Intelligent judgment quality control system and method for surface defects of terminal product

Abstract

The invention discloses a terminal product surface defect intelligent judgment quality control system and method, and relates to the technical field of industrial automatic detection and intelligent quality control, and the system comprises the following steps: a data acquisition module generates an original detection signal containing environmental interference compensation; the dynamic detection module receives an original detection signal, performs illumination invariance processing through a self-optimization feature extraction network, and outputs a defect feature vector with confidence rating; the quality association module receives the defect feature vector and constructs a three-dimensional association map with real-time equipment state data, and generates a tracing analysis signal containing root cause probability distribution; and the feedback control module analyzes the key process parameter offset in the tracing analysis signal, generates an equipment adjusting instruction and feeds back the equipment adjusting instruction to the production line. The terminal product surface defect intelligent judgment quality control system and method can solve the problems of insufficient surface defect detection precision, difficulty in quality tracing and lack of process closed-loop control in industrial production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation detection and intelligent quality control, in particular to a terminal product surface defect intelligent judgment quality control system and method. BACKGROUND

[0002] In industrial production, product surface defect detection is a key link of quality control. Traditional manual inspection is inefficient and easily affected by subjective factors, which is difficult to meet the needs of modern high-speed production. With the development of machine vision technology, automatic detection systems based on image processing gradually replace manual detection, but there are still significant limitations in complex industrial scenes: first, the imaging quality is unstable due to the changeable lighting environment and material reflection characteristics, and conventional image enhancement algorithms cannot effectively eliminate environmental interference, resulting in insufficient defect feature extraction accuracy; second, existing detection systems focus on defect recognition itself and lack deep correlation with production process parameters, making it difficult to trace the equipment or process root cause of defects; third, most quality control systems use an open-loop architecture, and the detection results are disconnected from the production line control link, which cannot achieve defect prevention and real-time process optimization.

[0003] In recent years, the introduction of deep learning technology has improved the accuracy of defect recognition, but new problems have emerged: first, network models rely on a large amount of labeled data, but in actual production, there are few defect samples and the classes are not balanced; second, the fluctuation of equipment state in dynamic production environment leads to defect pattern drift, and static models are difficult to adaptively adjust; third, there is a lack of data exchange between the detection system and the production equipment, and the process parameter adjustment lags behind the occurrence of defects. In addition, existing quality traceability systems are mostly based on independent databases, with defect features, equipment states, and process parameters stored separately, weak spatiotemporal correlation analysis capabilities, and difficulty in quickly locating defect causes. SUMMARY

[0004] In view of the shortcomings of the above prior art, the purpose of the present application is to provide a terminal product surface defect intelligent judgment quality control system and method to solve the problems of insufficient surface defect detection accuracy, difficult quality traceability, and missing process closed-loop control in industrial production. The present application generates an environment compensation signal by synchronously collecting multi-spectral imaging and production process parameters, dynamically adjusts the defect recognition area using a self-optimizing feature network combined with process parameters, extracts defect features with confidence ratings; by constructing a three-dimensional correlation graph to match historical defect patterns, analyzing defect causes and generating traceability signals; based on parameter offset analysis, generate process adjustment instructions feedback to the production line, and simultaneously optimize the attention weight of the detection module with the corrected parameters, forming a closed-loop system with dynamically improved detection accuracy with process adjustment. Through the synergistic effect of multi-modal data fusion, spatiotemporal feature alignment, and parameter feedback compensation, full-link closed-loop control of defect recognition, cause tracing, and process optimization is achieved.

[0005] The application provides a terminal product surface defect intelligent judgment quality control system, comprising: A data acquisition module, which acquires product surface images through a multispectral imaging unit, synchronously acquires production process parameters, and generates original detection signals containing environmental interference compensation; A dynamic detection module, which receives the original detection signals, performs illumination invariance processing through a self-optimizing feature extraction network, generates an attention weight matrix in combination with the process parameters, dynamically adjusts a defect recognition area, and outputs a defect feature vector with a confidence rating; A quality association module, which receives the defect feature vector and constructs a three-dimensional association graph with real-time equipment state data, matches historical process defect modes through a time sequence backtracking algorithm, and generates a traceability analysis signal containing root cause probability distribution; A feedback control module, which analyzes key process parameter offsets in the traceability analysis signal, generates equipment adjustment instructions, and feeds back to the production line; The defect feature vector generated by the dynamic detection module is spatiotemporally aligned with the equipment state data in the quality association module.

