Kohonen's Learning Vector Quantization Research Articles

Hydroelectric power plant management relying on neural networks and expert system integration

The use of Neural Networks (NN) is a novel approach that can help in taking decisions when integrated in a more general system, in particular with expert systems. In this paper, an architecture for the management of hydroelectric power plants is introduced. This relies on monitoring a large number of signals, representing the technical parameters of the real plant. The general architecture is composed of an Expert System and two NN modules: Acoustic Prediction (NNAP) and Predictive Maintenance (NNPM). The NNAP is based on Kohonen Learning Vector Quantization (LVQ) Networks in order to distinguish the sounds emitted by electricity-generating machine groups. The NNPM uses an ART-MAP to identify different situations from the plant state variables, in order to prevent future malfunctions. In addition, a special process to generate a complete training set has been designed for the ART-MAP module. This process has been developed to deal with the absence of data about abnormal plant situations, and is based on neural nets trained with the backpropagation algorithm.

EMPIRICAL COMPARISON TO PREDICTIVE CLUSTERING BETWEEN A BIO-MIMETIC METHODOLOGY AND A NON CONNECTIONIST APPROACH

An empirical comparison is carried out on two real data sets between a bio-mimetic clusterwise algorithm, the Kohonen's Learning Vector Quantization, and a non-connectionist approach- to clusterwise regression, the fuzzy clustering approach Glimmix based on regression mixture model.

Encoded pattern classification using constructive learning algorithms based on learning vector quantization

A novel encoding technique is proposed for the recognition of patterns using four different techniques for training artificial neural networks (ANNs) of the Kohonen type. Each template or model pattern is overlaid on a radial grid of appropriate size, and converted to a two-dimensional feature array which then acts as the training input to the ANN. The first technique employs Kohonen's self-organizing network, each neuron of which is assigned, after training, the label of the model pattern. It is found that a graphical plot of the labels of the neurons exhibits clusters (which means in effect that the feature array pertaining to distorted versions of the same pattern belongs to a specific cluster), thereby justifying the coding strategy used in this paper. When the new, unknown pattern is input to the network, it is classified to have the same label of the neuron whose corresponding model pattern is closest to the given pattern. In an attempt to reduce the computational time and the size of the network, and simultaneously improve accuracy in recognition, Kohonen's learning vector quantization (LVQ) algorithm is used to train the ANN. To further improve the network's performance and to realize a network of minimum size, two constructive learning algorithms, both based on LVQ, are proposed: (1) multi-step learning vector quantization (MLVQ), and (2) thermal multi-step learning vector quantization (TLVQ). When the proposed algorithms are applied to the classification of noiseless and noisy (and distorted) patterns, the results demonstrate that the pattern encoding strategy and the suggested training techniques for ANNs are efficient and robust. For lack of space, only the most essential results are presented here. For details, see Ganesh Murthy and Venkatesh (1996b).

Fire Detection Using Quartz Crystal Microbalance Chemical Sensors

The new application of piezoelectric quartz-crystal microbalance (QCM) sensors to fire detection is described. The QCM sensors are coated with plasma organic films which are sensitive to vapors emitted from burning polyvinyl chloride (PVC) cables and epoxy resin circuit boards. These vapors must be distinguished because these sensors also respond to humidity. Principal component analysis reveals that this system can discriminate between vapors emitted from burning materials and humidity. Vapor discrimination is also described using the Kohonen's Learning Vector Quantization (LVQ).

Learning Vector Quantization with Training Count (LVQTC)

Kohonen's learning vector quantization (LVQ) is modified by attributing training counters to each neuron, which record its training statistics. During training, this allows for dynamic self-allocation of the neurons to classes. In the classification stage training counters provide an estimate of the reliability of classification of the single neurons, which can be exploited to obtain a substantially higher purity of classification. The method turns out to be especially valuable in the presence of considerable overlaps among class distributions in the pattern space. The results of a typical application to high energy elementary particle physics are discussed in detail. © 1997 Elsevier Science Ltd.

A comparison of neural network and Bayes recognition approaches in the evaluation of the brainstem trigeminal evoked potentials in multiple sclerosis

This article describes the application of Multi-Layer Perceptron (MLP), Probabilistic Neural Network and Kohonen's Learning Vector Quantization to the problem of diagnosing Multiple Sclerosis. The classification information is obtained from brainstem trigeminal evoked potential. The performance of the neural networks based classifiers is compared with that of the human experts and the Bayes classifier. The ability of the MLP classifier to generalize is far better than that of the Bayes classifier. The efficiency of the neural network based classifiers in conjunction with several types of well-known evoked potential features, such as Fourier transform space, latency and temporal wave, is examined. Although a large clinical data base would be necessary, before this approach can be fully validated, the initial results are promising.

