Research On Injection Molding Product Quality Prediction Method Based On Improved CNN-LSTM

Injection molding processing has been widely used in many traditional and emerging fields due to its cost-efficiency and high scalability production characteristics,including equipment manufacturing,fine chemicals,semiconductor manufacturing,and other industries.However,during the injection molding production process,there are inevitable variables such as material properties and processing conditions that can introduce uncertainties,resulting in a more challenging task to maintain consistent levels of product quality control.Although manual inspection can be used for quality control of injection molded products during production,it has issues of low accuracy and inability to provide real-time feedback.Thus,using machine learning methods to analyze the injection molding process data to improve production qualification rate has become a current research focus.This thesis focuses on the serious lag in traditional injection molding process quality feedback and builds a quality prediction model based on long short-term memory(LSTM)network and convolutional neural network(CNN)by improving the sparrow search algorithm model parameters and designing an injection molding product quality prediction prototype system,which realizes real-time prediction of injection molding product quality to solve the problem of difficulty in accurately predicting the quality of injection molded products.The research content of this thesis is as follows:Based on analysis of injection-molded product production process and data characteristics,this study focuses on high-dimensional and time-series datasets from quality inspection.Key technologies covered include data preprocessing methods,selection of key quality features,handling of time-series data,and optimization methods for quality prediction model parameters.The study constructs a quality prediction framework for injection molding products using an improved CNN and LSTM,describing the involved modules in detail.In order to enhance the accuracy of predicting injection-molded product quality,this study investigates dataset preprocessing for such products.Based on characteristics of these datasets,multi-stage analyses are conducted to extract features from high-frequency sensor data in realtime,with efforts made to achieve more comprehensive processing information.After cleaning and normalization of data,a comprehensive sampling method that combines over-sampling and under-sampling is introduced to address an imbalanced dataset of small samples.The method balances positive and negative samples prior to reducing data dimensionality.This approach lays a foundation for subsequent research aimed at predicting injection-molded product quality.To address the high-dimensional time-series characteristics of the injection molded product quality dataset,a prediction method combining random forest algorithm,LSTM neural network,and CNN is proposed.After selecting important features using the random forest algorithm,redundant features are further removed using the maximum information coefficient.A CNN is leveraged to extract localized features from the data input,then a LSTM neural network is employed for sequence modeling of these extracted features in order to capture the salient timebased patterns within the input data.An applied engineering research was conducted and the results demonstrated that this proposed method exhibits a superior predictive performance compared to three other benchmark algorithms.Specifically,the root mean square error reached 0.1.To address the challenge of how hyperparameters impact the precision of quality prediction models,the Sparrow Search Algorithm(SSA)is introduced for hyperparameter optimization.To overcome potential issues of SSA getting trapped in local optima,a Fusion strategy Sparrow Search algorithm(FSSA)is proposed for further improvement.By conducting simulations with various test functions in low-and high-dimensional data spaces,the results demonstrate that this method achieves higher convergence speed and more accurate solutions,thereby representing a significant enhancement to optimization performance.The quality prediction model of CNN and LSTM network optimized by FSSA algorithm is established.The engineering application results show that the error of the quality prediction model of injection molding products optimized by the proposed method is smaller than that optimized by particle swarm optimization and sparrow search algorithm,and the root mean square error reaches 0.02.In order to facilitate practical application,a prototype system for quality prediction of injection molding products was designed and developed.The front end and back end were written using Vue.js and.NET Core3.1 frameworks respectively.The B/S architecture design was used to realize the functions of quality prediction,data visualization,equipment management and basic information management,which provided an effective means for real-time quality prediction of injection molding product manufacturing process.