is used to acquire a sequence of surface images of aluminum metal decorative strips during continuous transmission on the production line;

[0016] The surface defect preliminary detection submodule is used to perform frame-by-frame analysis of the surface image sequence based on a preset defect feature library, identify potential defect areas, and generate initial defect location marking information.

[0017] The defect feature depth analysis submodule is used to extract image data of the area corresponding to the initial defect location marking information, perform multi-scale feature fusion calculation, and output the defect type feature vector and defect degree feature value of the aluminum metal trim surface defect;

[0018] The surface state dynamic tracking submodule is used to perform temporal alignment and feature difference calculation on the surface state feature map of the current frame and the surface state feature map of the historical frame in the surface image sequence to generate a surface state offset feature matrix.

[0019] The defect handling requirement assessment submodule is used to fuse the defect type feature vector, the defect degree feature value, and the surface state offset feature matrix to obtain defect handling requirement data for aluminum metal trim surface defects.

[0020] Preferably, the preliminary surface defect detection submodule performs frame-by-frame analysis of the surface image sequence based on a preset defect feature library, including:

[0021] Extract the grayscale distribution information and texture gradient information of each frame in the surface image sequence;

[0022] The grayscale distribution information and texture gradient information are input into a pre-trained defect feature matching network;

[0023] Output the similarity matching value with the standard defect features in the preset defect feature library;

[0024] When the similarity matching value exceeds the preset defect detection threshold, initial defect location marking information containing location coordinates is generated.

[0025] Preferably, the defect feature depth analysis submodule extracts image data of the region corresponding to the initial defect location marker information and performs multi-scale feature fusion calculation, including:

[0026] A local image block of the defect is extracted from the area corresponding to the initial defect location marking information;

[0027] Gaussian pyramid decomposition is performed on the local image block of the defect to generate multi-scale defect image sub-blocks;

[0028] Deep convolutional neural network features were extracted from defect image sub-blocks at each scale.

[0029] The deep features of the defect image sub-blocks at each scale are fused to generate the defect type feature vector;

[0030] The spatial aggregation degree of the defect type feature vector is calculated as the defect severity feature value.

[0031] Preferably, the surface state dynamic tracking submodule performs temporal alignment and feature difference calculation between the surface state feature map of the current frame and the surface state feature map of historical frames, including:

[0032] Extract surface state feature maps from N consecutive frames of the surface image sequence;

[0033] Optical flow motion compensation calculations are performed on the surface state feature maps of adjacent frames;

[0034] Align the spatial positions of the defect features in the current frame with those in historical frames based on the compensation results;

[0035] Calculate the difference between the cosine similarity and the Euclidean distance of the aligned feature maps;

[0036] The surface state offset feature matrix is ​​generated by fusing the cosine similarity and the Euclidean distance difference.

[0037] Preferably, the defect handling requirement assessment submodule integrates the defect type feature vector, defect severity feature value, and surface state offset feature matrix to calculate defect handling requirement data, including:

[0038] The defect type feature vector is subjected to category one-hot encoding mapping, and the mapping result is weighted and fused with the defect severity feature value;

[0039] Based on the weighted fusion results, the temporal variation gradient values ​​of the surface state offset feature matrix are superimposed, and the defect processing requirement data are generated through a normalized exponential function.

[0040] Preferably, the defect handling decision generation module queries a preset defect handling strategy space based on defect handling requirement data, including:

[0041] Construct a defect handling strategy space that includes mechanical polishing, laser repair, and chemical treatment;

[0042] Establish a mapping relationship between each strategy in the defect handling strategy space and the defect handling requirement data;

[0043] When the defect handling requirement data falls into the preset strategy trigger range, the corresponding first defect handling decision scheme is activated.

[0044] The configuration information of the execution parameters required for the first defect handling decision scheme is associated with it.

[0045] Preferably, the quality and safety monitoring module monitors the changing trend of defect severity characteristic values ​​in real time and triggers safety control commands, including:

[0046] Record the change curve of the defect severity characteristic value within a continuous time window;

[0047] Calculate the absolute value of the second derivative of the aforementioned curve;

[0048] When the absolute value of the second derivative exceeds a preset characteristic mutation threshold, the security control command is generated.

[0049] The safety control commands include emergency shutdown commands or processing priority escalation commands.

[0050] Preferably, the defect feature depth optimization module performs feature energy distribution spectrum analysis on the defect type feature vector, including:

[0051] Calculate the energy contribution rate of each feature dimension in the defect type feature vector;

[0052] Filter the feature dimensions whose energy contribution rate exceeds a preset threshold.

[0053] Reconstruct the high-confidence defect feature vector based on the screening results;

[0054] The high-confidence defect feature vector is added to the preset defect feature library.

[0055] Preferably, the decision-making dynamic compensation module predicts the defect evolution trend of future frames based on the surface state offset feature matrix, including:

[0056] Extract the temporal variation features of the surface state offset feature matrix;

[0057] Predict the defect feature offset of the next K frames using an autoregressive integral moving average model;

[0058] Based on the prediction results, the processing parameters of the first defect handling decision scheme are corrected to generate the optimized defect handling decision scheme that includes the parameter correction amount.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This automated surface defect detection and processing system for aluminum trim strips demonstrates numerous advantages in practical applications through the collaborative work of multiple modules. The defect data acquisition module accurately captures defect information on the aluminum trim strip surface, providing comprehensive and reliable raw data support for subsequent processing. Processing requirement data acquired based on actual defect conditions avoids errors in processing direction caused by incomplete or biased data collection in traditional detection methods, making subsequent decision-making more targeted.