Comparison of the BP training algorithm and LVQ neural networks for e, μ, π identification

Two different kinds of neural networks, feed-forward multi-layer mode with Back-Propagation training algorithm (BP) and Kohonen's Learning Vector Quantization networks (LVQ), are adopted for the identification of e, μ, π particles in Beijing Spectrometer (BES) experiment. The data samples for training and test consist of μ from cosmic ray, e and π from experimental data by strict selection. Although their momentum spectra are non-uniform, the identification efficiencies given by BP are quite uniform versus momentum, and LVQ is little worse. At least in this application BP is shown to be more powerful in pattern recognition than LVQ.

Neural 2.00 — A program for neural net and statistical pattern recognition

A neural net program for pattern classification is presented, which includes: i) an improved version of Kohonen's learning vector quantization (LVQ with training count); ii) feed-forward neural networks with back-propagation training; iii) Gaussian (or Mahalanobis distance) classification; iv) Fisher linear discrimination. Back-prop trainings with emulations of Intel's ETANN and Siemens' MA16 neural chips are available as options. The program has been developed for high energy physics applications.

Fuzzy Model Using a Modified LVQ Network for Sinter Strand Process

Rule-based fuzzy systems can be structurally mapped into Kohonen's Learning Vector Quantization (LVQ) network. However the structure of the Kohonen layer is fixed in standard LVQ and the recall is in the manner of winner-takes-all. This is not suitable for the self-organizing of rule-based fuzzy systems. Modifications are made to the LVQ's learning algorithm and the activity function of the Kohonen node. With the modified LVQ network the fuzzy model of an iron-ore sinter strand process is constructed. The resulting fuzzy model is dynamically self-organized from input-output observation data with very little a priori knowledge about the sinter strand process and is adaptive to the change in the environment in which the sinter strand process operates. The results of the simulation show that the model posseses very good performance in terms of prediction accuracy, adaptive ability and robustness.

Automated feature selection with a distinction sensitive learning vector quantizer

An extended version of Kohonen's Learning Vector Quantization (LVQ) algorithm, called Distinction Sensitive Learning Vector Quantization (DSLVQ), is introduced which overcomes a major problem of LVQ, the dependency on proper pre-processing methods for scaling and feature selection. The algorithm employs a weighted distance function and adapts the metric with learning. Highest weights are assigned to components in the input vectors which are most informative for classification; non-informative components are discarded. The algorithm is applied to the analyses of multi-channel EEG data and compared with experienced methods.

Neural network-based inspection of solder joints using a circular illumination

In this paper, we describe an approach to inspection of solder joints on printed circuit boards by using a circular illumination technique and a neural network classifier. The illumination technique, consisting of three tiered circular colour lamps and one colour camera, gives good visual cues to infer 3D shape of the solder joint surface. A general aspect of this inspection is that the shape of the solder joint tends to greatly vary according to soldering conditions. Due to this, a neural network classifier based on a supervised version of Kohonen learning vector quantization (LVQ) is proposed to automatically and efficiently make classification criteria of the solder joint shapes according to their quality. The practical feasibility of the proposed approach is demonstrated by building a prototype inspection machine and testing its performance.

Fuzzy self-organizing map

Kohonen's Self-Organizing Map is one of the best-known neural network models. In this paper, we introduce a fuzzy version of the model called: Fuzzy Self-Organizing Map. We replace the neurons of the original model by fuzzy rules, which are composed of fuzzy sets. The fuzzy sets define an area in the input space, where each fuzzy rule fires. The output of each rule is a singleton. The outputs are combined together by a weighted average, where the firing strengths of the fuzzy rules act as the weights. The weighted average gives a continuous valued output for the system. Thus the Fuzzy Self-Organizing Map performs a mapping from a n-dimensional input space to one-dimensional output space. The learning capability of the Fuzzy Self-Organizing Map enables it to model a continuous valued function to an arbitrary accuracy. The learning is done by first self-organizing the centers of the fuzzy sets according to Kohonen's Self-Organizing Map learning laws. After that, the fuzzy sets and the outputs of the fuzzy rules are initialized. Finally, in the last phase of the new learning method, the fuzzy sets are tuned by an algorithm similar to Kohonen's Learning Vector Quantization 2.1. Simulation results of a two-dimensional sinc function show good accuracy and fast convergence.

Constrained learning vector quantization.

Kohonen's learning vector quantization (LVQ) is an efficient neural network based technique for pattern recognition. The performance of the method depends on proper selection of the learning parameters. Over-training may cause a degradation in recognition rate of the final classifier. In this paper we introduce constrained learning vector quantization (CLVQ). In this method the updated coefficients in each iteration are accepted only if the recognition performance of the classifier after updating is not decreased for the training samples compared with that before updating, a constraint widely used in many prototype editing procedures to simplify and optimize a nearest neighbor classifier (NNC). An efficient computer algorithm is developed to implement this constraint. The method is verified with experimental results. It is shown that CLVQ outperforms and may even require much less training time than LVQ.

