A knowledge graph-based rubber surface defect real-time detection system and method

Abstract

The application discloses a kind of based on knowledge graph's rubber surface defect real-time detection system and method.The system includes defect fingerprint extraction module, defect dynamics modeling module, hierarchical decision graph module, adaptive strategy optimization module, defect field analysis module and intelligent compensation control module.Through multiscale feature extraction, construct defect fingerprint and pedigree network, based on physical and chemical principles, predict defect evolution, use four-layer knowledge structure to realize multi-hop graph reasoning, use Bayesian optimization and meta-learning to dynamically adjust detection strategy, model rubber surface as continuous field for anomaly positioning, and generate hierarchical compensation strategy according to impact score.The application realizes the integration of detection, prediction, reasoning and control, and improves detection accuracy and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of rubber product quality inspection technology, specifically relating to a real-time detection system and method for rubber surface defects based on knowledge graphs. Background Technology

[0002] Various surface defects, such as bubbles, cracks, impurities, and deformation, are inevitably generated during the production of rubber products. These defects seriously affect the performance and safety of the products. Traditional manual inspection methods suffer from low efficiency, high subjectivity, and a high rate of missed detections.

[0003] While existing machine vision inspection technologies have improved inspection efficiency to some extent, they still have the following shortcomings: First, they have poor adaptability to new types of defects and require a large number of labeled samples for training; second, they lack a deep understanding of the causes and evolution of defects, making it difficult to provide effective prevention and repair suggestions; third, the interpretability of inspection results is insufficient, and the formation process of defects cannot be traced; fourth, the correlation between different types of defects has not been fully explored and utilized.

[0004] Therefore, there is an urgent need for an intelligent detection technology that can integrate domain knowledge, possess reasoning capabilities, and support few-shot learning. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a real-time detection system and method for rubber surface defects based on knowledge graphs. By constructing a multi-level defect knowledge system, it achieves intelligent identification, evolution prediction, and decision support for defects.

[0006] The technical solution of the present invention is as follows: A real-time rubber surface defect detection system based on knowledge graph, characterized in that it includes: The defect fingerprint extraction module is used to extract multi-scale features from rubber surface images. This includes extracting directional texture features at the microscale using Gabor filter banks, extracting rotation-invariant features at the mesoscale using local binary mode operators, and extracting statistical texture features at the macroscale using gray-level co-occurrence matrices. Simultaneously, morphological features and spectral response features are extracted to construct defect fingerprint vectors. A defect family tree relationship network is established by calculating the similarity of fingerprint vectors. The defect dynamics modeling module is used to construct defect evolution models based on physicochemical principles, including stress-driven crack propagation models, diffusion-driven concentration field evolution models, and chemical reaction-driven aging models. It predicts the growth rate, propagation direction, and evolution trajectory of defects by solving a set of multiphysics coupling equations. The hierarchical decision graph module is used to construct a four-layer knowledge structure including a perception layer, a cognition layer, a reasoning layer, and a decision layer. It maps defect features into semantic representations, aggregates features through graph convolutional networks, performs multi-hop graph reasoning based on attention mechanisms, and generates a decision chain that includes cause analysis, impact prediction, and solutions. The adaptive strategy optimization module is used to dynamically adjust the detection parameters according to product characteristics and environmental conditions, construct an optimization space including illumination parameters, imaging parameters and algorithm parameters, search for the optimal parameter combination using Bayesian optimization method, and achieve rapid adaptation of new products through meta-learning mechanism; The defect field analysis module is used to model the rubber surface as a continuous field, solve the distribution of stress field, temperature field and chemical concentration field, calculate field gradient to identify abnormal regions, detect field singularities through divergence and curl analysis, and predict defect propagation trends. The intelligent compensation control module is used to assess the impact of defects on product quality, generate graded compensation strategies, including process parameter adjustment, material ratio optimization, and equipment parameter correction. It calculates the compensation amount through control algorithms and predicts the compensation effect.

[0007] Furthermore, in the defect fingerprint extraction module, the morphological features include skeleton features extracted by a morphological thinning algorithm, contour features encoded by the Freeman chain code method, and fractal dimension calculated by box counting; the spectral response features include near-infrared spectral response, fluorescence excitation mode, and polarization characteristics characterized by Stokes parameters.

[0008] Furthermore, in the defect dynamics modeling module, the stress-driven model uses the Paris formula to describe fatigue crack propagation, the diffusion-driven model uses Fick's second law to describe the concentration field evolution, and the chemical reaction model uses first-order reaction kinetics to describe the aging process; the diffusion coefficient and reaction rate constant follow the Arrhenius temperature dependence.

[0009] Furthermore, in the hierarchical decision graph module, knowledge representation adopts the form of triples, and vector representations of entities and relations are learned through the TransE model; multi-hop reasoning includes causal reasoning from defects to causes, propagation reasoning from causes to effects, and strategy reasoning from effects to solutions.

