A steel cord drawing production defect on-line detection system

Abstract

The application discloses a kind of steel cord drawing production defect on-line detection systems, including following steps: by optical, electromagnetic and ultrasonic testing mode obtains multimodal data of steel cord, and carries out fusion processing, based on variational autoencoder and denoising processing method optimization defect data, dynamic adjustment receptive field is combined with time series analysis, improve defect area identification ability, optimize defect features by feature alignment and modal complementary strategy, utilize generation model to amplify defect sample, improve the training quality of detection model, adopt diffusion flow denoising and attention flow identification, ensure that detection result is accurate, combined with reinforcement learning dynamic adjustment detection strategy.The application can effectively improve the precision and stability of defect detection, realize the intelligent quality control of steel cord production process.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and in particular to an online defect detection system for steel cord drawing production. Background Technology

[0002] With the advancement of Industry 4.0 and intelligent manufacturing, the quality control requirements for steel cord drawing, as a high-precision production process, are becoming increasingly stringent. During the drawing and stranding process, various defects often occur, such as surface cracks, bubbles, and surface unevenness. If these defects are not detected in a timely manner, they will directly affect product quality and production efficiency. Therefore, online detection of defects in the steel cord drawing process has become a key technology for improving production quality and reducing resource waste.

[0003] In existing technologies, traditional defect detection methods mostly rely on a single sensor or detection technique, such as visual inspection or sensor data acquisition. While these methods can detect some defects, they are limited by a single data source and have relatively low detection accuracy, making it difficult to accurately and comprehensively identify and locate various types of defects. Specifically, existing methods have significant shortcomings in the following aspects:

[0004] 1. Dependence on Single Modal Data: Existing defect detection systems typically rely on a single sensor, making it difficult for the system to acquire comprehensive defect information in complex production environments. Since single-modal data may not be able to fully capture the characteristics of different types of defects, detection accuracy is reduced.

[0005] 2. Lack of multimodal feature fusion: Although multimodal data can provide information from different perspectives, existing technologies often neglect the deep fusion of features between modalities, resulting in the information from different data sources not being fully utilized, thus affecting the accuracy and robustness of defect detection.

[0006] 3. The scarcity of defect samples: Existing defect detection methods often face the problem of scarce defect samples, especially in high-precision production processes, where the number of real defect samples is far lower than that of normal samples, resulting in unbalanced training data, which in turn affects the training effect and detection capability of the model.

[0007] 4. Low model training efficiency: Due to the lack of effective sample augmentation and data amplification techniques, traditional defect detection methods often struggle to build effective models when faced with a small number of training samples, thus failing to improve the model's generalization ability and robustness.

[0008] Therefore, how to provide an online detection system for defects in steel cord drawing production is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] One objective of this invention is to propose an online defect detection system for steel cord drawing production. This invention achieves efficient and accurate defect detection through multimodal data fusion, defect feature extraction and enhancement, and generative model optimization. Specifically, this invention utilizes data from multiple sensors, including optical imaging, electromagnetic detection, and ultrasonic detection, to perform multi-level feature extraction, alignment, and fusion. Combined with a variational autoencoder, it generates synthetic defect samples, thereby constructing an optimized training set. This method comprehensively improves detection accuracy, effectively solves the problem of scarce defect samples, and possesses the advantages of high efficiency, accuracy, and robustness.

[0010] An online detection system for defects in steel cord drawing production according to an embodiment of the present invention includes the following steps:

[0011] S1, Data Acquisition Module, is used to acquire data on the surface morphology, internal conductivity changes and structural integrity of steel cord through optical imaging, electromagnetic detection and ultrasonic detection, forming a multimodal defect data stream;

[0012] S2, Defect Feature Reconstruction Module: Based on multimodal defect data stream, it models the distribution of latent variables of defects through variational autoencoder, removes noise interference using denoising autoencoder, and outputs optimized defect features.

[0013] S3, the spatiotemporal feature optimization module, extracts the temporal features of defects in the steel cord drawing and twisting process based on the optimized defect features, and adjusts the receptive field by combining time series mapping to generate the spatiotemporal attention weight matrix of the defect area.

[0014] S4, Cross-modal feature fusion module, adopts feature mapping alignment method to perform multi-level feature extraction and information association on multi-modal defect data stream, and generates fused defect features through cross-modal feature fusion network;

[0015] S5, Defect Sample Enhancement Module: Based on fused defect features, it generates multi-class defect samples through variational autoencoder, and combines probability distribution to augment the dataset and construct an optimized training set;

[0016] S6, Dual-stream defect detection module, constructs a dual-stream architecture of denoising stream and attention stream based on the optimized training set, and outputs the detection results;

[0017] S7, Detection Parameter Optimization Module: Based on the detection results, reinforcement learning is used to adjust the spatiotemporal attention weight matrix of the defect region and fuse defect features, and the detection results are optimized by combining dynamic feedback mechanism.

[0018] Optionally, S2 specifically includes:

[0019] S21, Defect data preprocessing unit, receives multimodal defect data stream generated by data acquisition module, extracts optical imaging, electromagnetic detection and ultrasonic detection data, normalizes and scales the data, and performs noise filtering to form a standardized defect feature matrix X;

[0020] S22, Defect Latent Variable Modeling Unit: Based on the standardized defect feature matrix X, a latent variable space is constructed. A variational autoencoder is used to map the standardized defect feature matrix X to the latent variable space. The latent variable z is set to follow a standard normal distribution. The posterior distribution is calculated, and the optimized latent variable matrix is ​​output.

