Predictive Maintenance for Ground Terrain Vehicles _all projects

Predictive maintenance is the practice of scheduling maintenance of a system in anticipation of predicted problems. It is not simply the servicing of a system (car, hydraulic circuit, aircraft, etc...) after a set passage of time or use, it is the active monitoring of data generated from onboard sensors and the prediction, from that data, of degradation in the system and the faults that would arise were it left unchecked. Our team has been working on a collection of projects aimed at implementing predictive maintenance methods on ground terrain vehicles. We received massive amounts of data generated from nearly 2000 different US Army vehicle and are using it in our research focused on the following key areas of predictive maintenance:

• Prediction of critical faults in advance

Personnel:
Mohan Kopuru (Principal Investigator)
Somayeh Bakhtiari Ramezani
Alexander Sommers
Dr. Shahram Rahimi

This project aims to create a sophisticated machine learning framework that predicts the occurrence of several faults that occur in the vehicles, due to their daily wear-and-tear, and provides the users with metrics for the making of maintenance decisions. PAL, in collaboration with The Institute for Systems Engineering Research at Mississippi State University and ERDC at Vicksburg, have managed to develop a machine learning system that predicts the occurrence of some key faults [Editor's note: We have? I thought our results weren't that good?], thereby allowing maintenance to be scheduled before a vehicle is unexpectedly disabled. This machine learning system was able to predict the occurrence of an Injector [Editor's note: Maybe people want to know what this is, a fuel injector?] fault in our use case vehicle model at least 12 hours in advance [Editor's note: We say "key faults", plural, but list one, singular.].

• Multivariate Long Time Series Segmentation and Motif Discovery

Personnel:
Somayeh Bakhtiari Ramezani(Principal Investigator)
Dr. Shahram Rahimi

Complex systems such as vehicles, or machinery comprised of multiple subsystems, can perform operations (Ex: Turn left, change gear, drive up a hill, drive down a hill) as sensor data is being captured. The captured "time series" can be seen as a combination of several building blocks (Motifs), each related to a unique behavior the system had while it was being measured (the behavior of: braking, driving up a hill, making a U turn, having an engine problem). Several studies have introduced segmentation algorithms to predict the upcoming behaviors of a system given some prior behaviors, or determine to just identify the motifs the system exhibits. Time series segmentation is a computationally expensive operation and implementing it on GPUs and Quantum computers are considered parts of emerging fields of computing. Efficient time series segmentation algorithms can benefit several domains such as predictive maintenance, self driving cars, and motif discovery and gene prediction in DNA sub-sequences.

• Multi-level Clustering for Identification of Degradation Patterns

Personnel:
Somayeh Bakhtiari Ramezani(Principal Investigator)
Dr. Shahram Rahimi

An analyzed time-series may be found to have one or more motifs in it which foretell some negative even in the future of the measured system. This would be of great value if the system was a vehicle, or a person being measured by medical sensors. In vehicles, for example, overdue maintenance can result in severe damage if a system is allowed to fully fail but this can be prevented through monitoring , detection of fault predicting motifs, and scheduling maintenance events before failure occurs. In this study we use a combination of Toeplitz Inverse Covariance-Based Clustering (TICC) and Hidden Markov Model (HMM) to extract the states/clusters in vehicular sensory data and detect the healthy/unhealthy state sequences to forecast fault occurrences.

Unsupervised Feature Analysis using Self Organizing Map and Auto-encoders
Personnel:
Somayeh Bakhtiari Ramezani(Principal Investigator)
Alexander Sommers
Dr. Shahram Rahimi

Key to the success of any machine learning algorithm is providing the algorithm with a representative set of features by which to model some target. Sometimes a problem arises not when there are too few features to use, but too many, and of unknown worth. Choice of feature reduction methods can vary depending on the supervised or unsupervised nature of the algorithm and thus the existence, or absence, of a target attribute. Feature selection refers to a process where a desirable subset of the input features is selected for further processing. This can lower the computational load on the system and increase model accuracy. Many feature selection methods depend on “High Correlation Filtering”, which work when the prediction of some target variable(s) is to be optimized. An algorithm that is independent of a target value is preferred to distinguish between features that uniquely describe the data, and the redundant ones. This type of feature reduction is fundamentally different from feature projection methods, such as PCA, as it keeps attributes as they are, without changing or transforming them into a new combination. Algorithms such as Backward Feature Extraction (BFE) or Forward Feature Selection (FFS) are only efficient for small databases, which makes them very time consuming and computationally expensive for larger databases. Self-Organizing Maps (SOM), on the other hand, can facilitate feature selection via extracting similar maps for similar attributes. This work uses a convolutional auto-encoder to cluster the SOMs for each attribute to find the similar attributes without using any target value for feature selection.

• Applying Life Analysis to CBM+: Using Data to Assist Experts in Decisions

Personnel:
Ben Wiggins
Dr. Shahram Rahimi

In the field of medical research the method known as "life analysis", or sometimes called "time before event analysis", has been employed to assist in understanding how time series data correlates to events occurring within a given time-frame. A statistical model is built from such datasets which can be used to pronounce the likelihood of a target event occurring for a subject within a given time span.Such analysis of the likely length of time before a medical event occurs in a patient is highly useful and the same method of thought can be applied to CBM+(Condition Based Maintenance +). By using Life analysis on data that involves the maintenance of vehicles, experts can gain a better understanding of the probability of a failure level event during a given vehicle run-time. In this paper, we will explore the methods and mindset behind how we can modify and expand life analysis to assist in performing CBM+.