[0061] The defect handling decision generation module uses a pre-defined defect handling strategy space for matching, enabling it to quickly generate initial handling decision solutions. This rule-based and experience-based matching method integrates effective handling experience from past production, reduces the randomness of manual decision-making, and ensures that the handling solution has a certain degree of rationality from the initial stage, which helps to shorten the decision-making cycle and improve processing efficiency.

[0062] The real-time monitoring function of the quality and safety monitoring module can continuously track the changing trends of defect severity characteristic values. When a sudden change occurs in defect characteristics, safety control commands are triggered in a timely manner, avoiding the problem of defect escalation caused by the lack of real-time monitoring in traditional processing. Through rapid response to abnormal situations, intervention measures can be taken when the impact of defects is small, reducing quality risks.

[0063] The defect feature depth optimization module analyzes the energy distribution spectrum of defect type feature vectors to select high-confidence feature vectors and update the defect feature library. This process enables the feature library to continuously absorb new and effective feature information, overcoming the problem of fixed and rigid feature libraries in traditional systems. As the production process progresses, the adaptability of the feature library gradually increases, helping to improve the accuracy of defect identification and reduce missed detections or misjudgments caused by feature mismatches.

[0064] The dynamic compensation module predicts the future defect evolution trend based on the surface state offset feature matrix and dynamically optimizes the initial decision scheme. This dynamic adjustment mechanism enables the treatment scheme to adapt to uncertainties in the production process, such as changes in raw material properties and fluctuations in the processing environment. By continuously optimizing the decision scheme, it ensures that the treatment measures always match the actual defect state, improving the stability of the treatment effect.

[0065] The execution control module accurately analyzes the optimized decision-making scheme and drives the execution mechanism to locate and handle defects. This step achieves seamless integration of detection, decision-making, and execution, reducing errors and delays caused by manual intervention, improving the automation level of the entire process, helping to increase production efficiency, and ensuring the surface quality stability of aluminum metal trim strips. Attached Figure Description

[0066] Figure 1 This is a timing diagram of the automatic detection and processing system for surface defects of aluminum metal trim as described in this invention;

[0067] Figure 2 A flowchart of the defect data acquisition module;

[0068] Figure 3 Workflow diagram for the defect feature depth analysis submodule;

[0069] Figure 4 This is a flowchart of the surface state dynamic tracking submodule. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Please see Figure 1 The present invention provides an automatic detection and processing system for surface defects of aluminum metal trim strips. The system includes: a defect data acquisition module, a defect processing decision generation module, a quality and safety monitoring module, a defect feature depth optimization module, a decision dynamic compensation module, and an execution control module.

[0072] The defect data acquisition module acquires real-time image sequences of the aluminum metal trim surface, performs preliminary detection and in-depth analysis using a pre-defined defect feature library, generates defect type feature vectors, defect severity feature values, and surface state offset feature matrices, and ultimately outputs defect handling requirement data. The defect handling decision generation module queries a pre-defined defect handling strategy space based on the defect handling requirement data and generates a first defect handling decision scheme. The quality and safety monitoring module monitors the changing trends of defect severity feature values ​​in real time and triggers safety control commands when a sudden feature change is detected. The defect feature depth optimization module performs feature energy distribution spectrum analysis on the defect type feature vectors, filters high-confidence feature vectors, and updates the defect feature library. The decision dynamic compensation module predicts the defect evolution trend in future frames based on the surface state offset feature matrix, dynamically corrects the first defect handling decision scheme, and generates an optimized defect handling decision scheme. The execution control module analyzes the optimized decision scheme and drives the execution mechanism to complete defect location and handling.

[0073] Example 1: See Figure 2 The implementation of the defect data acquisition module in the automatic detection and processing system for aluminum metal trim surface defects includes the functional implementation of five sub-modules: real-time acquisition of surface images, preliminary detection of surface defects, in-depth analysis of defect features, dynamic tracking of surface state, and assessment of defect processing requirements.

[0074] The real-time surface image acquisition submodule uses a high-resolution industrial line scan camera as the core acquisition device. This camera is mounted vertically above the aluminum trim strip along the production line's transmission direction. The camera lens is positioned at a fixed distance from the trim strip surface to ensure full coverage of the trim strip's width. The camera trigger mode is set to encoder triggering; the encoder pulse signal is synchronized with the production line's transmission speed, ensuring an acquisition action is triggered every millimeter of displacement. The acquired image resolution is set to a minimum of 5000 pixels in width, with a pixel depth of 12-bit grayscale. The acquisition frequency is dynamically adjusted according to the production line's maximum transmission speed, supporting continuous acquisition up to 30 frames per second. Acquired images are transmitted in real-time to a dedicated image processing server's memory buffer, with a buffer queue depth set to 300 frames. Each frame is appended with a timestamp and encoder position information, constructing a sequential data stream of surface images.

