Research On Quality Prediction Method Of Injection Molding Process Based On Recurrent Neural Network

The value of the product in the injection molding process is high,but the quality is difficult to control.The quality inspection is often carried out by manual sampling,resulting in a delay in the discovery of waste products and a high rejection rate,etc.With the development of enterprise informatization,a large number of sensors are deployed in the injection molding process,production data can be collected and stored,and process information can be mined to achieve quality prediction by using deep learning methods.In this paper,based on the recurrent neural network model,the research on the quality prediction method of injection molding process is carried out from three aspects: variable selection in complex industrial environment,extraction of key sequence features and model transfer in a new working condition.(1)Aiming at the problem that the high-dimensionality of the injection molding process data leads to the complexity of the quality prediction model and the decrease of the prediction accuracy,a sensitive variable selection method based on the latent space representation of the sequence-tosequence(Seq2Seq)model is proposed.Using the information in the time dimension of the injection molding process variables,the Seq2 Seq model is used to reconstruct the variables? the latent space representation is extracted? and based on the correlation between the latent space representation and the quality variable,the screening of sensitive variables for quality prediction in the injection molding process is realized.This method innovatively realizes the solution of the correlation between two-dimensional sequence variables and one-dimensional variables.(2)Aiming at the problem that the quality prediction is difficult due to the complex temporal relationship in the injection molding process,a quality prediction method based on the key sequence feature attention(KSFA)double-layer LSTM(DLLSTM)is proposed.Establish a dual-level LSTM to fit the strong correlation within the multi-phase and the weak correlation between the phases in the injection molding process,so as to avoid the different correlation between the phases and within the phase from affecting the prediction accuracy? and extract the quality-related sequence features within the phase through KSFA,and obtain the critical moment that contributes greatly to the quality prediction? the quality-related sequence features are integrated through the top LSTM of the dual-level LSTM to perform the quality prediction.This method can simulate the way of production information transmission in the actual injection molding process,and visualize the key moments,which has certain interpretability.(3)Aiming at the small sample quality prediction problem caused by new working conditions in the injection molding process,a new working condition transfer learning quality prediction model based on joint Y(JY)multi-kernel maximum mean difference(MKMMD)is proposed.For the source domain KSFA-DLLSTM model,the top LSTM parameters are transferred to the target domain model? then,the distribution of the top LSTM output feature layer of the target domain and the source domain are aligned by the JYMKMMD method? at the same time,the MKMMD method is used to align the distribution of the top LSTM input feature layer of the target domain with the one of the corresponding part of the source domain.The method can quickly deploy the quality prediction model for new working conditions in the injection molding process by mixing model transfer and feature transfer.