BR0116453B1 - APPARATUS AND METHOD FOR CONVERTING THE SIGNALS OF A REGISTRATION TOOL - Google Patents

Description

“APPARATUS AND METHOD FOR CONVERTING SIGNALS OF A REGISTRATION TOOL” Priority This application claims the benefit of US Provisional Application No. 60 / 256,679, filed December 19, 2000; and Provisional Application No. 60 / 299,002, filed June 18, 2001. 1. Field of the Invention This invention relates to well logging, and more particularly to a neural network trained to process signals from a logging tool in a representation of training parameters. 2. Background of the Invention Modern oil drilling and production operations require a great deal of information regarding the parameters and conditions below. Such information typically includes characteristics of the well cavity-traversed earth formations, as well as data regarding the size and configuration of the well cavity itself. Oil well registration has been known in the industry for many years as a technique for providing information to a formation appraiser or driller regarding the particular formation of land being drilled. Gathering information regarding the conditions below, which is often referred to as "logging", can be performed by various methods. These methods include measurement while drilling, MWD, and logging while drilling, LWD, in which a logging tool is carried on a drill chain during the drilling process. The methods also include wire line registration.

In conventional oil well wire recording, a probe is lowered into the borehole after some or all of the well has been drilled, and is used to determine certain characteristics of the formations traversed by the well cavity. The probe may include one or more sensors to measure downhole parameters and is typically constructed as a hermetically sealed cylinder to house the sensors, which hang at the end of a long cable or "wire line". The cable or wire line provides mechanical support to the probe and also provides electrical connections between the sensors and associated instrumentation within the probe, and electrical equipment located at the well surface. Typically, the cable supplies operational energy to the probe and is used as an electrical conductor to transmit probe information signals to the surface. According to conventional techniques, various parameters of the earth formations are measured and correlated with the position of the probe in the drillhole when the probe is pulled up the hole.

A map or graph of a ground parameter or a log tool signal against the position or depth in the drillhole is called a "log". Depth can be the distance from the ground surface to the location of the tool in the borehole or it can be true depth, which is the same only for a perfectly vertical straight well cavity. Recording of the tool signal or raw data often does not provide a clear representation of the ground parameter that the training assessment or punching professional needs to know. The tool signal should normally be processed to produce a register that most clearly represents a desired parameter. The record is usually first digitally created by a computer and stored in computer memory, tape, disk, etc., and can be displayed on a computer screen or printed in hard copy.

Sensors used in a wire line probe typically include a source device for transmitting energy to the formation, and one or more receivers for detecting reflected energy from the formation. Several sensors were used to determine particular features of the formation, including nuclear sensors, acoustic sensors, and electrical sensors. See generally J. Lab, "A Practical Introduction to Borehole Geophysics" (Society of Exploration Geophysicists 1986); D.R. Skinner, "Introduction to Petroleum Production", Volume 1, 54-63 (Gulf Publishing Co. 1981).

For a formation to contain oil, and for the formation to allow oil to flow through it, the rock including the formation must have certain well-known physical characteristics. A feature is that the formation has a certain measurable resistivity range (or conductivity), which in many cases can be determined by inducing an alternating electromagnetic field in the formation by a transmitter coil arrangement. The electromagnetic field induces alternating (or whirling) electric currents in formation on paths that are substantially coaxial with the transmitter. These currents in turn create a secondary electromagnetic field in the middle, inducing an alternating electrical voltage in the receiver coil. If the current in the transmitter coil is kept constant, the eddy current intensity is generally proportional to the conductivity of the formation. Consequently, the conductivity of the formation determines the intensity of the secondary electromagnetic field, and thus the amplitude of the electrical voltage in the receiver coil. See generally, James R. Jordan, et al., "Well Logging II-Electric And Acoustic Logging," SPE Monograph Series, Volume 10, 71-87 (1986).

An exemplary induction tool is shown in the prior art drawing of Figure 1, in which one or more transmitters (T) and a plurality of receivers (Ri) are shown in a record probe. Each transmitter or receiver may be a set of coils, with modern arrangement induction tools having various receivers, for example Ri, R2, R3, and R4, of increasing transmitter to receiver spacing to progressively measure deeper into the formation.

In a conventional induction tool such as that shown in Figure 1, the coils are coaxially wrapped around a cylindrical mandrel. Both transmitter coils and receiver coils are solenoid shaped, and are coiled coiled with the mandrel. Such coils would therefore be aligned with the main axis of the recording tool, which is usually also the central axis of the borehole and is commonly referred to as the z axis. That is, the magnetic moments of the coils are aligned with the spindle axis on which they are wound. The number, position and number of turns of the coils are arranged to nullify the signal in a vacuum due to the mutual inductance of transmitters and receivers.

During operation, an oscillator supplies alternating current to the transmitter coil or coils, thereby inducing current in the coil or receiver coils. The induced current voltage in the receiver coils results from the sum of all induced eddy currents in the surrounding formations by the transmitter coils. Phase Sensitive Electronics measures the receiver electrical voltage that is in phase with the transmitter current divided by the magnitude of the transmitter current. When normalized to the correct scaling factor, this provides signals that represent the apparent conductivity of that part of the formation through which the transmitted signal has passed. The out-of-phase or quadrature component may also be useful because of its sensitivity to skin effect, although it is less stable and adversely affected by contrasts in magnetic permeability.

As noted, induced eddy currents tend to flow in circular paths that are coaxial with the transmitter coil. As shown in Figure 1, for a vertical drill hole that traverses horizontal formations, there is general symmetry for the induced current around the recording tool. In this ideal situation, each current flow line remains in the same formation along its entire flow path, and never crosses a bed boundary.