[0010] Furthermore, the adaptive policy optimization module uses a Gaussian process as a surrogate model and selects experimental points through the expected improvement criterion; the meta-learning adopts a model-independent meta-learning algorithm to achieve rapid adaptation through a small number of gradient updates.

[0011] Furthermore, in the defect field analysis module, the nonlinear elastic behavior of the rubber material is described using the Mooney-Rivlin constitutive model, and the field equations are solved using the finite element method; when the field gradient amplitude exceeds a set threshold, it is determined to be a potential defect region.

[0012] Furthermore, the intelligent compensation control module calculates the impact score based on the size, location, type, and evolution risk of the defect, calculates the adjustment amount of process parameters using a PID control algorithm, and optimizes the material ratio parameters using the gradient descent method.

[0013] A real-time detection method for rubber surface defects based on knowledge graphs, characterized by comprising the following steps: S1. Defect multidimensional fingerprint extraction and genealogy construction: Collect rubber surface images, extract texture features at three scales: micro, meso and macro, and extract morphological features and spectral response features at the same time to construct defect fingerprint vectors. Build a defect genealogy relationship network by calculating similarity. S2. Defect Evolution Dynamics Modeling and Prediction: Based on the physicochemical properties of rubber materials, stress-driven, diffusion-driven, and chemical reaction-driven defect evolution models are constructed. By solving the multiphysics coupling equations, the growth rate and evolution trajectory of defects are predicted. S3, Hierarchical Knowledge Graph Reasoning and Decision Generation: Construct a four-layer knowledge structure, map defect features to the perception layer, perform semantic understanding through the cognition layer, and perform multi-hop graph reasoning in the reasoning layer to generate a decision chain that includes cause analysis, impact prediction, and treatment suggestions; S4. Adaptive optimization of detection strategy: Based on product characteristics and environmental conditions, a parameter optimization space is constructed, and the optimal combination of detection parameters is searched using the Bayesian optimization method. The new product can be quickly adapted through the meta-learning mechanism. S5. Multiphysics Coupling Analysis and Anomaly Location: The rubber surface is modeled as a continuous field, the distribution of stress field, temperature field and chemical concentration field is solved, the field gradient is calculated to identify abnormal areas, and the defect propagation trend is predicted. S6. Intelligent Compensation Strategy Generation and Execution: Assess the impact of defects on product quality, formulate graded compensation strategies, and generate compensation schemes that include process parameter adjustments, material ratio optimization, and equipment parameter corrections.

[0014] Furthermore, in step S1, directional texture features are extracted at the microscale using Gabor filter banks, rotation-invariant features are extracted at the mesoscale using local binary mode operators, and contrast, correlation, energy, and uniformity features are extracted at the macroscale using gray-level co-occurrence matrix; morphological features include skeleton, contour chain code, and fractal dimension; spectral response features include near-infrared spectrum, fluorescence mode, and polarization characteristics.

[0015] Furthermore, the multi-hop reasoning in step S3 includes: the first hop reasoning from the defect type to the cause, the second hop reasoning from the cause to the potential impact, and the third hop reasoning from the potential impact to the solution; the reasoning path with the highest confidence is selected through a reasoning chain scoring mechanism.

[0016] Furthermore, in step S6, the compensation level is determined based on the impact score: when the impact score is below the first threshold, a monitoring strategy is adopted; when the impact score is between the first and second thresholds, preventive compensation is implemented; and when the impact score is above the second threshold, immediate compensation is performed.

[0017] The beneficial effects of this invention are mainly reflected in the following aspects: 1. Small sample learning and rapid adaptation capability: By constructing a defect fingerprint and defect family tree network, newly emerging defects can be quickly classified into existing families based on their fingerprint features, inheriting the processing experience of similar defects, which greatly reduces the large number of labeled samples required for model training; combined with the meta-learning mechanism, the system can achieve rapid transfer and adaptation of detection strategies through a small number of gradient updates when facing new product types.

[0018] 2. Physical interpretability and predictive capability: By introducing defect dynamics modeling based on physicochemical principles, including stress, diffusion, and chemical reaction-driven models, and multiphysics field coupling analysis including stress field, temperature field, and concentration field, the detection results are no longer black box outputs, but have clear physical meanings. The system can not only identify current defects, but also predict their growth rate and evolution trajectory, achieving a leap from passive detection to proactive prevention.

[0019] 3. End-to-end intelligent decision support: A four-layer knowledge graph structure including perception, cognition, reasoning, and decision-making is constructed. Through multi-hop graph reasoning: causality, propagation, and countermeasures, a complete decision chain including defect cause analysis, potential impact prediction, and specific solutions is automatically generated, providing operators with comprehensive decision support that goes beyond simple alarms.

