Improving rail network velocity: A machine learning approach to predictive maintenance 提高铁路网络速度：用于预测性维护的机器学习方法
Hongfei Li ^("a,* "){ }^{\text {a,* }}, Dhaivat Parikh ^("b "){ }^{\text {b }}, Qing He ^("c "){ }^{\text {c }}, Buyue Qian ^(a){ }^{a}, Zhiguo Li ^("a ")^{\text {a }}, Dongping Fang ^("a "){ }^{\text {a }}, Arun Hampapur ^("a "){ }^{\text {a }} 李 ^("a,* "){ }^{\text {a,* }} 洪飞 ， Dhaivat Parikh ^("b "){ }^{\text {b }} ， Qing He ^("c "){ }^{\text {c }} ， Buyue Qian ^(a){ }^{a} ， 李 ^("a ")^{\text {a }} 志国 ， 方东平 ^("a "){ }^{\text {a }} ， Arun Hampapur ^("a "){ }^{\text {a }}^("a "){ }^{\text {a }} IBM T.J. Watson Research Center, Yorktown Heights, NY, United States ^("a "){ }^{\text {a }} IBM T.J. Watson Research Center， Yorktown Heights， NY， 美国^(b){ }^{\mathrm{b}} IBM Global Business Services, Dallas, TX, United States ^(b){ }^{\mathrm{b}} IBM 全球业务服务部，美国德克萨斯州达拉斯^("c "){ }^{\text {c }} The State University of New York at Buffalo, Buffalo, NY, United States ^("c "){ }^{\text {c }} The State University of New York at Buffalo， 美国， 纽约州 布法罗市
ARTICLE INFO 文章信息
Article history: 文章历史：
Received 8 May 2013 接收日期 2013 年 5 月 8 日
Received in revised form 31 March 2014 2014 年 3 月 31 日以修订形式接收
Accepted 21 April 2014 接受日期 2014 年 4 月 21 日
Available online 16 May 2014 2014 年 5 月 16 日在线提供
Keywords: 关键字：
Big data 大数据
Condition based maintenance 基于状态的维护
Multiple wayside detectors 多个轨旁探测器
Information fusion 信息融合
Predictive modeling 预测建模
Rail network velocity 铁路网速度
Abstract 抽象
Rail network velocity is defined as system-wide average speed of line-haul movement between terminals. To accommodate increased service demand and load on rail networks, increase in network velocity, without compromising safety, is required. Among many determinants of overall network velocity, a key driver is service interruption, including lowered operating speed due to track/train condition and delays caused by derailments. Railroads have put significant infrastructure and inspection programs in place to avoid service interruptions. One of the key measures is an extensive network of wayside mechanical condition detectors (temperature, strain, vision, infrared, weight, impact, etc.) that monitor the rolling-stock as it passes by. The detectors are designed to alert for conditions that either violate regulations set by governmental rail safety agencies or deteriorating roll-ing-stock conditions as determined by the railroad. 铁路网络速度定义为航站楼之间线路运输的整个系统的平均速度。为了满足铁路网络不断增长的服务需求和负载，需要在不影响安全性的情况下提高网络速度。在整体网络速度的众多决定因素中，一个关键驱动因素是服务中断，包括由于轨道/火车状况而导致的运行速度降低以及因脱轨造成的延误。铁路已经实施了重要的基础设施和检查计划，以避免服务中断。其中一项关键措施是建立广泛的路边机械状态检测器网络（温度、应变、视力、红外线、重量、撞击等），在机车车辆经过时对其进行监控。这些检测器旨在提醒违反政府铁路安全机构制定的法规或铁路确定的机车车辆状况恶化的情况。
Using huge volumes of historical detector data, in combination with failure data, maintenance action data, inspection schedule data, train type data and weather data, we are exploring several analytical approaches including, correlation analysis, causal analysis, time series analysis and machine learning techniques to automatically learn rules and build failure prediction models. These models will be applied against both historical and real-time data to predict conditions leading to failure in the future, thus avoiding service interruptions and increasing network velocity. Additionally, the analytics and models can also be used for detecting root cause of several failure modes and wear rate of components, which, while do not directly address network velocity, can be proactively used by maintenance organizations to optimize trade-offs related to maintenance schedule, costs and shop capacity. As part of our effort, we explore several avenues to machine learning techniques including distributed learning and hierarchical analytical approaches. 利用大量历史探测器数据，结合故障数据、维护行动数据、检查计划数据、列车类型和天气数据，我们正在探索多种分析方法，包括相关性分析、因果分析、时间序列分析和机器学习技术，以自动学习规则并构建故障预测模型。这些模型将应用于历史和实时数据，以预测未来导致故障的情况，从而避免服务中断并提高网络速度。此外，分析和模型还可用于检测多种故障模式的根本原因和组件磨损率，虽然这些方法不能直接解决网络速度问题，但维护组织可以主动使用它来优化与维护计划、成本和商店容量相关的权衡。作为我们工作的一部分，我们探索了机器学习技术的多种途径，包括分布式学习和分层分析方法。
