تشخیص ناهنجاری در نظارت بر داده های حسگر برای تعمیر و نگهداری پیشگیرانه
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|22549||2011||13 صفحه PDF||سفارش دهید||محاسبه نشده|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 38, Issue 6, June 2011, Pages 7003–7015
Today, many industrial companies must face problems raised by maintenance. In particular, the anomaly detection problem is probably one of the most challenging. In this paper we focus on the railway maintenance task and propose to automatically detect anomalies in order to predict in advance potential failures. We first address the problem of characterizing normal behavior. In order to extract interesting patterns, we have developed a method to take into account the contextual criteria associated to railway data (itinerary, weather conditions, etc.). We then measure the compliance of new data, according to extracted knowledge, and provide information about the seriousness and the exact localization of a detected anomaly.
Today, many industrial companies must face problems raised by maintenance. The most common solution is called curative maintenance, i.e., equipment is replaced or repaired after the appearance of obvious failures, once the damage is occurred. This solution poses many problems. Curative maintenance is too belated and is particularly costly on several aspects. On the financial side first, for many companies, a few hours of downtime can result in millions of dollars in losses. It is generally much less expensive to make predictive maintenance to prevent a serious breakdown. In addition, the corrective maintenance is also a problem for security aspects. In many sensitive areas, equipment failures can cause death. For example, it is estimated that approximately 5% of motor vehicle accidents are caused by equipment malfunction or a lack of maintenance.1 Another aspect is related to the environment and energy saving. Indeed, equipment that is worn or subject to malfunctions often consumes more energy than equipment that operates optimally. In addition, a systematic maintenance planning is not a satisfactory solution as too expensive compared to real needs. To reduce the problems of equipment maintenance, to propose ways to make maintenance both faster and more effective by anticipating serious breakdowns represents a particularly critical issue. Such a preventive maintenance consists in detecting anomalous behavior in order to prevent further damages and avoid more costly maintenance operations. To this end, it is necessary to monitor the working equipment. Usually, monitoring data is available through embedded sensors and provides us with important information such as temperatures, humidity rates, etc. Nevertheless, data collected by sensors are difficult to exploit for several reasons. First, a very large amount of data usually available at a rapid rate must be managed to provide a relevant description of the observed behaviors. Furthermore, they contain many errors: sensor data are very noisy and sensors themselves can become defective. Finally, when considering data transmission, very often lots of information are missing. In this paper, we focus on the field of train maintenance. Trains monitoring is also ensured by sensors positioned on the main components (wheels, motors, etc.) to provide much information (temperature, acceleration, velocity). In this context, we are subject to the difficulties we have described above: voluminous and noisy data, information transmission problems, etc. Moreover, it is important to take into account the different types of data available. We therefore wish to propose a method to exploit this information in order to assist the development of an effective predictive maintenance. The needs in the context of train maintenance are twofold. First, it is important to provide a better understanding of monitored systems. Indeed, as they are often complex and contain many components, the experts have little knowledge about their actual behavior. This lack of knowledge makes the problem of maintenance very difficult. From another point of view it could be interesting to get an overview of the normal behavior (e.g., in case of monitoring) and then it is necessary to propose a way for characterizing such normal behaviors from a huge amount of historical data. Another challenging point that we must consider is that normal behavior strongly depends on the context. For example, a very low ambient temperature will probably affect a train behavior. Similarly, each itinerary with its own characteristics (slopes, turns, etc.) influences a journey. Consequently it is essential, in order to efficiently characterize the behavior of trains as well as to detect anomalies, to consider the surrounding context. In this paper, we will show how these elements can be directly addressed by data mining techniques and how they can be used to design a system for detecting anomalies in train behavior and help experts so that a detected problem can be dealt as promptly as possible, and the right decisions can be made. Our approach follows the framework presented in Fig. 1. We address the two issues mentioned above: (i) the knowledge discovery process about normal train behavior and (ii) the anomaly detection in new data.The characterization of normal behavior is divided into three steps. First, we consider the data describing the normal behavior (i.e., containing no anomalies) that were recorded in the past. These so-called historical data are segmented into different classes, which are defined by the context in which the data were recorded. This organization brings together all trips that were conducted under similar conditions. The criteria used to create the classes are, for example, the outside temperature, the itinerary, etc. Then, we extract the most frequent behaviors to characterize each of these classes. We thus obtain knowledge classes that describe very precisely normal train behavior and also provide us with essential information about the impact of contextual criteria. For example, we can answer questions such as “What are the behaviors that are specific to a high outside temperature?”. After having characterized the train behavior in the first step of our framework, we wish to detect anomalies in newly recorded data. To this end, we use the previously obtained knowledge. We have developed a method to compare new monitoring data with a class of knowledge. Thus, we can detect critical events and notify experts that a maintenance operation may be necessary. This paper is organized as follows. Section 2 describes the data representation in the context of train maintenance. Section 3 shows the characterization of normal behaviors by discovering sequential patterns. Then we present the anomaly detection for predictive maintenance approach in Section 4. Experiments conducted with real and simulated data are described in Section 5 and related work is presented in section 6. Finally, we conclude in Section 7.
نتیجه گیری انگلیسی
In this paper, we have addressed a problem involved in many areas: the maintenance of complex systems. There are several solutions to perform maintenance of such systems. Firstly, corrective maintenance, consisting in making the necessary maintenance operations after the occurrence of a failure, is not suited to this contexte as too costly and too dangerous. A planned and systematic maintenance is too expensive, although it can usually avoid serious failures. Thus, we addressed the problem of developing an approach to allow preventive maintenance, which is a good compromise between the two previous solutions. Preventive maintenance consists in detecting abnormal behavior that may be harbingers of major failures, to perform only the necessary preventive maintenance operations. However, the complexity of such systems (e.g., trains, industrial machinery, etc.) makes their behavior difficult to understand and interpret. In these circumstances it is particularly difficult to detect abnormal behavior that often herald significant and costly failures. The problem that we address is particularly challenging. Indeed, errors in diagnosis may cause many inconveniences. We have therefore developed an approach to use data collected by sensors in order to analyze behaviors and allow the detection of anomalies. Our contribution is divided into three parts. First, we have proposed an adequate representation of data in order to extract knowledge based on sensor data describing the past normal behavior of systems. Secondly, we studied the possibility to extract sequential patterns in historical data both to improve the understanding of systems for experts, but also to provide a knowledge database used to develop a method for detecting anomalies. To characterize normal behavior adequately, we also took into account the context. Indeed, the context in which a system is running often has an influence on its behavior, and on the definition of an anomaly. Finally, we proposed an approach to compare new patterns with all sequential patterns describing normal behavior. We provide experts with an anomaly score for each sensor and each component of the systems studied. The approach allows to locate precisely the anomalies and to quantify the extent to which the anomaly seems problematic. One main advantage of the presented approach is that all obtained results are easily interpretable for decision makers. This is particularly important because users must be able to make decisions with certainty. Although the presented work meets our expectations in terms of results, it opens interesting perspectives. In particular, the development of a graphical user interface will allow users to access results quickly and efficiently. Another important aspect may be to adapt the approach to a real-time context. This will detect the anomalies on the running trains. Furthermore, an interesting question may be addressed: how to manage the evolution of normal behavior over time? This will conduct the knowledge database to be incrementally updated in order to ensure handled knowledge to be valid over time