[0006] In an embodiment of the application, the data acquisition module further comprises an environmental interference compensation unit, which is configured to, when generating the original detection signals, establish an ambient light intensity distribution model based on the reflectivity features collected by the multispectral imaging unit, construct a virtual lighting environment through a generative adversarial network, map the actually collected images to a standard lighting space, and the multispectral imaging unit comprises a visible light and infrared waveband cooperative collection device, the imaging parameters of which are dynamically adjusted according to the product material characteristics, the production process parameter collection synchronously records the equipment motion trajectory, and establishes a spatiotemporal correspondence relationship with the image collection timestamp at millisecond level precision, and the original detection signals contain texture-enhanced images processed through frequency domain filtering and process parameter time sequence.

[0007] In an embodiment of the application, the self-optimizing feature extraction network of the dynamic detection module comprises a double-channel feature learning structure, the first channel extracts surface topography features through a convolutional neural network, the second channel models the topological relationship between process parameters through a graph neural network, the illumination invariance processing adopts feature decoupling technology to separate material reflection characteristics and environmental illumination components, the generation of the attention weight matrix introduces mutual information measurement of process parameters and defect features, dynamically constructs a probability heat map of defect sensitive areas, the confidence rating is based on clustering density analysis in the feature space, the defect feature vector is projected into a manifold space constructed by historical defect samples, and the similarity distance with known defect modes is calculated.

[0008] In an embodiment of the present application, the three-dimensional correlation map construction process of the quality correlation module contains a spatio-temporal feature fusion mechanism, which tensor splices the spatial position information of the defect feature vector in the device coordinate system with the production beat time axis, the time sequence backtracking algorithm adopts a long short-term memory network to model the process parameter evolution path, the periodic characteristics of the device state data are extracted through a sliding time window, the matching process of the historical process defect mode contains feature dimension reduction processing, the multi-dimensional process parameters are compressed into latent space representation using an autoencoder, the calculation of the root cause probability distribution introduces a Bayesian inference framework, and a prior probability distribution model is established in combination with the device maintenance records.

[0009] In an embodiment of the present application, the feedback control module contains a parameter offset analysis unit and an instruction generation unit, the parameter offset analysis unit calculates the influence weight of the key process parameters on the defect probability through a gradient backpropagation algorithm, the generation of the device adjustment instruction adopts a fuzzy control strategy, a multi-level adjustment rule library is constructed according to the parameter offset direction and amplitude, the feedback channel of the production line control system contains a bidirectional verification mechanism, the process parameter convergence situation is continuously monitored after sending the adjustment instruction, the compensation coefficient calculation is automatically triggered when the parameter adjustment lag is detected, and the updated adjustment instruction is superimposed on the original control signal.

[0010] In an embodiment of the present application, a feature enhancement channel is arranged between the dynamic detection module and the quality correlation module, the feature enhancement channel is configured to perform multi-scale feature fusion processing on the defect feature vector, defect morphology features under different receptive fields are extracted through a hollow convolution, and cross-modal correlation analysis is performed on the enhanced feature vector and the device vibration frequency spectrum data, the cross-modal correlation analysis adopts an attention mechanism to align the time-frequency features of the visual features and the vibration signals, and a joint representation vector is constructed and input to the three-dimensional correlation map construction process of the quality correlation module.

[0011] In an embodiment of the present application, the data acquisition module contains a surface pretreatment unit, the surface pretreatment unit projects a structured light field to the product surface before image acquisition, calculates the surface curvature distribution through a grating phase shift method, generates a material adaptive lighting scheme, the multi-spectral imaging unit dynamically adjusts the intensity proportion of each waveband light source according to the lighting scheme, and the environmental interference compensation includes correction of imaging distortion caused by air turbulence, an optical flow algorithm is used to estimate the medium disturbance mode, and the original surface texture information is recovered through inverse filtering.