[0075] The surface defect preliminary detection submodule initiates a parallel detection process after receiving the image sequence. This module loads a pre-set defect feature library, which stores grayscale and texture feature templates of typical defects. The detection process first extracts the grayscale distribution histogram of each pixel region in the current frame image and performs contrast expansion processing. Simultaneously, it calculates the texture gradient magnitude map of the region and uses an edge direction detection operator to analyze the local texture direction. The processed grayscale and texture data are input into a pre-trained feature matching network, which adopts a three-layer convolutional layer and a two-layer fully connected layer design. The first convolutional layer uses a 5×5 convolutional kernel to extract basic features, the second convolutional layer refines the features using a 3×3 convolutional kernel, and the third convolutional layer uses dilated convolution to expand the receptive field. The fully connected layer unfolds the feature map into a feature vector, activates it using the ReLU function, and outputs a feature matching score. This score is compared with the cosine similarity of each standard defect template in the feature library. When the similarity value exceeds a preset threshold range, the region is determined to have a potential defect. The threshold range is dynamically adjusted based on the complexity of the background on the trim surface, using a lower limit of 0.7 in smooth areas and automatically decreasing to 0.6 in areas with complex textures. The detection results generate a bounding box marker matrix containing the defect locations. The matrix stores the vertex coordinates and confidence scores of the bounding boxes and is output to the next processing stage through a data structure.

[0076] The defect feature depth analysis submodule performs precise analysis of the marked regions. After receiving the initial defect location marking data, the module extracts image patches from the original image, with the bounding box region size set to 20% larger than the bounding box. Multi-scale decomposition is performed on the image patches, using a Gaussian pyramid algorithm for scale transformation. Five scale images are generated: original resolution, 1 / 2 downsampled, 1 / 4 downsampled, 1 / 8 downsampled, and 1 / 16 downsampled. Each scale image patch is input into a pre-trained convolutional neural network to extract deep features. The network structure uses an improved VGG-16 model, removing fully connected layers while retaining convolutional feature maps. The feature maps extracted at each scale are sized uniformly through a spatial pyramid pooling layer and then concatenated along the channel direction to form a multi-scale fused feature tensor. The fused tensor undergoes 1×1 convolution dimensionality reduction, compressing it to a 128-dimensional feature vector. Each dimension in the feature vector corresponds to the response intensity of a specific defect type, constituting the defect type feature vector. The defect severity characteristic value is obtained by calculating the variance of each component in the characteristic vector. The variance value reflects the spatial consistency of the defect characteristics. The analysis results include an array of characteristic vectors and real values ​​of characteristic values, stored in a structured data format.

[0077] The surface state dynamic tracking submodule establishes a continuous temporal surface state model. This module maintains a circular buffer of the surface state feature maps from the most recent 30 frames, with the feature map data generated by the aforementioned defect feature depth analysis module. The processing flow first performs inter-frame alignment, using a dense optical flow estimation algorithm to calculate the pixel displacement vectors of adjacent feature maps. The displacement vector field is fitted with least squares to construct global motion parameters, establishing a spatial mapping relationship between adjacent frames. Based on this mapping relationship, the feature maps undergo affine transformation correction to eliminate positional shifts caused by the mechanical vibration of the trim. The aligned feature maps then enter the feature comparison process: extracting the gradient direction histogram descriptor and local binary pattern feature descriptor for the defect region in the feature map. The current frame feature descriptor and historical frame descriptors are compared using cosine similarity and Euclidean distance measurement, respectively. The calculation results are filled into a 31×31 difference matrix, with the main diagonal recording the difference between the current frame and all previous frames, and the secondary diagonal recording the reverse temporal difference. The difference matrix is ​​then decomposed into eigenvalues, retaining the eigenvector corresponding to the largest eigenvalue as the offset feature. The final output of the surface state offset feature matrix is ​​a 3×3 double-precision floating-point matrix, where the matrix elements represent the offset weights of each dimension.

[0078] The defect handling requirement assessment submodule completes data fusion and requirement quantification. The input data includes three parts: defect type feature vector, defect severity feature value, and surface state offset feature matrix. The defect type feature vector is first encoded by querying a pre-set defect type mapping table to generate an eight-dimensional one-hot encoded vector representing the current defect category. This vector is then weighted with the defect severity feature value, with the weighting coefficient configured according to the defect type importance weight table. The feature value data of the surface state offset feature matrix participates in the calculation, extracting the component with the largest rate of change among the time-series offset components. All processed data is input into a three-layer fully connected neural network, with the number of nodes in the input layer equal to the sum of the total dimensions of the input data. The first hidden layer has 128 nodes to achieve non-linear feature combination, and the second hidden layer is compressed to 32 nodes for information compression. The output layer has three nodes, with output values ​​corresponding to the requirement intensity of the three basic processing strategies. The output values ​​are normalized to a probability distribution, and the final generated defect handling requirement data includes the probability values ​​of the three types of processing requirements and their timestamp information, which is transmitted to the decision module via a data bus. The data processing follows a real-time priority scheduling strategy, and each evaluation cycle is completed within 50 milliseconds.

[0079] Example 2: See Figure 3The implementation of the surface defect preliminary detection submodule and defect feature depth analysis submodule in the automatic detection and processing system for aluminum metal trim surfaces involves the specific implementation of a dual-channel feature extraction architecture and multi-scale feature fusion technology. The surface defect preliminary detection submodule adopts a parallel processing architecture, with the first channel processing grayscale distribution features and the second channel processing texture gradient features. The grayscale distribution channel performs adaptive histogram equalization on the input image to enhance the detail representation in low-contrast areas. The equalization parameters are dynamically adjusted according to the overall grayscale distribution of the image to avoid local over-enhancement. The texture gradient channel uses an improved Sobel operator to calculate edge responses, with the operator template size set to 5×5 pixels. Direction detection covers four main directions: 0°, 45°, 90°, and 135°. The gradient magnitude in each direction is calculated using a nonlinear combination formula.