Rail network operators across the world are seeing an increase in demand for services driven by increased global trade and increasing cost of fuel. Accommodating this increased load on relatively fixed rail networks requires increase in network velocity without compromising safety. Network velocity is defined as system-wide average speed of line-haul movement 由于全球贸易的增加和燃料成本的增加，世界各地的铁路网络运营商对服务的需求都在增加。在相对固定的铁路网络上适应这种增加的负载需要在不影响安全性的情况下提高网络速度。网络速度定义为全系统的平均线路运输速度
between terminals and is calculated by dividing total train-miles by total hours operated. Higher network velocity represents efficient utilization of capital-intensive assets, and it is one of the most important metrics to measure performance of a railroad. Among many determinants of overall network velocity, service interruption is a key driver, which includes lowered operating speed due to track/train condition and delays caused by derailments. Railroads have put significant infrastructure and inspection programs in place to avoid service interruptions. One of the key measures is an extensive network of rolling stock or train monitoring detectors. Each of these detector systems consists of multiple sensors (temperature, strain, vision, infrared, weight, impact, etc.) and processing software and hardware. Detectors are installed along rail tracks and inspect the rail cars and locomotives passing over them to monitor and detect equipment health conditions (Ouyang et al., 2009). The detectors are primarily designed to alert for conditions that violate regulations set by government rail safety agencies. 的计算方法是将火车总里程数除以总运营小时数。更高的网络速度代表对资本密集型资产的有效利用，它是衡量铁路绩效的最重要指标之一。在整体网络速度的众多决定因素中，服务中断是一个关键驱动因素，其中包括由于轨道/火车状况而导致的运行速度降低以及因脱轨造成的延误。铁路已经实施了重要的基础设施和检查计划，以避免服务中断。关键措施之一是广泛的机车车辆或列车监控探测器网络。这些探测器系统中的每一个都由多个传感器（温度、应变、视觉、红外、重量、影响等）和处理软件和硬件组成。探测器安装在铁路轨道沿线，检查经过它们的轨道车和机车，以监测和检测设备健康状况（Ouyang 等人，2009 年）。这些探测器主要用于对违反政府铁路安全机构制定的规定的情况发出警报。
Reducing the number of derailments attributed to mechanical faults in car and locomotive as primary cause and reducing intermediate maintenance calls due to false alarms can significantly improve rail network velocity. The extensive sensor network provides information sources to enable the solutions. One approach is to use machine learning techniques to automatically “learn rules” from historical sensor measurements and then better predict which rail cars are likely to have problems and thus maximize the hit rate of rail operator’s setouts. Machine learning techniques allow systematic classification of patterns or relationships contained in data and identification of the attributes, containing information about condition of physical assets that contribute associated failure mode, or class (Hastie et al., 2001). 减少归因于汽车和机车机械故障的主要原因脱轨次数，并减少由于误报而导致的中间维护呼叫，可以显著提高铁路网络速度。广泛的传感器网络为实现解决方案提供了信息源。一种方法是使用机器学习技术从历史传感器测量中自动“学习规则”，然后更好地预测哪些轨道车可能存在问题，从而最大限度地提高铁路运营商出发的命中率。机器学习技术允许对数据中包含的模式或关系进行系统分类，并识别属性，包含有关导致相关故障模式或类别的物理资产条件的信息（Hastie et al.， 2001）。
Considering the complexity of sensor network, there are several challenges in developing machine-learning techniques for predictive maintenance in railway operations. 考虑到传感器网络的复杂性，开发用于铁路运营预测性维护的机器学习技术存在一些挑战。