In many situations, as shown for example in Figure 2, the well cavity is not vertical and the bed boundaries are not horizontal. The well cavity in Figure 2 is shown with a tilt angle Θ measured relative to the true vertical. A bed boundary between formations is shown as a slope angle α. The inclined well cavity reaches the inclined bed at an angle β. As a result, induced parasitic currents flow through more than one medium, finding formations with different resistive properties. The resulting records are distorted, especially as the slope angle α of the bed boundaries increases. If the registration tool crosses a thin bed, the problem becomes even more exaggerated.

As shown in the graph of Figure 3A, an induction probe traversing an inclined bed produces a distorted register commonly referred to as "horns". The more severe the inclination angle, the less accurate the depth measurement is. Figure 3A represents a computer simulation of a register that would be generated during registration of a 3m thick bed (at real depth), with different graphs for different inclination angles. Figure 3B shows a computer simulation of a record that would be generated if the bed thickness were true vertical depth with different graphs for different inclination angles. As is evident from these simulated records, when the inclination angle increases, the accuracy and significance of the register decreases. In cases of high inclination angles, the graphs become virtually meaningless in the vicinity of the bed boundaries.

Figures 3A and 3B also illustrate that even for a vertical well traversing horizontal formations, the actual electrical signal or data produced by an induction logging tool is quite different from an accurate formation resistivity graph. In these figures, the desired representations of forming resistivity are the dotted line waveforms 10 and 20. The actual resistivity within a layer is generally uniform, so that there are abrupt changes in resistivity at the interfaces between layers. However, logging tools have limited resolution and do not directly measure these abrupt changes. When the transmitter coil T in Figure 1 is near an interface, as illustrated, its transmitted signal is divided between layers of different resistivity. As a result, the raw data or sign of the logging tool is a combination or average of the actual values of the adjacent layers. This effect is referred to as the shoulder effect. Even in the case of 0 ° shown in Figure 3A and 3B, where the tool is vertical and the formation is horizontal, the measured data is quite different from the desired representation of resistivity. As the slope increases, the effect is increased.

Much work has been done on methods and equipment for processing data or recording tool signals to produce an accurate representation of forming parameters. This data process is usually called inversion. Inversion is usually performed on some type of computer. In the prior art system of Figure 1, a block labeled "compute module" may perform some type of inversion process. The methods currently available to perform this processing are iterative in nature. Standard iterative methods have the disadvantage of being computationally intensive. As a result, inversion should typically be performed at computer centers using relatively large computers, which can deliver inversion results in a reasonable amount of time, and typically cannot be performed on computers suitable for well site use.

An alternative processing method is the reverse convolution method. This method is very fast and can be implemented at the well site, for example in the computation module of Figure 1. However, this method is based on linear filter theory, which is an approximation that is not always accurate. In deviated drillholes, the nonlinearity of the tool response becomes manifest, making the problem difficult for the reverse convolution method to operate. Reverse convolution methods do not generate actual representations of the formation parameters, so they cannot be correctly called inversion methods.

Previous attempts to solve log data problem inversion have used the parametric inversion method. This method is an iterative method that uses a front solver and criteria, such as least square inversion, to determine the best fit for the parameters of a predefined formation, usually a model with a step profile. However, if the actual formation does not conform to the predefined model, the output parameters determined by this method may be too far from the actual formation parameters. This is a consequence of the misplaced nature of the inversion problem that makes it highly nontrivial.

A more current method for reversing resistivity log data is the "Maximum Entropy Method", MEM. In this iterative method, a proposed test or formation model is modified to maximize functional entropy, which depends on the formation parameters. This method does not use predefined training and produces better quality solutions. It is more efficient than parametric approaches, but is still computationally intensive. It can be applied to any type of tool for which a front resolver is available. An example of the MEM method is set forth in US Patent No. 5,210,691, entitled "Method and Apparatus for Producing a More Accurate Resistivity Log from Data Recorded by an Induction Sonde in a Borehole".

In general, all iterative inversion schemes have essentially two parts. The first part is a front resolver that generates a synthetic record of a synthetic test formation that is a reasonable representation of a real formation. The test formation is a generally assumed square wave graph of a formation parameter, eg resistivity, against depth, such as graphs 10 and 20 in Figures 3A and 3B. The front resolver simulates the response of a selected recording tool to the test formation to generate the synthetic registration. If the registration tool has multiple transmitter-receiver sets or arrays, as illustrated in Figure 1, a separate front solution is needed for each set as long as each set responds differently. The second part of the iterative method is a criterion for modifying test formation. The criterion is based on the difference between the synthetic record that corresponds to the test formation and the actual record data measured by the tool. After the test formation has been modified, a new synthetic record is generated by the front resolver. This process is repeated iteratively until the difference between the synthetic register and the actual register is less than a predefined tolerance. The output of the inversion algorithm is the parameters of the final test formation. These parameters are plotted against depth to produce the desired record. It is the iterative nature of these methods that make them computationally intensive. Several efforts have been made to use Artificial Neural Networks, ANN, as part of inversion processes. For example, in the document entitled "Detection of Layer Boundaries of Array Induction Tool Responses Using Neural Networks", 69th Annual SEG Annual Meeting (Houston, 1999). Expanded Summary, VI, pp. 140-143, authors Srinivasa V. Chakravarthy, Raghu K. Chunduru, Alberto G. Mezzatesta, and Otto Fanini use a trained radial base function neural network to identify induction well record bed boundaries. The network is trained using the logarithmic derivative of both measured and synthetic log data. As a result, actual log data to be processed by the trained neural network must also be first processed by taking the logarithmic derivative. The detected bed limits are then used in known inversion processes.