[0012] In an embodiment of the present application, the quality correlation module comprises a defect evolution prediction unit, which constructs a defect growth dynamics model based on a three-dimensional correlation graph, predicts the expansion trend of the defect morphology through a spatiotemporal graph convolution network, and considers the coupling effect of material stress distribution and process temperature gradient in the dynamics model, performs residual analysis on the prediction result and real-time detection data, triggers the feature extraction network retraining process of the detection module when the residual exceeds the set threshold, and updates the historical process defect mode database.

[0013] In an embodiment of the present application, a safety check interface is provided between the feedback control module and the production line control system, which is configured to perform feasibility verification on the equipment adjustment instruction, including process parameter safety boundary check, equipment response characteristic matching degree analysis and production rhythm compatibility evaluation, the verification process adopts digital twin technology to construct a virtual production line simulation environment, performs multi-condition simulation test before the instruction is actually issued, and automatically generates a correction scheme when parameter overrun risk is detected, the correction scheme includes the combined application of gradual adjustment strategy and emergency stop protocol.

[0014] The present application also provides an intelligent terminal product surface defect judgment quality control method, comprising the following steps: S1: acquiring product surface images through a multispectral imaging unit, synchronously collecting production process parameters, and generating original detection signals containing environmental interference compensation; S2: receiving the original detection signals, performing illumination invariance processing through a self-optimizing feature extraction network, generating an attention weight matrix in combination with the process parameters, dynamically adjusting the defect recognition area, and outputting a defect feature vector with confidence rating; S3: receiving the defect feature vector and constructing a three-dimensional correlation graph with real-time equipment state data, matching historical process defect modes through a time series backtracking algorithm, and generating a traceability analysis signal containing root cause probability distribution; S4: analyzing the key process parameter offset in the traceability analysis signal, generating an equipment adjustment instruction, and feeding back to the production line.

[0015] The intelligent terminal product surface defect judgment quality control system and method provided by the present application generate environmental compensation signals through synchronous collection of multispectral imaging and production process parameters, dynamically adjust the defect recognition area using a self-optimizing feature network in combination with process parameters, and extract defect features with confidence rating; match historical defect modes by constructing a three-dimensional correlation graph, analyze defect root causes and generate traceability signals; generate process adjustment instructions based on parameter offset analysis and feed back to the production line, while simultaneously optimizing the attention weight of the detection module with the corrected parameters, forming a closed-loop system in which detection accuracy is dynamically improved with process adjustment. Through the synergistic effect of multi-modal data fusion, spatiotemporal feature alignment and parameter feedback compensation, full-link closed-loop control of defect recognition, cause tracing and process optimization is achieved. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 System architecture diagram of the intelligent quality control system for determining surface defects in end products; Figure 2 A flowchart of a quality control method for intelligent determination of surface defects in end products. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0021] Please see Figures 1-2The terminal product surface defect intelligent judgment quality control system and method are shown in the figure. The terminal product surface defect intelligent judgment quality control system of the application comprises a data acquisition module, a dynamic detection module, a quality correlation module and a feedback control module. The data acquisition module acquires product surface images through a multispectral imaging unit, synchronously acquires production process parameters, and generates original detection signals containing environmental interference compensation; the dynamic detection module receives the original detection signals, performs illumination invariance processing through a self-optimizing feature extraction network, generates an attention weight matrix in combination with the process parameters, dynamically adjusts the defect recognition area, and outputs a defect feature vector with a confidence rating; the quality correlation module receives the defect feature vector and constructs a three-dimensional correlation graph with real-time equipment state data, matches historical process defect patterns through a time sequence backtracking algorithm, and generates a backtracking analysis signal containing root cause probability distribution; the feedback control module analyzes the key process parameter offset in the backtracking analysis signal, generates equipment adjustment instructions and feeds back to the production line; wherein the defect feature vector generated by the dynamic detection module is spatiotemporally aligned with the equipment state data in the quality correlation module.