Web process inspection using neural classification of scattering light

Web process inspection requires rapid examination of vast amounts of data. The two resulting issues, the sensory system and the computer processing required to detect faults in the sheet material accurately, were examined. It was found that scattered coherent light from the surface of the material being processed could be directly conditioned by a photodetector so as to produce a small set of features which are then examined by a neural network trained to find unsatisfactory surface conditions. A surface inspection system using measurement of the angular distribution over a 25 degrees C cone angle of the scattering was constructed, calibrated, and evaluated for inspection of coated sheet and steel samples. Features, created by a simulated segmented photodetector, were inputs to a neural network which used classification based upon T. Kohonen's (1989) learning vector quantization (LVQ2). The system was evaluated with CrO/sub 2/ coated steel samples. Classification by fault or no-fault categorized 133 samples corrected out of 135, while there were seven errors in one attempt at classification by the various common types of surface fault out of the same number of test samples and nine in another.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Discrimination of $$p\bar p \to t\bar t$$ events by a neural network classifier

Neural network and conventional statistical techniques are compared in the problem of discriminating\(p\bar p \to t\bar t\) events, with top quarks decaying into anything, from the associated hadronic background at the energy of the Fermilab collider. The NN we develop for this sake is an improved version of Kohonen's learning vector quantization scheme. Performance of the NN as a\(t\bar t\) event classifier is found to be less satisfactory than that achievable by statistical methods. We conclude that the probable reasons for the are: i) the NN approach presents advantages only when dealing with event distributions in the feature space which substantially differ from Gaussians; ii) NN's require much larger training sets of events than statistical discrimination in order to give comparable results.

A self-learning neural net for ultrasonic signal analysis

This paper deals with ultrasonic signal analysis using artificial neural nets. In particular, it investigates the classification of ultrasonic inspection data using a backpropagation (BP) neural network. The traditional BP learning algorithm requires a set of signals that have been classified a priori for training the net. To eliminate the uncertainties involved in selecting the training set, a self-learning algorithm is developed. A BP net is trained using the self-learning algorithm and it is used to classify ultrasonic inspection data. The classification is performed in both the time and the frequency domains. The classification results are compared with the results obtained by conventional C-scan methods, and it is found that 95% of all the signals with flaws are classified correctly by the unsupervised BP net. The self-learning algorithm developed for classification of ultrasonic data is also detailed in this paper. It utilizes the correlation information computed by the first hidden layer of the BP net for the generation of the initial training set consisting of one signal per each class. Additional signals in each class are selected by the net using the nearest neighbour approach. An estimate of the number of different classes present in the data is made using the damage profile of the sample being investigated by Kohonen's learning vector quantization (LVQ) algorithm.

SPEAKER-INDEPENDENT RECOGNITION OF ISOLATED WORDS USING CONCATENATED NEURAL NETWORKS

A speaker-independent isolated word recognizer is proposed. It is obtained by concatenating a Bayesian neural network and a Hopfield time-alignment network. In this system, the Bayesian network outputs the a posteriori probability for each speech frame, and the Hopfield network is then concatenated for time warping. A proposed splitting Learning Vector Quantization (LVQ) algorithm derived from the LBG clustering algorithm and the Kohonen LVQ algorithm is first used to train the Bayesian network. The LVQ2 algorithm is subsequently adopted as a final refinement step. A continuous mixture of Gaussian densities for each frame and multi-templates for each word are employed to characterize each word pattern. Experimental evaluation of this system with four templates/word and five mixtures/frame, using 53 speakers (28 males, 25 females) and isolated words (10 digits and 30 city names) databases, gave average recognition accuracies of 97.3%, for the speaker-trained mode and 95.7% for the speaker-independent mode, respectively. Comparisons with K-means and DTW algorithms show that the integration of the splitting LVQ and LVQ2 algorithms makes this system well suited to speaker-independent isolated word recognition. A cookbook approach for the determination of parameters in the Hopfield time-alignment network is also described.

Neural networks for enhanced human-computer interactions

The use of neural networks to create adaptive models of users and communications channels for use in designing system response characteristics is discussed. Two types of neural networks that will be useful for this type of task are considered. One, the Kohonen learning vector quantization (LVQ) network, is a clustering network. It can adjust the vector element values of a set of quantizing vectors in order to create exemplar vectors which represent clusters in a set of data. The other, the Kohonen self-organizing topology-preserving map (SOTPM), is a more advanced and powerful network that uses similar principles as the LVQ. It can create a topographic mapping of a set of vector data which creates a data clustering visible in a reduced dimensionality space from the original data. This will facilitate interpretation of the data clusters describing different user types. The use of a rather highly evolved form of neural networks to create more powerful system models is discussed.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>