[0021] D KL (q(z|x)||p(z));

[0022] Among them, D KL Kullback-Leibler divergence measures the difference between the latent variable distribution q(z|x) and the standard normal distribution p(z);

[0023] S23, the defect feature reconstruction unit, receives the optimized latent variable matrix, uses the decoder of the variational autoencoder to reconstruct the features, restores the latent variable z to the original feature space, generates the reconstructed defect data matrix X', and calculates the reconstruction error L. recon ;

[0024] S24, Defect Denoising Unit, receives the reconstructed defect data matrix X', applies Gaussian noise perturbation to the reconstructed defect data matrix X', performs denoising processing through a denoising autoencoder, and generates a denoised defect data matrix by using a learning mapping function;

[0025] S25, Defect Feature Optimization Unit: Receives the denoised defect data matrix, constructs a joint loss function, adjusts the weight ratio of reconstruction error and latent variable regularization term, and generates optimized defect features X. * :

[0026] L = L recon +βD KL (q(z|x)||p(z));

[0027] Where L is the joint loss function, L recon This represents the reconstruction error, and β is the regularization coefficient.

[0028] Optionally, S3 specifically includes:

[0029] S31, Defect temporal feature extraction unit, receives the optimized defect feature X output by the defect feature reconstruction module. * Defect temporal feature matrix T is constructed based on time series modeling methods. s Ts It consists of N time steps t, and the feature vector of each time step is represented as:

[0030]

[0031] in, d represents the state of the defect feature at time step t. s Dimensions of defect characteristics;

[0032] S32, Spatiotemporal mapping calculation unit, based on the defect temporal feature matrix T s Construct a temporal relationship mapping and define the hidden state h at time step t. s,t To form the hidden state matrix H s :

[0033]

[0034] Among them, h s,t The hidden state at the current time step t. h is the time series mapping function. s,t-1 For the hidden state of the previous time step, calculate the time series correlation and adjust the time weight matrix to keep the information dependence of defect features consistent in the time dimension.

[0035] S33, Dynamic Receptive Field Adjustment Unit: Based on the calculated time-series correlation, it calculates the receptive field range of defect features at different time steps and defines the receptive field weight matrix.

[0036]

[0037] Among them, A s For the receptive field weight matrix, a s,t This represents the receptive field weight at time step t. The receptive field calculation function dynamically adjusts the receptive field size based on the time-series correlation changes of defect features;

[0038] S34. The spatiotemporal attention generation unit for the defect region calculates the spatiotemporal correlation of defect features based on the dynamic adjustment of the receptive field size, and constructs the spatiotemporal attention weight matrix W for the defect region. s :

[0039]

[0040] Among them, W s Let w be the spatiotemporal attention weight matrix for the defect region. s,t The spatiotemporal attention weights represent time step t. This is the function for calculating spatiotemporal attention.

[0041] Optionally, S33 specifically includes:

[0042] S331, Time Series Correlation Calculation Unit, receives the hidden state matrix H output by the Spatiotemporal Mapping Calculation Unit. s Construct the time series feature matrix:

[0043]

[0044] Among them, h s,t d represents the hidden state at time step t. h As the hidden layer feature dimension, the similarity between time steps is calculated based on the correlation of time series data, and a correlation matrix C between time steps is constructed. s :

[0045]

[0046] Among them, C s (t1,t2) represents the correlation matrix between time steps t1 and t2, W c Given a trainable relevance transformation matrix, the weight matrix is ​​calculated based on the inner product of time-step features, h. s,t1 This represents the hidden state at time step t1. Show the hidden state at time step t2;

[0047] S332, Receptive Field Weight Calculation Unit, based on the correlation matrix C between time steps. s Construct time-step dependency weights and calculate the receptive field weight 'a' for each time step. s,t :

[0048]

[0049] in, The receptive field calculation function adaptively adjusts the receptive field range based on the correlation distribution between time steps, and the weight calculation considers the local patterns and global dependency information of the time series features.

[0050] S333, Dynamic receptive field adjustment unit, receives receptive field weight matrix A s ={a s,1 ,a s,2 ,...,a s,N The receptive field coverage r at each time step t is dynamically calculated. s,t :

[0051] r s,t =∑ t' I(C s (t,t')> );

[0052] Where, r s,tThis represents the size of the receptive field at time step t. To set a preset correlation threshold, I(·) is an indicator function, when C s (t,t') exceeds the threshold τ s At that time, time step t' is incorporated into the receptive field range of time step t, generating a dynamically adjusted receptive field size, and outputting it to the spatiotemporal attention generation unit of the defect region.

[0053] Optionally, S4 specifically includes:

[0054] S41, Multimodal Feature Extraction Unit: Receives the multimodal defect data stream output from the data acquisition module, extracts features from the optical imaging, electromagnetic detection, and ultrasonic detection data respectively, and constructs a multimodal feature matrix M. f :

[0055] M f ={M f,o M f,e M f,u};

[0056] Among them, M f,o M represents the feature matrix extracted from optical imaging data. f,e M represents the feature matrix extracted from the electromagnetic detection data. f,u This represents the feature matrix extracted from ultrasonic testing data;

[0057] S42, Feature Mapping Alignment Unit, based on multimodal feature matrix M f A modal feature transformation function is constructed to map features from different modalities to a unified feature representation, generating an aligned feature matrix.

[0058]

[0059] in, The feature mapping function normalizes the scale, distribution, and dimension of different modal features through nonlinear transformation;

[0060] S43, Feature Association Construction Unit, based on the aligned feature matrix Calculate the correlation between different modal features to form the intermodal feature correlation matrix C. f :

[0061]

[0062] Among them, C f (i,j) represents the feature correlation matrix between mode i and mode j, W c Given a trainable modal feature transformation matrix, calculate the similarity weights between cross-modal features. This represents the feature matrix corresponding to mode i. This represents the feature matrix corresponding to mode j;

[0063] S44, Cross-modal feature fusion unit, based on inter-modal feature correlation matrix C f A cross-modal feature fusion network is used to calculate the fused defect feature F. f :

[0064]

[0065] in, The cross-modal feature fusion function utilizes the similarity weights between cross-modal features to perform a weighted summation of features from different modalities, fusing local and global information to construct a global feature representation and generate a fusion defect feature matrix F. f .

