**1. Introduction**

By the end of 2018, the total operating mileage of urban rail transit (URT) in China exceeded 5700 km, including 4350 km of subway lines, and it is expected to double in the next 3 to 5 years [1]. With the rapid extension of the URT network, the current maintenance mode relies on humans, and it is challenging to ensure the safe and stable operation of trains. Therefore, intelligent URT maintenance work should be promoted for higher e fficiency.

As one of the most prevalent kinds of URT, subways are increasingly essential in people's daily lives. However, abnormal vibration and noise significantly a ffect passengers' riding experience. Moreover, these abnormalities provide information about wheel-rail interactions and degradation of the track structures. Generally, train-induced noise can be categorized as external or interior noises [2]. Vehicle interior noise which is pertinent to this study mainly consists of noise from electrical equipment, aerodynamic noise, and wheel-rail noises [3]. Usually, the aerodynamic noise is dominant when the train speed exceeds 250 km/h, and electrical equipment noise dominates for speeds slower than 35 km/h [4]. As the subway trains usually run at 30–80 km/h, the wheel-rail noise is the main component of vehicle interior noise [5]. The wheel-rail interaction significantly influences the wheel-rail noise. Therefore, we assumed that there exists a mapping relationship between vehicle interior noises and wheel-rail interactions. This mapping relationship provides an approach to monitor track conditions through vehicle interior noise. Moreover, it would be convenient to develop a simple onboard interior noise monitoring system that contributes to the safety and reliability of the railway system.

Regarding vehicle interior noise, past studies have mainly focused on the generation mechanism, transmission characteristics, and control strategies [6–10]. Typical study topics, such as noise characteristics analysis [11], sound quality evaluation [12], and noise level prediction [13], can be attributed to the above research fields. However, because the vehicle-track coupling system consists of a large number of components, the interior noise is affected by numerous factors, such as track slab [14], rail roughness, wheel out-of-roundness [9], and car body structure [15]. These factors may interact with each other and influence the characteristics of vehicle interior noise. Therefore, researchers generally choose one or two factors, such as rail fastener stiffness [7] and wheel polygonal wear [9], to perform their analysis at a lower complexity.

Among related studies, the prediction of vehicle interior noise is one of the most prevalent topics because it benefits the design and construction of track-vehicle systems at the early stages. Methods such as the boundary element method (BEM) [16], finite element method (FEM) [17], and statistical energy analysis method (SEAM) [15] are commonly used in this. However, their effectiveness relies significantly on the selected boundary conditions and model parameters. Thus, these numerical models are generally applied for specific problems. Moreover, the results of field tests are also often used for model verification. Despite the effectiveness of the method combining analytical models, numerical simulation, and field tests in the study of vehicle interior noise, the difficulty to obtain model parameters limits its application. Moreover, field tests may also interfere with daily operations. Overall, these studies do not make the best use of data collected during the daily operation and maintenance of the railway system.

In this context, the railway transportation industry is at the forefront of implementing analytics and big data [18]. Machine learning (ML) and artificial intelligence (AI) are two concepts at the leading edge of information technology, both of which contribute to big data technology. In recent years, the implementation of ML in the railway industry has been widely studied, for example in the prediction of passenger flow [19], delay events [20], and railway operation disruptions [10]. Moreover, many cases have been reported for railway infrastructure managemen<sup>t</sup> and maintenance, including the detection and diagnosis of defects [21–23], prediction of failure events [24,25], and forecast of remaining useful life of devices [26]. These studies indicate that ML technologies have a promising prospect in promoting intelligent railway maintenance, thus ensuring the safety of the railway transit system.

As for data on vehicle interior noise, users require automatic methods to segment, label, and store the increasing amount of acoustic data from monitoring systems. The major challenge in this field is the automatic classification of audio [27]. Recent studies on the classification of traffic noise have been conducted, for example, to identify the type of vehicle through roadside noise [28,29] and evaluate passengers' subjective experience by categorizing the cabin's interior noise [30]. However, compared with traffic noise, the factors influencing vehicle interior noise of subway trains are considerably more complicated.

For collecting track conditions, the railway industry has employed various dedicated devices, such as track inspection vehicles [31] and visual inspection systems [32]. Although these devices perform well in detecting track conditions, the expensive cost and the interference for regular operation limit their usage in urban rail transit systems. There are also some on-board devices being developed to monitor track conditions using in-service vehicles [33–35]. However, the installation of these devices may change the design characteristics of cars and cause potential safety issues. As of now, these novel on-board monitoring devices have not been widely used. As an integrated platform, a smartphone can achieve data collection, storage, and transmission individually. Besides, the smartphone is mature, cost-effective, and easy to use, promoting its application in various fields. Studies using the embedded

accelerometers of smartphones to monitor road conditions and evaluating the ride quality have been reported [36,37]. These research works inspired the authors to investigate the feasibility of using smartphones to collect multi-source data about subway vehicles.

According to the above literature review, current studies about vehicle interior noise mainly focus on its generation mechanism and influencing factors through analytical models, numerical simulations, and field tests. To the best of our knowledge, only a few studies have analyzed vehicle interior noise using data-driven methods. Therefore, this study aims to advance data mining of vehicle interior noise for decision making in rail maintenance, such as for rail grinding. In this context, there are two significant challenges. First, despite sensing technologies being well developed now, it is still di fficult to establish an onboard data collection framework that is easy to deploy, cost-e fficient, and reliable. Moreover, the simultaneous collection of dynamic responses from the car body and interior noise is essential because these two datasets are connected to each other. Second, due to the complexity of vehicle interior noise, the extraction of useful features and correct labeling of noise classes remain challenging.

The goal of this study is to mine useful information from the vast amount of interior noise data using ML methods. To pursue this goal, onboard smartphone data were collected, including dynamic responses and noises. Further, a series of analyses were performed to classify the noises and clarify the influencing factors. The novel contributions of this paper are summarized as follows:


This paper is organized as follows. Section 2 briefly illustrates the research methodology. Section 3 introduces the data utilized in this study and its collection framework. Section 4 describes the modeling approaches, including data segmentation and time windows, and establishes the multi-classification model with the Extreme Gradient Boosting (XGBoost) method. Furthermore, Section 5 presents the analysis results and discussions. Finally, in Section 6, conclusions are drawn according to the relevant analysis.