The first challenge is caused by spatio-temporal incompatible information collected through multiple detectors, which are not always co-located. The detector system consists of multiple detectors measuring temperature, strain, vision, weight/impact, etc. Existing systems issue alerts primarily using one or two types of detectors at a time and only partial information is used. For example, when the bearing temperature is above 170^(@)F170^{\circ} \mathrm{F}, the system issues an alert to request an immediate train-stop. The rule is simple but does not consider measurement errors or impact of environmental variables on detectors, which may lead to high false alarms and lower hit rate of rail operator’s setouts. To better understand the conditions of a railcar, it is essential to integrate the information collected from various detectors. Since these detectors are not co-located, the measurements coming out of them are spatially and temporally incompatible, posing challenges when combining the information. We use bad truck/bad wheel prediction model as an example to show how we address this issue in Section 3.2. 第一个挑战是由通过多个探测器收集的时空不兼容信息引起的，这些信息并不总是位于同一位置。探测器系统由多个探测器组成，用于测量温度、应变、视力、重量/撞击等。现有系统发出警报，一次主要使用一种或两种类型的探测器，并且只使用部分信息。例如，当轴承温度高于 170^(@)F170^{\circ} \mathrm{F} 时，系统会发出警报，请求立即停车。该规则很简单，但没有考虑测量误差或环境变量对探测器的影响，这可能导致高误报和铁路运营商出发的命中率降低。为了更好地了解轨道车的状况，必须整合从各种探测器收集的信息。由于这些探测器不在同一位置，因此它们得出的测量值在空间和时间上不兼容，在组合信息时带来了挑战。我们以坏卡车/坏车轮预测模型为例，在第 3.2 节中展示我们如何解决这个问题。
The second challenge is big data. The ubiquity of connectivity and the growth of sensors have opened up a large storehouse of information. The bearing temperature detectors, for Class I railroad under this study, generate 3 terabytes of data in a year. Other industries offer possibilities of even larger amount of data generated under normal operating conditions, e.g., a Boeing jet generates 10 terabytes of information per engine every 30 min of flight-time (Higginbotham, 2010). The amount of data is only going to continue to rise. There has been a lot of recent progress in big data warehousing to manage, store and retrieve this information including Netezza and Teradata, but the true value is realized only if we are successful in mining and using the signals contained in the information effectively. In this paper, we will show a customized support vector machine (SVM) technique that effectively utilizes large-scale data and provides valuable tools for operational sustainability as described in the scenario of alarm prediction in Section 3.1. 第二个挑战是大数据。无处不在的连接和传感器的增长打开了一个庞大的信息仓库。本研究中用于 I 类铁路的轴承温度检测器在一年内产生 3 TB 的数据。其他行业提供了在正常运行条件下生成更大数据量的可能性，例如，波音喷气式飞机每 30 分钟飞行时间每个发动机生成 10 TB 的信息（Higginbotham，2010 年）。数据量只会继续增加。大数据仓储在管理、存储和检索这些信息（包括 Netezza 和 Teradata）方面最近取得了许多进展，但只有我们成功地挖掘和有效地使用信息中包含的信号，才能实现真正的价值。在本文中，我们将展示一种定制的支持向量机 （SVM） 技术，该技术可以有效地利用大规模数据，并为运营可持续性提供有价值的工具，如第 3.1 节中的警报预测场景所述。
The third challenge comes from the need to learn and create alarm rules in the context of industry operations, so that the rules generated can be interpreted by operators easily leading to efficient operational decision support. On one hand, subject matter experts (SME) can derive rules based on their knowledge and expertise in concert with industry standards. Those rules are easy to interpret, but do not accommodate the complexity required for accurate prediction based on large, spatially and temporally incompatible and dirty, heterogeneous data coming from multiple detectors. On the other hand, machine learning techniques provide good approaches to learn