[0022] As Figure 1As shown, the present application relates to a terminal product surface defect intelligent judgment quality control system, which contains four core components: data acquisition module, dynamic detection module, quality correlation module and feedback control module. The data acquisition module, as the perception layer of the system, uses multispectral imaging technology to obtain product surface image data, while simultaneously collecting real-time process parameter information during production. The multispectral imaging unit works in coordination with visible light and infrared bands, which can adapt to the optical properties of different material surfaces, such as the high-reflectivity area of metal parts or the texture details of plastic parts. During data acquisition, the module has a specially designed environmental interference compensation mechanism, which establishes an ambient light intensity distribution model by analyzing multispectral reflectance characteristics, uses a generative adversarial network to construct a virtual standard lighting space, and maps the actual collected images to standard images under ideal lighting conditions, effectively eliminating the impact of complex lighting environments in the production site on imaging quality. The collection of production process parameters not only includes temperature, pressure and other conventional parameters, but also synchronously records equipment trajectory data, and through high-precision timestamps, it achieves millisecond-level synchronization between image acquisition and process parameter recording, ensuring accurate correspondence of time and space data in subsequent analysis. The generation process of the original detection signal includes frequency domain filtering processing, which uses adaptive wavelet transform to remove image noise, and at the same time highlights the surface micro features through a texture enhancement algorithm, finally outputting a fusion signal containing preprocessed images and process parameter time series. The dynamic detection module, as the intelligent analysis center of the system, receives the multi-modal fusion signal from the data acquisition module and performs deep processing through a self-optimizing feature extraction network. This network uses a dual-channel architecture design, where the first channel uses a convolutional neural network to extract local features of the surface topography, capturing multi-scale defect features from micro-texture to macro-morphology through multiple levels of convolution kernels; the second channel uses a graph neural network to model the topological relationship between process parameters, converting discrete parameters such as temperature and pressure into node features, and constructing a relationship graph of parameter interactions. During feature extraction, the network implements illumination invariance processing, separates the material's inherent reflection characteristics from the ambient lighting components through feature decoupling technology, and uses an adversarial training strategy to eliminate the impact of lighting condition changes on feature expression. The dynamic fusion of process parameters is represented as the generation of attention weight matrices, which introduces a mutual information measurement algorithm to calculate the correlation strength between process parameter changes and defect features, dynamically adjusts the perception area weight of the convolution kernel according to real-time parameter values, and realizes adaptive focusing on defect-sensitive areas. The confidence rating mechanism is based on feature space clustering analysis, which projects the extracted defect feature vector into the manifold space constructed by historical defect samples, evaluates the reliability of the detection results by calculating the geodesic distance from known defect categories, and effectively distinguishes between suspected defects and normal texture fluctuations.

[0023] Further, the quality correlation module undertakes the functions of defect root cause analysis and process traceability, the core of which is to build a three-dimensional correlation graph to realize the deep fusion of multi-source data. After receiving the defect feature vector output by the dynamic detection module, the module first registers the spatial coordinate information of the feature vector with the device operation coordinate system, and integrates the defect position, device state parameters and production tempo time axis to form a spatio-temporal cubic data structure through tensor concatenation technology. The implementation of the time sequence backtracking algorithm relies on the long short-term memory network. When modeling the process parameter evolution path, the network adopts a sliding time window mechanism to extract the periodic characteristics of the device state data, capturing the gradual trend and sudden events of the parameters in the production process. The matching process of historical process defect patterns introduces feature dimension reduction technology, which uses variational autoencoder to compress multi-dimensional process parameters to latent space, reducing data dimension while preserving key features and improving pattern matching efficiency. The calculation of root cause probability distribution uses the Bayesian inference framework to build a prior probability model combined with device maintenance records, and updates the posterior probability through the evidence propagation algorithm, finally generating a traceability analysis signal containing the probability of occurrence of each potential cause. The innovation of this module lies in breaking through the limitations of traditional single-dimensional analysis and realizing the three-dimensional correlation of defect features, device states and process parameters in the space-time dimension. The technical details of the data acquisition module focus on the synergistic optimization of the environmental interference compensation mechanism and multispectral imaging. The environmental interference compensation unit establishes an environmental light intensity distribution model based on the radiation transfer theory using the reflectance characteristics collected by the multispectral imaging unit, and uses a generative adversarial network to build a virtual lighting environment. The network consists of a generator and a discriminator to form an adversarial training framework: the generator learns the mapping relationship from the actual environment to the standard lighting conditions, and the discriminator evaluates the consistency of the generated image and the real standard image, and through iterative optimization, the lighting interference is effectively removed. The multispectral imaging unit includes a cooperative acquisition device for visible light and infrared bands. The visible light band uses a high-resolution area array camera to capture surface topography details, and the infrared band uses a line array scanning camera to obtain thermal distribution information. The imaging parameters are dynamically adjusted according to the optical properties of the material, such as automatically reducing the exposure intensity and enabling polarization filtering for high-reflectivity metal surfaces. The acquisition process of production process parameters uses a distributed sensing network. When acquiring the device motion trajectory, it combines inertial measurement unit and encoder data, and eliminates measurement noise through Kalman filtering algorithm to ensure the accuracy of motion trajectory reconstruction. The generation process of the original detection signal implements frequency domain filtering processing, which uses an adaptive wavelet threshold denoising algorithm to dynamically adjust the filtering parameters according to the local frequency characteristics of the image, removing noise while preserving defect edge features. The construction of the process parameter time sequence uses a sliding window mechanism, and the window size is dynamically adjusted according to the production process period to ensure the effective extraction of parameter fluctuation characteristics.

