No description available.

<p>Neural networks are mathematical formulations that can be "trained" to perform certain functions. One particular application of these networks of interest in this thesis is to "model" a physical system using only input-output information. The physical system and the neural network are subjected to the same inputs. The neural network is then trained to produce an output which is the same as the physical system for any input. This neural network model so created is essentially a "blackbox" representation of the physical system. This approach has been used at the University of Saskatchewan to model a load sensing pump (a component which is used to create a constant flow rate independent of variations in pressure downstream of the pump). These studies have shown the versatility of neural networks for modeling dynamic and non-linear systems; however, these studies also indicated challenges associated with the morphology of neural networks and the algorithms to train them. These challenges were the motivation for this particular research.</p> <p>Within the Fluid Power Research group at the University of Saskatchewan, a "global" objective of research in the area of load sensing pumps has been to apply dynamic neural networks (DNN) in the modeling of loads sensing systems.. To fulfill the global objective, recurrent generalized neural network (RGNN) morphology along with a non-gradient based training approach called the complex algorithm (CA) were chosen to train a load sensing pump neural network model. However, preliminary studies indicated that the combination of recurrent generalized neural networks and complex training proved ineffective for even second order single-input single-output (SISO) systems when the initial synaptic weights of the neural network were chosen at random.</p> <p>Because of initial findings the focus of this research and its objectives shifted towards understanding the capabilities and limitations of recurrent generalized neural networks and non-gradient training (specifically the complex algorithm). To do so a second-order transfer function was considered from which an approximate recurrent generalized neural network representation was obtained. The network was tested under a variety of initial weight intervals and the number of weights being optimized. A definite trend was noted in that as the initial values of the synaptic weights were set closer to the "exact" values calculated for the system, the robustness of the network and the chance of finding an acceptable solution increased. Two types of training signals were used in the study; step response and frequency based training. It was found that when step response and frequency based training were compared, step response training was shown to produce a more generalized network.</p> <p>Another objective of this study was to compare the use of the CA to a proven non-gradient training method; the method chosen was genetic algorithm (GA) training. For the purposes of the studies conducted two modifications were done to the GA found in the literature. The most significant change was the assurance that the error would never increase during the training of RGNNs using the GA. This led to a collapse of the population around a specific point and limited its ability to obtain an accurate RGNN.</p> <p>The results of the research performed produced four conclusions. First, the robustness of training RGNNs using the CA is dependent upon the initial population of weights. Second, when using GAs a specific algorithm must be chosen which will allow the calculation of new population weights to move freely but at the same time ensure a stable output from the RGNN. Third, when the GA used was compared to the CA, the CA produced more generalized RGNNs. And the fourth is based upon the results of training RGNNs using the CA and GA when step response and frequency based training data sets were used, networks trained using step response are more generalized in the majority of cases.</p>

<p>Neural networks are mathematical formulations that can be "trained" to perform certain functions. One particular application of these networks of interest in this thesis is to "model" a physical system using only input-output information. The physical system and the neural network are subjected to the same inputs. The neural network is then trained to produce an output which is the same as the physical system for any input. This neural network model so created is essentially a "blackbox" representation of the physical system. This approach has been used at the University of Saskatchewan to model a load sensing pump (a component which is used to create a constant flow rate independent of variations in pressure downstream of the pump). These studies have shown the versatility of neural networks for modeling dynamic and non-linear systems; however, these studies also indicated challenges associated with the morphology of neural networks and the algorithms to train them. These challenges were the motivation for this particular research.</p> <p>Within the Fluid Power Research group at the University of Saskatchewan, a "global" objective of research in the area of load sensing pumps has been to apply dynamic neural networks (DNN) in the modeling of loads sensing systems.. To fulfill the global objective, recurrent generalized neural network (RGNN) morphology along with a non-gradient based training approach called the complex algorithm (CA) were chosen to train a load sensing pump neural network model. However, preliminary studies indicated that the combination of recurrent generalized neural networks and complex training proved ineffective for even second order single-input single-output (SISO) systems when the initial synaptic weights of the neural network were chosen at random.</p> <p>Because of initial findings the focus of this research and its objectives shifted towards understanding the capabilities and limitations of recurrent generalized neural networks and non-gradient training (specifically the complex algorithm). To do so a second-order transfer function was considered from which an approximate recurrent generalized neural network representation was obtained. The network was tested under a variety of initial weight intervals and the number of weights being optimized. A definite trend was noted in that as the initial values of the synaptic weights were set closer to the "exact" values calculated for the system, the robustness of the network and the chance of finding an acceptable solution increased. Two types of training signals were used in the study; step response and frequency based training. It was found that when step response and frequency based training were compared, step response training was shown to produce a more generalized network.</p> <p>Another objective of this study was to compare the use of the CA to a proven non-gradient training method; the method chosen was genetic algorithm (GA) training. For the purposes of the studies conducted two modifications were done to the GA found in the literature. The most significant change was the assurance that the error would never increase during the training of RGNNs using the GA. This led to a collapse of the population around a specific point and limited its ability to obtain an accurate RGNN.</p> <p>The results of the research performed produced four conclusions. First, the robustness of training RGNNs using the CA is dependent upon the initial population of weights. Second, when using GAs a specific algorithm must be chosen which will allow the calculation of new population weights to move freely but at the same time ensure a stable output from the RGNN. Third, when the GA used was compared to the CA, the CA produced more generalized RGNNs. And the fourth is based upon the results of training RGNNs using the CA and GA when step response and frequency based training data sets were used, networks trained using step response are more generalized in the majority of cases.</p>