[0066] Optionally, S5 specifically includes:

[0067] S51, Defect Feature Extraction Unit, receives the fused defect features F output by the cross-modal feature fusion module. f The fusion defect features are normalized and transformed to construct a normalized defect feature matrix S. d ;

[0068] S52, Defect Sample Generation Unit, based on the normalized defect feature matrix S d A variational autoencoder is used to map and normalize defect features to a latent variable space, and the latent variable z is set. s Following a standard normal distribution, calculate the posterior distribution and generate the defect latent variable matrix Z. s :

[0069] Z s ={z s,1 ,z s,2 ,...,z s,n},z s,n ~N(μ(s) d,n ),Σ(s d,n ));

[0070] Among them, z s,n μ(s) represents the nth normalized defect feature. d,n ) and Σ(s d,n ) represent the mean and variance parameters of the n normalized defect features, respectively;

[0071] S53, Defect Data Augmentation Unit, receives the defect latent variable matrix Z s A variational autoencoder decoder reconstructs defect samples from the latent variable space to generate a synthetic defect sample matrix.

[0072]

[0073] in, The decoder mapping function maps the defect latent variable matrix Z. s Mapping regression simplification of defect feature matrix S d The feature space in which it is located is used to generate multi-class synthetic defect samples;

[0074] S54, Training set optimization unit, based on the synthetic defect sample matrix. and normalized defect feature matrix S d The probability distribution matching degree between the synthetic defect samples and the normalized defect features is calculated, and an optimized training set is constructed.

[0075]

[0076] Where λ is a weighting coefficient used to control the proportion of normalized defect features and synthetic defect samples in the training dataset, and the final output is the optimized training set. And provide it to the dual-flow defect detection module.

[0077] Optionally, S53 specifically includes:

[0078] S531, Latent Variable Decoding Unit, receives the defective latent variable matrix Z. s The latent variables are mapped using the decoder of the variational autoencoder to generate a reconstructed defect sample matrix.

[0079]

[0080] in, For decoder mapping function, This represents the reconstructed defect sample matrix after initial decoding;

[0081] S532, Defect Sample Screening Unit: Based on the sample consistency measurement method, calculates and reconstructs the defect sample matrix. With the normalized defect feature matrix S d The similarity between samples is used to filter high-confidence defect samples and generate a filtered defect sample matrix. The screening method eliminates or retains samples based on the matching degree of the sample distribution;

[0082] S533, Multi-category Sample Amplification Unit, based on the screened defective sample matrix A class balance optimization strategy is employed to probabilistically amplify defect samples from different categories, generating the final synthetic defect sample matrix.

[0083]

[0084] in, Let α be the sample generation function, and let α be the sample amplification ratio control parameter.

[0085] The beneficial effects of this invention are:

[0086] (1) This invention solves the problems of accuracy and sample scarcity in defect detection during steel cord drawing production by combining multimodal data fusion technology, deep learning models, and variational autoencoders. Through feature extraction and fusion from multiple data sources such as optical imaging, electromagnetic detection, and ultrasonic detection, this invention can more comprehensively and accurately identify and analyze various defects in the production process, thereby improving the robustness and accuracy of the detection system.

[0087] (2) This invention ensures the consistency of data from different sensors in the same feature space through feature mapping alignment and cross-modal feature fusion, greatly enhancing the comparability and processing capability of the data. In addition, the multi-class defect samples generated based on the variational autoencoder further optimize the training set, solve the problem of insufficient defect samples, and improve the generalization ability and training effect of the model.

[0088] (3) This invention further improves detection accuracy and real-time response capability by utilizing techniques such as dynamic receptive field adjustment, spatiotemporal feature optimization, and defect data amplification. In particular, it can flexibly respond to the detection needs of various defect types in complex and ever-changing production environments. This method not only achieves efficient defect detection, but also further enhances the system's adaptability and intelligence level through continuous learning and data optimization. Attached Figure Description

[0089] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0090] Figure 1 This is a general framework diagram of an online detection system for defects in steel cord drawing production proposed in this invention. Detailed Implementation

[0091] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0092] refer to Figure 1 An online detection system for defects in steel cord drawing production includes the following steps:

[0093] S1, Data Acquisition Module, is used to acquire data on the surface morphology, internal conductivity changes and structural integrity of steel cord through optical imaging, electromagnetic detection and ultrasonic detection, forming a multimodal defect data stream;

[0094] This implementation combines optical imaging, electromagnetic detection, and ultrasonic testing to comprehensively acquire information on the surface morphology, internal conductivity changes, and structural integrity of steel cord, forming a multimodal defect data stream. This method provides a more complete defect characterization, improves the accuracy and stability of data acquisition, and provides a reliable data foundation for subsequent defect detection and analysis, thereby enhancing the quality control capabilities of the steel cord production process.

[0095] S2, Defect Feature Reconstruction Module: Based on multimodal defect data stream, it models the distribution of latent variables of defects through variational autoencoder, removes noise interference using denoising autoencoder, and outputs optimized defect features.

[0096] S3, the spatiotemporal feature optimization module, extracts the temporal features of defects in the steel cord drawing and twisting process based on the optimized defect features, and adjusts the receptive field by combining time series mapping to generate the spatiotemporal attention weight matrix of the defect area.

[0097] S4, Cross-modal feature fusion module, adopts feature mapping alignment method to perform multi-level feature extraction and information association on multi-modal defect data stream, and generates fused defect features through cross-modal feature fusion network;

[0098] S5, Defect Sample Enhancement Module: Based on fused defect features, it generates multi-class defect samples through variational autoencoder, and combines probability distribution to augment the dataset and construct an optimized training set;

[0099] S6, Dual-stream defect detection module, constructs a dual-stream architecture of denoising stream and attention stream based on the optimized training set, and outputs the detection results;

[0100] This implementation constructs a dual-stream architecture of denoising flow and attention flow. The denoising flow enhances defect features, while the attention flow accurately identifies defect regions, achieving efficient and precise defect detection. Based on the optimized training set, the system maintains stable detection performance in complex noise environments, improving the accuracy and robustness of defect identification and providing a reliable intelligent detection method for quality control in steel cord production.