[0080] Where: G represents the magnitude of the synthesized gradient. Let w represent the gradient component in the k-th direction. k The directional weight coefficients range from [0.8, 1.2]. After spatial alignment, the dual-channel feature maps are input into the feature matching network for joint processing. The network structure consists of two convolutional layer groups, each processing feature data from different modalities. The first branch uses three 3×3 convolutional kernels to extract grayscale features, followed by batch normalization and LeakyReLU activation at each layer. The second branch uses dilated convolutional structures to process texture features, with the kernel dilation rate set to an increasing sequence of [1, 2, 4]. The dual-branch features are fused in the third layer through a cross-modal attention mechanism, with attention weights calculated from the correlation matrix between channels. The fused feature maps are then input into a fully connected layer after global average pooling, and the number of output nodes matches the number of categories in the defect feature library. The matching score calculation uses an improved similarity metric, introducing a weighted coefficient of feature importance based on standard cosine similarity.

[0081] The defect feature deep analysis submodule performs multi-scale feature extraction on the initially detected and marked regions. The input image patch is first decomposed into a pyramid. The Gaussian pyramid is constructed using an interleaved downsampling strategy, with each layer's scale reduced to 0.6 times that of the previous layer. During scale transformation, the aspect ratio of the image patch remains unchanged, and boundary regions are padded with mirror images to maintain size consistency. Image patches at each scale are then processed by the feature extraction network. The network structure is an improvement on the ResNeXt architecture, employing a design combining grouped convolutions and residual connections. The basic convolutional block contains 32 groups, each with 4 kernels, all uniformly 3×3 in size. Each scale feature extraction path contains four convolutional blocks, with the number of channels adjusted between blocks using 1×1 convolutions. The multi-scale feature fusion stage uses a feature pyramid structure, where high-level semantic features are upsampled and fused step-by-step with low-level detail features. The fusion operation uses a learnable weighted summation method, with weight parameters automatically optimized through backpropagation. The final feature vector generation process incorporates dual selection using spatial and channel attention mechanisms to highlight the feature responses of key regions.

[0082] Feature vector dimensionality compression employs a hierarchical dimensionality reduction strategy. First, principal component analysis reduces the feature dimension to 256, then an autoencoder network further compresses it to 128 dimensions. The encoder portion of the autoencoder consists of three fully connected layers with 192, 160, and 128 nodes respectively, each followed by LayerNormalization. The decoder portion features a symmetrical design for supervised training. Defect severity features are calculated based on the spatial distribution characteristics of the feature vectors, using a sliding window statistical method to analyze the spatial consistency of the feature response map. The calculation process is performed at multiple scales, with the final value being the geometric mean of the results at each scale. Feature analysis results are stored as structured data records, containing feature vector arrays, feature value scalars, and timestamp information. These data records are transmitted to subsequent processing modules via a message queue, employing a zero-copy transmission protocol to reduce memory copying overhead.

[0083] The parameter configuration in the multi-scale processing follows a dynamic adjustment principle. The number of pyramid decomposition layers is automatically determined based on the input image patch size, supporting a maximum of seven layers. The convolutional kernel parameters of the feature extraction network are loaded from the pre-trained model during system initialization, supporting online fine-tuning to adapt to the process characteristics of different production lines. The training data for the autoencoder network comes from the feature vector set of historical defect samples, and the training dataset is updated using a sliding window approach. The dynamic adjustment algorithm for the feature matching threshold is based on the statistical characteristics of recent detection results, maintaining a circular buffer of detection results with a length of 100, and automatically adjusting the threshold parameters according to the ratio of true positives to false positives in the buffer. During system operation, the processing latency of each submodule is continuously monitored. When the processing time of a single frame exceeds the set threshold, the image resolution is automatically reduced or the number of pyramid layers is decreased to ensure real-time requirements.

[0084] Hardware acceleration technology is applied to computationally intensive processes. The forward inference process of the feature matching network is deployed on a dedicated neural network accelerator, utilizing a parallel computing architecture to accelerate matrix operations. Convolution operations in the multi-scale feature extraction process are implemented using the Winograd fast algorithm, reducing computational complexity. Matrix operations in the feature vector dimensionality reduction process are optimized using the BLAS library, fully utilizing the CPU's SIMD instruction set. Data transmission employs RDMA technology to achieve direct memory access between processing nodes, avoiding the overhead of traditional network protocol stacks. The system resource management module dynamically allocates computing resources, reserving sufficient CPU cores and memory bandwidth for critical path tasks, while non-critical tasks are scheduled using a time-slice round-robin approach. Error detection and recovery mechanisms cover all submodules; when data anomalies or computational errors are detected, a retry process is automatically triggered, and if a retry fails, it is reported to the central monitoring system. The status monitoring interface displays the operating metrics of each submodule in real time, including parameters such as processing frame rate, resource utilization, and error count, supporting remote diagnostics and maintenance.