The purpose of this article is try to find the efficient factor that affect corporate's financial structure.

In our research, we proposed a novel invariant in 2-D image contour recognition based on Hopfield-Tank neural network. At first, we searched the feature points, the position of feature points where are included high curvature and corner on the contour. We used polygonal approximation to describe the image contour. There have two patterns we set, one is model pattern another is test pattern. The Hopfield-Tank network was employed to perform feature matching. In our results show that we can overcome the test pattern which consists of translation, rotation, scaling transformation and no matter single or occlusion pattern.

This work presents a new approach for detecting and isolating faults in nonlinear processes using independent neural network models. In this approach, an independent neural network is used to model the proton exchange membrane fuel cell nonlinear systems using a multi-input multi-output structure. This research proposed the use of radial basis function network and multilayer perceptron network to perform fault detection. After training, the neural network models can give accurate prediction of the system outputs, based on the system inputs. Using the residual generation concept developed in the model-based diagnosis, the difference between the actual and estimated outputs are used as residuals to detect faults. When the magnitude of these residuals exceed a predefined threshold, it is likely that the system is faulty. In order to isolate faults in the system, a second neural network is used to examine features in the residual. A specific feature would correspond to a specific fault. Based on features extracted and classification principles, the second neural network can isolate faults reliably and correctly. The developed method is applied to a benchmark simulation model of the proton exchange membrane fuel cell stacks developed at Michigan University. One component fault, one actuator fault and three sensor faults were simulated on the benchmark model. The simulation results show that the developed approach is able to detect and isolate the faults to a fault size of ±10% of nominal values. These results are promising and indicate the potential of the method to be applied to the real world of fuel cell stacks for dynamic monitoring and reliable operations.

