Den här sidan är utskriven från Högskolan i Halmstads webbplats (www.hh.se). Texten uppdaterades senast den 2018-08-16. Besök webbplatsen om du vill vara säker på att läsa den senaste versionen.
Predictive maintenance aims at identifying imminent failures and intervening before they happen, early enough to avoid downtime and reduce the consequential costs of the breakdown, both in terms of physical damage and lost business opportunity. Good understanding of the current health condition of the assets also leads to shorter and more efficient repairs, since less time needs to be spent on diagnostics and fault tracing. At the same time, it is important that predictive maintenance methods are accurate in suggesting operations to be performed, since failures that are predicted too early lead to wasted lifetime of the components and lower confidence in the method.
There has for many years been continuous successful collaboration regarding uptime, quality and predictive maintenance between CAISR researchers and different partners within Volvo Group through several research projects, including ReDi2Service, In4Uptime, ARISE and HEALTH, most of them funded through Vinnova (Sweden’s Innovation Agency). Product quality and customer satisfaction is top priority for Volvo, and thus Volvo Trucks has been the first major brand to launch the “uptime promise” service. Volvo is continuously investing into further improvements, new technologies and better services in this area, and CAISR researchers are an important part of this work.
Every day, thousands of sensor measurements stream on-board of millions of vehicles, in addition to information created off-board, in workshops or factories. This huge amount of data contains clues as to what is happening to the vehicles, and how will they operate in the future. Scanning this information, deciding on what to store and analyze, and how to represent it, is crucial for successful failure predictions. However, doing it manually is overly expensive and time consuming, in addition to being depending on the knowledge and availability of individual experts – and since experts come with their own bias and preconceptions, the solutions based only on their expectations often do not capture all the different aspects of the problem. Predictive analytics has the potential to provide value by combining expert knowledge with insights obtained from the data, and using real-world results to validate and corroborate expert’s knowledge.
There are many possible approaches for this. Our approach is largely based on the ``wisdom of the crowd’’; looking for consensus among distributed self-organized agents, and highlighting the odd-one-out. Our solutions combine several techniques, allowing for life-long learning under computational and communication constraints. It is a step towards autonomous knowledge discovery in domains where data volumes are increasing, the complexity of systems is growing, and dedicating human experts to build fault detection and diagnostic models for all relevant faults is not economically viable.
In our solution, embedded, self-organized agents operate on-board the vehicles, modelling their states and comparing across vehicles, to find systems that deviate from the consensus. The group (e.g., a fleet of vehicles) is used to provide a continuous standard that automatically deals with e.g. varying ambient conditions. The intention is to detect faults and learn the characteristics of faults without the need for human experts to anticipate and define them beforehand. This can be used to build up a knowledge base that accumulates over the life-time of the systems. Furthermore, automatically learned data representations are a valuable tool to monitor the health status of various components.
The overall scientific goal of the intelligent predictive maintenance research effort is to construct algorithms that build up knowledge about a fleet of vehicles, over time, so that the fleet becomes “self-monitoring”. Such monitoring goes beyond just the predictive maintenance, although it is probably the most obvious and clear benefit. Self-monitoring systems can detect malfunctions and estimate the remaining useful life of different parts, but also identify other issues, such as inefficient or suboptimal usage patterns. Finally, they should be able to describe the normal operation based on operations of fleet of vehicles and learn from other vehicles’ malfunctions. Ideally, they should be able to present surprising findings (novel knowledge) to the human operators. The ability to construct such self-monitoring systems will be necessary in a smart society with “internet of things”.
Vehicles that are “self-aware”, i.e., possess the capability to monitor their own operation, allow for many new services. Better service planning can be done not only by identifying the necessary repair with good lead time, but also by directly informing the workshops about the upcoming needs, to schedule the repair. Remote diagnostics, ordering the necessary spare parts, allocation of resources, all while taking into account customer preferences, location and available resources – all of this leads to improved customer satisfaction.
Many companies are working on new technologies for predictive maintenance and uptime. This includes the automotive sector but also most other industries. We have adapted the methods developed based on the Volvo use case to other domains, for example with Getinge Sterilization we are working on predictive maintenance of medical equipment, and within the SeMI project we collaborate with six different companies on creating methods for self-monitoring for district heating, heat pumps and industrial communication. Moreover, we are using many similar ideas and techniques in the healthcare area, within our growing collaboration with Region Halland and the Halmstad Hospital.
From the machine learning and data mining perspective, predictive maintenance is a uniquely interesting application domain. It showcases a number of issues that still remain a challenge, especially in combination, for state-of-the-art algorithms. Therefore, it is a good benchmark and validation domain for new methods and algorithms. Many of the issues here translate to other domains. For example, predictive maintenance deals with big streaming data that include concept drift due to both changing external conditions, but also normal wear of the equipment. It requires combining multiple data sources, and the resulting datasets are often highly imbalanced. The knowledge about the systems is detailed but in many scenarios there is large diversity in both vehicle model configurations, as well as their usage, additionally complicated by low data quality and high uncertainty in the labels.