[0085] Example 3: See Figure 4 The implementation of the surface state dynamic tracking submodule and the defect processing requirement assessment submodule in the automatic detection and processing system for aluminum metal trim surface defects involves the technical implementation of temporal feature alignment and multimodal data fusion. The surface state dynamic tracking submodule uses a motion compensation algorithm based on optical flow field to process continuous frame image sequences and establish a cross-frame surface state association model. The module maintains a circular buffer of length N to store historical frame feature maps; the buffer capacity is dynamically adjusted according to the system memory configuration, with a default setting of 30 frames. After the current frame feature map is input, key point detection is performed first, using an accelerated segmented test feature extraction algorithm to locate stable feature points in the image, with the number of feature points controlled within the range of 200-300. Adjacent frame feature point matching uses dual verification of descriptor distance and geometric constraints; the descriptors use an improved binary robust independent basic feature representation.

[0086] Where: D represents the distance between two feature point descriptors, p i and p j b represents the feature points of the current frame and the reference frame, respectively. kThis represents the k-th binary descriptor bit, ⊕ represents the XOR operation, and δ(·) is the sign function. Successfully matched feature point pairs are used to calculate inter-frame motion model parameters, and a robust estimation method is used to eliminate outlier pairs. The motion model is an affine transformation with six degrees of freedom, obtained through least squares fitting. The compensated feature map enters the difference calculation stage, comparing the current frame feature map with the reference frame feature map in the frequency domain. The difference in discrete cosine transform coefficients reflects subtle changes in surface condition. After dimensionality reduction using principal component analysis, the difference features generate three sets of core difference indices, representing the intensity of texture change, the magnitude of structural change, and the trend of defect expansion, respectively.

[0087] The defect handling requirements assessment submodule implements feature-level fusion and decision-level transformation of multi-source data. The module receives three types of input data: one-hot encoded representations of defect type feature vectors, logarithmic transformation results of defect severity feature values, and principal component projections of the surface state offset feature matrix. During data preprocessing, all input types are normalized to eliminate the influence of dimensional differences. Normalization parameters are derived from the statistical characteristics of the training dataset and are updated online using a moving average method. The feature fusion network adopts a three-branch input architecture, with each branch containing an independent feature transformation layer. The defect type branch maps the one-hot encoding to a low-dimensional continuous space through an embedding layer, with the embedding dimension set to 16. The defect severity branch applies a cubic spline interpolation function to expand the feature dimension, generating a 32-dimensional smooth feature representation. The state offset branch uses matrix factorization to extract latent factors, with the number of factors set to 8. The three features undergo cross-attention calculation in the fusion layer, generating a 128-dimensional joint feature representation. The attention weights are determined jointly by the learnable parameter matrix and feature correlation, dynamically adjusting the contribution ratio of each feature source.

[0088] The decision generation process employs a hierarchical neural network structure. The first layer maps fused features to the policy space, with output nodes corresponding to the initial scores of three basic processing policies. The second layer introduces temporal context information, combining the current score with historical scores into a time series, and capturing the changing trends of decision preferences through gated recurrent units. The final decision probability is generated by a weighted combination of the current score and the trend prediction result, with the weight coefficients dynamically adjusted according to the decision stability index. The probability output is converted into a standardized demand intensity value, limited to the range [0,1], representing the urgency and applicability of each processing policy. The demand assessment results are encapsulated as a structured message, including a policy type identifier, demand intensity value, timestamp, and confidence index, and broadcast to relevant modules of the system via a publish-subscribe pattern.

[0089] Time-series data processing employs a combination of a sliding window mechanism and an incremental update strategy. Surface state feature maps are stored using a lossy compression format to reduce memory usage while preserving key features. A forgetting factor mechanism is introduced for feature difference calculation, with recent frame differences receiving higher weight than historical frame differences, reflecting the time-sensitive nature of defect evolution. Regularization constraints are incorporated into the motion compensation parameter estimation process to prevent overfitting due to uneven feature point distribution. The difference feature matrix is ​​updated using a sparse representation method, modifying only significantly changing elements to reduce computational overhead. The online parameter learning mechanism of the demand assessment model supports incremental updates; new samples are used to gradually adjust model weights through mini-batch gradient descent, maintaining decision consistency.

[0090] In terms of system resource management, the computational tasks of the surface state dynamic tracking submodule are divided into multiple priorities with different real-time requirements. Feature extraction and motion compensation are high-priority tasks, allocated dedicated computing cores to ensure timely processing. Difference analysis and state updates are medium-priority tasks, using an elastic scheduling strategy to allocate resources. Historical data maintenance and compression are low-priority tasks, executed during system idle periods. The neural network inference of the defect handling requirement assessment submodule is deployed on a heterogeneous computing platform, with matrix operations handled by the GPU and logic control executed by the CPU. The memory access mode has been optimized, with feature data organized and stored according to the principle of temporal locality to improve cache hit rate. Inter-module communication adopts a combination of shared memory and message queues; critical data is quickly exchanged through memory-mapped files, while non-critical data is asynchronously transmitted via message middleware.

[0091] Error handling and anomaly detection mechanisms are implemented throughout the entire processing flow. When motion compensation fails, the system automatically switches to a block-matching-based fallback algorithm to ensure system robustness. When data anomalies occur during feature fusion, a missing data compensation algorithm is activated, utilizing the statistical characteristics of historical data to fill the current gaps. When neural network inference results exceed a reasonable range, a review process is triggered, recalculating and validating the results by simplifying the model. System operation status monitoring covers all key indicators, including processing latency, memory usage, and CPU / GPU load; when thresholds are exceeded, resource reallocation or degraded operation mode is automatically triggered. The debugging interface supports real-time visualization of feature matching results, motion compensation effects, and key intermediate data from the decision generation process, assisting technical personnel in analyzing system behavior. A version management mechanism records the history of changes to algorithm parameters and model structure, supporting rapid rollback to a stable version.

