Advances in Adaptive Data Analysis

An important perspective in electric consumption obviously is the forecasting. A crucial step in the forecasting process is the modeling. It is commonly admitted that many variables are influential for the prediction in this context. On the other hand, a prediction, to be robust and efficient, has necessarily to rely on a small number of well chosen predictors. We are typically in a situation where sparse multidimensional modeling can bring an essential input, and this paper is an attempt to prove it. In this perspective, we shall address the question of providing a sparse representation of the intraday load curves, with good approximation properties.

One difficulty is that we have here a large set of potential predictors among climate variables and shape "patterns", and even if high dimensional sparse methods have clearly as objective to select among a large number of covariates, they are especially efficient when the predictors are not too correlated. So our task is twofold: first we need to operate a preselection of the predictors, and then use an appropriate sparse method.

The first part and especially when shape patterns are concerned is not simple. We solve this aspect using a pre-processing in three steps process. We first provide a sparse modeling of the intraday load curves as functions of the time. Then, using the sparse representation of each load curve, we define clusters of consumption, which yield finally typical profiles, the group centroids or patterns of consumption.

Then we investigate the modeling of the intraday load curves using both exogenous climate variables and endogenous consumption variables as previously designed. We exhibit, as a result, sparse models relying just on few variables and also characterized by very good approximation performances. We end by explaining how sparse approximation is a first step towards clustering. The sparse approximation is here obtained by performing a robust and simple algorithm, LOLA, presented at the end of the paper.

The authors and their research team developed several nondestructive inspection-based structural health assessment and monitoring procedures to detect defect at the local element level, representing structures by finite elements. They are based on time domain nonlinear system identification-based concept. By tracking changes in the identified system parameters of the elements, the location and severity of defect can be assessed. To increase their implementation potential, only limited numbers of very short durations measured noise-contaminated acceleration time-histories are used in the identification algorithms. Numerous multi-disciplinary advanced data processing schemes used in developing these algorithms are briefly discussed. Presence of uncertainties in data processing and mitigation strategies are emphasized.

Traditional engine waveform analysis in time-domain fails to perform an accurate fault diagnosis when the fault waveform is very close to the normal waveform in time domain. A novel engine waveform analysis method is presented. In this paper, the aim is to perform fault diagnosis efficiently under such circumstances. This method is proposed by combining a new technique, called sensitive frequency band (SFB) selection, with the developed Hilbert–Huang transform (HHT). This can alleviate "mode mixing" by removing noise from the engine waveforms and reveal the time–frequency characteristics for a signal by deriving its time–frequency spectrum (TFS) distribution. The method is then applied to analyze the engine injector-pulse-width waveforms, and it works well for signal noise reduction and fault diagnosis.

Empirical mode decomposition (EMD) lacks theoretical support. We propose a piecewise monotonous model for EMD, and prove that the trend-subtracting iteration converges and IMF-separating procedure ends up in finite steps under mild conditions. Experiments are implemented and compared with the classical EMD.