[0020] 4. Dynamic closed-loop quality control: The system employs a Bayesian optimization method to search for optimal detection parameters in real time and continuously optimizes them through meta-learning, ensuring that the system always maintains the best detection performance. At the same time, the intelligent compensation control module generates graded compensation strategies based on defect impact scores, such as monitoring, prevention, and immediate response, and directly feeds the detection results back to the production process, forming a closed-loop quality management system of detection, decision-making, control, and optimization, which significantly improves the yield and consistency of rubber product production. Attached Figure Description

[0021] Figure 1 A schematic diagram of the system architecture described in this invention is shown; Figure 2 A flowchart illustrating the steps of the method described in this invention is shown. Detailed Implementation

[0022] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] Combination Figure 1 This invention provides a real-time detection system for rubber surface defects based on knowledge graphs. The system constructs a multi-level defect knowledge system to achieve intelligent identification, evolution prediction, and decision support for surface defects of rubber products.

[0024] The system comprises six parallel functional modules: a defect fingerprint extraction module, a defect dynamics modeling module, a hierarchical decision graph module, an adaptive strategy optimization module, a defect field analysis module, and an intelligent compensation control module. These modules work together to form a complete detection and decision-making closed loop.

[0025] The defect fingerprint extraction module is responsible for constructing a unique identification system for defects. This module performs multi-scale analysis on the acquired rubber surface images, with an image resolution of 2048×2048 pixels and a sampling precision of 0.1 mm / pixel. At the microscale, Gabor filter banks are used to extract texture features. The kernel function of the Gabor filter is defined as: in: The horizontal coordinate of the image is in pixels. The vertical coordinate of the image is in pixels; The x-coordinate after rotation; The y-coordinate is the result of rotation; The wavelength is a sine wave, and its value ranges from 2 to 1 / 5 of the image width. The orientation of the Gabor nucleus, from 0 to Eight directions are evenly distributed between them; For phase offset, the value is 0 and... ; The standard deviation of the Gaussian envelope is calculated based on the wavelength. This proportional relationship is determined based on the visual perception theory commonly used in this field; The aspect ratio is 0.5, which is chosen based on the anisotropic characteristics of the rubber surface texture. The imaginary unit; It is an exponential function.

[0026] By using Gabor filter banks with different parameter combinations, a 32-dimensional micro-texture feature vector can be extracted. At the meso-scale, the defect fingerprint extraction module employs the Local Binary Mode (LBP) operator to extract texture features. The calculation formula for the LBP operator is as follows: in: The x-coordinate of the center pixel; The ordinate of the center pixel; The grayscale value of the center pixel, ranging from 0 to 255; For radius The circle on the first The grayscale value of each sampling point; For sampling point index; The total number of sampling points is 8 in this embodiment; The sampling radius is 1, 2, and 3 pixels respectively; For a sign function, when hour ,when hour .

[0027] At a macroscopic scale, the defect fingerprint extraction module extracts statistical texture features using a gray-level co-occurrence matrix (GLCM). The formula for calculating the contrast feature Contrast is: The formula for calculating the correlation feature is as follows: The formula for calculating the energy characteristic Energy is as follows: The formula for calculating the homogeneity feature is as follows: in: The normalized gray-level co-occurrence matrix is ​​the first... Line number Column elements, satisfying ; For matrix row index; Matrix column index; The gray level of the image is 256. This is the row mean; The column mean; The standard deviation is the row number. To list the standard deviation.

[0028] To avoid division by zero errors, when or When the relevance feature is set to 0, the relevance feature is set to 0.

[0029] The defect fingerprint extraction module also extracts morphological features, including skeleton features, contour chain codes, and fractal dimension. Skeleton extraction employs a morphological thinning algorithm, obtaining the central axis representation of the defect through iterative erosion operations. The iteration terminates when the results of two consecutive iterations remain unchanged. Contour chain codes utilize the Freeman chain code method, encoding the defect boundary as an 8-directional sequence. The fractal dimension is calculated using box counting. in: It is the fractal dimension, which is dimensionless; This refers to the box dimensions, in pixels. The number of boxes required to cover the defect; It is the natural logarithm function; This represents the limit operation. In actual calculations, it is performed at different scales. and For linear fitting, the slope is the fractal dimension, and the coefficient of determination is required. .

[0030] Spectral response characteristics were acquired using a multispectral imaging system, including near-infrared spectral response, fluorescence excitation mode, and polarization characteristics. The near-infrared spectral response was acquired in the 900 nm to 1700 nm band, with a band spacing of 10 nm, totaling 81 spectral channels. Principal component analysis was used to extract the top five principal components as features, requiring a cumulative contribution rate of over 95%. Fluorescence excitation employed a 365 nm ultraviolet light source, recording the fluorescence intensity distribution in the 450 nm to 650 nm band. The polarization characteristics were calculated using Stokes parameters from images at four polarization angles: 0°, 45°, 90°, and 135°. in: This refers to the total light intensity parameter; This refers to the horizontal-vertical polarization difference parameter. This refers to the diagonal polarization difference parameter; The light intensity at a polarization angle of 0°; The light intensity at a polarization angle of 45°; The light intensity at a polarization angle of 90°; The light intensity is at a polarization angle of 135°; all light intensity values ​​are in units of relative intensity.