In the publication entitled "Artificial Neural Networks And High Speed Resistivity Modeling Software Speeds Reservoir Characterization", Jeff S. Arbogast and Mark H. Franklin, 'Petroleum Engineer International', pp. 57-61, the authors describe use of a trained neural network in real well records of various types. By correctly selecting available records for training, it is reported that it is possible to synthesize lost records or fill in bad data for other wells in the same field.

In US Patent No. 5,251,286, entitled "Method for Estimating Formation Permeability of Wireline Logs Using Neural Networks," inventors Jacky M. Wiener, Robert F. Moll, and John A. Rogers, disclose the use of a neural network to determine permeability. . The network is trained with resistivity, neutron porosity, volumetric density, interval transit time, and other measurements and real measured core permeability. It is then able to use the same wire record measurements from other wells in the same area to estimate formation permeability in wells from which cores were not actually taken and measured.

In US Patent No. 5,862,513 entitled "Systems and Methods for Forward Modeling of Well Logging Tool Responses" inventors Alberto G. Mezzatesta, Michael A. Jervis, David R. Beard, Kurt M. Strack, and Leonty A. Tabarovsky, expose use of a neural network to produce synthetic tool responses for a well logging tool. The neural network is trained to simulate the response of a particular recording tool to landform models. The trained network is intended for use as the front resolver in an iterative inversion process.

In US Patent No. 6,044,325 entitled "Conductivity Anisotropy Estimation Method for Inversion Processing of Measurements Made by a Transverse Electromagnetic Induction Logging Instrument", inventors Srinivasa V. Chakravarthy, Pravin Gupta, Raghu Chunduru, Berthold G. Kriegshauser, and Otto N. Fanini, teach a method of using a trained neural network to improve initial estimates of training parameters. The network is trained by synthesizing the tool's response first to ground formation models. Then initial estimates of the ground parameters are calculated from the synthesized responses. Initial estimates and known ground models are used to train a neural network. To use the trained network with actual data, actual tool signals are first processed to produce an initial estimate of ground parameters. These processed signals are then introduced to the trained neural network to produce improved parameter estimates.

While these references showed improvements in well log inversion using trained neural networks, none of them taught a method for direct inversion of log tool signals to produce a log of formation parameters. Direct inversion would be faster than prior art methods and would allow real time generation of well logs at the well site. It would also allow real-time processing of LWD or MWD logging tool signals. This would be very useful to the drilling engineer during the drilling process. For example, in inclined well drilling, well logs could be used when guiding the drilling system.

SUMMARY OF THE INVENTION The present invention provides an improved method for training a neural network to process register signals to produce representative records of a grounding parameter and an improved trained neural network. Synthetic or artificial models of ground formation parameters are generated to train a neural network. Models are selected to cover the entire operating range of a selected tool based on its sensitivity and resolution characteristics and based on a realistic range of forming parameters. In each model, parameter contrasts on layer interfaces are limited to realistic values that are within the operational range of the tool. Selected models include models that have minimum parameter values at the lower limit of the tool operating range and models that have maximum parameter values at the upper limit of the tool operating range. A front resolver is used to simulate tool response for models. Simulated responses and models are then used to train a neural network to produce models as outputs in response to simulated responses as inputs. Actual data collected by the logging tool can then be processed by the neural network to produce ground parameter records.

BRIEF DESCRIPTION OF DRAWINGS

For a more detailed understanding of the invention, reference is now made to the drawings, wherein: Figure 1 is an illustration of a prior art induction recording system;

Figure 2 is an illustration of a well cavity being drilled by an inclined bed formation according to typical drilling practices;

Figures 3A-3B depict resistivity data obtained from inclined bed formations using the induction probe of Figure 1;

Figure 4 is a flow chart illustrating training of a neural network;

Figure 5 is a depth formation resistivity plot for Oklahoma type earth formations;

Figure 6 is a depth formation resistivity plot for a chirp type earth formation;

Figure 7 is a performance demonstration of a neural network trained in zero slope record data inversion;

Figure 8 is a performance demonstration of a neural network trained in reversing 55 degree slope log data;

Figure 9 is an illustration of a two-dimensional test formation;

Figure 10 is an illustration of the performance of a neural network trained in processing two-dimensional formation record data of Figure 9;

Figure 11 is a typical neural network inversion fiuxogram of induction registers;

Figure 12 is an illustration of an advanced multilayer multi-outlet power supply network forming a subjected network joint;

Figure 13 is an illustration of averaging outputs of a network joint submitted across a diagonal;

Figure 14 is a flow chart illustrating arrangement induction register processing with a submitted network joint;

Figure 15 are inversion outputs providing a performance comparison of a submitted network joint and a single output network;

Figure 16 are inversion outputs for an inclined bed application providing a performance comparison of a submitted mesh joint and a single output mesh;

Figure 17 is a standard boot training path graph for a submitted network joint;

Figure 18 is a standard boot training path graph for a network joint submitted with the quick inspection method; and Figure 19 is a standard boot training path chart for a submitted network joint with an option other than the quick inspection method.

DETAILED DESCRIPTION OF THE INVENTION

As can be seen from the references cited above, attempts to use artificial neural networks, ANNs, in the various inversion processes for well log data have had limited success. Using large amounts of existing data to train from neural networks increases the cost and expense of generating trained ANN and has not improved speed or accuracy or made ANN widely applicable outside the area from which data was selected. Some developers have suggested that it is important to choose only those parts of the available data that are clearly accurate and not contaminated by poor drillhole conditions, poor tool response, or other sources of error. If an ANN is taught to reverse bad data in a good record, it is likely to reverse good data in a bad record.

The present inventors have found that a viable way to avoid using bad or inaccurate data to train an ANN is to use only synthetic data. In addition, training data selection should be based on the behavior of the selected tool, not ground formations in any particular area. The training set should cover the operating range of the tool, but interface contrasts should be limited to realistic parameter ranges. Using these simple rules to select only a few landform models and to grade models, trained ANNs were generated, which provided good inversion of log data without geographical area limitation.

