Abstract:

To process multi-dimensional time-series data collected from aero-engine sensors， previous methods tend to focus on the identification of a single sample point and the proposed models often contain a large number of parameters to be tuned， while the single-variable mining approach often ignores the influence of the delayed effect of the state parameters on the identification of working models. In addition， these methods poorly consider the case of large-scale processing of data containing transition state patterns. This paper proposes a new method， named AutoMiner， that can quickly and automatically mine the implied working patterns in multi-dimensional time-series data. The method solves for the combination of breakpoints， minimizing the encoding cost by constructing the encoding cost of the model. In the process of the encoding cost construction， the method of marker propagation is adopted to identify the working mode of each segment， generating soft labels to successfully solve the problem of identifying the transition state mode. Multiple experiments are conducted on the sensor data collected from an engine and presented visually. It is demonstrated that the AutoMiner method is superior to the comparison algorithms in terms of evaluation metrics for both temporal segmentation and pattern mining. In addition， the AutoMiner method supports parallelized computation with good transferability in more complex working pattern mining scenarios.

Key words: working pattern mining, multi-dimensional time series, non-parametric model, cost of coding, labeling propagation, aero-engines