Research related to CBM (condition monitoring maintenance) and PHM (prognostics health and management) is being conducted worldwide in the railway sector. In the case of the Republic of Korea, a significant amount of research related to smart maintenance is being carried out as a part of a national R&D project due to the rapid decrease in population. Currently, the preventive maintenance method is the most commonly used in the field of railway vehicles [
1]. In addition, during the stage of light maintenance prior to heavy maintenance, faults are often visually inspected by the mechanics themselves. As a result, there are instances where indicators of potential failures are overlooked, or defects remain undetected. Furthermore, a disadvantage exists in that the preventive maintenance method requires the premature replacement of parts that are still available. For this reason, maintenance costs increase. Therefore, it was determined that applying CBM technology to the current maintenance method, preventive maintenance could reduce maintenance costs. For example, in the case of wheels of railway vehicles, which is the subject of this paper’s research, wheel turning is required according to the safety standard of
Table 1 in the Republic of Korea data from [
2]. However, due to mechanics’ work errors, wheel turning is frequently performed, even when it is not necessary. Therefore, this suggests that the introduction of CBM technology could reduce unnecessary maintenance costs. In addition, in the case of Korea, the railway vehicle maintenance market is about USD 400 million. Of these, light maintenance is USD 280 million, accounting for about 70% [
3]. For this reason, this indicates that the railway sector could develop further if it invested elsewhere by reducing unnecessary maintenance costs.
Chenyi Zhou et al. proposed a method of detecting the location of wheel flats by installing sensors on rails. At this time, a method of finding a location using the response results through the rail strain was proposed through a sensor installed on the rail [
4]. Stasys Steišūnas et al. modeled the railway vehicle and wheel flats using SIMPACK, a multi-body dynamics software. They also proposed a method to define the parameters of the suspension using acceleration data generated by wheel flats [
5]. Xinyu Peng et al. conducted research by modeling the railway vehicle and wheel through SIMPACK, as in this paper and Stasys Steišūnas’ papers. At this time, Xinyu Peng compared and verified the SIMPACK model and the data of the actual railway vehicle. As such, in the field of railways, many researchers, including this paper, use SIMPACK, a dynamics software. Therefore, it is possible to prove the reliability of SIMPACK through many papers [
6]. Bo Liang et al. conducted research to detect surface defects on wheels and rails. They applied STFT (short-time Fourier transform), WVT (Wigner–Ville transform), and WT (wavelet transform) for their study [
7]. Mohammadreza Mohammadi et al. used unsupervised learning to detect wheel flats. At this time, the data were obtained through 113 simulations without using an actual vehicle. Also, the data were obtained from the rail, not from the railway vehicles. In addition, a total of four unsupervised learning models were used [
8]. Chuncheng Yang et al. used supervised learning methods to detect defects in rails. Unlike this paper, a wavelet was selected for signal processing at this time. And the deep learning models used were ResNet and FCN [
9]. Long Zhang et al. proposed a method for diagnosing defects in a rotating machine by applying a CNN algorithm. At this time, the model was evaluated by proposing an improved model based on the LeNet-5 model, similar to this paper [
10]. Run Gao et al. proposed a method of detecting wheel flats using a reflective optical position sensor. The research was conducted by making a multi-body dynamics model. At this time, unlike this paper, sensors were installed on a rail, not a wheel [
11]. Gabriel Krummenacher et al. proposed two methods to detect anomalies in wheel flats. First, unlike this paper, they used the SVM (support vector machine) model based on wavelets to detect wheel flats in time series data. Second, they used the deep neural network (DNN) algorithm to detect wheel flats. The results of the two methods were then compared and analyzed [
12]. Yunguang Ye et al. first discussed the severity of shocks resulting from faults in rotating machinery. Subsequently, they conducted research on shock detection to diagnose these faults. They proposed an adaptive feature called ATDI (activated time–domain image) and utilized a hybrid approach, ATDI-NN, combining DNN for wheel flat detection [
13]. Yunguang Ye et al. proposed a novel entropy called MTFIE (multislice time-frequency image entropy) for wheel fault diagnosis. They utilized time-frequency images incorporating machine fault information [
14]. J. Brizuela et al. introduced the use of ultrasound techniques for wheel flat detection. They employed Rayleigh wave ultrasonic pulses for detection. Finally, they successfully validated the detection of wheel flats using this technology [
15]. Araliya Mosleh et al. utilized envelope spectrum analysis for the detection of wheel flats. They tested and analyzed the sensitivity of the proposed method. Finally, they concluded that it is an efficient approach for wheel flat detection [
16]. Esteban Bernal et al. presented a dynamic verification of an onboard wheel flat detection technique using analog signal processing. And this was shown to reduce the power consumption and hardware costs of condition-monitoring sensor nodes [
17]. Yunguang Ye et al. conducted research that considered not only the detection of wheel flats but also the estimation of wheel flat length. They utilized a dynamics model for simulation-based research. They constructed a KSM (Kriging surrogate model) and applied a PSO (particle swarm optimization)-based algorithm. Furthermore, they verified their findings through field tests, ultimately demonstrating the ability to estimate wheel flat length [
18]. Yunguang Ye et al. investigated the induction and exacerbation of wheel polygonization caused by wheel flats. They also examined the impact of the speed of railway vehicles and the length of wheel flats on wheel polygonization. This was not within the scope of the research direction and objectives of our paper. However, it is considered a valuable resource for future research on wheel flats [
19]. Dachuan Shi et al. proposed a machine fault diagnostic model in the railway domain through data augmentation. They introduced a data augmentation framework called MBS-FWFSA (multibody dynamic simulation-fast-weighted feature-space averaging) and used it to enhance the robustness of machine-learning-based fault diagnosis methods as proven in their study [
20].
In this paper, further research was conducted based on the knowledge that the researchers of this paper learned through prior research. First of all, in the prior paper, vibration signals caused by wheel flats were measured, and raw data and signal-processed data were used for learning, leading to a comparison of accuracy [
21]. In this paper, to differentiate it from the related studies mentioned above, consideration was given not only to simple wheel flat anomaly detection but also to the number and size of wheel flats. In fact, not only the size of wheel flats but also the number of wheel flats and the wheel diameter affect the vibrated acceleration signal. Therefore, we conducted the research, taking all of these aspects into consideration. At this point, the research conducted anomaly detection by distinguishing between anomaly data and normal data based on the fault regulations outlined in
Table 1. Next, the signal-processed data were compared through the supervised and semi-supervised methods. The reason for this comparison is written in
Section 4. And, in this paper, deep learning was based on CNN (convolution neural network). Subsequently, the results of the deep learning process were evaluated by means of accuracy, recall, and the ROC curve.