[0031] Based on the extracted multidimensional features, the defect fingerprint extraction module constructs a defect fingerprint vector. ,in This is a texture feature vector with a dimension of 40; This is a morphological feature vector with a dimension of 15; The spectral feature vectors are 8-dimensional. The feature vectors are normalized to eliminate the influence of dimensions. Defect genealogy relationships are established by calculating the cosine similarity between fingerprint vectors. in: This is the similarity value, which ranges from [-1, 1]. For the first The fingerprint vector of a defect; For the first The fingerprint vector of a defect; This represents the vector dot product operation; For vectors The modulus length; For vectors The similarity is determined by the modulus. When the similarity exceeds the threshold of 0.85, the two defects are considered to belong to the same family.

[0032] The crack length and fractal dimension in the morphological features output by the defect fingerprint extraction module are used as the initial condition inputs of the defect dynamics modeling module, and the evolution prediction results of the defect dynamics modeling module are used to update the fingerprint features in the defect fingerprint extraction module.

[0033] The defect dynamics modeling module constructs a defect evolution model based on physicochemical principles. For stress-driven crack propagation, this module uses the Paris equation to describe the fatigue crack propagation rate: in: The crack length is in mm. The stress cycle number is dimensionless. This represents the crack propagation rate, expressed in mm / cycle. For natural rubber, the constant is taken as determined by standard fatigue testing. (mm / cycle) / (MPa·√m)^m; The Paris index is dimensionless and is determined to be 3.5 using the same experimental method. The stress intensity factor amplitude is expressed in MPa·√m and is calculated using the following formula: in: This is a geometric correction factor, dimensionless, and is set to 1.12 for surface semi-elliptical cracks; This represents the stress amplitude, in MPa. Pi; The crack length is converted to a value in meters, i.e. m, to ensure dimensional consistency.

[0034] For diffusion-driven defect evolution, the defect dynamics modeling module uses Fick's second law to describe the spatiotemporal distribution of the concentration field: in: This represents the concentration of the diffusing substance, expressed in mol / m³. Time, in seconds; This is the partial derivative of concentration with respect to time, in units of mol / (m³·s); The diffusion coefficient is expressed in m² / s. The Laplace operator is used, with units of 1 / m². The temperature dependence of the diffusion coefficient follows the Arrhenius relation: in: As a frequency factor, for the diffusion of oxygen in natural rubber, based on experimental data known in the art, it is taken as... m² / s; To obtain the activation energy, take J / mol; The gas constant is taken as 8.314 J / (mol·K); This is absolute temperature, measured in Kelvin (K).

[0035] The chemically driven aging process is modeled using a first-order reaction kinetics model: in: This represents the concentration of rubber molecular chains, expressed in mol / m³. This is the derivative of concentration with respect to time, in units of mol / (m³·s); is the reaction rate constant, in units of 1 / s, which also follows the Arrhenius relation.

[0036] The defect dynamics modeling module describes multi-field interactions through a set of coupled equations: in: This represents the stress field variable, with units of MPa. The temperature field variable is expressed in Kelvin (K). For concentration field variables, the unit is mol / m³; , , The coupling function is defined by the material constitutive relation and transport equations. The coupling equations are solved using the finite element method with a time step of 0.01 s and a mesh size of 0.1 mm to ensure numerical stability.

[0037] The output of the defect dynamics modeling module serves as the input to the hierarchical decision graph module. When the dynamics model predicts that the defect will expand to a critical size within a predetermined time, the decision graph module generates corresponding intervention measures.

[0038] The hierarchical decision graph module constructs a four-layer knowledge structure to realize the reasoning process from perception to decision. The perception layer maps the original image features into semantic vectors, and the cognition layer aggregates features through a graph convolutional network. in: For nodes In the The feature vector of the layer has a dimension of ; For nodes In the The feature vector of the layer has a dimension of ; This is a network layer index, with values ​​ranging from 0 to... ,in Total number of floors; For nodes The set of neighboring nodes; This represents the number of neighboring nodes. Index the neighboring nodes; Neighboring nodes In the The feature vector of the layer; For the first The weight matrix of the layer has dimensions of ; For the first The layer's bias vector, with dimension ; The ReLU function is used as the activation function. This indicates a summation operation.

[0039] The inference layer performs multi-hop graph inference, where the inference process for each hop is implemented through an attention mechanism: in: For nodes To the node Attention weights, satisfying ; The unnormalized attention score; Index the neighboring nodes; For nodes To the node The unnormalized attention score; A learnable attention vector with dimension . ,in For the hidden layer dimension; This is the transpose of the attention vector; The node feature transformation matrix has a dimension of . ; For nodes eigenvectors; For nodes eigenvectors; The relation feature transformation matrix has dimension 1. ; For the edge Relationship characteristics; This represents a vector concatenation operation; It is an exponential function.

[0040] The knowledge representation in the hierarchical decision graph module adopts the form of triples. ,in For the head entity, For the relationship, This is the tail entity. The knowledge graph stores domain knowledge related to defects, and the vector representations of entities and relations are learned through the TransE model: Training is performed by minimizing the loss function: in: The set of correct triples; For the set of negative sampled triples; The interval parameter is set to 1.0; The distance function is defined using the L2 norm. This indicates taking the positive part.

