Kohonen Neural Network Research Articles

Estimation of flow in ungauged catchments by coupling a hydrological model and neural networks: case study

The prediction of hydrological variables for ungauged basins is still a big challenge. Regionalization is the most widely used method to date, which relates parameters of watershed models to catchment characteristics. Relating catchment characteristics to watershed model parameters is too difficult for distributed hydrological models, due to the heterogeneous nature of catchments. A regional model was proposed by coupling a Kohonen neural network (KNN) and distributed Water Balance Simulation Model (WaSiM-ETH) to estimate flow in ungauged basin. KNN was used to delineate a hydrological homogeneous group based on predefined physical characteristics of catchments and WaSiM-ETH was applied to generate daily stream flow. Twenty-six subcatchments of the Blue Nile River basin, Ethiopia, were grouped into five hydrological homogenous groups, each with its own full set of optimized WaSiM-ETH parameters. In the regional model, the KNN assigned the ungauged catchment into one of the five hydrological homogenous groups. The whole set of optimized WaSiM parameters from the homogeneous group (which the ungauged river belongs to) were transferred to the ungauged river and WaSiM-ETH was used to compute the flow for this ungauged river. The regional model generally overestimated the low flow. In general, the results for validation subcatchments showed the regional model is satisfactory in transferring information from data-rich to data-poor catchments.

THE APPLICATION OF SELF-ORGANIZING NEURAL NETWORKS TO REMOTE SENSING IMAGERY CLASSIFICATION

In comparison with the Kohonen neural networks, the structure of the compete study networks is relatively simple because it does not consider neighboring neural units. In experiments, the authors adopted this simplified neural structure and improved its study algorithm by using max-min distance means. Experimental results show that the classification accuracy and efficiency of the improved compete study networks are remarkably raised in unsupervised classification of remote sensing imagery, and hence the technique of compete study networks has the practical application value.

A NEW ADAPTIVE CLASSIFICATION FUSION METHOD BASED ON DEMPSTER-SHAFER THEORY IN REMOTE SENSING IMAGE

In this paper, a new adaptive classification fusion method was proposed based on the Dempster-Shafer's theory of evidence and fuzzy Kohonen neural network in remote sensing image. The new method incorporates ideas from unsupervised neural network model and uses neighborhood information in the framework of the Dempster-Shafer theory of evidence. This approach mainly consists in considering each neighbor of a pattern to be classified as an item of evidence supporting certain hypotheses concerning the class membership of that pattern. This evidence is represented by basic probability assignment, with pooled utilization of the Dempster's rule of combination. Experiments with SPOT remote sensing image demonstrate the excellent performance of this classification scheme as compared with the existing neural network techniques.

A fuzzy/Bayesian approach for the time series change point detection problem

This paper addresses the change point detection problem in time series. A methodology based on the Metropolis-Hastings algorithm applied to time series modeled as a process with Beta distribution is discussed. In order to make this methodology useful in practice, a fuzzy cluster technique is applied to the initial time series at first, generating a new data set with Beta distribution. Bayesian procedures are considered for inference and the Metropolis-Hastings algorithm is used to sample from the posteriors. In the clustering process, a Kohonen neural network is used having as objective to find the best centers of the time series to be used in the fuzzyfication process. Finally, it will be presented a simulation results in the series of the electric energy consumption in Brazil, between January of 1976 and December of 2000, five months before the blackout occurred in 2001. Such result illustrates the efficiency of the proposed methodology for change point detection in time series.

Automated multi-level pathology identification techniques for abnormal retinal images using artificial neural networks

AimTo automatically classify abnormal retinal images from four different categories using artificial neural networks with a high degree of accuracy in minimal time to assist the ophthalmologist in subsequent treatment...

Artificial neural networks to evaluate the boron concentration decreasing profile in Blood-BPA samples of BNCT patients

For the prediction of decay concentration profiles of the p-boronophenylalanine (BPA) in blood during BNCT treatment, a method is suggested based on Kohonen neural networks. The results of a model trained with the concentration profiles from the literature are described. The prediction of the model was validated by the leave-one-out method. Its robustness shows that it is mostly independent on small variations. The ability to fit retrospective experimental data shows an uncertainty lower than the two compartment model used previously.

