Recently, the “4th Industrial Revolution” has emerged as a major keyword for economic growth and has had a great effect in various fields, including manufacturing. In particular, the concept of Industry 4.0, which enables factories to become independent and self-adaptive depending on input from data that are gathered, is known in manufacturing as smart manufacturing. Smart factories are production systems wherein factory devices and parts are connected and interact with each other by combining existing production manufacturing technologies with technologies such as the Internet of things, big data, and cloud computing. A key feature of smart manufacturing is to assess and extract relevant information from collected data using deep learning [
1,
2,
3].
Because deep learning can analyze raw data and automatically identify feature representations of data across several levels of abstraction, it has attracted interest as a tool in smart manufacturing. The application of deep learning is not limited to process fault monitoring [
4,
5,
6] or state estimation [
7,
8]; several studies have explored its potential for various other manufacturing applications [
9,
10]. In deep learning, artificial neural networks (ANNs) and convolutional neural networks (CNNs) are widely acknowledged as the leading technologies for pattern recognition from tabular and image data, respectively.
Each layer of an ANN is made up of a collection of several perceptrons or neurons. Because an ANN only processes inputs in a forward manner, it is often referred to as a feedforward neural network. It can easily be used to process image, textual, and tabular data. Such neural networks are among the simplest variants. They pass information in one direction through various input nodes until sending it to an output node. The network may or may not have hidden node layers, rendering their functions more interpretable. Several studies have shown that ANNs can implicitly detect complex nonlinear relationships between dependent and independent variables. However, proper feature selection is crucial when applying an ANN. The features input into the model must be well designed according to the problem at hand. A CNN comprises convolution, pooling, and fully connected layers. A CNN is best used when millions of features need to be retrieved, since the convolutional layer generates feature maps that capture an area of an image that is then divided into rectangles and transmitted for nonlinear processing. The CNN automatically aggregates these characteristics rather than measuring each one separately. The fully connected layers use the extracted features to identify the input picture after the pooling layer reduces the number of the collected features [
11,
12]. CNNs based on auto-feature extraction have been used in various systems for fault detection, material degradation, and other applications. Glaeser et al. developed a fault-detection algorithm for industrial cold forging. Based on a CNN, the algorithm can detect faults with 99.02% accuracy, and classify each fault with 92.66% accuracy [
13]. Nakazawa and Kulkarni proposed a CNN with a SoftMax activation function to classify 22 WM defect patterns [
14]. Saqlain et al. proposed a deep learning-based CNN for automatic wafer defect identification (CNN-WDI) in semiconductor manufacturing processes [
15]. However, CNNs are better suited for processing image data rather than tabular data. Accordingly, several studies have utilized the conversion of tabular data into image data to leverage the advantages of CNNs, such as automatic feature extraction. Numerous time–frequency analysis techniques, including short-time Fourier transform (STFT), continuous wavelet transform (CWT), and wavelet packet transform (WPT), were combined with CNNs to convert tabular time-series data [
16]. These techniques use deep learning techniques to extract discriminative features from time–frequency representations rather than the time domain and convert continuous time-series data to two-dimensional (2D) representations using time–frequency analysis [
17,
18,
19,
20,
21]. The second method involves the conversion of numbers into images for noncontinuous time-series data. Sezer et al. [
22] generated 15 × 15 pixel images using 15 technical indicators related to stock prices. A CNN was adopted as the classification and prediction model to classify financial data as images and predict buy, sell, or hold signals for stocks. They evaluated the performance of their proposed model on Dow 30 stocks. In addition, Lee et al. [
23] converted tabular data, such as vehicle spare parts, into 3D voxel images and applied them to a 3D CNN to perform demand forecasting for spare parts. By comparing them with other methods, they concluded that the proposed method exhibited good prediction performance. However, there has been no research related to the application of CNNs using the dataset conversion of numbers into images for manufacturing processes. In addition, many manufacturing process data are recorded as noncontinuous time-series and tabular data types.
Consequently, the main contribution of this study is to detect defects in manufactured products by applying data obtained from the seat foam manufacturing process to the CNN algorithm. Since the data obtained from the manufacturing process are numerical data in tabular form, they were normalized, converted to gray images, and applied to the CNN algorithm. To solve the imbalanced data problem, data augmentation and hyperparameter optimization were also performed. In order to confirm the excellence of the proposed method, defect detection was performed by applying the features extracted from tabular numerical data to the ANN algorithm and then comparing the results with the results of the proposed method. Consequently, it was possible to develop a defect detection model with an accuracy of 98.33%, and the results confirmed the effectiveness of the proposed technique.