The above cited references also illustrate that the structure of ANNs and methods for training ANNs are well known. The aforesaid U.S. Patent Nos. 5,251,286 and 5,862,513 are hereby incorporated by reference for all purposes and particularly for their disclosures regarding the structure and training of ANNs.

Referring to Figure 4, training of an ANN according to the present invention will be described. A first step, represented by box 30, is to produce various synthetic training models or target formations for the training process. Formations that have been selected for a preferred embodiment will be described below with reference to Figures 5 and 6. All selected formation models are introduced to a front solver to generate a synthetic conductivity record in box 32. The front solver is a computer program. which simulates the response of a selected recording tool to any selected grounding model. Modern induction tools have multiple transmitter to receiver spacings and a separate ANN can be trained for each or any combination of them. A separate front resolver is used for each transmitter to receiver spacing. In developing the present invention, the inventors based their work on the HRAI induction registration tool developed by the representative of this patent application. That tool is the subject of a paper entitled "The New High Resolution Array Induction Tool" by Randy Beste, T. Hagiwara, George King, Robert Strickland, and GA Merchant, presented at the 41st Annual Registration Symposium SPWLA, 4-7. June 2000. The log of the synthetic conductivity record generated in step 32 is fed into the ANN 34 input in the form of small data windows. The input window is selected to include multiple samples that correspond at least to the well cavity depth range that a selected tool transmitter-receiver set would be measuring at once in a high resistivity zone. The window should include multiple samples that correspond to a range of several times the distance from transmitter to receiver. Making the window wider generally improves the inversion scheme but also increases training time and cost.

In the preferred embodiment, ANN was an advanced multilayer perception neural network with fifty-one entrances. The input window included fifty-one samples representing tool signals spaced 15 cm apart. These samples therefore represented 7.6 m drillhole data. In this embodiment, the network had two hidden layers with eight neurons in the first and twenty in the second and had one exit. Good results were also achieved with networks having three hidden layers. For each input data window, ANN 34 generated an output value centered on the input window. The process continued to introduce fifty-one sample rolling windows. That is, a sample was taken from one end of the window and one was added to the other end. At the end of the process, an output value was generated for each 15 cm spacing in the drillhole.

In another assay, a window size of 101 samples was used, with ANN again producing an output value centered on the input window. In this test, the inlet sample spacing was 7.6 cm, so the inlets also represented 15.2 m borehole. The rolling input window process was again used to generate an output value for every 7.6 cm of drillhole. However, this requires an ANN with 101 entries and increases training time and expense. The results achieved were not sufficiently different to justify the extra cost. Various commercially available software can be used to build, train and test neural networks. For example, the 'Neural NetWork Toolbox' for MATLAB was used in development and testing of the present invention.

When the logarithm of the synthetic log data is introduced to ANN 34, it produces a representation of the ground formation models at its output 36. Since the input data was the logarithm of the synthetic tool conductivity signals, the output is the logarithm. of a representation of the conductivity of formation. The exponential output of ANN is therefore used as the representation of formation. Both output representation 36 and input models 30 are fed to a comparison step 38, which determines the difference between the two. The difference is fed back to ANN 34 to adjust ANN parameters or coefficients and the process is repeated. When the difference detected in comparison step 38 is below an acceptable error level, the process is stopped. When ANN has thus been trained with all selected training models, the final ANN coefficients are stored for future use. ANN with these coefficients is then ready for use in processing actual data produced by the selected recording tool. The process can be repeated for any selected tilt angle. The same selected forming models are used in step 30. The front solver is adjusted to simulate the selected inclination angle so that the synthetic register produced at 32 precisely simulates tool response at the selected inclination angle. As shown in Figures 3A and 3B, the tool response to the same earth formation can change significantly when the slope angle changes. When ANN is trained with this data, it can be used to process actual data from wells that have the selected slope angle. In practice, a separate ANN does not need to be trained for every degree of possible inclination. Instead, an ANN can be trained for every five to ten degrees of inclination. When actual data is to be processed, the ANN with the closest slope can be selected and used. If the slope of the actual data is, for example, midway between two of the slope angles selected for training ANNs, both can be used to process the actual data and the result can be extrapolated from both outputs.

Referring now to Figure 5, one of two basic ground models used in the preferred embodiment is shown. This formation model is referred to as an Oklahoma type formation because it is similar to the actual land formations that occur in Oklahoma. The particular shown in Figure 5 is however only a synthetic model selected to have certain characteristics. The formation is represented by the square wave graph 40 of resistivity against depth. Since each earthbed layer typically has uniform resistivity between its upper and lower interfaces with adjacent layers, the model has a square waveform. Models used in the preferred embodiment assume that the tool signal will be preprocessed to correct well cavity effect, which is a function of drilling mud conductivity and well cavity diameter. The model includes relatively thick layers such as layer 42. It also includes thin layers, such as 44, which are close to the minimum bed thickness that can be resolved by the selected recording tool. The model includes a variety of contrasts, which are changes in resistivity between adjacent layers. Contrasts range from below 10 to 1, for example 46, to above 100 to 1, for example 48. This contrast range has been chosen to be realistic in terms of which contrasts are most common in real earth formations. . There are cases where contrasts from 1000 to 1 or higher can be found, but these are considered unusual or extreme and are not considered realistic as that term is used in the present invention. When they occur, they are likely to extend above or below the limits of the tool operating range, and therefore will not be measured precisely anyway. In the unusual case, where such a 1000 to 1 contrast occurs within the tool operating range, ANN can interpret it as two closely spaced contrasts. None of the interface contrasts extend from the minimum to maximum tool sensitivity values, which for the tool type and transmitter-receiver spacings used in the test were from about 0.2 ohm-m to about 2000 ohm-m for a total range of about 10,000 to 1. Contrast changes occur across low resistivity ranges, for example at 50, and across high ranges, for example at 46. Various versions of the model in Figure 5 are used when training the ANN. Two blush the extreme upper and lower resistivity ranges. In Figure 5, the resistivity at 42 is the maximum for the entire model. At least one version was selected by placing this maximum near the maximum sensitivity of the selected tool, which in this embodiment was about 2000 ohm-m. Similarly, a low range version was selected by placing the lowest resistivity at 52 near the lowest expected range of about 0.2 ohm-m. One or more intermediate range versions as shown in Figure 5 are also used. The inventors thought that ANNs do a good job of interpolating between tracks they were trained on, but they don't do a good job of extrapolating beyond those tracks. Including maximum and minimum levels based on tool operating range and realistically occurring contrasts in nature, ANN receives all the training it needs to interpolate any reasonable signal.