efficient rules to predict failures. But those rules are usually complicated and thus not easily understood by human operators. To facilitate predictive maintenance operations in railway, it is important to set proper trade-off in learning system to create efficient but human-interpretable rules. We address this issue by using two different approaches. One is to extract logicalized rules from the complex machine-learning based algorithm outputs, and the other is to use certain machine learning techniques, such as decision tree, to derive rules that are easily to understand and implement. The approaches are described with more details in the two scenarios in Section 3.1 and 3.2, respectively. 第三个挑战来自需要在行业运营环境中学习和创建警报规则，以便作员可以轻松解释生成的规则，从而提供高效的运营决策支持。一方面，主题专家 （SME） 可以根据他们的知识和专业知识得出规则，并与行业标准保持一致。这些规则易于解释，但无法适应基于来自多个检测器的大型、空间和时间不兼容以及脏的异构数据进行准确预测所需的复杂性。另一方面，机器学习技术提供了很好的方法来学习有效的规则来预测故障。但这些规则通常很复杂，因此人类作员不容易理解。为了促进铁路的预测性维护作，在学习系统中进行适当的权衡以创建高效但人类可解释的规则非常重要。我们通过使用两种不同的方法来解决这个问题。一种是从基于机器学习的复杂算法输出中提取逻辑化规则，另一种是使用某些机器学习技术（例如决策树）来派生易于理解和实施的规则。 这些方法分别在 Section 3.1 和 3.2 中的两个场景中进行了更详细的描述。
In the area of condition monitoring and predictive maintenance, some work has been done to provide failure predictions using statistical and machine learning approaches. Lin and Tseng et al. (2005) presents reliability modeling to estimate machine failures. Saxena and Saad (2007) uses neural network classifier for condition monitoring of rotating mechanism systems. In railway applications, Yella et al. (2009) adopts a pattern recognition approach to classify the condition of the sleeper into classes (good or bad). Yang and Létourneau (2005) proposes an approach to predict train wheel failures but only using one type of detectors, Wheel Impact Load Detector (WILD), without considering the impacts of multiple detectors. Recently, Hajibabai et al. (2012) develops a logistic regression model to classify wheel failures based on WILD and Wheel Profile Detector (WPD). They claim that the classification accuracy is 90%90 \% with 10%10 \% false alarm rate. However, only two detectors are taken into account in that study. The problems that those papers have worked on are not as complicated as what we face and none of them has addressed all the challenges we describe above. 在状态监测和预测性维护领域，已经做了一些工作，使用统计和机器学习方法提供故障预测。Lin 和 Tseng 等人（2005 年）提出了可靠性模型来估计机器故障。Saxena 和 Saad （2007 年）使用神经网络分类器对旋转机构系统进行状态监测。在铁路应用中，Yella 等人（2009 年）采用模式识别方法将轨枕的状态分类为几类（好或坏）。Yang 和 Létourneau （2005） 提出了一种预测火车车轮故障的方法，但只使用一种类型的检测器，即车轮碰撞载荷检测器 （WILD），而不考虑多个检测器的影响。最近，Hajibabai 等人（2012 年）开发了一种逻辑回归模型，基于 WILD 和车轮轮廓检测器 （WPD） 对车轮故障进行分类。他们声称分类精度与误报率有关 90%90 \%10%10 \% 。 然而，该研究只考虑了两个检测器。这些论文解决的问题并不像我们面临的那么复杂，也没有一个能解决我们上面描述的所有挑战。
In this paper, we develop machine-learning approaches to predict impending failures and alarms of critical rail car components. The prediction will drive proactive inspections and repairs, reducing operational equipment failure. This work delivers significant gains in rail network velocity and safety. The rest of the paper is organized as follows. Section 2 gives the overview of sensor network operations in a US Class I railroad and various data sources provided. Section 3 describes the composite detector analysis, using two specific use cases as examples, including problem formulation, methodologies and results. Section 4 summarizes the work. 在本文中，我们开发了机器学习方法来预测关键轨道车组件即将发生的故障和警报。预测将推动主动检查和维修，减少运营设备故障。这项工作在铁路网络速度和安全性方面取得了显着进展。本文的其余部分组织如下。第 2 节概述了美国 I 类铁路中的传感器网络作以及提供的各种数据源。第 3 节描述了复合探测器分析，以两个特定的用例为例，包括问题制定、方法和结果。第 4 节总结了这项工作。
2. Overview of sensor network in a US Class I railroad 2. 美国 I 类铁路中的传感器网络概述
