A Study On Intelligent Quality Detection Tech-Nology Of Injection Molding Products

Automatic product quality detection technology is beneficial to improve the automation level of plastic injection molding production and reduce labor cost.It also provides important feedback and decision-making basis for process parameter optimization and process control.In view of the current situation that product quality testing relies heavily on labor,this paper proposes four basic objectives of intelligent detection technology,namely,non-destructiveness,high efficiency,robustness and universality.Based on in-depth study of injection molding process and product quality characteristics,two important development directions of intelligent quality inspection technology for injection molding products have been established: direct detection based on machine vision and indirect detection based on process data.The two methods have been studied and have achieved a series of results:An efficient product image feature learning and defect recognition method is proposed.Aiming at the problem that traditional methods cannot effectively extract product defect information,improved image feature learning methods are proposed,including multichannel convolution,multi-scale convolution,deep convolution,nonlinear convolution and nonlinear transformation.Based on these methods,a product defect recognition model of deep convolutional neural network is constructed.The results show that the accuracy of defect recognition is improved from 88% to 97% comparing with traditional methods,and it is robust to the problems of positional shift,angular rotation,illumination variation and image scaling.In addition,this method does not require complicated parameter settings,and the model training and defect recognition processes do not require manual intervention.A method for automatic localization and detection of product defects is proposed.Aiming at the problem of multi-defect area localization in the quality detection,a defect region proposal network is proposed to realize the automatic localization of the defect regions.Furthermore,the defect detection method of deep convolutional neural network is proposed,which realizes the unification of three steps of product image feature learning,defect region localization and defect type identification.The results show that the proposed method reduces the average time of defect localization from 4s to 35ms;meanwhile,defect region proposal recall increases from 80% to 100%.The method has good universality for defects of different products,different types and forms,and can not only locate the position of the defect efficiently,but also accurately determine the type of defects,and the average accuracy of defect detection reaches 98%.A convolution-deconvolution auto-encoder feature learning method for the injection molding process is proposed.Since the traditional method can not effectively express the non-Gaussian,nonlinear,time series autocorrelation and dynamic cross-correlation of the process variables,based on the in-depth study of the characteristics of the injection molding process,this work first proposes the 4-D representation method of the injection molding process data,which effectively preserves the characteristics of the injection molding process.Further,the process feature learning method of convolution-deconvolution auto-encoder is proposed,and the basic characteristics of the injection molding process can be learned unsupervised from the historical trajectories of the process variables.The experimental results show that the proposed method is more effective than the traditional feature extraction method while retaining the same feature dimension.The process data reconstruction error is less than 56% of the latter.In addition,under a variety of different process conditions,it has superior model generalization ability and feature extraction ability,which is a powerful manifestation of the robustness and universality of the proposed method.A quality detection method for injection molding products based on process feature is proposed.Aiming at the problem of process condition identification,product quality anomaly detection and product quality prediction during injection molding,a unified online quality inspection method based on injection molding process features is proposed.Aiming at the problem of working condition identification,a multi-class classification model between process characteristics and forming conditions is established.The average accuracy of working condition recognition is improved by 3.5% compared with the traditional SVM method.Aiming at the problem of product quality anomaly detection,a binary classification model between process characteristics and quality anomaly is established.When the traditional MPCA method cannot effectively detect the quality anomaly(AUC=0.5),the abnormal detection recall rate of this method reaches 0.76(AUC=0.713).Aiming at the problem of product quality prediction,a regression model between process features and product weight is established.The correlation coefficient of product weight prediction results is 35.33% higher than the traditional method,and the average relative error is reduced to 49.64%.The intelligent detection system for injection molding products is designed and implemented.Considering the industry characteristics of injection molding,an efficient and flexible intelligent detection system is developed and implemented,including flexible communication based on MQTT,reliable data service based on MongoDB,and efficient computing service based on TensorFlow.Moreover,the intelligent detection platform based on machine vision is built independently,and the intelligent detection terminal based on process data is developed;and they are applied in the enterprise cooperation projects.