Master of Science / Department of Geology / Matthew W. Totten / The Meramec member of Mississippian age is a proficient oil and gas producing formation within the midcontinent region of the United States. It is produced in Kansas, Oklahoma, and Texas. In Kansas, 12% of the state’s petroleum production comes from Mississippian-aged rocks. Bindley Field, located in central west Kansas, has produced 3,669,283 barrels of oil from one facies within the M2 interval of the Meramec formation. This facies is a grain-supported echinoderm/bryozoan dolostone, of variable thickness. Its sporadic occurrence in the subsurface has made exploring Bindley Field and the surrounding area difficult. The challenge in finding oil in this area is in locating a producible zone of this productive facies. Previously, Bindley Field has been the subject of detailed reservoir characterization studies (Ebanks et al., 1977; Johnson, 1990; Johnson, 1994). These studies helped to contribute to a better understanding of Meramecian stratigraphy in Kansas. The Meramec was divided into four major depositional sequences, with some of those sequences nonexistent in the subsurface, due to aerial exposure and erosion post-deposition. The Meramecian units were further separated into parasequence-scale chronostratigraphic units based on marine flooding events. The primary producing interval in Bindley Field is the Meramec 2 interval which consists of seven lithotypes, and is recognized to have six, meter-scale depositional cycles (Johnson, 1990). As production from this interval increased, more information became available about controls on reservoir quality. There are still areas, however, where core data do not exist, and predicting the productive facies remains challenging. The aim of this study is to create a workflow for evaluating the subsurface using regional core and log data from Bindley Field to create a model of the subsurface distribution of the reservoir facies, which could be extended to data poor areas. Geophysical logs (neutron, gamma ray, guard) along with an artificial neural network (ANN), was used to create an accurate prediction of producing intervals within the subsurface. Values are derived from wire line log data and used to develop the ANN definition of facies distribution within Bindley Field. The ANN model was examined for accuracy and precision using core description and well cuttings from wells within Bindley Field and the surrounding area. Correlations were found between the subsurface geometry of the study area, and the production of oil and gas within the study area. An ANN model with an accuracy of 72% was achieved and applied to wells surrounding the Bindley Field, where reservoir intervals have not been as extensively studied. A total of 87 wells in Bindley Field and the surrounding 50 square mile area where applied to the ANN model. The model predicted that the productive facies thickens gradually to the northwest of Bindley Field. Cross sections as well as an isopach map were created using the prediction data from the ANN. Finally, an analysis for the accuracy of the ANN and the predicted facies was created. The productive facies yielded an accuracy value of 77%.

Spiking neural networks (SNNs) are a more biologically realistic model of the brain than traditional analog neural networks and therefore should be better for modelling certain functions of the human brain. This thesis uses the concept of deriving an SNN from an accepted non-spiking neural network via analysis and modifications of the transmission function. We investigate this process to determine if and how the modifications can be made to minimize loss of information during the transition from non-spiking to spiking while retaining positive features and functionality of the non-spiking network. By comparing combinations of spiking neuron models and networks against each other, we determined that replacing the transmission function with a neural model that is similar to it allows for the easiest method to create a spiking neural network that works comparatively well. This similarity between transmission function and neuron model allows for easier parameter selection which is a key component in getting a functioning SNN. The parameters all play different roles, but for the most part, parameters that speed up spiking, such as large resistance values or small rheobases generally help the accuracy of the network. But the network is still incomplete for a spiking neural network since this conversion is often only performed after learning has been completed in analog form. The neuron model and subsequent network developed here are the initial steps in creating a bidirectional SNN that handles hetero-associative and auto-associative recall and can be switched easily between spiking and non-spiking with minimal to no loss of data. By tying everything to the transmission function, the non-spiking learning rule, which in our case uses the transmission function, and the neural model of the SNN, we are able to create a functioning SNN. Without this similarity, we find that creating SNN are much more complicated and require much more work in parameter optimization to achieve a functioning SNN.

To understand remotely sensed data, one must understand the relationship between radiative transfer models and their predictions of the interaction of solar radiation on geophysical media. If it can be established that these models are indeed accurate, some form of evaluation has to be performed on these models, for users to choose the model that suits their requirements. This thesis focuses on the implementation of a variable linear kernel model, its validation, and to study its application in the prediction of BRDF effects using two different neural networks-- the backpropogation and the radial basis function neural network and finally to draw conclusions on which neural network is best suited for this model. Based on these results the optimum number of kernels for this model is derived.

Ferrous objects in earth's magnetic field cause distortion in the surrounding ambient field. This distortion is a function of the object's material properties and geometry, and is known as the magnetic signature. As a precursor to first principle modeling of the phenomenon and a proof of concept, the goal of this research is to predict offboard magnetic signatures from on-board sensor data using a neural network. This allows magnetic signature analysis in applications where direct field measurements are inaccessible. Simulated magnetic environments are generated using MATLAB's Partial Differential Equation toolbox for a 2D geometry, specifically for a rectangular shell. The resulting data sets are used to train and validate the neural network, which is configured in two layers with ten neurons. Sensor data from within the shell is used as network inputs, and the off-board field values are used as targets. The neural network is trained using the Levenberg-Marquardt algorithm and the back propagation method by comparing the estimated off-board magnetic field intensity to the true value. This research also investigates sensitivity, scalability, and implementation issues of the neural network for signature estimation in a practical environment. / Electrical and Computer Engineering