Railroad operators are keen to move from a reactive (respond to alerts from detectors and act) to a proactive stance around safety and network velocity. In this context, IBM works with a US Class I railroad to develop predictive failure analytics based on large-scale data from multiple-detector systems. The railroad manages about 20 K miles of tracks and has about 1000 detectors installed along their railway network. A huge amount of data has been collected over the past several years. For example, they manage about 800 hot box detectors (HBD) which comprises around 900 million temperature records collected from 4.5 million bearings just within 3 months in early 2011. They also have 90 hot wheel detectors (HWD), which contain 500 million records collected from 4.5 million wheels within 6 months in 2011. The above are examples of the scale of the data that needs to be analyzed in order to develop models. We expect the data volume to grow by a 100 fold as new detectors come online. 铁路运营商热衷于从被动（响应来自探测器的警报并采取行动）转变为围绕安全和网络速度的主动立场。在这种情况下，IBM 与美国 I 级铁路合作，根据来自多探测器系统的大规模数据开发预测性故障分析。该铁路管理着大约 20 K 英里的轨道，并在其铁路网络上安装了大约 1000 个探测器。在过去几年中，已经收集了大量数据。例如，他们管理着大约 800 个热轴探测器 （HBD），其中包括在 2011 年初的 3 个月内从 450 万个轴承收集的大约 9 亿条温度记录。他们还有 90 个热轮探测器 （HWD），其中包含在 2011 年的 6 个月内从 450 万个车轮收集的 5 亿条记录。以上是开发模型所需分析的数据规模的示例。 我们预计，随着新探测器的上线，数据量将增长 100 倍。
The detectors are installed along the wayside. When a rail car passes an equipped detector, the detector will report mechanical condition observations of the equipment or components including truck, wheel or bearing. Now we give a brief description of the detectors used in our analysis (Ameen, 2006). Hot Box Detector (HBD) is a pyrometer-based technology used to monitor the temperature of bearings and wheel (if equipped to do so). Wheel Impact Load Detector (WILD) is built into the track to detect defective wheels based on the impact they generate on the track structure. The system reports the features to describe dynamic impact load at wheel level, such as maximum force (KIP) and max lateral force (KIP). In addition, the detector collects observations at car level, such as speed and weight. Truck Performance Detectors (TPD) are usually built into SS curves of the track to monitor performance of trucks (the assemblies that hold railcar wheels). The key feature collected through TPD is ratio of lateral force over vertical force. Both WILDs and TPDs use strain-gauge-based technologies to measure the performance of a freight car, its wheels and trucks in a dynamic mode, which tells us more about a car than a static inspection. Machine Vision (MV) technology uses computer algorithms to process digital image data of railcar underframes and side-frames into diagnostic information. The key wheel features captured by MV include flange height, flange thickness, rim thickness, diameter, etc. Optical Geometry Detector (OGD) uses angle of attack and tracking position to calculate the following features for each set of axles on the same truck - tracking error, truck rotation, inter axle misalignment and shift. It is a laser-based system with cameras mounted on tangent track. Acoustic Bearing Detectors (ABD) capture the noise signature emitted by bearing in motion and detector processes this information internally and issues alarms when anomalous acoustic signature is detected. 探测器安装在路边。当轨道车通过配备的探测器时，探测器将报告设备或部件（包括卡车、车轮或轴承）的机械状态观察结果。现在我们简要介绍一下分析中使用的探测器（Ameen，2006 年）。热轴探测器 （HBD） 是一种基于高温计的技术，用于监测轴承和车轮的温度（如果配备）。车轮碰撞载荷探测器 （WILD） 内置在轨道中，根据车轮对轨道结构产生的冲击来检测有缺陷的车轮。该系统报告描述车轮水平动态冲击载荷的特征，例如最大力 （KIP） 和最大侧向力 （KIP）。此外，探测器还收集汽车级别的观察结果，例如速度和重量。卡车性能探测器 （TPD） 通常内置在轨道的曲线中 SS ，以监测卡车（固定轨道车车轮的组件）的性能。通过 TPD 收集的关键特征是侧向力与垂直力的比率。 WILD 和 TPD 都使用基于应变片的技术来测量货车、车轮和卡车在动态模式下的性能，这比静态检查更能告诉我们关于汽车的信息。机器视觉 （MV） 技术使用计算机算法将轨道车底架和侧架的数字图像数据处理成诊断信息。MV 捕获的关键车轮特征包括翼缘高度、翼缘厚度、轮辋厚度、直径等。光学几何检测器 （OGD） 使用攻角和跟踪位置来计算同一辆卡车上每组车轴的以下特征 - 跟踪误差、卡车旋转、轴间错位和偏移。它是一个基于激光的系统，摄像头安装在切线轨道上。 声学轴承探测器 （ABD） 捕获轴承运动中发出的噪声特征，探测器在内部处理此信息，并在检测到异常声学特征时发出警报。
