| Summary: | Manufacturing industries across the world witnessing emergence of data driven decision making fueled by advanced sensor technology, efficient and robust back end analytics, and availability of powerful computing options. Prognostics and Health Management (PHM) is gaining traction for machine component diagnostics and prediction of remaining useful life (RUL). Most often diagnostic and prediction applications involve sensors that assess state of the system followed by
analytics to gain insights from the data. The sensors generate time series data which have nonlinear, nonstationary characteristics with noise and drifts. Traditional time series representation techniques involve time, frequency or time-frequency analysis. These approaches are effective however with increasing complexity, many methods fail to fully capture the underlying dynamics of the process form the time series. Moreover, such computation methods are mostly suitable for offline or
cloud computing lacking real time analysis. With emergence of edge computing, data analytics and inference are carried at node (machine) in real time. This requires effective signal representation techniques that can best represent the dynamical process underlying the time series. In this work we investigate the suitability of three signal representation methods used for analyzing nonlinear time series in financial and health care applications. In Chapter 2 we investigate
Complexity-entropy causality plane (CECP), a parsimonious representation space for time series that two dimensions: normalized permutation entropy and Jensen-Shannon complexity . We demonstrate the effectiveness of CECP on three rotary component condition assessment applications. We use CECP representation of vibration signals to differentiate health conditions of machine components (bearings, and gears). The results confirm that the CECP representation is able to detect, with high
accuracy, changes in underlying dynamics of machine component degradation states. The CECP representation can be used for both predictive analytics and visual monitoring of changes in component condition. This method requires minimal pre-processing of raw signals. Furthermore, it is insensitive to noise, stationarity, and trends. In Chapter 3 we investigate Recurrence Quantification Analysis (RQA), and CECP for estimation of tool flank wear using multi-modal sensor signals. Two flank
wear datasets, one containing 15 tools and another containing 9 tools are created with 168 and 95 flank wear records respectively. We observe that the features generated by the CECP method yields lower values of prediction errors as compared to RQA method. Among the sensor signals, the feature representation of the force in cutting direction, using the CECP method, yields the highest value of classification accuracy (88.54%) In Chapter 4 we investigate visibility graph (VG) method and
its variant horizontal visibility graph (HVG) for estimation of flank wear for turning operation. The sensor signals are converted to complex network using VG and HVG method. For the reconstructed network we compute two measures, namely graph energy and average local clustering coefficient. The computed measures are then used in estimation of flank wear. HVG generates the highest value of classification accuracy (88.17%) for force in the feed direction. We observe that both VG and HVG
are suitable methods for representing sensor signals and the computed network measures exhibit sensitivity towards flank wear.
|
|---|