Tool operating range is limited by several factors. In very high resistivity formations, the induced current is too small and produces very small signals in the receiving coils. Receiving electronics have some inherent electrical noise that limits the system's ability to analyze signals received above some resistivity level. At very low resistivity, the skin effect becomes strong, causing the response to become nonlinear as well as preventing actual formation resistivity measurement. The response of interface registration tools is affected by more than just the contrast value to that interface. It is also affected by the range at which contrast occurs. That is, your response to a 10 to 1 contrast between 100 ohm-m and 1000 ohm-m is not the same as it is to a 10 to 1 contrast between 10 ohm-m and 100 ohm-m. Response is affected by thickness relative to adjacent layers equally. That is, the response to a given contrast between two thick layers is different from the response to the same contrast between two thin layers or between a thin layer and a thick layer. The Oklahoma type was selected because it provides examples across a range of these possible interface conditions.

Figure 6 illustrates a twitter formation that was the other type of model used when training the ANN. This model is again represented by a square waveform for the same reasons as apply to the Oklahoma type model. The main difference is that for any single version, the resistivity contrast for each interface is the same. The model contains a series of layers, 62, 64, etc., of varying thicknesses, starting with the thickest at 64 and ending with the thinnest at 66. The rules for selecting parameter values and limits for the twitter model are: essentially the same as for the Oklahoma model. In Figure 6, each contrast is illustrated as having a 100 to 1 contrast. Other versions of this model having a 10 to 1 contrast were also used to train ANN. Two extreme track versions have been selected. One had a maximum resistivity value of about 2000 ohm-m. The other had a minimum resistivity value of about 0.2 ohms. These values were again selected to cover the limits of the tool operating range while not exceeding realistic values of contrasts found in land formations.

In Figures 5 and 6, the synthetic recording tool signals 54 and 68, respectively, are also illustrated. These are the signals generated in step 32 of Figure 4. Signals 54 and 68 are the type of signals actually produced by an induction logging tool used to measure resistivity. The inversion process is the process of converting these signals to square wave signals 40 and 60 which more accurately represent the actual resistivity profile of formations through which the drillhole is drilled. As discussed with reference to Figure 4, signals 54 and 68 are the signals fed to the ANN 34 input during the training process.

Figure 6 also provides homogeneous formation information for ANN. A homogeneous formation is thick enough that the tool measures only the formation at a particular point in the well cavity. That is, the field of the transmitter coil actually or effectively passes only through an essentially uniform resistivity formation. In Figure 6, this is simulated in the depth ranges from -3 m to +3 m and from 42 m to 48 m. The model has a resistivity of 10 ohm-m in these ranges. But the tool signal 68 indicates a resistivity of about 11 ohm-m. The difference is caused by the skin effect. In a previous test of the present invention, homogeneous models separated at various resistivity levels were used in the training set as a way of teaching ANN the skin effect at various resistivity levels. This was found to be nonessential, probably because ANN learns a lot of information regarding homogeneous formations of model part types of Figure 6. ANN has been trained as discussed above for various inclination angles. It was then tested by introducing other registration signals, both synthetic and real. As was done during the training process, tool conductivity signal logarithms were introduced to the trained ANN, and the exponential of the ANN output was taken as the formation conductivity representation. Good inversion results were achieved essentially in all cases, including synthetic data representing formations not in the training set as well as actual record data from different geographical areas. Figures 7 and 8 are representative of the results achieved. In both Figures 7 and 8, the same synthetic formation profile 70 was used. In Figure 7, register signal 72 was for a zero slope case. In Figure 8, the registration data 74 was for a 55 ° slope. The difference between tool responses is apparent. However, in both cases, the ANN inversion, 76 and 78, of the tool signal closely matches the forming model.

A more complex synthetic test formation is illustrated in Figure 9. This is a two-dimensional model in which true resistivity is shown by the solid line graph 80. In a two-dimensional case, parts of the formation near the well cavity were invaded by drilling that changes the resistivity in the invaded zone. In the dotted line graph 82, part of the formation was invaded by drilling mud at a depth of 0.76 m from the borehole lowering the resistivity from 60 ohm-m to 10 ohm-m. In the dotted line graph 84, part of the formation was invaded by drilling mud at a depth of 1m from the well cavity lowering the resistivity from 50 ohm-m to about 15 ohm-m. This model was used to produce synthetic records for shallow, medium, and transponder spacing. The synthetic records were then processed by trained ANNs as published above, one ANN for each recipient. The results of this test are shown in Figure 10. The shallow, medium, and bottom results are shown by curves 86, 88, and 90. The shallow curve 86 is close to the resistivity of the invaded zones, because most of the signal measured by the shallow receiver originates. in the region near the drillhole. The deep curve 90 is close to the resistivity of the uninvaded formation, because much of its signal measurements go deeper into the formation. The average depth curve 88 is between as expected. Known methods of further processing curves 86, 88 and 90 may provide a good estimate of the depth of the invaded zone. This test illustrates that a trained ANN as specified herein is suitable for processing two-dimensional formation recording tool signals, although no two-dimensional data has been used in the training model data set. However, if desired, ANN can be trained using two-dimensional models. Front loggers for logging tools can produce the synthetic answer to the two-dimensional models that are needed for the training process. Otherwise, the training process is the same as for one-dimensional models. In some cases, ANNs trained with two-dimensional and three-dimensional models may give better results.

