`(EXCERPTED)
`
`
`
`
`
`
`
`LEAST SQUARES FILTERING AND TESTING FOR
`POSITIONING AND QUALITY CONTROL DURING 3D
`MARINE SEISMIC SURVEYS
`
`
`
`VASSILIS N GIKAS
`Surveying Engineering N.T.U.A.
`
`
`
`
`
`Thesis submitted for the Degree of
`Doctor of Philosophy
`
`Department of Surveying
`University of Newcastle upon Tyne
`August 1996
`
`Ex. PGS 1041
`
`
`
`Introduction
`
`RESEARCH OBJECTIVES AND SCIENTIFIC RESULTS EXPECTED
`
`The overall aim of the project has already been outlined, namely to develop a general,
`integrated and rigorous approach to the positioning and quality control in real time of
`marine seismic networks. In order to achieve this, emphasis has been placed on a
`number of objectives
`
`
`
` Determination of an optimum general mathematical description of the streamer
`shape by preliminary fitting of streamer models to compass data.
` Acquisition of a formal description of the geometry of the whole configuration by
`integrating all positioning data types into a single functional model.
` Computation of the real-time position and quality measures of any point deployed
`in a seismic network by adopting a Kalman (or other) filter as the basic stochastic
`process.
` Test the integrated model for appropriateness and for its sensitivity to detect and
`identify expected biases in the raw data by incorporating a uniform testing
`procedure.
` Assessment and testing of the correctness of the mathematics and the feasibility of
`the associated algorithms in terms of convergence, solubility and computational
`efficiency by preparing software for the various parts of the process and testing
`with real offshore data.
` Refinement of functional and stochastic models based on detailed analysis using
`alternative model hypotheses.
`
`
`The results are tested mathematical models, in the form of computational algorithms,
`for the following
`
`
`
` The shape of the seismic streamers.
` The dynamics of 3D seismic configurations during data collection.
` The real-time positions and quality measures for offshore seismic surveys.
` The effect of the network geometry and the relative stochastic properties on seismic
`network positioning and quality control.
`
`
`
`
`- 3 -
`
`Ex. PGS 1041
`
`
`
`Chapter One: Acquisition and Positioning 3D Marine
` Seismic Surveys - An Overview
`
`issues such as multi-source, multi-streamer acquisition. Four types of records have
`been defined (CENSUS User’s Guide, 1994)
`1. Header records - This type of records holds similar information as the header files
`of UKOOA P2 formats.
`2. Point position records - These records are used to identify the point being
`positioned. The most common are, source fired (S), vessel (V, P1/90 only), and
`tailbuoy (T, P1/90 only). The source records contain also information such as line
`number, shotpoint number, date/time and water depth.
`3. Receiver records - Receiver records contain information such as receiver ID flag,
`receiver position (easting and northing), and cable depth.
`4. Relation records - This type of records is an extension to the format and is used to
`prevent the pointless repetition of unchangeable information for different shots.
`
`
`In addition to the UKOOA format other exchange data formats have been developed
`such as, SEG P1 (1983) as well as industry standard formats such as, the Shell’s SPS
`format, the Advance Geophysical’s ProMAX database format and the Green
`Mountain’s MESA format.
`
`1.5.2 Geophysical Contractors’ Navigation and Binning Systems
`
`It is a general conclusion from the discussion so far that the trend seems to be a
`movement of the seismic industry towards faster multi-tasking integrated software and
`central processing units (UNIX based workstations). Almost all major geophysical
`contractors/companies have developed (and continuously
`improve)
`their own
`navigation and binning/processing systems to meet this demand. The main
`characteristics of these systems are outlined bellow
`1. During acquisition usually some of the data are synchronized with shot time (as
`compass azimuths and network acoustics), and some are recorded at the sensor time
`(Syledis, GPS, RGPS).
`2. Storage at the UKOOA P2/91/94 formats and real-time graphic display of
`acquisition is a common practice.
