Manufacturing systems of the injection molding industry often have quality regarding problems caused by the absence of the quality inspection process that eventually leads to considering all manufactured products as non-defective. Since products manufactured in injection molding manufacturing systems are usually small and have low prices, the cost of inspecting the quality of all the products could take charge of a relatively high proportion of the total cost when the total inspection process is conducted. Thus, quality inspection process is often omitted, and only sampling-based quality investigation is carried out occasionally in the real field of the injection molding industry. In fact, several injection molding machines provide a function of discarding products manufactured in certain conditions that exceed lower and upper limits of processing data (e.g., injection pressure, switch over position, switch over pressure, etc.); however, filtering out defective products only utilizing simple logic has high limits. Therefore, this can lead to several problems as follows: 1) Service level of customer can be decreased by delivering defective products, 2) Total cost can be eventually increased by penalty of defective product delivery and urgent production or change of production plan to reproduce the faulty products, 3) Machines of the following process of the injection molding such as assembly process can be broken by processing defective products, and 4) Waste of plastic might occur by producing more quantities of products than the required quantities because of unawareness of quantities of defective products. So far, the quality acceptance range of injection molding products has been broad, since the products are usually utilized as small parts of huge end items such as automobiles. However, as safety becomes the most important issue in the automobile industry, which consumes a vast number of injection molding products, the quality requested for the injection molding industry begins to narrow. Moreover, as the shapes of injection molding products are designed in more complicated forms compared with the past, low-quality products cause a worse influence on machines of the following process of the injection molding process. Besides, as responsibility of companies for protecting the environment increases, wastes of plastic generated by over-production should be managed. Accordingly, four chronic problems that we stated are being magnified as important issues that should be addressed by predicting the quality of the injection molding products.
1.2 Related works
To address the problems that we mentioned, research on analyzing the quality of the injection molding products have been conducted actively. Sadeghi  developed a neural-network model to predict the soundness of the injected plastic parts utilizing melt flow rate, injection pressure, mold temperature, and melt temperature as independent variables. Kim et al.  investigated the pressure-volume-temperature (PVT) relationship realized in the injection molding process utilizing the ultrasonic technique and tried to predict products’ weights using the PVT relationship. As a matter of the quality prediction method, they considered a feed-forward back-propagation neural network. Bataineh and Barney  focused on predicting three forces: 1) local part-mold force, 2) local ejection force, and 3) total ejection force in the injection molding process with numerical simulation. Utilizing the experimentally derived coefficients of friction, they developed a simulation system that could forecast forces and insisted that the quality of the injected products could be estimated by the results. Chen et al.  developed a self-organizing map-based back-propagation neural-network model for predicting injection molding product’s quality considering parameters such as injection stroke, injection velocity, injection pressure, and switch over time as input variables of the algorithm. They utilized weights of products as indicators of quality. Tsai et al.  examined how the eight process parameters of temperature and pressure affect the quality of the optical lenses’ surface. They especially considered waviness, transmission rate, and roughness as quality factors. They developed three regression models that predicted each quality factor and compared the accuracy of the model with numerical experiments. Lee et al.  took the approach of a rule-based quality prediction framework. They proposed system architecture consists of three layers: 1) the application layer, 2) the engine layer, and 3) the database layer, but did not provide any experimental results. Yin et al.  tried to predict the warpage of plastic products produced by the injection molding process and developed a back-propagation neural network. They mainly considered five process parameters: mold and melt temperature, packing pressure, packing time, and cooling time of the injection molding process. Moayyedian et al.  evaluated the effects of diverse processing parameters (e.g., filling time, part cooling time, pressure holding time, and melt temperature) on short shot defect. They insisted that melt temperature mostly affects the possibility of short shot defects, with a contribution of 74.25%. Ramana et al.  also considered short shot defects and developed a statistical automated neural networks algorithm, a general chi-square automatic interaction detector algorithm, and an association rule algorithm to predict the quality of the injection molding products. Ramana et al.  proposed three algorithms based on a decision tree, k-nearest neighbor, and polynomial by binomial classification methodologies to predict short shot of injection molding product considering processing parameters such as injection speed, nozzle temperature, and injection time, etc. Kwon et al.  especially focused on injection speed among the processing parameters. They analyzed how injection speed in the injection molding process affects physical variations of the products, such as shrinkage and deformation, utilizing a simulation-based mold-flow analysis program. Ogorodnyk et al.  proposed a multi-layer perceptron-based algorithm to predict the quality of thermoplastic products. Besides, many other studies, such as [13–21], have developed machine learning-based quality prediction algorithms for injection molding products.
These previous studies have common limitations. First, they only considered the process parameters provided by injection molding machines regarding injection pressure, temperature of nozzles and barrels, switch-over point, and approximate mold temperature. Although they commonly insisted that mold data are important factors that affect the quality of the injection molding process, they did not consider in detail how temperature and pressure change inside the molds. Moreover, data that could be collected by attaching additional sensors, such as vibration of the injection molding machines did not be considered. Second, they did not consider practical types of defects that could be utilized in the real field. Most previous research has dealt with defects regarding the weights of products or the approximate transform of the shapes. However, in the real field, more specific types of defects regarding injection molding products’ shapes should be considered. There are some studies considering defects of the injection molding products detected by vision images (e.g., [22–24]); however, they only concentrated on detecting flaws on the surface of the injection molding products, and the detected defects were not analyzed with the processing parameters. Third, previous studies have conducted experiments in the laboratory, not in the real field. This research examines the quality prediction problem of injection molding products, considering the limitations of previous research.
1.3 Overview of the research
The purpose of this research is to develop a neural-network-based quality prediction algorithm for injection molding products targeting a real injection molding factory in Korea. We consider a Fanuc injection molding machine (Roboshot S-2000I100B), and the considered machine of the real field is shown in Fig. 1. Target product of this research analyzed is a printed-circuit-board (PCB) connector, which is a part of automobile.
As independent variables of the algorithm, we used the online temperature and pressure data of the mold collected by cavity sensors that we installed inside the mold of the machines. By utilizing the data collected from additionally installed sensors inside the mold, we can consider more directly related data when the process of a product is injected. We also utilized vibration data collected by sensors installed on the cylinder and motor of the injection molding machine as independent variables. Dependent variables, which are classifications of injection molding products’ quality, are derived from vision-image data of the products collected by the vision camera that we installed. We classified the quality of the products into three grades with two indicators regarding the shape of the products: 1) flection of the housing and 2) alignment of pinholes. The overview of this research is presented in Fig. 2.
The rest of this paper is organized as follows. Section 2 explains how we constructed an infrastructure for collecting data for the injection molding machine. Section 3 presents in detail how the collected raw data are processed into the independent and dependent variables and then proposes a neural-network algorithm for quality prediction. Section 4 shows the performance of the developed quality prediction algorithm and discuss the results. Finally, Section 5 concludes the paper and discusses future studies.