ANN outputs, such as those shown in Figures 7, 8 and 10, are not perfect representations of the desired formations. In an effort to reduce errors, several ANNs were trained with the same training set to produce several different sets of coefficients. This can be done by training ANNs having the same structure with identical data sets, but with different initial conditions for network coefficients. Alternatively, the ANN structure may be modified slightly prior to training. In either case, the final coefficients are different and the inherent errors are different. Multiple trained ANNs were used to process the same test data and the results were combined. The errors inherent in the outputs were different and therefore canceled out to some extent. The combined outputs provided a more accurate representation of the actual forming parameters. This multiple ANN process also illustrates that there is no better ANN structure or set of initial conditions. Many combinations of structure and initial conditions will result in coefficients that provide acceptable inversion processing.

While particular synthetic formation models were used in the embodiment set forth herein, it is apparent that other models could be used with similar success if the basic selection rules were followed. The training set should include synthetic models that have upper and lower parameter values that span the upper and lower operating range limits of the selected recording tool. Multiple versions of each model should be included with at least one having a higher parameter value near the upper operating range limit of the tool and at least one having a lower parameter value near the lower operating range limit of the registration tool. Parameter contrasts in bed layer interfaces should vary across a range that is realistic in terms of what usually happens in actual ground formations, which for the preferred embodiment was about 10 to 1 to about 100 to 1. Models should include bed layer thicknesses ranging from a maximum approximately corresponding to the area measured by the tool to the minimum thickness that the selected tool can analyze. These conditions can be met with a relatively small set of synthetic models, so that training time is reasonable. Using only synthetic models, all the "rules" that ANN "learns" during the training process are accurate, not contaminated by measurement or other errors. As a result, ANNs trained with these types of training sets are able to perform direct inversion of register signals on the desired formation parameter records for data from essentially any area.

In developing the present invention, ANNs were initially trained with both phase and quadrature signals of the recording tool and good results were achieved. However, using both signals slowed the development project because it doubles the inputs to ANN during the training process and therefore increases the time and expense involved. Generally, the quadrature signal for actual recording tool signals is noisier than the phase signal. It was decided to use only the in-phase component to simplify the development and testing process and to avoid the noisiest part of the actual data. The results indicate that this was a good choice for induction tools such as HRAI. There may be some unusual or extreme training for which you may need to use both components, and therefore you will need to train an ANN with both. In LWD or MWD processes, it is customary to use both phase and quadrature signals to evaluate phase and attenuation, and ANN should be trained for both signals to do this.

As noted above, separate ANNs can be trained for each transmitter-receiver set in a registration tool. In the HRAI tool there are six different spacings and ten different receiver arrangements. In the preferred embodiment, ANNs have been generated for several different spacings as illustrated in Figure 10. It is also possible, and often desirable, to use the signals from two or more of the spacings to produce a record. This is often done to improve the quality of inversion. By combining signals from two or more receivers, the uncorrelated part of the noise tends to cancel out while the desired signals reinforce. An ANN can be trained for this purpose. That is, the front resolver outputs for two or more transponder spacings can be used to train an ANN, which can then be used to process corresponding actual signals to generate a record of the desired parameter. The present invention was developed and demonstrated using HRAI induction recording tool conductivity data. The invention is equally applicable to other data produced by induction logging tools such as slope, incidence and anisotropy measurements. It is also applicable to data from other types of recording tools, such as electrical, acoustic, magnetic, gravity, and nuclear tools (eg, neutron radius or gamma). It is applicable to tools carried on drill chains as well as those suspended by wire lines. Submitted Joint Grid Realization The common practice for ANN inversion of array induction registers is to use a sequence of apparent resistivity measurements as input to retrieve true forming resistivity at the center point of the input window. We have demonstrated that this single multiple input / output architecture works well for multiple short transceiver spacing sub-arrangements. For deep spacing subarrangements, however, the predicted variance with a single ANN is relatively high. This is due to the fact that background arrangement measurements have less high frequency content, greater shoulder bed effect, and increased nonlinearity. Typically, regions surrounding bed boundaries in formation become more difficult to recover precisely. To improve the accuracy of ANN inversion models for deep-spaced subarrangements, we expose a new neural network architecture, called the submitted network joint, along with a "quick inspection" method to significantly reduce the computational load of train the proposed network board.

Figure 11 shows a typical flow diagram of the use of ANN inversion algorithm in the processing of registration data. In this flowchart, a single multiple input / output network is used to search for the transformed correlation between window apparent sensitivities and true referential forming resistivity, where the input window contains the same measurement points above and below the reference center. at true vertical depth with the training error being evaluated at a single position, the position of the output resistivity that corresponds to the position of the center of the input window. Being a more difficult problem, due to its lower frequency content and increased nonlinearity, the inversion of deeper subarray measurements is less accurate. Uncertainty in modeling input / output relationship using fixed window structure is one of the main reasons that cause variance in the interpretation of log data. To solve this complex problem, current technology allows us to combine multiple networks together to form a joint. Board members could be selected from many different networks with various frame / initial weights and trained with different data sets. The joint output can be taken as the average output of the various single networks that include the joint. The problem with this approach in interpreting well log data is processing time. The total number of parameters in such a joint is the sum of the parameters of each individual network, and dealing with networks with different architectures requires more matrix manipulations that take more time in serial processing, and need more processors in parallel processing. Also, training and keeping track of many different networks will increase the computational and administrative burden.