[0101] S7, Detection Parameter Optimization Module: Based on the detection results, reinforcement learning is used to adjust the spatiotemporal attention weight matrix of the defect region and fuse defect features, and the detection results are optimized by combining dynamic feedback mechanism.

[0102] This implementation uses reinforcement learning to dynamically adjust the spatiotemporal attention weight matrix of the defect region and fuse defect features, enabling the system to adaptively optimize the detection strategy and improve the accuracy of defect identification. Combined with a dynamic feedback mechanism, the system can adjust parameters in real time based on the detection results, enhancing its ability to identify complex defects, effectively improving detection stability and generalization ability, and ensuring more intelligent and efficient quality control in the steel cord production process.

[0103] In this embodiment, S2 specifically includes:

[0104] S21, Defect data preprocessing unit, receives multimodal defect data stream generated by data acquisition module, extracts optical imaging, electromagnetic detection and ultrasonic detection data, normalizes and scales the data, and performs noise filtering to form a standardized defect feature matrix X;

[0105] S22, Defect Latent Variable Modeling Unit: Based on the standardized defect feature matrix X, a latent variable space is constructed. A variational autoencoder is used to map the standardized defect feature matrix X to the latent variable space. The latent variable z is set to follow a standard normal distribution. The posterior distribution is calculated, and the optimized latent variable matrix is ​​output.

[0106] D KL (q(z|x)||p(z));

[0107] Among them, D KL Kullback-Leibler divergence measures the difference between the latent variable distribution q(z|x) and the standard normal distribution p(z);

[0108] S23, the defect feature reconstruction unit, receives the optimized latent variable matrix, uses the decoder of the variational autoencoder to reconstruct the features, restores the latent variable z to the original feature space, generates the reconstructed defect data matrix X', and calculates the reconstruction error L. recon ;

[0109] S24, Defect Denoising Unit, receives the reconstructed defect data matrix X', applies Gaussian noise perturbation to the reconstructed defect data matrix X', performs denoising processing through a denoising autoencoder, and generates a denoised defect data matrix by using a learning mapping function;

[0110] S25, Defect Feature Optimization Unit: Receives the denoised defect data matrix, constructs a joint loss function, adjusts the weight ratio of reconstruction error and latent variable regularization term, and generates optimized defect features X. * :

[0111] L = L recon +βD KL (q(z|x)||p(z));

[0112] Where L is the joint loss function, L recon This represents the reconstruction error, and β is the regularization coefficient.

[0113] This implementation achieves high-precision reconstruction of defect features through multimodal data preprocessing, latent variable modeling, feature reconstruction, and denoising optimization. A variational autoencoder is used to construct the latent variable space and optimize defect features, making the defect data more stable and distinguishable. By adjusting the weight ratio of reconstruction error to regularization term using a joint loss function, the robustness and generalization ability of defect detection are improved, ensuring accurate defect identification even in complex environments and providing reliable support for quality control in the steel cord drawing process.

[0114] In this embodiment, S3 specifically includes:

[0115] S31, Defect temporal feature extraction unit, receives the optimized defect feature X output by the defect feature reconstruction module. * Defect temporal feature matrix T is constructed based on time series modeling methods. s T s It consists of N time steps t, and the feature vector of each time step is represented as:

[0116]

[0117] in, d represents the state of the defect feature at time step t. s Dimensions of defect characteristics;

[0118] S32, Spatiotemporal mapping calculation unit, based on the defect temporal feature matrix T s Construct a temporal relationship mapping and define the hidden state h at time step t. s,t To form the hidden state matrix H s :

[0119]

[0120] Among them, h s,t The hidden state at the current time step t. h is the time series mapping function. s,t-1 For the hidden state of the previous time step, calculate the time series correlation and adjust the time weight matrix to keep the information dependence of defect features consistent in the time dimension.

[0121] S33, Dynamic Receptive Field Adjustment Unit: Based on the calculated time-series correlation, it calculates the receptive field range of defect features at different time steps and defines the receptive field weight matrix.

[0122]

[0123] Among them, A s For the receptive field weight matrix, a s,t This represents the receptive field weight at time step t. The receptive field calculation function dynamically adjusts the receptive field size based on the time-series correlation changes of defect features;

[0124] S34. The spatiotemporal attention generation unit for the defect region calculates the spatiotemporal correlation of defect features based on the dynamic adjustment of the receptive field size, and constructs the spatiotemporal attention weight matrix W for the defect region. s :

[0125]

[0126] Among them, W s Let w be the spatiotemporal attention weight matrix for the defect region. s,t The spatiotemporal attention weights represent time step t. This is the function for calculating spatiotemporal attention.

[0127] This implementation extracts the temporal features of defects through time series modeling, and combines spatiotemporal mapping calculation and dynamic receptive field adjustment methods to construct the correlation of defect features at different time steps, thereby improving the spatiotemporal consistency of defect detection. By calculating the spatiotemporal attention weight matrix, the spatiotemporal feature changes of defect areas can be accurately identified, improving the positioning accuracy of dynamic defects and providing a more efficient and stable spatiotemporal feature optimization method for online defect detection in the steel cord production process.