[0041] The detection strategy suggestions generated by the hierarchical decision graph module are input into the adaptive strategy optimization module. The adaptive strategy optimization module sends the optimized optimal detection parameter combination, such as illumination angle and exposure time, to the image acquisition device in real time to realize the dynamic adjustment of the detection strategy. At the same time, the optimized detection results and their corresponding parameter configurations are fed back into the knowledge graph of the hierarchical decision graph module to form empirical knowledge for rapid matching and strategy reuse in subsequent cases.

[0042] The adaptive strategy optimization module dynamically adjusts the detection parameters based on product characteristics and environmental conditions. This module constructs a parameter optimization space comprising three dimensions: illumination parameters, imaging parameters, and algorithm parameters. The illumination parameters include illumination angle (0° to 90°), light intensity (100 to 1000 lux), and light source wavelength (visible, ultraviolet, near-infrared). The imaging parameters include image resolution (0.01 to 1.0 mm / pixel), exposure time (0.1 to 10 ms), gain coefficient (1 to 10), and depth of focus (1 to 50 mm). The algorithm parameters include detection sensitivity (0.1 to 1.0), minimum defect size threshold (0.1 to 5.0 mm), and confidence threshold (0.5 to 0.99).

[0043] The adaptive strategy optimization module uses a Bayesian optimization method to search for the optimal parameter combination, and the objective function is defined as: in: For parameter vectors; The detection rate has a value range of [0,1]. For accuracy, the value range is [0,1]. The normalized processing time has a value range of [0,1]. As the detection rate weight, it is set to 0.5; As a precision weight, it is set to 0.3; The time cost weight is set to 0.2, and the weight satisfies... .

[0044] In the Bayesian optimization process, a Gaussian process is used as a surrogate model, with its mean function and covariance function as follows: in: To predict the mean; To predict variance; Points to be predicted The covariance vector between the observed points and the observed points; The covariance matrix between the observed points; This is the vector of observed objective function values; For kernel functions; This is a regularization term to prevent numerical instability. A radial basis function (RBF) kernel is used: in: and It consists of two parameter vectors; Let be the signal variance, set to 1.0; The length scale parameter is set to 0.5; This represents the Euclidean norm.

[0045] The criteria for selecting the acquisition function based on Expected Improvement (EI) are as follows: in: The current known optimal objective function value; Let EI represent the expectation operator. Under the Gaussian process assumption, EI has an analytical expression: in: ,Add to Prevent division by zero; The cumulative distribution function of the standard normal distribution; It is the probability density function of the standard normal distribution.

[0046] The adaptive policy optimization module also includes a meta-learning mechanism for rapid adaptation to new product types. The meta-learning employs a Model-Independent Meta-Learning (MAML) algorithm, achieving rapid adaptation through a small number of gradient update steps. in: These are the initial model parameters; For the adapted parameters; The learning rate for the inner loop is set to 0.01. For the task The loss function; For parameterized models; Indicates the parameter The gradient.

[0047] The high-risk area information identified by the defect field analysis module is input into the adaptive strategy optimization module to adjust the detection parameters of the corresponding area.

[0048] The defect field analysis module models the rubber surface as a continuous field and analyzes the coupling relationship of multiple physics fields. The stress field is solved using the finite element method, and the governing equations are: in: This is the stress tensor, with units of MPa. This is a vector of body force density, in N / m³. Let represent the divergence operator. For rubber materials, the Mooney-Rivlin constitutive model is used: in: is the strain energy density function, with units of J / m³; and For material constants, 0.293 MPa and 0.177 MPa are taken respectively for natural rubber; and For the first and second invariants of the right Cauchy-Green deformation tensor, which are dimensionless; is the determinant of the deformation gradient tensor, which is dimensionless; These are parameters related to bulk modulus, in units of 1 / Pa. For approximately incompressible rubber, take... ,in The bulk modulus is taken as 2000 MPa.

[0049] The temperature field is described by the heat conduction equation: in: For the material density, take 1200 kg / m³ for rubber; For specific heat capacity, take 1800 J / (kg·K); Temperature, in Kelvin (K). Time, in seconds; The thermal conductivity is taken as 0.13 W / (m·K); This is the internal heat source term, with units of W / m³. This is the gradient operator.

[0050] The chemical concentration field follows the reaction-diffusion equation: in: This refers to the concentration of chemical substances, expressed in mol / m³. The chemical diffusion coefficient is expressed in m² / s. This is the reaction rate constant, with units of 1 / s; For the Laplace operator.

[0051] The defect field analysis module identifies abnormal regions through field gradient calculation, where the gradient magnitude is defined as: in: It represents any field variable in the stress field, temperature field, or chemical concentration field; , , The coordinates are spatial coordinates in meters (m). The unit of gradient amplitude is determined according to the type of field variable. When the stress field gradient amplitude exceeds three times the average value, the temperature field gradient amplitude exceeds 5 K / mm, or the chemical concentration field gradient amplitude exceeds 0.1 mol / (m³·mm), the region is determined to be a potential defect region.

