Traitement des diagraphies acoustiques. Première partie : application de techniques issues de l'intelligence artificielle au pointe des diagraphies acoustiques Full Waveform Acoustic Data Processing. Part One: an Artificial Intelligence Approach for the Picking of Waves on Full-Waveform Acoustic Data

Abstract

Les enregistrements des données acoustiques en champ total (fuit waveform) ont conduit le géophysicien et le diagraphiste à utiliser des techniques de traitement du signal pour séparer les différentes ondes observées sur les enregistrements. L'une des tâches importantes du traitement des diagraphies acoustiques est le pointé des temps d'arrivée des différentes ondes enregistrées. Une démarche de type système expert a été utilisée pour mettre au point un algorithme multicanaux qui réalise le pointé des différentes ondes, à l'aide de règles faisant intervenir les caractéristiques ou attributs de chaque onde. Une onde est caractérisée par sa vitesse, sa fréquence, son amplitude et sa cohérence latérale. L'algorithme fournit un ensemble de logs accompagnés d'une estimation de la dispersion des mesures à chaque cote profondeur. Les logs fournis sont les logs de lenteur et les logs de fréquence. Les résultats obtenus sur un ensemble de diagraphies acoustiques enregistrées dans un puits vertical du Bassin parisien montrent que la dispersion des mesures reste faible en comparaison des pas d'échantillonnage en temps et profondeur. Les logs de dispersion peuvent aussi permettre de détecter des phénomènes physiques tels que caves, fractures, conversions d'ondes ou interférences, reliés à la lithologie. Dans une deuxième partie, nous montrerons différentes techniques de séparation d'ondes. La troisième partie illustrera, sur un cas particulier, l'utilisation des logs issus des diagraphies acoustiques pour caractériser les formations. The full waveforms recorded by an array of recievers in a borehole sonic tool contain a set of waves that can be fruitfully used to obtain detailed information about the nearborehole lithology and structure. The different waves that can be observed by full-waveform sonic data are described in this article. The main tools used in the recording of full-waveform data are then reviewed. The different waves and the effects of interferences, lithological variations and reflections of waves, are illustrated for common-offset trace collections (Figs. 4 to 8). Full waveform acoustic data processing mainly involves wave arrival-time picking and wavefield separation. The first part of this paper is devoted to arrival-time picking. The second part will be devoted to wave separation. The third part will present a case history on the use of acoustic logs. The great amount of full-waveform sonic data leads geophysicists and log analysts to implement automatic algorithms for picking the wave arrival times. After a review of the main conventional picking methods (Figs. 9 to 12), an automatic routine based on an Artificial Intelligence approach is described. In this routine, which is a stand-alone multichannel algorithm, the reasoning of the geophysicists picking a particular wave on common-offset trace collections of fullwaveform data is expressed in the form of rules. Identification of arrivals is based on criteria of similarity of shape and lateral continuity of the waves in contiguous traces. The main parameters used are amplitude, frequency and lateral correlation. A positive or negative peak is picked, depending on the polarity of the wave. The geophysicist indicates on the first record the particular wave to be picked by giving its approximate arrival time. Then the algorithm automatically picks the wave arrival-times on all the other records. When picking is performed on one common-offset trace collection, the A algorithm is used to search for the optimal path in the graph where the nodes are extrema on the traces and rows are the possible links between the extrema on contiguous traces. The principle of the A algorithm is shown in Fig. 13 for a simple example. Since picking for one common-offset trace collection must be coherent with picking for the others, rules expressing the coherency conditions are added to the algorithm for a more robust selection of the optimum path, leading to picking for common-shot records. The algorithm was first implemented using OPS5. OPS5 is a rule-based programming language with a first-order inference engine. OPS5 allows declarative programming, which makes it easy to introduced new rules corresponding to knew knowledge and/or a modification of the heuristics. This facilitates the evolution and incremental development of the algorithm. Then, to optimize computing time, the algorithm was implemented in FORTRAN. The wave arrival times given by the automatic routine were used to determine, at each depth, the slowness of the wave in the formation. The routine also delivers an apparent frequency curve. At each depth the redundancy of measurements (depending on the geometry of the sonic tool array and on the depth sampling rate) was used to compute a mean value of each parameter as well as its standard deviation. The standard deviation gives the measurement dispersion and an estimate of the error. The error log can be used to detect mispicks (due to phase shifts) as well as interferences of waves. When the standard deviation is low, it is related to the accuracy of the measurement, which depends, for the time value, on the time sampling rate and on the signal-to-noise ratio. The standard deviation increases when direct waves and converted waves interfere. Examples of results obtained with the picking method described are shown for a borehole experiment performed in sandstone reservoir layers. Full-waveform sonic data were acquired in a vertical well drilled in an anticline structure used for underground gas storage by Gaz de France. The sonic tool used was a Schlumberger Dipole Sonic Imaging tool. This tool is an eight-receiver/three-transmitter device. The receiver section contains eight dipole monopole stations spaced 6 inches apart. The distance between tho monopole transmitter and the first recolver was 9 ft. The source was fired at equal spacings of 6 inches throughout the open-hole section. The monopole transducer can be driven either by a low frequency pulse for Stoneley wave excitation or by a high frequency pulse for head waves. Each dipole transducer was driven by a low frequency pulse. In the studied part of the well, the reservoir layers are clean sandstones or shaly sandstones overlain by impervious shale cap rock. The well is completed by a casing down to the top of the reservoir layers. The slowness and frequency logs with the corresponding error logs are shown in Figs. 16 and 17 for Stoneley waves, in Figs. 19 and 20 for P waves, and in Figs. 22 and 23 for shear waves. The average standard deviation is very low campared with the time sampling rate, thus indicating the good quality of the measurements. Table 1 gives the mean standard deviations for the slowness logs. Typical increases in the standard deviation of the slownes log due to interferences can be observed in Figs. 15 and 16 at 888, 898 and 903 m. We can observe good correlation between the Stoneley frequency log and lithology derived from an independent analysis. In clean sandstone, the Stoneley waves have an apparent frequency of about 1200 Hz. In shaly sandstone, the frequency ranges from 800 to 1100 Hz. The value o Stoneley frequency is directly related to shaliness. Maximum shaliness is observed in Fig. 17 in the 898-903 depth interval.