[0092] Example 4: The implementation of the defect processing decision generation module and quality and safety monitoring module in the automatic detection and processing system for aluminum metal trim surface defects involves the technical implementation of strategy space construction and real-time risk control. The defect processing decision generation module maintains a structured strategy configuration database, storing the core parameters and technical specifications of three basic processing strategies. The mechanical grinding strategy records a matching table of grinding wheel material and grit size specifications. When the defect type is scratch, the optimal grinding wheel model is automatically matched according to the scratch depth index. The laser repair strategy includes a spot focusing parameter comparison table, dynamically adjusting the lens focal length for oxide spot defects of different sizes. The chemical treatment strategy stores a solvent ratio scheme library, calling the corresponding solvent combination according to the chemical composition analysis results of oil stains. The strategy activation logic implements multi-level judgment based on the probability distribution vector of defect processing demand data: when the probability value of a single strategy exceeds a set threshold, the strategy is directly activated; when the probability values ​​of multiple strategies are in the intersection interval, a collaborative processing mode is started, and the execution order between strategies is determined by calculation based on the material removal rate model.

[0093] The quality and safety monitoring module establishes a cyclically covering monitoring network deployed at key inspection nodes on the production line. The module continuously receives defect severity characteristic value stream data from the feature analysis stage, using a circular buffer to store the latest 10 sets of readings. The safety assessment algorithm comprises a three-level processing flow: the first level calculates the first-order difference value of the continuous readings to determine if the defect change rate is within the normal range; the second level calculates an acceleration index based on the difference value sequence, triggering an early warning state when the absolute value of the acceleration exceeds a critical value; the third level analyzes the overall trend of the reading sequence and generates a timestamped status score report. When the system determines a feature mutation, the safety control command generation mechanism matches a preset response plan according to the mutation mode: for a continuous growth trend, an emergency stop signal is immediately sent to the equipment and the power supply to the actuator is cut off; for oscillating changes, the processing task priority queue is readjusted, and the current defect is marked as the highest processing level.

[0094] Table 1: Dynamic Matching Table for Defect Handling Strategies.

[0095]

[0096] The system detected an abnormal response in area S001 on the surface of the aluminum trim strip. The defect feature depth analysis module provided the following processing requirements: mechanical grinding probability 0.82, laser repair probability 0.15, and chemical treatment probability 0.03. The decision generation module queried the strategy database and determined that the conditions for independent mechanical grinding were met. Based on the defect location coordinates, the processing parameter configuration file was called: diamond grinding wheel material was selected, grit size was set to P800, spindle speed was configured to 4200 rpm, and initial feed rate was 0.3 m / s. The quality and safety monitoring module simultaneously monitored the feature value changes in this area, recording the feature values ​​within a continuous time window as [0.48, 0.52, 0.59, 0.68, 0.81]. The second-order difference sequence was calculated to obtain [-0.01, 0.06, 0.13, 0.19]. The third set of difference values ​​exceeded the threshold, triggering a sudden change alarm. Since the characteristic value shows a monotonically increasing trend, the system immediately sends an emergency stop command sequence to the motion control unit: servo drive torque reduction and deceleration command, pneumatic clamp pressure holding command, and conveyor belt braking command. The safety interlock status is simultaneously displayed on the monitoring interface.

[0097] Another defect case, S002, demonstrates a multi-strategy collaborative scenario. The assessed probability distribution of processing requirements is: mechanical grinding 0.41, laser repair 0.38, and chemical treatment 0.21. The system activates the collaborative processing mode, first performing a mechanical grinding process to remove surface protrusion defects, with parameters configured as a grinding wheel speed of 3500 rpm and a feed rate of 0.25 m / s. After grinding, it switches to the laser repair process, reducing the power to 70% of the standard value and shortening the action time to 1.2 seconds. During monitoring, oscillating changes were observed in the characteristic value sequence [0.33, 0.29, 0.41, 0.37, 0.52], and second-order difference sequence analysis showed a single mutation amplitude of 0.22. The system judged this to be an intermittent expanding defect, automatically elevating the task to the top of the emergency queue, and simultaneously increasing the number of laser scans to 2 to ensure complete elimination of the potential hazard.

[0098] In terms of data processing workflow configuration, the strategy decision-making module adopts a three-layer caching architecture. The first layer cache stores the latest five sets of demand probability vectors for analyzing short-term decision stability; the second layer cache retains historical matching records, forming a case database; and the third layer cache records parameter adjustment trajectories for easy problem tracing. The data transmission channel of the security monitoring module is independently configured, and a real-time operating system is configured to ensure command timeliness. The monitoring signal sampling period is fixed at 100 milliseconds, and the data processing pipeline includes three parallel computing units: signal filtering, feature extraction, and trend prediction. The system maintains abnormal event logs, storing detailed feature value sequences, device status snapshots, and environmental parameters at the time of alarm triggering. The debugging interface supports a simulated test mode for security responses, and preset abnormal data patterns can be injected to verify system behavior.