[0052] Field singularities are detected by calculating the curl and divergence of the field: in: It is the field vector; , , These are the components of the field vector in the three directions; Indicates divergence; Indicates curl; This represents the cross product operation.

[0053] The stress concentration area information identified by the defect field analysis module is input into the intelligent compensation control module to formulate corresponding compensation strategies.

[0054] The intelligent compensation control module generates a compensation strategy based on the detection results. This module assesses the impact of defects on product quality, and the impact scoring function is as follows: in: The overall impact score ranges from [0,1]. The defect size is scored based on the ratio of the defect area to the product surface area, with a value range of [0,1]. The importance of a location is scored as follows: 0.3 for the edge area, 0.7 for the middle area, and 1.0 for the key functional area. The severity of the defect type is scored as follows: surface blemishes are scored as 0.2, cracks as 0.6, and through-hole defects as 1.0. The evolution risk score is determined based on the propagation rate predicted by the defect dynamics model, and the value range is [0,1]. , , , The weighting coefficients, determined using the analytic hierarchy process (AHP), are set to 0.25, 0.20, 0.30, and 0.25 respectively, satisfying the following conditions: .

[0055] In this embodiment, the defect type has the most critical impact on the final performance of the product, so it is given the highest weight of 0.30; the defect location is the next most important, and is given a weight of 0.20; the size and evolution risk have intermediate weights, both at 0.25.

[0056] Based on the impact score, the intelligent compensation control module generates a tiered compensation strategy. When, a monitoring strategy is adopted, and no parameter adjustments are made; when At that time, preventative compensation is implemented, and production parameters are fine-tuned; when Immediate compensation will be provided.

[0057] The process parameter adjustment amount is calculated using a PID control algorithm: in: To control the output, the unit is determined based on the specific parameters; This is the error signal, which is the difference between the actual defect rate and the target value, and is dimensionless. The gain is set to 0.8 according to the Ziegler-Nichols adjustment rule. This is the integral gain, in units of 1 / s, and is set to 0.2. This is the differential gain, expressed in seconds (s), and is set to 0.1. The variable is the integral variable, and the unit is seconds (s). The time unit is seconds (s).

[0058] The material mix design is optimized using the gradient descent method, with the objective function being: in: Let be the matching parameter vector, with each component taking values ​​in the range [0,1], satisfying... ; For the first Expected quality index for each sample; These are the predicted values ​​from the matching model; For the first Input features of each sample; The number of samples; is the regularization coefficient, set to 0.01; Let L2 be the norm of the parameter vector.

[0059] The parameter update rules are as follows: in: For the first The parameter values ​​for the next iteration; The updated parameter value; For the first The learning rate for this iteration adopts an adaptive learning rate strategy: in: The initial learning rate is set to 0.01. It is an exponential moving average with squared gradient; The attenuation rate is set to 0.9; To prevent small amounts from being divided by zero; For the objective function in The gradient at that point.

[0060] The intelligent compensation control module also includes a predictive compensation mechanism, which adjusts production parameters in advance by predicting the defect development trend within a future time window. The prediction model uses a Long Short-Term Memory (LSTM) network, with an input layer, hidden layers, and an output layer. The hidden layer contains 128 LSTM units, and a dropout strategy is used to prevent overfitting, with a dropout rate set to 0.2. Input features include historical defect rates, production parameters (temperature, pressure, time), and environmental conditions (temperature, humidity), and the output is a prediction of the defect rate within the future time window.

[0061] The collaborative working mechanism between the modules is realized through information flow transmission and feedback loops. The output of the defect fingerprint extraction module serves as the input of the hierarchical decision graph module; the prediction results of the defect dynamics modeling module are input to the intelligent compensation control module; the high-risk area information identified by the defect field analysis module is fed back to the adaptive strategy optimization module; and the inference results of the hierarchical decision graph module guide the intelligent compensation control module in generating a repair plan.

[0062] During system operation, intermediate results and final decisions generated by each module are stored in a unified knowledge base, forming a case library and an experience library. Through a case-based reasoning (CBR) mechanism, the system retrieves similar historical cases and reuses successful detection and processing strategies. Case similarity is calculated using weighted Euclidean distance. in: and Two cases; and For the case in the first The normalized values ​​on each feature range from [0,1]. For the first The weights of each feature satisfy... ; The total number of features.

[0063] Combination Figure 2 The present invention also provides a real-time detection method for rubber surface defects based on knowledge graphs, comprising the following steps: S1. Defect Multidimensional Fingerprint Extraction and Genealogy Construction: The defect fingerprint extraction module acquires rubber surface images and extracts texture features at three scales: micro, meso, and macro. This includes extracting directional texture features using a Gabor filter bank, extracting rotation-invariant features using a local binary mode operator, and extracting statistical texture features through a gray-level co-occurrence matrix. Simultaneously, morphological features are extracted, including skeleton features, contour chain codes, and fractal dimension. Further, spectral response features are extracted, including near-infrared spectral response, fluorescence excitation modes, and polarization characteristics. The extracted multidimensional features are fused to construct a defect fingerprint vector. By calculating the similarity between fingerprint vectors, a defect genealogy network is established, enabling defect family classification and rapid matching of similar defects.