[0128] In this embodiment, S33 specifically includes:

[0129] S331, Time Series Correlation Calculation Unit, receives the hidden state matrix H output by the Spatiotemporal Mapping Calculation Unit. s Construct the time series feature matrix:

[0130]

[0131] Among them, h s,t d represents the hidden state at time step t. h As the hidden layer feature dimension, the similarity between time steps is calculated based on the correlation of time series data, and a correlation matrix C between time steps is constructed. s :

[0132]

[0133] Among them, C s (t1,t2) represents the correlation matrix between time steps t1 and t2, W c Given a trainable relevance transformation matrix, the weight matrix is ​​calculated based on the inner product of the time-step features. This represents the hidden state at time step t1. Show the hidden state at time step t2;

[0134] S332, Receptive Field Weight Calculation Unit, based on the correlation matrix C between time steps. s Construct time-step dependency weights and calculate the receptive field weight 'a' for each time step. s,t :

[0135]

[0136] in, The receptive field calculation function adaptively adjusts the receptive field range based on the correlation distribution between time steps, and the weight calculation considers the local patterns and global dependency information of the time series features.

[0137] S333, Dynamic receptive field adjustment unit, receives receptive field weight matrix A s ={a s,1 ,a s,2 ,...,a s,N The receptive field coverage r at each time step t is dynamically calculated. s,t :

[0138] r s,t =∑ t' I(C s (t,t')> );

[0139] Where, r s,t This represents the size of the receptive field at time step t. To set a preset correlation threshold, I(·) is an indicator function, when C s (t,t') exceeds the threshold τ s At that time, time step t' is incorporated into the receptive field range of time step t, generating a dynamically adjusted receptive field size, and outputting it to the spatiotemporal attention generation unit of the defect region.

[0140] This implementation constructs a time-series feature matrix, calculates the correlation between time steps, and optimizes the time-step dependencies using a trainable transformation matrix, ensuring the integrity and accuracy of the time-series features. Simultaneously, a dynamic receptive field adjustment mechanism is employed to adaptively calculate the receptive field size based on the correlation between time steps, accurately identifying local patterns and global features, and increasing the spatiotemporal attention weight of defect regions, providing reliable support for the accurate detection of complex defects.

[0141] In this embodiment, S4 specifically includes:

[0142] S41, Multimodal Feature Extraction Unit: Receives the multimodal defect data stream output from the data acquisition module, extracts features from the optical imaging, electromagnetic detection, and ultrasonic detection data respectively, and constructs a multimodal feature matrix M. f :

[0143] M f ={M f,o M f,e M f,u};

[0144] Among them, M f,o M represents the feature matrix extracted from optical imaging data. f,e M represents the feature matrix extracted from the electromagnetic detection data. f,u This represents the feature matrix extracted from ultrasonic testing data;

[0145] S42, Feature Mapping Alignment Unit, based on multimodal feature matrix M f A modal feature transformation function is constructed to map features from different modalities to a unified feature representation, generating an aligned feature matrix.

[0146]

[0147] in, The feature mapping function normalizes the scale, distribution, and dimension of different modal features through nonlinear transformation;

[0148] S43, Feature Association Construction Unit, based on the aligned feature matrix Calculate the correlation between different modal features to form the intermodal feature correlation matrix C. f :

[0149]

[0150] Among them, C f (i,j) represents the feature correlation matrix between mode i and mode j, W c Given a trainable modal feature transformation matrix, calculate the similarity weights between cross-modal features. This represents the feature matrix corresponding to mode i. This represents the feature matrix corresponding to mode j;

[0151] S44, Cross-modal feature fusion unit, based on inter-modal feature correlation matrix C f A cross-modal feature fusion network is used to calculate the fused defect feature F. f :

[0152]

[0153] in, The cross-modal feature fusion function utilizes the similarity weights between cross-modal features to perform a weighted summation of features from different modalities, fusing local and global information to construct a global feature representation and generate a fusion defect feature matrix F. f .

[0154] This implementation achieves efficient integration of data from different modalities through multimodal feature extraction, feature mapping alignment, feature association construction, and cross-modal feature fusion. The system can extract key features from optical, electromagnetic, and ultrasonic detection data, and optimize feature alignment using nonlinear transformation and similarity weight calculation to ensure the consistency of information from different modalities in a unified feature space. This improves the accuracy and stability of defect identification and enhances the system's adaptability and detection capabilities in complex environments.

[0155] In this embodiment, S5 specifically includes:

[0156] S51, Defect Feature Extraction Unit, receives the fused defect features F output by the cross-modal feature fusion module. f The fusion defect features are normalized and transformed to construct a normalized defect feature matrix S. d ;

[0157] S52, Defect Sample Generation Unit, based on the normalized defect feature matrix S d A variational autoencoder is used to map and normalize defect features to a latent variable space, and the latent variable z is set. s Following a standard normal distribution, calculate the posterior distribution and generate the defect latent variable matrix Z. s :

[0158] Z s ={z s,1 ,z s,2 ,...,z s,n},z s,n ~N(μ(s) d,n ),Σ(s d,n ));

[0159] Among them, z s,n μ(s) represents the nth normalized defect feature. d,n ) and Σ(s d,n ) represent the mean and variance parameters of the n normalized defect features, respectively;

[0160] S53, Defect Data Augmentation Unit, receives the defect latent variable matrix Z s A variational autoencoder decoder reconstructs defect samples from the latent variable space to generate a synthetic defect sample matrix.

[0161]

[0162] in, The decoder mapping function maps the defect latent variable matrix Z. s Mapping regression simplification of defect feature matrix S d The feature space in which it is located is used to generate multi-class synthetic defect samples;

[0163] S54, Training set optimization unit, based on the synthetic defect sample matrix. and normalized defect feature matrix S d The probability distribution matching degree between the synthetic defect samples and the normalized defect features is calculated, and an optimized training set is constructed.

[0164]

[0165] Where λ is a weighting coefficient used to control the proportion of normalized defect features and synthetic defect samples in the training dataset, and the final output is the optimized training set. And provide it to the dual-flow defect detection module.

