Machine Learning-Driven RAM Analysis Using Multi-variate Sensor Data

Abstract

SUMMARY & CONCLUSIONSWith the proliferation of economical sensors and data collection devices, it is becoming easier to collect real-time usage data from equipment while in-service, resulting in an abundance of time-series data (such as temperature, flow, vibration, strain, etc.). Revisiting the traditional life data analysis generated using accelerated test data or time-of-failure service data can be enhanced with these new data to improve asset reliability, operational performance, and cost. The need is to incorporate time-series data into life data analysis, enabling continuous estimation of time-to-failure and associated maintenance planning. This paper will present a simple method to contain time-series data collected from a fleet of ground vehicles to update previously calculated reliability distributions for various vehicle subsystems, such as brake and engine subsystems. Through this continuous estimation, subsystem maintenance can be optimized. The gathered performance and time-series sensor data are modeled using Machine Learning (ML) models, which helps establish a vehicle/subsystem level performance profile.Additionally, the data clusters are augmented using maintenance text datasets or subject matter expert opinion. The vehicle/subsystem performance profile captures various components' mission and usage information. This profile information helps capture vehicle/subsystem performance for normal operation through machine learning methods. Features for the ML models are derived using higher-order moments and other descriptive statistics on time-series data. The ML models are developed to capture various vehicle/subsystem performance profiles. Using the predictions of ML models, a potential anomaly in subsystems, such as brake failure, can be detected. When detected, these anomalies help update the previously generated reliability score of the vehicle/subsystem. The failure prediction for the vehicle/subsystem is updated using this technique. This continuous loop of performance profiling, reliability estimation, and prognostics prediction helps to determine vehicle/subsystem maintenance needs proactively. This proposed methodology is demonstrated using real-life in-service data collected from a fleet of ground vehicles. The results show a significant improvement in detecting failure for the fleet of vehicles under investigation.