[0064] S2. Defect Evolution Dynamics Modeling and Prediction: Using the defect dynamics modeling module, a defect evolution model is constructed based on the physicochemical properties of the rubber material. Defect growth models driven by stress, diffusion, and chemical reaction are established respectively. By solving the multiphysics coupling equations, the growth rate, expansion direction, and evolution trajectory of defects in complex environments are predicted. Based on the evolution model, the time window for the defect to reach the critical state is determined, providing a time reference for subsequent compensation control.

[0065] S3. Hierarchical Knowledge Graph Reasoning and Decision Generation: Through the hierarchical decision graph module, a four-layer knowledge structure is constructed to realize the reasoning process from perception to decision; defect features are mapped to the perception layer to form semantic representations, feature aggregation and semantic understanding are performed in the cognition layer, and multi-hop graph reasoning is performed in the reasoning layer, including causal reasoning from defect to cause, propagation reasoning from cause to effect, and countermeasure reasoning from effect to solution; the reasoning results based on the knowledge graph generate a complete decision chain containing cause analysis, effect prediction, and treatment suggestions.

[0066] S4. Adaptive Optimization of Detection Strategy: The adaptive strategy optimization module dynamically adjusts the detection strategy based on product characteristics and environmental conditions; it constructs an optimization space containing illumination parameters, imaging parameters, and algorithm parameters, and uses a Bayesian optimization method to search for the optimal parameter combination; at the same time, it uses a meta-learning mechanism to transfer knowledge from the detection experience of similar products, thereby achieving rapid adaptation of the detection strategy for new products.

[0067] S5. Multiphysics Field Coupling Analysis and Anomaly Location: The defect field analysis module models the rubber surface as a continuous field for analysis; solves the distribution of stress field, temperature field and chemical concentration field, calculates the gradient and field anomaly of each physical field, identifies potential defect areas and high-risk locations; and predicts the expansion trend and impact range of defects through field evolution analysis.

[0068] S6. Intelligent Compensation Strategy Generation and Execution: The intelligent compensation control module assesses the comprehensive impact of defects on product quality; it formulates graded compensation strategies based on the degree of impact, including monitoring strategies, preventive compensation, and immediate compensation; it generates compensation schemes that include process parameter adjustments, material ratio optimization, and equipment parameter corrections, forming a compensation sequence with execution priorities; and it evaluates the compensation effect through a predictive model to ensure the effectiveness of the compensation strategy.

[0069] The steps S1 to S6 above form a complete detection and control closed loop. The defect features extracted in step S1 provide input data for steps S2 and S3; the evolution prediction results of step S2 are used in step S6 to determine the urgency of compensation; the reasoning results of step S3 provide decision support for step S6; step S4 optimizes the detection parameter configuration based on the abnormal areas identified in step S5; and the field analysis results of step S5 guide step S6 in formulating targeted compensation measures. The detection data, reasoning results, and processing experience generated in each step are stored in a knowledge base, and knowledge accumulation and reuse are achieved through a case-based reasoning mechanism.

[0070] In summary, this invention achieves rapid identification of novel defects by constructing a defect fingerprint family tree, predicts defect evolution through a multi-physics coupling model, generates a complete decision chain through knowledge graph reasoning, maintains optimal detection performance through adaptive optimization, accurately locates abnormal areas through field analysis, and forms closed-loop control through intelligent compensation. This upgrades passive detection to proactive prevention, reduces reliance on large amounts of labeled data, and provides a complete technical solution for intelligent quality management of rubber products.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time detection system for rubber surface defects based on knowledge graphs, characterized in that, include: The defect fingerprint extraction module is used to extract texture features of rubber surface defects at three scales: micro, meso, and macro. It also extracts morphological features and spectral response features, fuses these features to construct a defect fingerprint vector, and establishes a defect family tree network by calculating the cosine similarity between the defect fingerprint vectors, thereby achieving defect family classification. The defect dynamics modeling module is used to construct defect evolution models that include stress-driven, diffusion-driven, and chemical reaction-driven models. By solving the coupled equations of stress field, temperature field, and concentration field, it predicts the growth rate and evolution trajectory of defects. The hierarchical decision graph module is used to construct a four-layer knowledge structure including a perception layer, a cognition layer, a reasoning layer, and a decision layer. It performs multi-hop graph reasoning through graph convolutional networks and attention mechanisms to generate a complete decision chain from defect identification to cause analysis to solution. The adaptive strategy optimization module is used to search for the optimal combination of detection parameters in an optimization space containing illumination parameters, imaging parameters, and algorithm parameters using Bayesian optimization methods, and to achieve rapid adaptation to new product detection strategies through a meta-learning mechanism. The defect field analysis module is used to model the rubber surface as a continuous field, solve the multiphysics field distribution and calculate the field gradient. When the gradient magnitude exceeds a set threshold, it is identified as a potential defect area. The intelligent compensation control module is used to calculate the impact score based on the size, location, type and evolution risk of the defect, and generate a graded compensation strategy based on the impact score.