[0166] This implementation generates latent defect variables through a variational autoencoder and reconstructs defect samples using a decoder, thereby amplifying and optimizing the defect data. Based on a probability distribution matching strategy, an optimized training dataset is constructed to improve the generalization ability and accuracy of the defect detection model. This method effectively alleviates the problem of defect sample scarcity, improves the robustness of the detection system in complex environments, and provides reliable data support for high-precision online detection.

[0167] In this embodiment, S53 specifically includes:

[0168] S531, Latent Variable Decoding Unit, receives the defective latent variable matrix Z. s The latent variables are mapped using the decoder of the variational autoencoder to generate a reconstructed defect sample matrix.

[0169]

[0170] in, For decoder mapping function, This represents the reconstructed defect sample matrix after initial decoding;

[0171] S532, Defect Sample Screening Unit: Based on the sample consistency measurement method, calculates and reconstructs the defect sample matrix. With the normalized defect feature matrix S d The similarity between samples is used to filter high-confidence defect samples and generate a filtered defect sample matrix. The screening method eliminates or retains samples based on the matching degree of the sample distribution;

[0172] S533, Multi-category Sample Amplification Unit, based on the screened defective sample matrix A class balance optimization strategy is employed to probabilistically amplify defect samples from different categories, generating the final synthetic defect sample matrix.

[0173]

[0174] in, Let α be the sample generation function, and let α be the sample amplification ratio control parameter.

[0175] This implementation decodes the latent variable matrix of defects using a variational autoencoder to generate high-quality reconstructed defect samples. High-confidence samples are then selected based on a sample consistency metric to ensure stable data quality. A class balancing optimization strategy is employed to expand the sample size, improving the coverage of different defect categories, enhancing the diversity of training data and the model's generalization ability, thereby effectively improving the identification accuracy and reliability of the defect detection system.

[0176] Example 1:

[0177] To verify the feasibility of this invention, we deployed it in the online quality control system of a steel cord manufacturing company. The system covers the entire process of steel cord drawing and stranding, aiming to achieve high-precision defect identification and intelligent detection optimization through multimodal perception and deep modeling. This company has long faced the problem of mixed occurrence and missed detection of various types of defects such as surface cracks, bubbles, and structural faults. Especially in the high-speed drawing process, traditional detection methods are slow to react and cannot meet the needs of automated quality supervision.

[0178] First, the system acquires multimodal raw data of the steel cord in real time using three types of high-precision sensors: an optical imaging module acquires surface morphology images at a frequency of 0.01 mm / pixel and 100 frames / second; an electromagnetic detection module measures conductivity anomalies in the 10kHz–500kHz frequency band to monitor internal defect signals; and an ultrasonic detection module acquires structural reflection waveforms at a 5MHz pulse frequency to detect latent defects such as internal faults and microcracks. After bandpass filtering, normalization, and scale unification, all data are sent to the backend data reconstruction module through a unified interface.

[0179] In the data processing stage, the system introduces a deep reconstruction network based on a multimodal variational autoencoder to perform feature compression and latent variable modeling on the collected data. Each modal feature is mapped to a 128-dimensional latent variable space, with a normal distribution prior assumed. By constructing a joint loss function (a weighted combination of reconstruction error and KL divergence), the model can automatically learn the distribution patterns of potential defects in the data structure. Next, the system introduces a denoising autoencoder to apply Gaussian perturbation to the reconstructed features and denoise them, further improving the model's sensitivity and discriminative power against anomalous features.

[0180] Subsequently, the system constructs a time-series feature matrix based on the optimized defect features, with each time step consisting of 5 frames. A bidirectional temporal encoder is used to extract the variation patterns of defects over time. Combined with a spatiotemporal attention mechanism, the system calculates the correlation between each time step and the context, and constructs a dynamic receptive field weight matrix based on cosine similarity to dynamically adjust the feature focusing range. This mechanism effectively improves the localization accuracy of time-varying defects such as crack propagation and bubble movement.

[0181] To enhance the model's detection capabilities, the system further expands the defect sample set using a VAE generator. Based on the original data distribution, five typical defect samples (cracks, bubbles, scratches, delamination, and broken wires) are generated. After screening by the sample consistency evaluation module, high-confidence samples are retained. The synthetic training set is a mixture of original and synthetic samples in a 7:3 ratio, and a dual-stream detection network is constructed: one for noise-robust identification and the other for spatial focusing and defect classification.

[0182] In actual operation, the system was continuously deployed on three drawing production lines for 60 days, processing over 400,000 samples. The system achieved an average defect detection accuracy of 95.8% and a recall rate of 96.5%. Compared with a manual sampling inspection system, the system demonstrated a 34% improvement in accuracy for surface crack detection, a 70% reduction in false alarm rate, and a decrease in system response time from 30 seconds to 2.5 seconds, fully meeting the online quality control requirements under high-speed production conditions.

[0183] The data for "Fault Detection Accuracy," "False Alarm Rate," and "False Negative Alarm Rate" in the table are all derived from statistical results obtained by comparing the system's on-site operation records after deployment with manual verification. The specific calculation method is as follows:

[0184] Fault detection accuracy refers to the proportion of samples output as "defect" by the system that are actually defects. The calculation formula is:

[0185]

[0186] During the deployment period, the system detected a total of 32,157 defect samples. After manual review, 30,822 of these were confirmed as valid defect samples, resulting in an accuracy rate of 95.8%.

[0187] Recall (also known as detection coverage) refers to the proportion of samples with actual defects that are correctly identified by the system. The formula is:

[0188]

[0189] The actual number of defective samples was 31,949, of which the system successfully identified 30,822, resulting in a recall rate of 96.5%.

[0190] False alarm rate refers to the proportion of normal samples that are mistakenly identified as defects by the system. The calculation formula is:

[0191]

[0192] During the deployment period, the total number of normal samples was 179,632, of which 9,462 samples were misjudged as defective by the system, resulting in a false positive rate of 5.3%.