[0024] Specifically, the feature extraction network of the dynamic detection module and the decision mechanism. The double-channel structure of the self-optimizing feature extraction network realizes the deep fusion of multi-source information, wherein the convolutional neural network channel adopts a residual connection structure, and the relevance of the shallow texture features and the deep semantic features is preserved through the jump connection; when the graph neural network channel constructs the process parameter relationship graph, the importance weight of the parameter node is dynamically adjusted by using the graph attention mechanism. The illumination invariance processing establishes a decoupling representation space between the material reflection characteristics and the illumination conditions, and uses the domain adversarial training method to make the feature distribution extracted by the network irrelevant to the illumination conditions. In specific implementation, a domain classifier is connected after the feature extraction layer, and the feature distribution is aligned through the gradient inversion layer. The generation process of the attention weight matrix introduces a mutual information calculation module to quantify the correlation between each process parameter and the defect feature region, and a soft attention mechanism is used to generate a probability heat map to dynamically adjust the sampling weight distribution of the convolution kernel. When the confidence rating system constructs the defect feature manifold space, the isometric mapping algorithm is used for dimension reduction processing of historical defect samples, and the local linear embedding is used to maintain the topological structure between features, calculate the neighbor distance distribution of the to-be-tested feature vector in the manifold space, and output the confidence score by combining the kernel density estimation method. Through the synergistic effect of dynamic weight adjustment and confidence evaluation, the reliability and adaptability of defect detection in complex scenes are improved.

[0025] In an embodiment of the present application, the feedback control module realizes precise process regulation through the double-layer architecture of parameter offset analysis and instruction generation. The parameter offset analysis unit adopts the gradient backpropagation algorithm to construct a sensitivity model of defect probability to process parameters, calculates the gradient direction and amplitude of each process parameter in the feature space through automatic differentiation technology, and establishes a quantitative relationship matrix between parameter changes and defect occurrence. The unit innovatively introduces a causal reasoning mechanism to distinguish between directly related parameters and indirectly affecting factors, and to exclude false correlation interference. The instruction generation unit adopts a fuzzy control strategy to design a multi-level regulation rule library, divides the parameter offset direction into positive and negative offset, sets progressive regulation levels according to the offset amplitude, for example, a small offset of the temperature parameter triggers a proportional regulation mode, while a significant offset starts the integral regulation mechanism. When constructing the rule library, the differences in device response characteristics are considered, and adjustment coefficient conversion tables are customized for different production line devices. The two-way verification mechanism of the feedback channel includes real-time monitoring and lag compensation functions. After sending the adjustment instruction, the actual parameter change curve is continuously collected through the process sensor network, the expected adjustment trajectory and the actual response path are compared using the dynamic time warping algorithm, and when the adjustment lag is detected, the compensation coefficient calculation module is started. Based on the lag time and the deviation amplitude, a three-dimensional compensation surface is constructed to generate a secondary adjustment instruction that is superimposed on the original control signal, ensuring that the process parameters quickly converge to the target interval. The closed-loop control capability of this module is reflected in the conversion of abstract defect features into executable process parameter adjustment strategies, and the feedback verification mechanism overcomes the nonlinearity of device response.

