ASSESSMENT OF DE-SPIKING ALGORITHMS APPLIED TO LWD DATA

摘要

Dealing with noise in log data analysis plays an important role for understanding and interpreting various formation properties. A particular type of impulse or spike noise is often generated during Logging While Drilling (LWD) operations due to harsh vibration and uneven tool movement. Some sensor physics based studies have been done to reject spurious noise if the rate of change of an observed variable exceeds realistic changes. The diversity of measurements, on the one hand, equips us with many tools. On the other hand, the profusion of sensors may cause confusion. It would be very helpful, therefore, to have some general data screening procedures. Although perhaps not perfect, such procedures should be useful to detect troublesome spikes in the early phase of an analysis without prior knowledge of data or an over dependence on the tool physics. This paper reviews several general spike detection methods from the literature, applies and evaluates two of them using LWD data. In the first method we utilize a simple time series predictive/adaptive technique to determine if a data point is ‘questionable’ or ‘good’. In the second method we use neural networks to model depth series data and designate spikes where the predicted values are significantly deviated from the expected outputs. The possibility that the detected points are true spikes is assessed through visual inspection and through the use of correlatedvariables. The advantages and limitations of each method are also addressed in the paper. De-spiking involves spike detection and spike replacement or correction. Spike replacement is more arbitrary than spike detection, and can be performed using simple heuristic smoothing, sophisticated statistical multiple imputation, tool physics based data correction, and numeric model prediction. This paper compares a spike correction method in using the local cubic spline repairing with a method in using the best linear transformation between the measurement data and the prediction data. The results show that both approaches are cost-effective and less sensitive to erroneous misreplacement of the ‘good’ data due to the lack of information.