[0193] All data is automatically recorded by the system and verified by manual sampling and comparison of sample numbers, ensuring the objectivity and reliability of the statistics and meeting the performance evaluation requirements under industry standards.

[0194] Table 1. Effectiveness Report of the Online Detection System for Steel Cord

[0195]

[0196] As can be seen from Table 1, after deployment, the system significantly improves defect detection accuracy, response speed, and system robustness, effectively overcoming the problems of high detection error, slow response, and scarce samples in traditional methods. It provides strong support for online quality monitoring of steel cord drawing process and has significant practical value and promotion potential.

[0197] These data demonstrate that the present invention, through its intelligent fault perception and dynamic data scheduling mechanism, has successfully solved the problems of delayed equipment fault detection, high false alarm rate, high false alarm rate, and unbalanced system load in traditional industrial equipment monitoring systems. This significantly improves the operating efficiency and safety of enterprise equipment while reducing equipment downtime and fault repair costs.

[0198] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A steel cord drawing production defect on-line detection system, characterized in that, The method comprises the following modules: S1, a data acquisition module, for acquiring steel cord surface morphology, internal conductivity changes and structural integrity data through optical imaging, electromagnetic detection and ultrasonic detection to form a multi-modal defect data stream; S2, a defect feature reconstruction module, based on the multi-modal defect data stream, modeling the defect latent variable distribution through a variational autoencoder, removing noise interference using a denoising autoencoder, and outputting optimized defect features; S3, a spatio-temporal feature optimization module, based on the optimized defect features, extracting defect temporal features in the steel cord drawing and cabling process, and combining time series mapping to adjust the receptive field to generate a spatio-temporal attention weight matrix of the defect area; S4, a cross-modal feature fusion module, using a feature mapping alignment method to perform multi-layer feature extraction and information association on the multi-modal defect data stream, and generating fused defect features through a cross-modal feature fusion network; S5, a defect sample enhancement module, based on the fused defect features, generating multi-class defect samples through a variational autoencoder, expanding the data set combined with the probability distribution, and constructing an optimized training set; S6, a dual-flow defect detection module, based on the optimized training set, constructing a denoising flow and attention flow dual-flow architecture, and outputting a detection result; S7, a detection parameter optimization module, based on the detection result, adjusting the spatio-temporal attention weight matrix of the defect area and the fused defect features using reinforcement learning, and optimizing the detection result combined with a dynamic feedback mechanism.

2. A steel cord drawing production defect on-line detection system according to claim 1, characterized in that, The S2 specifically comprises: S21, a defect data preprocessing unit, receiving the multi-modal defect data stream generated by the data acquisition module, extracting optical imaging, electromagnetic detection and ultrasonic detection data, normalizing, scaling and standardizing the data, and performing noise filtering processing to form a standardized defect feature matrix X; S22, a defect latent variable modeling unit, based on the standardized defect feature matrix X, constructing a latent variable space, mapping the standardized defect feature matrix X to the latent variable space using a variational autoencoder, setting the latent variable z to follow a standard normal distribution, calculating the posterior distribution, and outputting an optimized latent variable matrix: D KL (q(z|x) || p(z)); where D KL is the Kullback-Leibler divergence, measuring the difference between the latent variable distribution q(z\x) and the standard normal distribution p(z). S23, the defect feature reconstruction unit receives the optimized latent variable matrix, uses the decoder of the variational autoencoder to perform feature reconstruction, restores the latent variable z to the original feature space, generates a reconstructed defect data matrix X', and calculates a reconstruction error L recon ; S24, a defect denoising unit, receiving the reconstructed defect data matrix X' and applying Gaussian noise disturbance to the reconstructed defect data matrix X', performing denoising processing through a denoising autoencoder, and generating a denoised defect data matrix using a learned mapping function; S25, a defect feature optimization unit, receives the denoised defect data matrix, constructs a joint loss function, adjusts the weight ratio of the reconstruction error and the hidden variable regularization term, and generates the optimized defect feature X * : L = L recon + βD KL (q(z\ x) \ \ p(z)); where L is the joint loss function, L recon denotes the reconstruction error, and β is a regularization coefficient.

3. A steel cord drawing production defect on-line detection system according to claim 1, characterized in that, The S3 specifically comprises: S31, a defect time sequence feature extraction unit, receives the optimized defect feature X output by the defect feature reconstruction module * , constructs a defect time sequence feature matrix T based on a time sequence modeling method s , wherein T s is composed of N groups of time steps, and the feature vector of each time step is represented as: wherein, denotes the state of the defect feature at time step t, d s is the dimension of the defect feature; S32, a space-time mapping calculation unit, based on the defect time sequence feature matrix T s , construct a time sequence relationship mapping, define the hidden state h of the time step t s，t , form a hidden state matrix H s : wherein h s，t is the hidden state of the current time step t, is a time series mapping function, h s，t-1 is the hidden state of the previous time step, the time series correlation is calculated, the time weight matrix is adjusted, and the information dependence of the defect features in the time dimension is kept consistent. S33, a dynamic receptive field adjustment unit, based on the calculated time series correlation, calculating the receptive field range of the defect features at different time steps, and defining a receptive field weight matrix: Wherein, A s is a receptive field weight matrix, a s，t represents the receptive field weight of time step t, is a receptive field calculation function, which dynamically adjusts the receptive field size according to the time series correlation change of defect features; S34, the defect region spatiotemporal attention generation unit, based on the receptive field size dynamic adjustment result, calculates the spatiotemporal correlation of the defect feature, and constructs a spatiotemporal attention weight matrix W of the defect region s : where W s is the spatio-temporal attention weight matrix for the defect region, w s，t denotes the spatio-temporal attention weight at time step t, is the spatio-temporal attention computation function.

