Sreerupa Das

An efficient way to enable prognostics in an onboard system

Prognostics and Health Management (PHM) systems are becoming increasingly important for monitoring and maintaining high value assets. In order to enable real time onboard diagnostic and prognostic capabilities, mechanisms for reading, manipulating and analyzing the data need to be architected into the onboard system. Machine learning and statistical algorithms provide tools to develop data models for enabling prognostics that are typically developed off-board by mining historical data. Once trained, the logic of processing real time data is then embedded on a real time onboard system. A straightforward approach for incorporating the knowledge and intelligence for real time data processing is to add the needed logic and algorithms as an integral part of the onboard software. While this method can serve the purpose of enabling real time health assessment and analysis, it is very restrictive in nature. Every time the analytics need to be updated or algorithms need refinement, it requires a refresh of the complete onboard software. The ability to fine tune onboard embedded logic for the purpose of making the analysis smarter is crucial for creating a successful and sound health monitoring system. In addition, it is desired that the process of encoding logic and algorithms should be simple and easy to incorporate into the system. User friendliness of the process of embedding intelligent logic is critical for long term maintenance of the system as well. This paper discusses an approach to build algorithms and logic into an onboard system such that they are programmatically decoupled from the onboard software. The approach described in this paper allows users the ease of use and flexibility in building knowledge into the system. In addition, as more historical data is collected and richer knowledge is discovered from mining the data, algorithms can be improved over time without having to update the onboard software.

Machine learning for improved diagnosis and prognosis in healthcare

Machine learning has gained tremendous interest in the last decade fueled by cheaper computing power and inexpensive memory — making it efficient to store, process and analyze growing volumes of data. Enhanced algorithms are being designed and applied on large datasets to help discover hidden insights and correlations amongst data elements not obvious to human. These insights help businesses take better decisions and optimize key indicators of interest. The growing popularity of machine learning also stems from the fact that learning algorithms are agnostic to the domain of application. Classification algorithms, for example, that could be applied to categorize faults in windmill blades can also be used for categorizing TV viewers in a survey. The actual value of machine learning however depends on the ability to adapt and apply these algorithms to solve specific real world problems. In this paper we discuss two such applications for interpreting medical data for automated analysis. Our first case study demonstrates the use of Bayesian Inference, a paradigm of machine learning, for diagnosing Alzheimers disease based on cognitive test results and demographic data. In the second case study we focus on automated classification of cell images to determine the advancement and severity of breast cancer using artificial neural networks. Although these research are still preliminary, they demonstrate the value of machine learning techniques in providing quick, efficient and automated data analysis. Machine learning offers hope with early diagnosis of diseases, help patients in making informed decisions on treatment options and can help in improving overall quality of their lives.

Adaptive automated scheduler in Prognostics Health Management

A mechanism for Automated Scheduler in maintenance management is explored in this paper where the duration of the maintenance tasks are not predefined, rather generated dynamically. An adaptive learning system is employed to determine the duration of a future task based on similar prior tasks. Durations of some maintenance tasks are calculated using statistical regression where nature and specification of the task are well defined. However, for other tasks where the durations are dependent upon the condition of several other parts and subsystems, they are derived through machine learning methodologies like Neural Network and Bayesian rule. Moreover, the duration of the tasks is further refined by Prognostics Health Management assessment that predicts impending failure based on near real time condition of vehicles and its subsystems. Determining the task durations dynamically based on prior knowledge or from prognostic data will make the schedule efficient, saving money and resources required for maintenance.