Research On Defect Detection Method Of Precision Parts Based On Deep Domain Adaptation Network

The high-end equipment manufacturing industry,represented by aerospace,energy and power,and defense and military industry,continuously improves the quality requirements for precision parts.Due to the influence of various complex factors such as production technology and manufacturing environment,various defects are easily generated on the machined surface of parts.These defects not only affect the appearance of products,but also affect the mechanical properties of parts themselves and their subsequent products to varying degrees,greatly reducing the quality of products and causing serious economic losses to manufacturing enterprises.Therefore,detecting surface defects of precision parts and improving the surface quality of terminal parts are of great significance for enhancing the competitive advantage of the manufacturing industry and promoting high-quality development of the manufacturing industry.With the rapid development of machine vision technology,intelligent detection technology based on deep learning has become a research hotspot in both academia and industry due to its advantages in automation and high-precision detection.However,the training samples used for intelligent detection in the actual industrial environment show the characteristics of little historical accumulation and difficult collection,and the existing data volume cannot match the intelligent model,resulting in weak model generalization.Moreover,the variable detection environment leads to poor adaptability of intelligent detection methods,which limits the application of intelligent detection technology in industrial detection.Aiming at the problems of incomplete detection information,poor adaptability and weak generalization of detection model in precision parts defect detection under unknown environment,this paper focuses on precision parts such as intermediate shells,impellers,and gears in automotive transmission systems,and studies deep domain adaptation technology to improve the detection performance of models for small sample defects in unknown environments,providing feedback information for discovering problems and adjusting processes in the production process of parts.The main research work of this paper includes the following parts:(1)To address the problem of small sample and distribution differences in defects faced by actual industrial inspection sites,a supervised domain adaptation defect classification method based on adaptive-adversarial learning is proposed.Firstly,the distribution characteristics of the data are analyzed,the uniform distribution strategy of domain labels is established,and the construction method of adversarial convolutional neural network model is studied to reduce the marginal distribution difference between domains.Then,a label constraint strategy is constructed in the category space,and a label probability alignment mechanism is proposed to reduce the conditional distribution difference between domains,thereby achieving defect feature knowledge transfer between different environments.Finally,the characteristics of nonlinear function are explored,the convex optimization method of objective function is studied,and the learning rate adaptive adjustment method based on loss and weight changes is constructed to improve the training performance of the model on small samples.The proposed model is applied to the detection of the published available defect datasets from Northeastern University and the actual collected defect data set of intermediate shell and impeller machining surface.The experimental results show that the proposed model has higher accuracy,with an impeller defect detection accuracy of 97.4%,which effectively solves the detection problem of the shortage of supervised defect samples.(2)To solve the detection problem of complex structures and diverse defects in precision parts,a semi-supervised domain adaptation defect classification method based on recalibration adversarial learning is proposed.Firstly,the adaptability of features is explored to establish feature calibration strategies,and the construction method of recalibration adversarial network model is studied to extract domain-invariant feature representations.Then,the coupling relationship between small-batch features and model parameters is analyzed,and the spatial semantic storage dictionary is constructed to decouple defect features.The semantic consistency is studied to construct task-oriented alignment module,extracting class-discriminability feature representations.Finally,the relationship between the defect feature and the gradient of the objective function is analyzed,and the objective gating function based on the unlabeled feature is established to balance the gradient between the features,fully mining the information of defect feature.The experimental analysis shows that the proposed algorithm achieves higher detection performance,especially when there is only one label sample for each type of defect,the defect detection accuracy of the impeller reaches96.0%,improving the adaptability of the model.(3)To solve the detection challenges of large intra-class difference and high inter-class fuzziness of precision parts defects,an unsupervised domain adaptation defect classification method based on dual-adversarial learning is proposed.Firstly,the relationship between data and decision boundary is analyzed,and the significance augmented mechanism is established to enrich the defect information of different environments.Then,the category and domain characteristics of features are studied in the augmented space,and semantic consistency strategies are established to generate clear class boundaries.Secondly,the transferability of features is analyzed,a dual-adversarial strategy based on features and their gradients is established,and a selective dual-adversarial network model construction method is studied to reduce the marginal distribution difference between domains.Finally,the model deviation is analyzed to study class confusion method,reducing the conditional distribution difference between domains,and improving model adaptability.The experiment shows that the proposed model achieves higher detection performance without labels,with a defect detection accuracy of 93.0% for the impeller,improving the generalization of the model.(4)To solve the problem of small defects in precision parts and strong background noise interference,a conditional domain adaptation defect localization method based on consistent adversarial learning is proposed.Firstly,the structural characteristics of features is explored,and the image-instance multi-level adversarial network is constructed to reduce the distribution difference between domains.The discrimination result of adversarial network is analyzed to establish a consistent adversarial strategy,strengthening the performance of the network.Then,the weak supervised location method is studied,and the image-level predictor is established to initially locate instance features of the interest region,solving the problem of background noise interference.Finally,the prediction similarity between the image-level predictor and the bounding box predictor is measured,and a conditional constraint strategy is established to adaptively adjust the proportion of difficult-to-align instances in the model training process to improve the location accuracy of small defects.The proposed model is used to detect the actual collected defect data set of impeller and gear machining surface,in which the mean average precision(m AP)of gear defect location of the proposed method is increased from 72.3% to 82.1% compared to the original YOLO model,improving the positioning accuracy of the model.(5)Based on the above research results,aiming at the small sample defect detection problem of intermediate shell,impeller and gear precision parts in automobile transmission system under unknown environment,a set of precision parts defect detection system is developed.The system architecture integrates data selection and calibration module,model training module and defect detection module,and effectively integrates three subsystems of data acquisition,data processing and data display,providing strong support for the development and on-site application of mechanical component defect detection technology based on deep domain adaptation networks.