In addition to the detector data, the railroad also stores historical equipment failures, maintenance records and alarms issued in recent past. That information gets connected to the detector readings to give a whole picture of the cause and effect. 除了探测器数据外，铁路还存储历史设备故障、维护记录和近期发布的警报。该信息与探测器读数相关联，以提供因果关系的全貌。
3. Composite-detector analysis 3. 复合探测器分析
The goal of this work is to predict equipment failures in advance and improve network velocity by reducing derailments or at minimum reducing intermediate maintenance calls due to false alarms. We achieve this by developing models to create “learned rules” automatically from historical data to better predict which rail cars are more likely to have problems and to predict the most severe existing alarms in advance of the actual alarm event to reduce immediate train stops and service interruptions. Four major pieces of analytics work are listed: 这项工作的目标是提前预测设备故障，并通过减少脱轨或至少减少由于误报而导致的中间维护呼叫来提高网络速度。我们通过开发模型来实现这一目标，从历史数据中自动创建“学习规则”，以更好地预测哪些轨道车更有可能出现问题，并在实际警报事件之前预测最严重的现有警报，以减少列车立即停止和服务中断。列出了四项主要的分析工作：
Alarm prediction: predict the alarms alerting catastrophic failures caused by hot bearings 7 days in advance of actual alarm event using multiple detectors, such as HBD, ABD and WILD. The significance of alarm prediction is to reduce the immediate train stops and service interruption. 警报预测：使用多个探测器（如 HBD、ABD 和 WILD）在实际警报事件发生前 7 天预测警报，以提醒由热轴承引起的灾难性故障。警报预测的意义在于减少列车立即停止和服务中断。
Bad truck prediction: detect truck performance issues earlier due to wear-out by identifying patterns in wheel movement error, wheel dimension and wheel impact load using multiple detectors, such as MV, OGD, TPD and WILD. 不良卡车预测：通过使用多个检测器（如 MV、OGD、TPD 和 WILD）识别车轮运动误差、车轮尺寸和车轮冲击载荷的模式，更早地发现由于磨损而导致的卡车性能问题。
Bad wheel prediction: predict wheel defects earlier by identifying abnormal patterns in wheel dimensions, movement errors and wheel impact load using multiple detectors, such as MV, OGD and WILD. 车轮不良预测：通过使用多个探测器（如 MV、OGD 和 WILD）识别车轮尺寸的异常模式、运动误差和车轮冲击载荷，更早地预测车轮缺陷。
Asymmetric wheel wear detection: detect one particular wheel defect, asymmetrically wearing wheels, which could lead to a shorter wheel lifespan and can also lead to truck performance issues, using MV and OGD. 不对称车轮磨损检测：使用 MV 和 OGD 检测一个特定的车轮缺陷，即不对称磨损的车轮，这可能导致车轮寿命缩短，还可能导致卡车性能问题。
In Sections 3.1 and 3.2, we describe in more details two of the above four applications - alarm prediction and bad truck/ bad wheel failure prediction. These two use cases are good examples to show how to apply machine learning approaches in railroad predictive maintenance applications from different perspectives. Alarm prediction focuses more on how to handle large-scale data and extract human interpretable rules from complex machine learning algorithm outputs. We present bad 在第 3.1 节和第 3.2 节中，我们更详细地描述了上述四种应用中的两种——警报预测和坏卡车/坏车轮故障预测。这两个用例是很好的例子，可以从不同的角度展示如何在铁路预测性维护应用中应用机器学习方法。警报预测更侧重于如何处理大规模数据，以及如何从复杂的机器学习算法输出中提取人类可解释的规则。我们介绍坏
Corresponding author. Address: IBM T.J. Watson Research, 1101 Kitchawan Rd., Yorktown Heights, NY 10510, United States. Tel.: +1 9149452656. 通讯作者。联系地址： IBM T.J. Watson Research， 1101 Kitchawan Rd.， Yorktown Heights， NY 10510， United States.电话：+1 9149452656。