In this embodiment, we present a new method for overcoming the limitations discussed above, and show how to improve the interpretation of log data. This embodiment provides: a new apparent true resistivity / resistivity mapping relationship; a submitted network joint to establish the assumed relationship and reduce the output uncertainty; and, a "quick inspection" training method to efficiently train the presented network joint. The method proposed in this embodiment can be applied to improve the processing of any record measurement with problems similar to those associated with the HRAI tool deep subsets. The proposed method should improve the processing of log data in each situation, although being relatively longer than the standard single-output procedure, we recommend the new method for relatively more difficult problems, where using the standard method leads to results. low quality. More generally, this embodiment is applicable to an uncertain non-causal system for which the output y (n0) depends on the input samples x (n), for n extended in both directions of n0. Many geophysical applications fall into this category.

In this embodiment, we have established a new input / output mapping relationship for background spacing arrangements when processing log data. For the given network inputs, the network output, which is the recovered resistivity of formation, is not necessarily mapped to the referential center only. It could be mapped to any point in the central neighborhood. By predicting resistivity at various positions, which form an exit window, and by evaluating training error across the length of this exit window, not only the output value error but also the output slope error are penalized, leading to less variation in prediction. In addition, the new procedure generates as many predictions at each position as the number of points in the output window. After averaging these outputs, the variance is further reduced. This relationship can be implemented with an advanced multilayer multi-output network, which is shown in Figure 12. Note that each output has shared network connections (weights) in the hidden layer. Shared connections put restrictions on the output set. This network architecture allows each output node to produce its interpretation sequence based on the slightly different mapping assumption. Figure 13 shows such sequences in an example of five outputs, with each sequence being one prediction shifted from the other. Models that simultaneously extract the forming resistivities in the central vicinity constitute a submitted grid joint (CNC), and outputs that refer to the same true vertical depth (TVD) index may exhibit observed variation, indicating uncertainty in interpretation of registration data. To reduce the uncertainty of interpretation, the CNC output can be computed by averaging its member outputs across the diagonal along which the outputs have the same TVD index (see Figure 13). The problem associated with CNC is the computational load when training the multi-input / multi-output neural network. Using an 11-member network joint for example, the total number of parameters can be up to 4,000, and the error terms involved in the optimization algorithm can exceed 500,000 with a moderate training set. A single training attempt for this problem using traditional method will cost more than one week with current computer capacity. In this embodiment, a "quick inspection" method has been developed to reduce the computational load. This method initializes network parameters in different ways when a new subarrangement model is ready for training. The main options of this "quick inspection" method are: Initialize a new subarray model with the net weights of the adjacent trained subarray.

Initialize a new sub-arrangement model of certain frequency with the weights resulting from the same different frequency sub-arrangement.

Initializing a new subarrangement model by matching your previous net weights with newly added near zero weights by increasing the number of input neurons, or hidden neurons, or output neurons is needed. The idea of using "quick inspection" approximation is based on the fact that some kind of similarity in tool response exists between adjacent subarrangements, or even subarrangements, but different excitation frequencies. Therefore, training started by option 1 and 2 somehow acts as an I / O mapping rescheduling, which typically creates a steeper gradient in error reduction during iterative learning. The approach described in option 3 can make the training start accuracy of the new model as close as the previous model, which will significantly improve training efficiency. The new approach has the following advantages: The CNC provides a promising method for reducing uncertainty caused by tool limitation, and environmental effects in interpreting registration data. Reduction in error can be seen as arising from reduced variance due to averaging across many solutions. The CNC can be manipulated using a single advanced multilayer neural network that is cost effective with competitive processing speed.

Because the CNC generates shared weights and averaged output, it usually produces uniform prediction. No other regularization technique is needed in this approach, which allows the joint network to be trained using the rapid training algorithm with less worry about overadaptation. The "quick inspection" method can reduce the computational load with the traditional boot method by up to 60 percent.

Processing array induction registers with the submitted network joint is illustrated in Figure 14. In this flowchart, the log data (normally apparent resistivity) of a certain subarray is formatted to fill the input window. Then, a preprocessing transformation to the input data is applied before it is presented to the CNC network. The CNC network acts as a nonlinear filter, and is trained with simulation data that is described in our previous publication. The CNC output feeds through a postprocessing procedure before providing it to the customer.

For the CNC network, the input window usually covers measurement information from 7.5m to 15m with respect to TVD. An 11 knot outlet (corresponding to 11 joint members) is preferred with a 7.5 cm gap in the middle, which constitutes a 0.75 m exit window. For high slope bed application, the similarly long exit window at true vertical distance is preferred. This window size provides a good compromise between prediction bias and variance, and allows for reasonable computational load during training. The CNC output can either be average or weighted sum through member outputs having the same TVD index.

In this section, we first illustrate the test results when processing simulated registers, for which true forming resistivities are known, with CNC and single output networks respectively. So, we demonstrate that the "quick inspection" method actually significantly reduces the computational load when training the CNC network.

Figure 15 shows a performance comparison of the CNC network and the standard single output network for subarray 1, using simulated 8 kHz frequency training data from zero slope geometry. The training profile used here is a test file that is excluded from training standards. It is not difficult to identify by visual inspection that the CNC network produces the prediction that generates better than the single output network. The other example is given in Figure 16 for sloped bed application, tested with subarray 1 models and 8 kHz data. The test file has the form Oklahoma formation, which is the standard benchmark for benchmarking. The same conclusion can be drawn from this test that the measurement interpretation using CNC lattice is less excursive within beds and more consistent with true forming resistivity.