High-resolution passive seismic tomography for 3D velocity, Poisson’s ratioν, and P-wave qualityQPin the Delvina hydrocarbon field, southern Albania

We have studied using traveltimes of P- and S-waves and initial seismic-pulse rise-time measurements from natural microearthquakes to derive 3D P-wave velocity VPinformation (mostly structural) as well as P- and S-wave velocity VP/VSand P-wave quality factor QPinformation (mostly lithologic) in a known hydrocarbon field in southern Albania. During a 12-month monitoring period, 1860 microearthquakes were located at a 50-station seismic network and were used to obtain the above parameters. The data set included earthquakes with magnitudes ranging from –0.1 to 3.0 R (Richter scale) and focal depths typically occurring between 2 and 10 km. Kohonen neural networks were implemented to facilitate the lithological classification of the passive seismic tomography (PST) results. The obtained results, which agreed with data from nearby wells, helped delineate the structure of the reservoir. Two subregions of the investigated area, one corresponding to an oil field and one to a gas field, were correlated with the PST results. This experiment showed that PST is a powerful new geophysical technique for exploring regions that present seismic penetration problems, difficult topographies, and complicated geologies, such as thrust-belt regions. The method is economical and environmentally friendly, and it can be used to investigate very large regions for the optimal design of planned 2D or 3D conventional geophysical surveys.

Color reduction using a multi-stage Kohonen Self-Organizing Map with redundant features

Reducing the number of colors in an image while preserving its quality, is of importance in many applications such as image analysis and compression. It also decreases the memory and transmission bandwidth requirements. Moreover, classification of image colors is applicable in image segmentation, and object detection and separation, as well as producing pseudo-color images. In this paper, the Kohonen Self-Organizing Map Neural Network is employed to form an adaptive color reduction method. To enhance the performance of this method, we have used redundant features obtained by one-to-one functions from three main components of the color image (e.g. Red, Green and Blue channels). Exploiting these features will increase the color discrimination and details illustration ability of the network compared to the conventional approaches. This method leads to satisfactory results in image segmentation, especially in small object detection problems. It is also investigated that if the number of features in Kohonen network grows even by using non-deterministic one-to-one functions, the network revenue considerably improves. Moreover, we will study the effect of various adaptation algorithms in Kohonen network training stage. Again, using a multi-stage color reduction procedure which employs both Kohonen neural networks and conventional vector quantization schemes improves the performance. Several experimental results are represented to illustrate the characteristics of different approaches.

On the Use of Kohonen Neural Networks for Site Effects Assessment by Means of H/V Weak-Motion Spectral Ratio: Application in Rio-Antirrio (Greece)

Abstract The investigated area, located around the Rio-Antirrio Strait, Central Greece, has been the target of a seismic microzonation campaign. Seventy seismic stations have been deployed for a period of 4 months, recording in continuous mode. Despite the high level of urban noise, we compiled a data set of 95 earthquakes recorded at most of the 70 sites. By employing the attributes of self-organizing maps (SOMs), a quality-control and signal-improving method is proposed. A SOM (Kohonen, 1997) is a type of unsupervised neural network. The main property of SOMs utilized is that while the competitive learning algorithm on whom this method is based maps the input data on an n -dimensional grid of neurons, the topological relations (proximity of patterns in input data) are preserved in the output space. SOM is applied to the horizontal-to-vertical spectral ratios (HVSR) of every weak event analyzed for each station separately and allows a better evaluation of the stability of the HVSR.

A novel self-organizing map (SOM) neural network for discrete groups of data clustering

The self-organizing map (SOM) neural network, also called Kohonen neural network, is an effective tool for analysis of multidimensional data. This network can be used for cluster analysis while preserving data structure (topology) in such a way that similar inputs (data) remain close together in the output layer of the network. However, no algorithm that can automatically cluster discrete groups of data is presented, and our simulation results show that the classic SOM algorithm cannot cluster discrete data correctly. In this paper, we present a novel SOM-based algorithm that can automatically cluster discrete groups of data using an unsupervised method. This method consists of three phases: at the first phase, an algorithm called “second winner” is performed, in which neurons in the competitive layer of the network find their initial location in the network space. At the second phase, a method called “batch learning” is employed, and at the end of this phase, training of the SOM network is finished. And finally at the third phase, data clustering is completed by removing the wrong links between neurons. Three real world data sets and an example of synthetic data are utilized to illustrate the accurateness and effectiveness of the proposed approach.