2. The system according to claim 1, characterized in that, In the defect fingerprint extraction module: At the microscale, Gabor filter banks are used to extract directional texture features; at the mesoscale, local binary mode operators are used to extract rotation invariant features; and at the macroscale, gray-level co-occurrence matrix is ​​used to extract four statistical features: contrast, correlation, energy, and uniformity. The morphological features include skeleton features extracted by a morphological thinning algorithm, contour features encoded using Freeman chain codes, and fractal dimension calculated using box counting. The spectral response characteristics include near-infrared spectral response, fluorescence excitation mode, and polarization characteristics characterized by Stokes parameters.

3. The system according to claim 1, characterized in that, In the defect dynamics modeling module: The stress-driven model uses the Paris formula to describe the fatigue crack propagation rate, which is proportional to the power of the stress intensity factor amplitude. The diffusion-driven model uses Fick's second law to describe the spatiotemporal distribution of the concentration field, and the diffusion coefficient follows the Arrhenius temperature dependence. The chemical reaction-driven model uses first-order reaction kinetics to describe the aging process of rubber molecular chains.

4. The system according to claim 1, characterized in that, The multi-hop reasoning of the hierarchical decision graph module includes: The first step in reasoning is to establish a causal relationship from the type of defect to its cause; The second step is to reason about the propagation relationship from the cause to the potential impact; The third step in reasoning moves from the potential impact to the strategy relationship of the solution; By learning the vector representations of entities and relations in the knowledge graph through the TransE model, the sum of the head entity vector and the relation vector is approximately equal to the tail entity vector.

5. The system according to claim 1, characterized in that, The adaptive strategy optimization module: A Gaussian process is used as a surrogate model, and the covariance between parameter points is calculated using radial basis function kernels. The expected improvement criterion is used as the acquisition function to select the next experimental point; The meta-learning mechanism employs a model-independent meta-learning algorithm, which achieves rapid adaptation to new tasks through inner loop gradient updates.

6. The system according to claim 1, characterized in that, The defect field analysis module: The Mooney-Rivlin constitutive model is used to describe the nonlinear elastic behavior of rubber materials; The stress field control equation, heat conduction equation, and reaction diffusion equation are solved using the finite element method. The location of the defect center is identified by detecting field singularities through the calculation of field divergence and curl.

7. A real-time detection method for rubber surface defects based on knowledge graphs, characterized in that, Includes the following steps: S1. Defect Multidimensional Fingerprint Extraction and Genealogy Construction: Extract texture features of rubber surface defects at three scales: micro, meso, and macro. Extract morphological features and spectral response features to construct defect fingerprint vectors. Build a defect genealogy relationship network by calculating similarity. S2. Defect Evolution Dynamics Modeling and Prediction: Construct defect evolution models driven by stress, diffusion, and chemical reaction, and predict the growth rate and evolution trajectory of defects by solving multiphysics coupling equations. S3, Hierarchical Knowledge Graph Reasoning and Decision Generation: Maps defect features to a four-layer knowledge structure, performs multi-hop graph reasoning, and generates a decision chain that includes cause analysis, impact prediction, and treatment suggestions; S4. Adaptive optimization of detection strategy: The Bayesian optimization method is used to search for the optimal combination of detection parameters, and the new product can be quickly adapted through the meta-learning mechanism. S5. Multiphysics Coupling Analysis and Anomaly Location: Solve for the distribution of stress field, temperature field and chemical concentration field, and calculate field gradient to identify anomaly regions; S6. Intelligent Compensation Strategy Generation and Execution: Based on the defect impact score, a graded compensation strategy is formulated to generate a compensation scheme that includes process parameter adjustment, material ratio optimization, and equipment parameter correction.

8. The method according to claim 7, characterized in that, In step S3, the multi-hop inference calculates the attention weights between nodes through an attention mechanism. The attention weights are equal to the normalized attention scores, which are calculated by the inner product of the learnable attention vector and the concatenated vector of node features and relation features.

9. The method according to claim 7, characterized in that, In step S4, the objective function is a weighted combination of detection rate, accuracy and processing time. The mean and variance of the parameter points are predicted by a Gaussian process, and the parameter point with the largest expected improvement value is selected as the next experimental point using the expected improvement criterion.

10. The method according to claim 7, characterized in that, In step S6: When the defect impact score is below the first threshold, a monitoring strategy is adopted and no parameter adjustment is made; When the defect impact score is between the first and second thresholds, preventative compensation is implemented, and production parameters are fine-tuned. When the defect impact score exceeds the second threshold, immediate compensation is performed, including process parameter adjustment, material ratio optimization, and equipment parameter correction. The process parameters are adjusted using a PID control algorithm, and the material ratio is optimized using a gradient descent method.

Images