Figure 17 illustrates a trajectory of the training error against the number of iterations for a CNC network using the subarray 2 training examples, 8 kHz excitation frequency, and zero tilt angle. The chosen network 101-20-40-11 is initialized with standard method and trained with graduated conjugate gradient algorithm (SCG). It takes about 2,000 iterations to reduce the mean square error (mse) to 0.006, and still needs more iterations to achieve proper training accuracy.

In comparison, Figure 18 shows the error reduction rate of the "quick inspection" method. The CNC network with the same architecture is trained with the same algorithm using the same examples. The difference here is the mode for initializing the network parameters. The initial weights used in this training attempt are the resulting CNC net weights for the same subarray but different excitation frequency (32 kHz). We can see that while the starting training error in Figure 18 is close to Figure 17, it only takes 200 iterations in Figure 18 to achieve much better training accuracy as illustrated in Figure 17.

Figure 19 presents another example using an option other than the "quick inspection" method. The CNC network is trained for subarray 1 under 32 kHz excitation frequency and zero tilt angle, but initialized with weights resulting from subarray CNC network 2. Training takes 400 iterations to achieve displayed accuracy, which typically needs more than 5,000 iterations to train using the default boot method. It is apparent that various changes may be made to the apparatus and methods set forth herein, without departing from the scope of the invention as defined by the appended claims.

Claims (10)

1. Apparatus for converting the output signals of a recording tool to a register representing a parameter of ground formations surrounding a drillhole, comprising: an artificial neural network trained with a set of synthetic grounding models selected for cover the operating range of a selected recording tool based on recording tool sensitivity and resolution limits and based on realistic ranges of forming parameters, where the output signals of the recording tool are a series of samples each representing the signal at a depth point in the drillhole, and where the neural network has a plurality of inputs that receive those samples from a depth range in the drillhole, characterized in that the artificial neural network has a plurality outputs, each producing an output signal representing a different depth point in the hole and a means for combining neural network outputs according to depth points to produce an average record of a forming parameter. Apparatus according to claim 1, characterized in that the artificial neural network trained with a set of synthetic grounding models comprises: a) a plurality of twitter models having continuously increasing layer thicknesses, each model of having parameter contrasts at layer interfaces limited to realistic contrasts found on actual ground formations, at least one model with an upper parameter limit near the upper limit of the selected tool's operating range, and at least one model with a limit of lower parameter near the lower limit of the operating range of the selected tool, and b) a plurality of Oklahoma-type models with parameter contrasts on layered interfaces limited to realistic contrasts found on actual ground formations, at least one model having an upper parameter limit. near the upper limit of the operating range of the selected tool, and eg lo minus a model having a lower parameter limit near the lower limit of the selected tool's operating range. Apparatus according to claim 2, characterized in that the recording tool is an induction recording tool having a maximum sensitivity to minimum sensitivity ratio of about 10,000 to 1 and the chirp models include at least one. model with parameter contrasts at layer interfaces of about 10 to 1 and at least one model with parameter contrasts at layer interfaces of about 100 to 1. Apparatus according to claim 2, characterized in that the recording tool is an induction recording tool having a maximum sensitivity to minimum sensitivity ratio of about 10,000 to 1 and Oklahoma type models have parameter contrasts. on layer interfaces from about 10 to 1 to about 100 to 1. A method for converting the output signals of a recording tool into a register representing a parameter of ground formations surrounding a drillhole, comprising: creating a set of synthetic grounding models, comprising: a) a plurality of chirp models having continually increasing layer thicknesses, each chirp model having parameter contrasts on layer interfaces limited to realistic contrasts found on real earth formations, at least one model with an upper parameter boundary near the upper boundary of the range. tool operation, and at least one model with a lower parameter limit near the lower limit of the selected tool operating range, and b) a plurality of Oklahoma-type models with parameter contrasts on layer interfaces limited to realistic contrasts found in formations. at least one model having a parameter limit upper near the upper limit of the operating range of the selected tool, and at least one model having a lower parameter limit near the lower limit of the operating range of the selected tool; generate synthetic responses of the selected tool for each of the artificial formation models; use synthetic responses and training models to train an artificial neural network to generate representations of training models in response to synthetic responses; and processing actual register signals from the selected tool with the neural network trained to produce a ground parameter register, characterized in that in processing the artificial neural network has a plurality of outputs, each producing an output signal representing a point. of different depth at the borehole, with the method further comprising: combining those neural network outputs according to depth points to produce an average record of a forming parameter. Method according to claim 5, characterized in that the recording tool is an induction recording tool having a maximum sensitivity to minimum sensitivity ratio of about 10,000 to 1 and the chirp models include at least one. model with parameter contrasts at layer interfaces of about 10 to 1 and at least one model with parameter contrasts at layer interfaces of about 100 to 1. Method according to claim 5, characterized in that the recording tool is an induction recording tool having a maximum sensitivity to minimum sensitivity ratio of about 10,000 to 1 and Oklahoma type models have parameter contrasts. on layer interfaces from about 10 to 1 to about 100 to 1. A method according to claim 5, further comprising: using synthetic responses and training models to train one or more additional artificial neural networks to generate representations of training models in response to synthetic responses; process the actual register signals of the selected tool with the additional trained neural network or networks to produce an additional ground parameter register or registers; and combining the ground parameter registers to produce an average ground parameter register. Method according to claim 5, characterized in that the selected recording tool is an induction recording tool with more than one transmitter-receiver pair and the synthetic responses of the selected tool include responses of more than one pair. transceiver. Method according to claim 5, characterized in that the selected recording tool is an induction recording tool having both phase and quadrature output signals and the synthetic responses of the selected tool include both signals.