Mapping patterns of pedestrian fatal accidents in Israel

This study intends to provide insight into pedestrian accidents by uncovering their patterns in order to design preventive measures and to allocate resources for identified problems. Kohonen neural networks are applied to a database of pedestrian fatal accidents occurred during the four-year period between 2003 and 2006. Results show the existence of five pedestrian accident patterns: (i) elderly pedestrians crossing on crosswalks mostly far from intersections in metropolitan areas; (ii) pedestrians crossing suddenly or from hidden places and colliding with two-wheel vehicles on urban road sections; (iii) male pedestrians crossing at night and being hit by four-wheel vehicles on rural road sections; (iv) young male pedestrians crossing at night wide road sections in both urban and rural areas; (v) children and teenagers crossing road sections in small rural communities. From the perspective of preventive measures, results suggest the necessity of designing education and information campaigns for road users as well as allocating resources for infrastructural interventions and law enforcement in order to address the identified major problems.

Classification of wavelet decomposed AE signals based on parameter-less self organised mapping

Composite materials are characterised by different types of failure mechanisms which are typically associated with matrix cracking, fibre-matrix debonding and fibre breakage. These three mechanisms result in a different AE signature, which can be often recognised. In certain cases it is necessary to classify and map the damage types so as to be able to evaluate the accumulated damage and remaining strength of the material. In this framework, neural networks are widely used for damage characterisation. The classical approach involves recording waveform features and tries to associate them with the underlying damage source. Nevertheless, very often, it is very hard to draw definite conclusions based on these features. In this study, we have used a new type of a neural network which is called parameter-less self organised mapping. It is based on Kohonen neural networks but it is not bound to the naturally subjective learning rate, neighbourhood function and their annealing with the training progress. Moreover, for the training of the NN and the subsequent classification, we have successfully used wavelet decomposed AE signals.

Recognition of the Arabic handwritten letters by the neural networks

In this paper, an Arabic character recognition system based on Artificial Neural Networks (ANN) is presented. In this system, each drawn Arabic character is represented by binary values that are used as input to a simple feature extraction system, whose output is fed to a Kohonen neural network that consists of two layers (first layer 99 neurons, second layer 28 neurons). Simulation results are provided and show that the proposed system always produces a lower Mean Squared Error (MSE) and higher success rates than the most of the other classifier method solutions, especially when the contaminating noise level is low.

Map Matching Using De-Noise Interpolation Kohonen Self-Organizing Maps

In this work, we have proposed a de-noise interpolation Kohonen Self-Organizing Maps(DNIKSOM) -based method for the Map matching(MM). It has been seen that there are some problems in the MM, such as large error range of the original position information, low match accuracy and so on. Therefore, in MM problem to achieve high accuracy, it is necessary to consider the topography of roads and the requirement for match accuracy lying within the original position information in the matching process. In the present study, Kohonen Self-Organizing Maps(KSOM) in the field of pattern recognition possesses good performance. Now to get more valuable position information, A kind of de-noise algorithm for Kohonen neural network is proposed to meet the case that neural network may not be trained sufficiently with consideration for the topography of roads. And a kind of Lagrange interpolation algorithm is also proposed to meet the requirements for matching accuracy. These processes make the amended position information closer to the true value. In this application to a city’s MM, we investigate DNIKSOM’s ,KSOM’s and Centroid localization algorithm’s location performance on a original position data set. Finally, the comparison of experimental results shows that DNIKSOM has better location performance than others.

Classification of microwatersheds based on morphological characteristics

Three classification techniques, namely, K-means Cluster Analysis (KCA), Fuzzy Cluster Analysis (FCA), and Kohonen Neural Networks (KNN) were employed to group 25 microwatersheds of Kherthal watershed, Rajasthan into homogeneous groups for formulating the basis for suitable conservation and management practices. Ten parameters, mainly, morphological, namely, drainage density ( D d), bifurcation ratio ( R b), stream frequency ( F u), length of overland flow ( L o), form factor ( R f), shape factor ( B s), elongation ratio ( R e), circulatory ratio ( R c), compactness coefficient ( C c) and texture ratio ( T) are used for the classification. Optimal number of groups is chosen, based on two cluster validation indices Davies–Bouldin and Dunn’s. Comparative analysis of various clustering techniques revealed that 13 microwatersheds out of 25 are commonly suggested by KCA, FCA and KNN i.e., 52%; 17 microwatersheds out of 25 i.e., 68% are commonly suggested by KCA and FCA whereas these are 16 out of 25 in FCA and KNN (64%) and 15 out of 25 in KNN and CA (60%). It is observed from KNN sensitivity analysis that effect of various number of epochs (1000, 3000, 5000) and learning rates (0.01, 0.1–0.9) on total squared error values is significant even though no fixed trend is observed. Sensitivity analysis studies revealed that microwatersheds have occupied all the groups even though their number in each group is different in case of further increase in the number of groups from 5 to 6, 7 and 8.