[0026] As shown in Figure 1 , the dynamic detection module uses historical adjustment records and root cause analysis results as training data to build a process parameter optimization strategy network using a deep reinforcement learning framework. A multi-dimensional state space is defined during network design, including the current process parameter state, equipment health index, and historical defect distribution characteristics. The action space is mapped to the continuous value range of the adjustment amplitude of each control parameter. The reward function design adopts a multi-objective optimization principle to balance the weight relationship between defect elimination efficiency and production capacity loss, and sets a dynamic adjustment coefficient to automatically switch the optimization center according to the production task priority. The dynamic adjustment mechanism of the parameter search space is realized through Bayesian optimization, which builds a Gaussian process model based on historical optimization results to predict the expected return of different parameter combinations and gradually narrow the search range to the efficient area. The strategy update package generation process includes policy distillation technology, which simplifies the complex strategy output by the reinforcement learning network into an interpretable rule set to ensure compatibility with existing control systems. The updated data pushed to the dynamic detection module is processed using differential privacy technology to improve the accuracy of attention weight generation while protecting process data security. The innovation value of this module lies in establishing a linkage optimization mechanism between detection accuracy and process control, enabling the system to evolve autonomously from historical experience and breaking through the limitations of traditional static parameter settings.

[0027] Further, the feature enhancement channel builds a cross-modal data analysis bridge between the dynamic detection and quality correlation module. When implementing multi-scale feature fusion, the channel adopts a parallel hollow convolution structure design, sets up convolution kernels with different expansion rates to synchronously extract local detail features and global context information, and realizes weighted fusion of multi-level features through a feature pyramid architecture. For the detection needs of micro defects, the channel integrates a sub-pixel level feature reconstruction algorithm to restore high-frequency detail information using the phase relationship between adjacent pixels. The core of cross-modal correlation analysis is to establish a joint representation model of visual features and vibration signals. A dual-flow attention network is used to process heterogeneous data: the visual flow network extracts the spatio-temporal features of defect morphology, the vibration flow network extracts the equipment operating state features through time-frequency analysis, and the two flow networks interact through cross-attention mechanisms at the high-level feature layer to calculate the correlation matrix between the visual feature map and the time-frequency spectrum and generate a joint feature vector that integrates the characteristics of the two modalities. When the vector is input into the quality correlation module, an adaptive weighting mechanism is used to dynamically adjust the contribution weights of visual and vibration features according to the current detection scene. The channel also includes an online feature selection function that dynamically masks redundant feature dimensions by calculating the information gain value of each feature dimension for defect classification, thereby improving the efficiency and accuracy of subsequent correlation analysis. The surface preprocessing unit of the data acquisition module improves imaging quality through active optical regulation. The structured light field projection system uses a digital micromirror device to generate a programmable grating pattern and calculates the three-dimensional surface curvature using the phase shifting method: it projects a sinusoidal fringe light field onto the measured surface, obtains a wrapped phase map using the three-step phase shifting method, and reconstructs the surface height distribution using a phase unwrapping algorithm. This process innovatively introduces a material reflectivity compensation algorithm that dynamically adjusts the projection light intensity and pattern frequency based on the surface reflectance characteristics obtained by the multi-spectral imaging unit to ensure that effective phase information is obtained for surfaces of different materials. When designing the material-adaptive lighting scheme, the surface is divided into specular reflection and diffuse reflection regions. For high-reflectivity areas, a ring-shaped LED array is used with a polarizing filter to suppress specular reflection, while for complex texture areas, coaxial lighting is used to enhance the contrast of surface concave-convex features. The imaging distortion correction in the environmental interference compensation uses an improved optical flow algorithm to establish an optical transmission model of the air turbulence field, estimate the medium disturbance pattern using particle image velocimetry technology, and construct an inverse filter in the frequency domain to restore the original texture. For motion blur interference, the preprocessing unit integrates an inertial measurement device to dynamically adjust the exposure time based on the product conveying speed, and uses a deblurring convolutional neural network to restore a clear image. The combination of active optical regulation and intelligent algorithm processing in the preprocessing unit significantly improves the imaging quality in complex industrial environments, providing a reliable data foundation for subsequent detection and analysis.

