Clays in Natural
and Engineered
Barriers for Radioactive
Waste Confinement
5
th
International
meeting
www.montpellier2012.com
Montpellier
October 22-25, 2012
© OT Montpellier Cecil Mathieu - © Phot
o d’ar
gile : LEM/Andr
Montpellier - October 22-25, 2012 5th International meeting
P/GPTP/11
INFORMATION CONTAINED WITHIN THE LARGE SCALE
GAS INJECTION TEST (LASGIT) DATASET EXPOSED USING A
BESPOKE DATA ANALYSIS TOOLKIT.
Bennett DP
1*, Cuss RJ
2, Vardon PJ
3, Harrington JF
2, Thomas HR
11.Geoenvironmental Research Centre (GRC), Cardiff School of Engineering, Cardiff University, Queen’s Buildings, The Parade, Cardiff, CF24 3AA, United Kingdom.
2.Transport Properties Research Laboratory, British Geological Survey (BGS), Keyworth, Nottingham, NG12 5GG, United Kingdom.
3.Geo-Engineering Section, Department of Geoscience and Engineering, Delft University of Technology, PO Box 5048, 2600 GA Delft, The Netherlands.
The Large Scale Gas Injection Test (Lasgit) is a field scale experiment run by the British Geological Survey (BGS) and is located approximately 420m underground at SKB’s Äspö Hard Rock Laboratory (HRL) in Sweden. It has been designed to study the impact on safety of gas build up within a KBS-3V concept high level radioactive waste repository. Lasgit has been in almost continuous operation for approximately seven years and is still underway†. An analysis of the dataset arising from the Lasgit experiment with particular attention to the smaller scale features and phenomenon recorded has been undertaken in parallel to the macro scale analysis performed by the BGS.
Lasgit is a highly instrumented, frequently sampled and long-lived experiment leading to a substantial dataset containing in excess of 14.7 million datum points. The data is anticipated to include a wealth of information, including information regarding overall processes as well as smaller scale or ‘second order’ features.
Due to the size of the dataset coupled with the detailed analysis of the dataset required and the reduction in subjectivity associated with measurement compared to observation, computational analysis is essential. Moreover, due to the length of operation and complexity of experimental activity, the Lasgit dataset is not typically suited to ‘out of the box’ time series analysis algorithms. In particular, the features that are not suited to standard algorithms include non-uniformities due to (deliberate) changes in sample rate at various points in the experimental history and missing data due to hardware malfunction/failure causing interruption of logging cycles (the probability of the latter occurring increases with experimental duration).
To address these features a computational toolkit capable of performing an Exploratory Data Analysis (EDA) on long-term, large-scale datasets with non-uniformities has been developed. Particular toolkit abilities include: the parameterisation of signal variation (noise) in the dataset with time, for use as a small scale event indicator; a non-parametric time-series component analysis technique (Singular Spectrum Analysis or SSA) for trend identification; and a unique Non-uniform Discrete Fourier Transformation (NDFT) technique that is suited to a non-uniformly sampled time-series input. Specific details of the implementations of these techniques are outlined.
As a result of the application of the developed toolkit a number of easily observable and quantified phenomena are revealed, for example:
• The location of a number of small scale anomalous behaviours of potential interest are highlighted; • Frequency, amplitude and phase of highly cyclic sensors (such as temperature) are deterministically
established;
• Long term trends in each sensor series are identified, revealing the residual forms of the sensor records without the long term behaviour superimposed.
*
Corresponding author ([email protected]).
†
More information about the Lasgit experiment can be found in SKB’s Technical Report TR-10-38 (http://www.skb.se/upload/publications/pdf/TR-10-38webb.pdf).
GPTP/11
GPTP/11
717
Clays in Natural and Engineered Barriers for Radioactive Waste Confinement • Poster Sessions
P/GPTP/11 Re-application of the toolkit when applied to the residual time series as determined by the initial application further reveals information of potential interest from the dataset. For example, small scale events as indicated by the noise parameterisation process and frequency information determined by the NDFT process are less likely to be masked by the long-term trend (variation) in the original sensor record. Results of these improvements are also presented within the manuscript. Figure 1 indicatively depicts the forms of various results produced by the toolkit, where from left to right in the top row is the original signal, the SSA derived trend, the residual signal and the comparison of the residual to the detected frequency content. The bottom row from left to right shows the peaks of noise (possible events) in the original signal, the frequency domain of the original signal and the noise and frequency domain of the SSA derived residual signal.
Figure 1: Example of results of toolkit application showing trends, residuals, frequency domains and event indicators. Occluded information about the experimental system is revealed on reapplication of the toolkit. Workflow of derivation is indicated by arrows.
Initial interpretation of the information exposed by the EDA performed through application of the developed toolkit is presented. The qualitative results, i.e. the event indicators (and to some extent the trend forms) are tentatively associated with experimental procedure or response e.g. changes in noise floor correlated with hydraulic over pressurisation down-hole. The quantitative results, i.e. the frequency information, are used to estimate the effect of environmental conditions on the experimental set-up.
While manipulation of a dataset to this extent can (and in the case of the Lasgit experiment, does) expose valuable information useful in further analysis, care must be given to ensure the phenomenon revealed are not a spurious result of the applied processing. A brief comparison of key observations made before and after trend removal are presented to evaluate the validity of the toolkit process with respect to information exposure.
A toolkit has been developed to perform an EDA on large-scale long-term datasets. An analysis on the Lasgit dataset successfully exposes, among other things: information regarding small scale events; non-parametric long-term trend identification; and deterministic quantification of frequency content. The uniformity and mechanical nature with which the information described above is exposed and quantified provides a level of rigour to and removes subjectivity from the resultant toolkit output by effectively turning observations into measurements. This exposed information can be used to guide and underpin investigation into scientific process within the experimental set-up. While developed specifically for the Lasgit experiment the toolkit is expected to be generally applicable to long-term, large-scale geotechnical or environmental experimental datasets with time series information.
Acknowledgements: The research leading to these results has received funding from the European Atomic Energy Community's Seventh Framework Programme (FP7/2007-2011) under Grant Agreement No. 230357, the FORGE project.