Kohonen Neural Network – Based Performance Improvements for Wafer Photolithography Process with CONWIP Control Strategy

Photolithography is usually the bottleneck process with the most expensive equipment in a semiconductor wafer fabrication system. To improve the performances of the photolithography area with dynamic combination rules, a method of Kohonen neural network (KNN)–based performance improvements is proposed. First, a dynamic scheduling framework based on a KNN model and scheduling rules is proposed. A KNN-based sample learning algorithm for improving the performances is presented. Finally, to demonstrate the validity and feasibility of the proposed method, data from a real wafer fabrication system are used to simulate the proposed method. Results of simulation experiments indicate that the proposed method can be used to improve a complex wafer photolithography performance.

Managing Internet routers congested links with a Kohonen-RED queue

The behaviour of the TCP AIMD algorithm is known to cause queue length oscillations when congestion occurs at a router output link. Indeed, due to these queueing variations, end-to-end applications experience large delay jitter. Many studies have proposed efficient active queue management (AQM) mechanisms in order to reduce queue oscillations and stabilize the queue length. These AQM attempt to improve the random early detection (RED) model. Unfortunately, these enhancements do not react in a similar manner for various network conditions and are strongly sensitive to their initial setting parameters. Although this paper proposes a solution to overcome the difficulties of configuring the RED parameters by using a Kohonen neural network model; another goal of this study is to investigate whether cognitive intelligence could be placed in the core network to solve such stability problem. In our context, we use results from the neural network area to demonstrate that our proposal, named Kohonen-RED (KRED), enables a stable queue length without complex parameters setting or passive measurements to obtain a correct configuration.

Selecting a reduced suite of diagnostic ratios calculated between petroleum biomarkers and polycyclic aromatic hydrocarbons to characterize a set of crude oils

A set of 34 crude oils was analysed by GC–MS (SIM mode) and a suite of 28 diagnostic ratios (DR) calculated. They involved 18 ratios between biomarker molecules (hopanes, steranes, diasteranes and triaromatic steroids) and 10 quotients between polycyclic aromatic hydrocarbons. Three unsupervised pattern recognition techniques (i.e., principal components analysis, heatmap hierarchical cluster analysis and Kohonen neural networks) were employed to evaluate the final dataset and, thus, ascertain whether the crude oils grouped as a function of their geographical origin. In addition, an objective variable selection procedure based on Procrustes Rotation was undertaken to select a reduced set of DR that comprised for most of the information in the original data without loosing relevant information. A reduced set of four DR (namely; TA21, D2/P2, D3/P3 and B(a)F/4-Mpy) demonstrated to be sufficient to characterize the crude oils and the groups they formed.

Electromagnetic Device Analysis Using a Meshless Approach Coupled to a Kohonen Network

This paper describes a novel adaptive solution system for electromagnetic field problems. The solution method is based on a meshless approach and thus the preprocessing effort is minimized. The positions of the nodes used for this solution are determined adaptively using a Kohonen neural network (NN). The objective of the proposed approach is to obtain an optimal placing of the nodes before adding any new degrees of freedom to reduce the error.

KOHONEN NEURAL NETWORK CLUSTERING FOR VOLTAGE CONTROL IN POWER SYSTEMS

Clustering a power system is very useful for the purpose of voltage stability control. However, the methods have developed usually have computational inefficiency. This paper presents a new cluster bus technique using Kohonen neural network for the purpose of forming bus clusters in power systems from the voltage stability viewpoint. This cluster formation will simplify voltage control in power system. With this proposed Kohonen algorithm, a large bus system will be partitioned into a small bus groups that have a coherence V, �, P and Q. The maximum number of area clusters will be formed need for voltage stability needed. The proposed technique was tested on IEEE 39 bus system by considering two contingency namely load increased and line outage by using voltage collapse analysis. This formation will be compared with the Learning Vector Quantization (LVQ) algorithm. The results showed the proposed technique produces four clusters on contingency load load increased and three clusters online outage contingency on IEEE 39 bus system as shown by the LVQ.