[0028] As Figure 2As shown, it is the terminal product surface defect intelligent judgment quality control method of the application.S1: obtain product surface image through multispectral imaging unit, synchronously collect production process parameters, and generate original detection signal containing environmental interference compensation;S2: receive the original detection signal, perform illumination invariance processing through self-optimizing feature extraction network, generate attention weight matrix combined with process parameters, dynamically adjust defect recognition area, and output defect feature vector with confidence rating;S3: receive the defect feature vector and construct three-dimensional correlation graph with real-time equipment state data, match historical process defect mode through time sequence backtracking algorithm, and generate traceability analysis signal containing root cause probability distribution;S4: analyze the key process parameter offset in the traceability analysis signal, generate equipment adjustment instruction and feedback to the production line.

[0029] The terminal product surface defect intelligent judgment quality control system and method of the application generate environmental compensation signals through synchronous collection of multispectral imaging and production process parameters, dynamically adjust the defect recognition area combined with process parameters through self-optimizing feature network, extract defect features with confidence rating; match historical defect mode through constructing three-dimensional correlation graph, analyze defect root cause and generate traceability signal; generate process adjustment instruction based on parameter offset analysis and feedback to the production line, and simultaneously optimize the attention weight of the detection module with the corrected parameters, forming a closed-loop system with dynamic improvement of detection accuracy with process adjustment.

[0030] Therefore, through the terminal product surface defect intelligent judgment quality control system and method of the application, the problems of insufficient surface defect detection accuracy, difficult quality traceability and missing process closed-loop control in industrial production are solved.

[0031] The above embodiments only exemplarily illustrate the principles and effects of the application, and are not used to limit the application. Any person skilled in the art can modify or change the above embodiments without departing from the spirit and scope of the application. Therefore, all equivalent modifications or changes completed by those skilled in the art without departing from the spirit and technical thought disclosed by the application should be covered by the claims of the application.

Claims

1. An intelligent quality control system for surface defect determination of end products, characterized by, Comprise: a data acquisition module that acquires product surface images through a multispectral imaging unit, synchronously acquires production process parameters, and generates original detection signals containing environmental interference compensation; a dynamic detection module that receives the original detection signals, performs illumination invariance processing through a self-optimizing feature extraction network, generates an attention weight matrix in combination with process parameters, dynamically adjusts a defect recognition area, and outputs a defect feature vector with a confidence rating; a quality association module that receives the defect feature vector and constructs a three-dimensional association graph with real-time equipment state data, matches historical process defect patterns through a time series backtracking algorithm, and generates a traceability analysis signal containing root cause probability distribution; a feedback control module that analyzes key process parameter offsets in the traceability analysis signal, generates equipment adjustment instructions, and feeds back to the production line; wherein the defect feature vector generated by the dynamic detection module is spatiotemporally aligned with the equipment state data in the quality association module.

2. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein, The data acquisition module further comprises an environmental interference compensation unit configured to, when generating original detection signals, establish an ambient light intensity distribution model based on reflectivity features acquired by the multispectral imaging unit, construct a virtual lighting environment through a generative adversarial network, map the actual acquired images to a standard lighting space, the multispectral imaging unit contains a visible light and infrared band cooperative acquisition device, the imaging parameters of which are dynamically adjusted according to the material characteristics of the product, the acquisition of the production process parameters synchronously records the equipment motion trajectory and establishes a spatiotemporal correspondence with the image acquisition timestamp at millisecond level accuracy, the original detection signal contains texture-enhanced images processed through frequency domain filtering and a time series sequence of process parameters.

3. The intelligent terminal product surface defect judgment quality control system according to claim 2, characterized in that, The self-optimizing feature extraction network of the dynamic detection module contains a dual-channel feature learning structure, the first channel extracts surface topography features through a convolutional neural network, and the second channel models the topological relationship between process parameters through a graph neural network, the illumination invariance processing uses feature decoupling technology to separate material reflection characteristics and environmental lighting components, the generation of the attention weight matrix introduces mutual information measurement of process parameters and defect features, dynamically constructs a probability heat map of defect sensitive areas, and the confidence rating is based on clustering density analysis in the feature space, projects the defect feature vector to the manifold space constructed by historical defect samples, and calculates the similarity distance with known defect patterns.

4. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein, The three-dimensional association graph construction process of the quality association module contains a spatiotemporal feature fusion mechanism, which tensor splices the spatial position information of the defect feature vector in the equipment coordinate system with the production beat time axis, the time series backtracking algorithm uses a long short-term memory network to model the process parameter evolution path, extracts the periodic features of the equipment state data through a sliding time window, the matching process of the historical process defect pattern contains feature dimension reduction processing, which uses an autoencoder to compress multi-dimensional process parameters into latent space representations, and the calculation of the root cause probability distribution introduces a Bayesian inference framework, and establishes a prior probability distribution model in combination with equipment maintenance records.

5. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein, The feedback control module includes a parameter offset analysis unit and an instruction generation unit, the parameter offset analysis unit calculates the influence weight of key process parameters on defect probability through gradient backpropagation algorithm, the generation of device adjustment instruction adopts fuzzy control strategy, and a multi-level adjustment rule library is constructed according to the parameter offset direction and amplitude, the feedback channel of the production line control system includes a two-way verification mechanism, which continuously monitors the process parameter convergence after sending the adjustment instruction, automatically triggers the compensation coefficient calculation when detecting parameter adjustment lag, and superimposes the updated adjustment instruction on the original control signal.

6. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein, The feature enhancement channel is provided between the dynamic detection module and the quality correlation module, the feature enhancement channel is configured to perform multi-scale feature fusion processing on the defect feature vector, extract defect morphology features under different receptive fields through hollow convolution, and perform cross-modal correlation analysis on the enhanced feature vector and device vibration frequency spectrum data, the cross-modal correlation analysis adopts an attention mechanism to align the time-frequency features of visual features and vibration signals, and a joint representation vector is constructed to input into a three-dimensional correlation graph construction process of the quality correlation module.

7. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein The data acquisition module includes a surface pretreatment unit, the surface pretreatment unit projects a structured light field to the product surface before image acquisition, calculates the surface curvature distribution through the grating phase shift method, generates a material adaptive lighting scheme, the multispectral imaging unit dynamically adjusts the intensity ratio of each waveband light source according to the lighting scheme, and the environmental interference compensation includes imaging distortion correction caused by air turbulence, estimates the medium disturbance mode by using an optical flow algorithm, and restores the original surface texture information through inverse filtering.

8. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein, The quality correlation module includes a defect evolution prediction unit, the defect evolution prediction unit constructs a defect growth kinetics model based on the three-dimensional correlation graph, predicts the expansion trend of the defect morphology through a spatio-temporal graph convolution network, the kinetics model considers the coupling effect of material stress distribution and process temperature gradient, performs residual analysis on the prediction result and real-time detection data, triggers the feature extraction network retraining process of the detection module when the residual error exceeds the set threshold, and updates the historical process defect mode database.

9. The intelligent terminal product surface defect judgment quality control system according to claim 1, wherein, The safety verification interface is provided between the feedback control module and the production line control system, the safety verification interface is configured to perform feasibility verification on the device adjustment instruction, including process parameter safety boundary check, device response characteristic matching degree analysis and production rhythm compatibility evaluation, the feasibility verification adopts digital twin technology to construct a virtual production line simulation environment, performs multi-condition simulation test before the instruction is actually issued, and generates a correction scheme automatically when detecting parameter out-of-limit risk, the correction scheme includes the combined application of a gradual adjustment strategy and an emergency stop protocol.

10. A method of intelligent quality control of surface defects of end products, using the intelligent quality control system of surface defects of end products according to any one of claims 1 to 9, characterized in that, The method comprises the following steps: S1: acquiring product surface images through a multispectral imaging unit, synchronously collecting production process parameters, and generating original detection signals including environmental interference compensation; S2: receiving the original detection signal, performing illumination invariance processing through a self-optimizing feature extraction network, combining process parameters to generate an attention weight matrix, dynamically adjusting the defect recognition area, and outputting a defect feature vector with a confidence rating; S3: receiving the defect feature vector and constructing a three-dimensional correlation graph with real-time equipment state data, matching historical process defect patterns through a time series backtracking algorithm, and generating a traceability analysis signal containing root cause probability distribution; S4: analyzing the key process parameter offset in the traceability analysis signal, generating equipment adjustment instructions, and feeding back to the production line.

Images