[0099] A version management mechanism ensures controllable changes to configuration parameters. Each update to the policy database generates a new version identifier, and the updated content includes modified fields, effective time, and responsible person information. Security threshold parameters employ a dual-review mechanism, and significant changes must be verified in a test environment. An online rollback function supports rapid restoration of policy configurations, retaining the 10 most recent valid versions for emergency use. The user interface provides a visual analysis tool for policy effectiveness, displaying the processing quality distribution under different parameter combinations to assist technical personnel in optimizing policy configurations.

[0100] Example 5: The implementation of the defect feature depth optimization module and decision dynamic compensation module in the automatic detection and processing system for surface defects of aluminum metal strips includes key processes of feature refinement and adaptive decision adjustment. The defect feature depth optimization module performs spectral feature analysis on the input defect type feature vector. The system establishes a feature evaluation system based on energy distribution and calculates the contribution weight of each dimension of the feature vector to the classification result. The analysis process first performs a fast spectral transformation to map the time-domain features to the frequency domain space. In the frequency domain, the energy proportion parameter of each frequency component is calculated, which reflects the discriminative strength of the corresponding feature dimension. The preset energy screening threshold is set as a variable and dynamically adjusted between 0.2 and 0.3 according to the number of samples of the current defect category. The screened feature dimensions enter the reconstruction process and are recombined into dimensionality-reduced feature vectors. The dimensionality reduction ratio of the new feature vectors is controlled within the range of 60% to 75% of the original dimensions. The screening mechanism for high-confidence feature vectors adopts a dual verification strategy: on the one hand, it checks the statistical distance with existing vectors in the feature library, and on the other hand, it evaluates the compactness of its spatial distribution. Feature vectors that meet the screening conditions trigger the feature library update process, and the update operation adopts an incremental addition mode. The feature library maintains a version control mechanism, generating new feature subset labels with each update while retaining historical version data to support rollback operations. Feature vector clustering employs a density peak detection algorithm to automatically identify and create new defect category groups. The feature library index structure uses a multi-level B-tree organization, supporting dual retrieval methods based on spatial location and feature similarity.

[0101] The decision-making dynamic compensation module establishes a dynamic prediction model to process time-series changing data. The module receives surface state offset feature matrix stream data and maintains the 20 most recent matrix data sets as modeling samples. Time series prediction employs an improved autoregressive integral moving average model structure, with the model order automatically determined based on information criteria. The prediction input data undergoes standard difference processing to eliminate non-stationarity, and the difference order is determined through unit root testing. The model parameter estimation process uses iterative reweighted least squares, assigning lower weights to outliers to ensure parameter robustness. The future prediction time span is configured with 5 sampling intervals, and the prediction results include point estimates and confidence intervals. Defect evolution trend analysis integrates the changing directions of multiple prediction dimensions to generate a three-dimensional trend description vector. The decision-making compensation mechanism differentiates strategies and implements varying correction schemes. For mechanical grinding strategies, a linear compensation function for the feed rate is set, with the function slope directly proportional to the predicted defect expansion rate. For laser repair strategies, a nonlinear compensation curve for power is set, triggering a quadratic function correction mode when the predicted feature offset exceeds a critical value. For chemical treatment strategies, the coupling parameters of solvent action time and spray pressure are adjusted. The compensation calculation takes into account the current equipment state constraints, and all corrected parameters must comply with the equipment's safe operating range. The optimization defect handling decision-making scheme is packaged according to a standardized protocol, including the original decision identifier, details of the corrected parameters, the correction basis code, and a valid timestamp. The scheme transmission uses a lightweight message queue protocol; the message header includes a decision version checksum, and the message body uses binary encoding to improve transmission efficiency.

[0102] During system implementation, the feature optimization and decision compensation modules share distributed computing resources. Feature spectrum analysis tasks are deployed on a mathematical accelerator card, utilizing hardware parallel architecture to accelerate matrix operations. Prediction model computation tasks are divided into two priorities: real-time prediction requests are handled by a dedicated processor, while model training tasks are executed in the background with low priority. Data exchange between modules uses a memory-mapped file mechanism, and feature library update operations are protected by read-write locks to ensure data consistency. The exception handling process includes numerical stability monitoring; when anomalies occur in feature energy analysis or the prediction model diverges, it automatically switches to a simplified algorithm mode. The resource monitoring subsystem records the computation time and memory usage of each submodule, triggering a dynamic degradation strategy when thresholds are exceeded. The version management service maintains the feature library change history, recording the operation time, feature vector fingerprint, and update type flag for each update. The debugging interface supports visual display of feature energy distribution, and can overlay historical feature vector spectrum comparison charts. The system provides feature library export and import functions; the standard exchange format includes feature dimension descriptions, version metadata, and integrity check codes. The decision compensation logic supports offline simulation testing, can load historical data to replay the decision adjustment process, and outputs parameter adjustment trajectory reports for analyzing the effectiveness of the decision logic.

[0103] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automatic detection and processing system for surface defects in aluminum metal trim strips, characterized in that, include: The defect data acquisition module acquires defect processing requirement data based on defects on the surface of aluminum metal decorative strips; The defect handling decision generation module is used to query a preset defect handling strategy space based on the defect handling requirement data and match and generate a first defect handling decision scheme. The quality and safety monitoring module is used to monitor the change trend of the defect degree characteristic value of the aluminum metal trim surface in real time, and trigger a safety control command when a sudden change in defect characteristics is detected. The defect feature depth optimization module is used to perform feature energy distribution spectrum analysis on the defect type feature vector, filter high-confidence defect feature vectors and update the defect feature library. The decision dynamic compensation module is used to predict the future frame defect evolution trend based on the surface state offset feature matrix, and to dynamically compensate and optimize the first defect handling decision scheme according to the prediction result, so as to generate an optimized defect handling decision scheme. The execution control module is used to analyze the optimized defect handling decision scheme and drive the execution mechanism to locate and process the surface defects of the aluminum metal trim.

2. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 1, characterized in that, The defect data acquisition module further includes: The real-time surface image acquisition submodule is used to acquire a sequence of surface images of aluminum metal decorative strips during continuous transmission on the production line; The surface defect preliminary detection submodule is used to perform frame-by-frame analysis of the surface image sequence based on a preset defect feature library, identify potential defect areas, and generate initial defect location marking information. The defect feature depth analysis submodule is used to extract image data of the area corresponding to the initial defect location marking information, perform multi-scale feature fusion calculation, and output the defect type feature vector and defect degree feature value of the aluminum metal trim surface defect; The surface state dynamic tracking submodule is used to perform temporal alignment and feature difference calculation on the surface state feature map of the current frame and the surface state feature map of the historical frame in the surface image sequence to generate a surface state offset feature matrix. The defect handling requirement assessment submodule is used to fuse the defect type feature vector, the defect degree feature value, and the surface state offset feature matrix to obtain defect handling requirement data for aluminum metal trim surface defects.

3. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 2, characterized in that, The preliminary surface defect detection submodule performs frame-by-frame analysis of the surface image sequence based on a preset defect feature library, including: Extract the grayscale distribution information and texture gradient information of each frame in the surface image sequence; The grayscale distribution information and texture gradient information are input into a pre-trained defect feature matching network; Output the similarity matching value with the standard defect features in the preset defect feature library; When the similarity matching value exceeds the preset defect detection threshold, initial defect location marking information containing location coordinates is generated.

4. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 3, characterized in that, The defect feature depth analysis submodule extracts image data of the region corresponding to the initial defect location marker information and performs multi-scale feature fusion calculation, including: A local image block of the defect is extracted from the area corresponding to the initial defect location marking information; Gaussian pyramid decomposition is performed on the local image block of the defect to generate multi-scale defect image sub-blocks; Deep convolutional neural network features were extracted from defect image sub-blocks at each scale. The deep features of the defect image sub-blocks at each scale are fused to generate the defect type feature vector; The spatial aggregation degree of the defect type feature vector is calculated as the defect severity feature value.

5. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 4, characterized in that, The surface state dynamic tracking submodule performs temporal alignment and feature difference calculation between the surface state feature map of the current frame and the surface state feature maps of historical frames, including: Extract surface state feature maps from N consecutive frames of the surface image sequence; Optical flow motion compensation calculations are performed on the surface state feature maps of adjacent frames; Align the spatial positions of the defect features in the current frame with those in historical frames based on the compensation results; Calculate the difference between the cosine similarity and the Euclidean distance of the aligned feature maps; The surface state offset feature matrix is ​​generated by fusing the cosine similarity and the Euclidean distance difference.

6. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 5, characterized in that, The defect handling requirement assessment submodule integrates the defect type feature vector, defect severity feature value, and surface state offset feature matrix to calculate defect handling requirement data, including: The defect type feature vector is subjected to category one-hot encoding mapping, and the mapping result is weighted and fused with the defect severity feature value; Based on the weighted fusion results, the temporal variation gradient values ​​of the surface state offset feature matrix are superimposed, and the defect processing requirement data are generated through a normalized exponential function.

7. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 6, characterized in that, The defect handling decision generation module queries a preset defect handling strategy space based on defect handling requirement data, including: Construct a defect handling strategy space that includes mechanical polishing, laser repair, and chemical treatment; Establish a mapping relationship between each strategy in the defect handling strategy space and the defect handling requirement data; When the defect handling requirement data falls into the preset strategy trigger range, the corresponding first defect handling decision scheme is activated. The configuration information of the execution parameters required for the first defect handling decision scheme is associated with it.

8. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 7, characterized in that, The quality and safety monitoring module monitors the changing trends of defect severity characteristic values ​​in real time and triggers safety control commands, including: Record the change curve of the defect severity characteristic value within a continuous time window; Calculate the absolute value of the second derivative of the aforementioned curve; When the absolute value of the second derivative exceeds a preset characteristic mutation threshold, the security control command is generated. The safety control commands include emergency shutdown commands or processing priority escalation commands.

9. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 8, characterized in that, The defect feature depth optimization module performs feature energy distribution spectrum analysis on the defect type feature vector, including: Calculate the energy contribution rate of each feature dimension in the defect type feature vector; Filter the feature dimensions whose energy contribution rate exceeds a preset threshold. Reconstruct the high-confidence defect feature vector based on the screening results; The high-confidence defect feature vector is added to the preset defect feature library.

10. The automatic detection and processing system for surface defects of aluminum metal trim strips according to claim 9, characterized in that, The decision-making dynamic compensation module predicts the defect evolution trend of future frames based on the surface state offset feature matrix, including: Extract the temporal variation features of the surface state offset feature matrix; Predict the defect feature offset of the next K frames using an autoregressive integral moving average model; Based on the prediction results, the processing parameters of the first defect handling decision scheme are corrected to generate the optimized defect handling decision scheme that includes the parameter correction amount.

Images