The degradation process of a complex system may be affected by many unknown factors, such as unidentified fault modes, unmeasured operational conditions, engineering variance, environmental conditions, etc. These unknown factors not only complicate the degradation behaviors of the system, but also lower the quality of the collected data for modeling. Due to lack of knowledge and incomplete measurements, certain important context information (e.g. fault modes, operational conditions) of the collected data will be missing. Therefore historical data of the system with a large variety of degradation patterns will be mixed together. With such data, learning a global model for Remaining Useful Life (RUL) prediction becomes extremely hard. This has led us to look for advanced RUL prediction techniques beyond the traditional global models.
In this thesis, a novel RUL prediction method inspired by the Instance Based Learning methodology, called Trajectory Similarity Based Prediction (TSBP), is proposed. In TSBP, the historical instances of a system with life-time condition data and known failure time are used to create a library of degradation models. For a test instance of the same system whose RUL is to be estimated, similarity between it and each of the degradation models is evaluated by computing the minimal weighted Euclidean distance defined on two degradation trajectories. Based on the known failure time, each of the degradation models will produce one RUL estimate for the test instance. The final RUL estimate can then be obtained by aggregating the multiple RUL estimates using a density estimation method.
A case study using the turbofan engine degradation simulation data supplied by NASA Ames is provided to study the performance of TSBP. In this study, the TSBP method has demonstrated significant improvement in performance over a Neural Network based prediction method.