`3. Some systems, as GIN 2000 developed by CGG, compute source and receiver
`positions based on least squares algorithms for the various networks of the spread
`- 28 -
`
`
`
`Ex. PGS 1041
`
`
`
`Chapter One: Acquisition and Positioning 3D Marine
` Seismic Surveys - An Overview
`
`(vessel-buoy, relative head and relative tail networks as well as streamer shape).
`Other systems, as TotalNet, developed by WESTERN ATLAS, implement
`integrated network solutions by means of a Kalman filter.
`4. Quality control, including monitoring of the quality of the recorded data (setup,
`configuration, spread geometry, data integrity, and statistical analysis) is an
`essential feature in today’s systems.
`5. On-board binning systems provide real time monitoring of CMP distribution
`throughout a 3-D survey. Also, most binning systems’ capabilities include, flex
`binning, editing and rebinning algorithms.
`
`
`In Table 1.1 a list of the navigation and binning systems of some major geophysical
`contractors is given.
`
`
`Contractor
`
`Navigation System
`
`WESTERN
`GECO / PRAKLA
`CGG
`DIGICON
`
`WISDOM II
`TRINAV
`GIN 2000
`MAGNAVOX 200 / SCOPE III
`
`Binning / Processing
`System
`FLEX QC / CNAVCHK
`TRINAV / QC
`GIN 2000
`BIRDOG
`
`
`Table 1.1 Contractors’ navigation and binning/processing systems
`
`- 29 -
`
`
`
`
`
`Ex. PGS 1041
`
`
`
`Chapter Three: The Kalman Filter
`
`CHAPTER THREE
`THE KALMAN FILTER
`
`
`
`INTRODUCTION
`
`3.1
`
`The Kalman filter is probably the best known of the commonly used recursive
`algorithms for the estimation of the parameters of time-varying systems. It has
`constituted the framework for a unified and concise treatment of a broad range of
`filtering problems from electronic engineering to surveying and geodesy. However,
`usually, the Kalman filter is perceived as a ‘black box’, into which measurements go in
`order to be converted into positions, since there still remains a certain amount of
`ignorance in the hydrographic surveying community with respect to Kalman filtering.
`Therefore, in the past, it has not proved popular with the offshore community and
`many offshore operators currently prefer simple and independent ‘epoch by epoch’
`least squares computations. This chapter aims to provide a brief description of the
`Kalman filter models and algorithms as well as to explain the meaning of the most
`commonly used terms associated with it.
`
`Kalman filter estimates have the advantage of being least squares estimators. This
`means, as can be shown (Cross, 1983), that they are the best in the minimum variance
`sense within the complete class of the linear unbiased estimators. For these reasons
`they are often referred to as Best Linear Unbiased Estimates (BLUEs). The basic
`difference between a simple least squares computation and Kalman filtering, is that, the
`Kalman filter comprises of the specification of a dynamic model in addition to an
`observation model that to together provide an optimal solution. The use of a dynamic
`system reveals, somehow, the amount of knowledge with respect to the system
`dynamics, i.e. the behavior of the system as it varies with time. For instance, in the
`case of a moving vessel, where its position and velocity are the desired results, the
`position fix measurements provided by a shore-based or satellite navigation system
`
`
`
`- 65 -
`
`Ex. PGS 1041
`
`
`
`Chapter Three: The Kalman Filter
`
`constitute the observation model while the dynamic model is expressed by the
`assumption of constant acceleration between the position fixes.
`
`3.1.1 Predicting, Filtering and Smoothing
`
`Three types of problems constitute the estimation problem associated with Kalman
`filtering. These are known as prediction, filtering and smoothing, and they are related
`to the estimation of the state vector parameters x, of a time-varying problem, computed
`at any instant with respect to the present time.
`
`The step of prediction is related to the computation of the filter estimates x(-), at time
`of interest tj that occurs after the last available measurement(s). In this case, only the
`state estimates and its associated covariance matrix computed from the previous epoch,
`as well as the dynamics of the system, are used to provide the state vector solution.
`
`Once a new measurement(s) is available the predicted state vector x(-) is used together
`with the new measurement(s) to solve for the state estimates. In this case, in which the
`time of the last measurement(s) coincides with the estimation time, the problem is
`referred to as filtering and the state vector denoted by x(+).
