Towards a Cloud-based Machine Learning for Health Monitoring and Fault Diagnosis Dr. Samir KHAN, Takehisa YAIRI and Mariam KIRAN (University of Tokyo, Lawrence Berkeley National Laboratory)
Complex engineered systems are often unable to effectively analyze data to diagnose, isolate and predict faults detected during operation. This is due to difficulties in processing large amounts of data, recognizing symptoms with standard testing tools, infer potential faults and eventually diagnose causes needing maintenance support. The aerospace industry is particularly concerned with analyzing and maintaining assets preventing potential failures. Automatically detecting faults during maintenance can help extend asset life, reduce costs, improving availability and reliability. Recent advances in Cloud computing has provided infinite computing resources to quickly process and troubleshoot, reducing ‘time-to-fix’ problems. Exploiting artificial intelligence algorithms, with Cloud resources, can help build an integrated fault diagnostic platform to provide resilient and scalable resources for data acquisition, processing and decision making. This paper discusses the current efforts in using machine learning methods for fault diagnosis, particularly using Cloud resources in the aerospace industry. Special attention is paid to the benefits; with potential future research on technical diagnosis being enumerated.
Requirements for Prognostics System to Improve Business Process of Machinery Maintenance Service Dr. Masakazu HORI, Takuya OYAMA and Makoto IMAMURA (INTEC Inc., Tokai University)
The number of companies that provide "predictive maintenance service” is increasing. If we can detect the sign of machinery failure from the collected data by IoT technologies, we exchange the parts or repair them in advance. By predictive maintenance service, people in a plant get merits to avoid great losses by unexpected troubles such as a sudden machinery failure, a manufacturing line stop and so on. As for maintenance service providers, they have such merits that they can maintain a machinery systematically and provide maintenance service by a limited number and skill of maintenance engineers.
In this paper, we firstly show the main processes for preventive maintenance service and predictive maintenance service, and then compare them and show the differences. In predictive maintenance, we have no regular manual inspection conducted in a preventive maintenance. We anticipate a maintenance task by the monitoring results of a machinery condition and make a plan for the task.
To solve the issues concerning preventive maintenance service provision, we propose not only the requirements for basic functionalities of prognostics system, but also the operational requirements for continuous service provision and the non-functional requirements to guarantee safe usage.
MCMC-based Efficient Maintenance Plan Decision Mr. Junya SHIMADA and Satoko SAKAJO (MITSUBISHI Electric Corporation)
In recent years, it has been an essential policy to monitor real-time health states of facilities and determine when to perform maintenance in order to ensure the high operation ratio and improve work efficiency. In this paper, target facilities diagnose their own health states by analyzing time-series sensor data and transmit warning data and failure data to the monitoring center. These data include date and time of occurrence and warning/failure code which identifies the factor. Utilizing these data, we propose an MCMC-based maintenance plan decision to reduce the failures and the workloads. Firstly, state-based warning patterns which are composed of several warning codes are extracted. At that time, to avoid the state explosion, only warning patterns which are closely related to failure occurrence are extracted based on the time interval from warning states to failure state. Secondly, warning patterns are modeled based on N-th order Markov model. Finally, maintenance plan is decided based on failure probability. Experiment to evaluate whether the facilities can be maintained before failure occurs proved that this approach could actually reduce the number of failures and the frequency of dispatches of maintenance workers.
Defect State and Severity Analysis Using the Discretized State Vectors Ms. Sujeong BAEK and Duck-Young KIM (UNIST)
The time series of sensor data for condition monitoring of a system is often characterized as very-short, intermittent, transient, highly nonlinear and non-stationary random signals, which hinders the straightforward pattern analysis. For discovering meaningful features from original sensor data, we transform continuous time series data into a set of contiguous discretized state vectors using a multivariate discretization approach. We then search for important patterns that are found only in the case of defective systems. We discuss how to measure the level of importance of each defect pattern and further how to assess the severity degree of a defective state. We consider that a defective state is more severe if various defect patterns are observed in the state. Likewise, if a particular defect pattern describes as many as defective states, the pattern will be treated as significant. The proposed procedure is applied to detecting defective car door trims that have the potential to generate small but irritating noises. We analyzed the datasets obtained from two different monitoring methods using a typical acoustic sensor array and acoustic emission sensors. Defective car door trims were efficiently identified with their severity degrees.