4. A steel cord drawing production defect on-line detection system according to claim 3, characterized in that, The S33 specifically comprises: S331, a time series correlation calculation unit, receiving the hidden state matrix H output by the spatio-temporal mapping calculation unit s , constructing a time series feature matrix: where h s，t denotes the hidden state at time step t, d h is the dimension of the hidden layer, the similarity between time steps is calculated based on the correlation of time series, and the correlation matrix C s between time steps is constructed where C s (t1, t2) denotes the correlation matrix between time steps t1 and t2, W c is a trainable correlation transformation matrix, the weight matrix is computed based on the inner product of the time step features, denotes the hidden state at time step t1, denotes the hidden state at time step t2; S332, a receptive field weight calculation unit, calculates the receptive field weight a based on the correlation matrix C between time steps s , constructs a time step dependent weight, and calculates the receptive field weight a of each time step s，t : wherein, For the receptive field calculation function, the receptive field range is adaptively adjusted according to the correlation distribution between time steps, and the weight calculation considers the local mode and global dependent information of the time sequence characteristics. S333, a dynamic receptive field adjusting unit, receives the receptive field weight matrix A s = {a s，1 , a s，2 , …, a s，N}, dynamically calculates the receptive field coverage r s，t at each time step t r s，t =∑ t' I(C s (t,t′) ); wherein r s，t represents the receptive field size of time step t, is a preset correlation threshold, and I(·) is an indicator function, when C s (t, t') exceeds the threshold τ s , the time step t' is included in the receptive field range of the time step t, a dynamically adjusted receptive field size is generated, and output to the defect region spatiotemporal attention generation unit.

5. A steel cord drawing production defect on-line detection system according to claim 1, characterized in that, The S4 specifically comprises: S41, a multi-modal feature extraction unit, receiving the multi-modal defect data stream output by the data acquisition module, performing feature extraction on the optical imaging, electromagnetic detection and ultrasonic detection data respectively, and constructing a multi-modal feature matrix M f : M f = {M f，o , M f，e , M f，u}; where M f，o represents a feature matrix extracted from optical imaging data, M f，e represents a feature matrix extracted from electromagnetic detection data, M f，u represents a feature matrix extracted from ultrasonic detection data; S42, a feature mapping alignment unit, based on the multi-modal feature matrix M f , constructs an inter-modal feature conversion function to map different modal features to a unified feature representation, generating an aligned feature matrix wherein, is a feature mapping function that normalizes the scale, distribution, and dimensionality of different modalities features through a non-linear transformation; S43, a feature correlation constructing unit, constructs a feature correlation matrix C based on the aligned feature matrix correlation between different modal features, forming an inter-modal feature correlation matrix C f : where C f (i, j) denotes the feature correlation matrix between modality i and modality j, W c is a trainable modality feature transformation matrix, and the similarity weight between cross-modal features is calculated, denotes the feature matrix corresponding to modality i, denotes the feature matrix corresponding to modality j; S44, a cross-modal feature fusion unit, based on the inter-modal feature correlation matrix C f , the cross-modal feature fusion network is used to calculate the fused defect feature F f : wherein, is a cross-modal feature fusion function, which utilizes the similarity weight between cross-modal features to perform weighted summation on different modal features, fuses local and global information, constructs a global feature representation, and generates a fused defect feature matrix F f .

6. A steel cord drawing production defect on-line detection system according to claim 5, characterized in that, The S5 specifically comprises: S51, a defect feature extraction unit, receives the fused defect features F output by the cross-modal feature fusion module f , and performs normalization and feature transformation on the fused defect features to construct a normalized defect feature matrix S d ; S52, a defect sample generation unit, generates a normalized defect feature matrix S based on the normalized defect feature matrix S d , maps the normalized defect feature to a latent variable space z using a variational autoencoder, sets the latent variable z s obeys a standard normal distribution, calculates a posterior distribution, and generates a defect latent variable matrix Z s : Z s = {z s，1 , z s，2 ,..., z s，n}, z s，n ~ N ( μ ( s d，n ), Σ ( s d，n ) ) ; where z s，n represents the nth normalized defect feature, μ(s d，n ) and Σ(s d，n ) represent the mean and variance parameters of the n normalized defect features, respectively. S53, a defect data augmentation unit that receives the latent variable matrix Z s reconstructs the defect samples from the latent variable space using a decoder of a variational autoencoder to generate a synthetic defect sample matrix wherein, For the decoder mapping function, the latent defect variable matrix Z s is mapped back to the normalized defect feature matrix S d to generate the multi-class synthetic defect samples. S54、training set optimization unit, based on the synthetic defect sample matrix and the normalized defect feature matrix S d , calculate the probability distribution matching degree of the synthetic defect sample and the normalized defect feature, and construct the optimized training set Wherein, λ is a weight coefficient, used to control the proportion of normalized defect features and synthetic defect samples in the training data set, and finally output the optimized training set And provide to the double-flow defect detection module.

7. A steel cord drawing production defect on-line detection system according to claim 6, characterized in that, The S53 specifically comprises: S531, a latent variable decoding unit receiving the defect latent variable matrix Z s , mapping the latent variable by using a decoder of a variational autoencoder to generate a reconstructed defect sample matrix wherein, is a decoder mapping function, denotes the reconstructed defect sample matrix after preliminary decoding; S532, a defect sample screening unit, calculates a reconstructed defect sample matrix based on a sample consistency measurement method and a normalized defect feature matrix S d between the two, screens high-confidence defect samples, and generates a screened defect sample matrix The screening method performs sample elimination and retention according to sample distribution matching degrees. S533, multi-class sample amplification unit, based on the screened defect sample matrix Adopting a category balance optimization strategy, the probability of different categories of defect samples is amplified to generate a final synthetic defect sample matrix wherein, is a function generating a sample, and a is a sample expansion ratio control parameter.

Images