`
`At a post-proccessing stage the state vector parameters can be computed at any time tj
`where information for some time interval prior and later to time tj is used. This part of
`the problem is known as smoothing and it denoted by x(s). Obviously, a solution of
`this type can only be available after some delay. Usually, in most real-time surveying
`applications, only the prediction and filtering steps are implemented since their
`implementation is straightforward. Although smoothing procedures can be executed in
`real time they are usually only used in post-processing because they require much more
`memory space.
`
`These three distinct estimation problems can be defined as
`
`
`
`
`
`
`
`
`
`
`
`
`
`3.1
`
`- 66 -
`
`ti > tj prediction
`filtering
`ti = tj
`ti < tj smoothing
`
`x(-)
`x(+)
`x(s)
`
`
`
`
`
`
`
`
`Ex. PGS 1041
`
`
`
`Chapter Three: The Kalman Filter
`
`The three types of the filtering problem are illustrated in Figure 3.1.
`
`
`
`
`Figure 3.1: Predicting, filtering and smoothing
`
`3.1.2 Kalman filtering versus Simple Least Squares
`
`Kalman filtering has the following specific advantages over simple ‘epoch by epoch’
`least squares and it is in order to exploit these fully that Kalman filtering is selected as
`the basic stochastic process for most offshore positioning applications.
`
`1. Simple least squares treats each epoch independently. This means that it does not
`use knowledge of the motion of the system. Often, and especially in seismic work,
`it is possible to make a very accurate prediction of where the network will be at any
`epoch using just the previous position and the estimated configuration motion. Not
`to use this ‘knowledge of motion’ is effectively throwing away information and
`must lead to poorer quality results than those obtainable from a properly tuned
`Kalman filter. In the past (and sometimes today) poorly tuned filters were used and
`in this case results might be worse - simply because the system motion may have
`not been well determined and/or not used properly in the estimation process. So
`simple least squares is a safe option - but it does not have the potential accuracy of
`Kalman filtering. The challenge, of course, is to tune the filter properly in real time
`- and the fact that some have failed to do this in the past has led to Kalman filtering
`gaining a poor reputation in some circles.
`2. The use of a Kalman filter for a highly complex seismic configuration enables a
`rigorous computation of precision and reliability measures such as error ellipses
`and marginally detectable errors respectively, (Cross et al, 1994a). If a step-by-step
`- 67 -
`
`
`
`Ex. PGS 1041
`
`
`
`CHAPTER FOUR
`QUALITY MEASURES IN OFFSHORE POSITIONING
`
`
`
`INTRODUCTION
`
`4.1
`
`4.1.1 Introduction to Quality Assessment
`
`In any measurement process all observations will invariably be erroneous, however
`sophisticated their measurement technology might be. Competent sensor calibration
`and careful measuring procedures are a first defense against observational errors that
`will further reduce these but not completely eliminate them. Therefore, it is essential to
`establish how ‘good’ the measurements are, in order to assess the quality of the results
`of an estimation process, i.e. to assess the size and nature of any undetected errors that
`might remain in these. The characteristic nature of the observational errors is usually
`described as being either random, or biased.
`
`Random errors are by definition unpredictable and unavoidable, caused by small
`fluctuations in the physical factors that constitute the measurement process (Cross et al,
`1994b). Random errors, that are usually small in size, are described by statistics and it
`can be shown, via the so-called central limit theorem (Cramer, 1946), that they are
`from a normally distributed population which allows statistical hypothesis to be tested.
`
`Biases can take the form of gross errors, known also as blunders or outliers, and
`systematic errors. Gross errors (often large in size) are due to erroneous observations
`on the part of the observer resulting of carelessness or confusion. Systematic errors
`refer to model misspecifications that follow some physical law and can be described by
`a mathematical function. This kind of errors can be eliminated by careful sensor
`calibration and design of the functional and dynamic models.
`
`
`
`
`- 85 -
`
`Ex. PGS 1041