1. Introduction
Injection molding is one of the most widely used plastic processing technologies—more than 30% of plastic products are produced by injection molding [
1]. Today injection molding is considered a highly automated mature technology. However, to stay competitive and adapt to the ever-changing market demands, injection molding companies must move towards smart manufacturing, or Industry 4.0 [
2]. Industry 4.0 involves the digitization of production, which inevitably leads to the generation of Big Data [
3]. The Big Data lifecycle includes the generation, acquisition, storage, processing, and analysis of data. During injection molding, a huge amount of data is generated and can be successfully collected by sensors installed in different units of the injection molding machine and the cavity [
4]. Although data is one of the most valuable assets of an intelligent company, many companies have difficulties selecting essential and useful data from the manufacturing process and processing and analyzing this data effectively [
5]. Therefore, many efforts have been made to adopt machine learning (ML) techniques for industrial application. Several authors are of the opinion that ML is one of the most important factors in upgrading a traditional manufacturing system to Industry 4.0 [
6].
According to Bertolini et al. [
6], ML is a set of methodologies and algorithms that can extract knowledge from data and continuously improve their capabilities by learning from experience (i.e., from data accumulating over time). All ML methods can be divided into three groups: supervised learning, unsupervised learning, and reinforcement learning (
Figure 1). In supervised learning, the feature acquires the relationship between the inputs and outputs using information contained in the dataset of training examples [
7]. All the output data is labeled or grouped. Based on the type of outputs, supervised learning can be divided into two categories: classification and regression. Classification algorithms are used for discrete outputs, while regression algorithms are used for continuous outputs [
8]. Unsupervised learning uses unlabeled datasets; therefore its goal is not to make a prediction, but rather to detect or extract patterns in the data, the nature of which may be partially or completely unknown [
6]. Reinforcement learning is not concerned with the specific form of the input, but focuses on the action that should be taken under the current state to achieve the final goal [
9]. The most explored ML methods are supervised learning, closely followed by unsupervised learning.
One of the main applications of ML algorithms in the injection molding industry is quality management. Product quality in injection molding is quite a complex issue, as quality can be interpreted in various ways [
11]. However, three important groups of quality indicators can be distinguished: (1) the stability of dimensions and weight of the produced parts [
12], surface properties (roughness, sink marks, weld lines, etc.) [
13], and physical properties (mechanical, optical, electrical, etc.) [
14]. In most cases, a combination of these criteria is understood as the quality of a part. Nevertheless, several studies have proved that weight is a reliable index characterizing the quality of an injection molded product and process stability, as variation in weight is inversely proportional to part quality [
15,
16].
Besides the direct measurement of the weight of the part, which is usually a quality control procedure, it is essential to find a reliable process parameter which will allow monitoring and predicting the weight of a part online. Changes in polymer properties, particularly the specific volume of the melt, clearly indicate changes in the weight of the part. Therefore, monitoring the specific volume of melt through controlling the pressure and temperature in a mold cavity is a reliable tool to predict weight variation [
17]. According to Zhou et al. [
18], the specific volume of melt is mainly affected by pressure. Therefore, they proposed a pressure integral as an effective process parameter to predict product weight variations and characterize the quality of the injection molded parts. The most relevant pressure data come from the runner and mold cavity [
19]. The importance of characteristics from pressure curves can have a complex relationship with the quality of products, which may not be described with simple linear functions. In this case, ML could help discover this relation, as ML algorithms work in multidimensional spaces by learning the structure in the dataset [
20].
Recent progress in the application of ML in injection molding is summarized by Selvaray et al. [
21]. Zhao et al. [
22] used the ML approach to optimize the processing parameters to achieve a target weight of an injection-molded product. The authors used the support vector machine (SVM) method together with the particle swarm optimization (PSO) algorithm. The proposed ML approach enabled stable injection molding. The deviation of product weight was only 0.0212%. Yin et al. [
23] proposed a back-propagation neural network to predict and optimize the warpage of injection-molded parts based on the main process variables, including mold temperature, melt temperature, packing pressure, packing time, and cooling time. The proposed method was able to predict the warpage of injection-molded parts within an error range of 2%. Ogorodnyk et al. [
24] applied four ML methods to create prediction models for the thickness and width of the injection-molded HDPE tensile specimens based on the injection molding process parameters. The authors found the best correlation coefficient was achieved with the random forest algorithm, while the second-best results were produced by the multilayer perceptron (MLP) neural network method. The reduced error pruning decision tree (REPTree) performs slightly better than the k-nearest neighbor (kNN) algorithm. The authors concluded that overall, all the four methods showed good prediction capabilities. Ke and Huang [
25] used an MLP neural network to evaluate the quality of injection-molded parts. Instead of using processing parameter settings as inputs, the authors used the so-called “quality indices” [
26] extracted from the system and cavity pressure curves. The maximum achieved accuracy of the proposed prediction method was 94%. Gülçür and Whiteside [
27] also used quality indexes connected with cavity and system pressure and the position of the injection piston to predict the quality of micro-injection-molded parts. They used a linear regression model, which predicted quality with an accuracy of 84%.
Several authors confirmed that cavity pressure is a valuable data source that can represent the quality of injection molded products [
27,
28,
29]. In a typical cavity pressure profile, several feature points can be extracted that define the characteristics of the injection molding conditions [
28,
29]. However, the question still remains: which and how many features should be selected from the cavity pressure profile to adequately characterize and predict the quality of injection-molded products? For example, Huang et al. [
30] used four features from the cavity pressure curve: the peak pressure, the pressure gradient, the viscosity index, and the energy index. Gim and Rhee [
28] proposed five features to be extracted from cavity pressure profiles: the starting point of the filling stage, the switchover point from filling to packing, maximum cavity pressure, the endpoint of the packing stage, and end of the cooling stage. Determining the optimal number of features is a difficult question. However, Hua et al. [
31] give some recommendations on how to define the optimal number of features based on sample size. The general rule is that the sample size must exceed the number of features [
31]. For example, for the LDA algorithm and a sample size of 30, the optimal number of features can vary from 3 to 12, depending on the correlation of the features. Jain and Waller [
32] claim that the optimal feature size is proportional to
, where n is sample size. One more rule that can help to define the optimal number of features is the Vapnik–Chervonenkis (VC) inequality [
33], which gives an upper bound for generalization [
34]. However, this generalization rule is only true if the VC dimension is finite, which is not the case for the kNN algorithm, for example, with k = 1 [
35,
36]. For discrete classifiers, there is a more accurate approach, which recommends a significantly smaller size of the learning sample [
37]. In summary, the optimal number of features can differ for different classification algorithms.
Many authors confirm that classification and regression ML algorithms can predict and control the quality of injection molding well. However, the great variety of the ML algorithms and the individual features of each production run requires the development of a new prediction method. In this study, we aim to compare the accuracy and effectiveness of four classification algorithms in predicting the quality of multi-cavity injection molding. We used pressure-based quality indexes as inputs for the classification algorithms.