Toru Fujinaka

Neuro-PID control for inverted single and double pendulums

The paper is concerned with a new architecture of a self-tuning neuro-PID control system and its application to stabilization of an inverted pendulum. A single-input multi-output system is considered to control the inverted pendulum by using the PID controller. The PID gains are tuned by using two kinds of neural network. The simulation results show effectiveness of the proposed approach.

Bill classification by using the LVQ method

For pattern classification problems the neuro-pattern recognition, which is the pattern recognition based on the neural network approach, has been increasingly popular since it can classify various patterns similar to human beings. In this paper we adopt the learning vector quantization (LVQ) method to classify the various bank notes. The reasons to use LVQ are that it can process the unsupervised classification and treat many input data with small computational burdens. We construct the LVQ network to classify the Italian Liras. Compared with a conventional pattern matching technique, which has been adopted as a classification method, the proposed method has shown excellent classification results.

Intelligent Electronic Nose Systems for Fire Detection Systems Based on Neural Networks

In this paper, an intelligent electronic nose (EN)system designed using cheap metal oxide gas sensors (MOGS) is designed to detect fires at an early stage. The time series signals obtained from the same source of fire are highly correlated, and different sources of fire exhibit unique patterns in the time series data. Therefore, the error back propagation (BP) method can be effectively used for the classification of the tested smell. The accuracy of 99.6% is achieved by using only a single training dataset from each source of fire. The accuracy achieved with the k-means algorithm is 98.3%, which also shows the high ability of the EN in detecting the early stage of fire from various sources.

Improvement of reliability in banknote classification using reject option and local PCA

The improvement of reliability in banknote neuro-classifier is investigated and a reject option is proposed based on the probability density function of the input data. The classification reliability is evaluated through two reliability parameters, which are associated with the winning class probability and the second maximal probability. Then a threshold value is set up to reject the unreliable classifications. As for modeling the non-linear correlation among the data variables and extracting the features, a local principal components analysis (PCA) is applied. The method is tested with a learning vector quantization (LVQ) classifier using 1440 data samples of various US dollar bills. The results show that by taking a suitable reject threshold value and also a proper number of regions for the local PCA, the reliability of system can be improved significantly.

Stabilization of double inverted pendulum with self-tuning neuro-PID

A self-tuning neuro-PID control architecture is proposed and applied to the stabilization of double inverted pendulum. The gain parameters of the PID controller are tuned using a neural network. The effectiveness of the proposed method is shown through simulation and experiment.

A reliable method for classification of bank notes using artificial neural networks

We present a method based on principal component analysis (PCA) for increasing the reliability of bank note recognition machines. The system is intended for classifying any kind of currency, but in this paper we examine only US dollars (six different bill types). The data was acquired through an advanced line sensor, and after preprocessing, the PCA algorithm was used to extract the main features of data and to reduce the data size. A linear vector quantization (LVQ) network was applied as the main classifier of the system. By defining a new method for validating the reliability, we evaluated the reliability of the system for 1200 test samples. The results show that the reliability is increased up to 95% when the number of PCA components as well as the number of LVQ codebook vectors are taken properly. In order to compare the results of classification, we also applied hidden Markov models (HMMs) as an alternative classifier.

A switching scheme for adaptive control using multiple models

Adaptive control of a linear plant using multiple models for the identification is considered. It is assumed that the plant parameters are either partially known or subject to abrupt changes. A switching scheme is proposed for improving the transient response of the plant output while maintaining the stability of the overall system established by Narendra et al. (1997). Models with fixed parameters as well as time-varying ones are used during operation, and the main idea is to start a new fixed model whenever the switching from a time-varying model to a fixed model occurs. The new fixed model will be effectively used if the plant parameters take the original values at a later time. The effectiveness of the proposed method is demonstrated through simulation studies.

Mixed Odor Classification for QCM Sensor Data by Neural Networks

Compared with metal oxide semiconductor gas sensors, quarts crystal microbalance (QCM) sensors are sensitive for odors. Using an array of QCM sensors, we measure mixed odors and classify them into an original odor class before mixing based on neural networks. For simplicity we consider the case that two kinds of odor are mixed since more than two becomes too complex to analize the classification results. We have used eight sensors and four kinds of odor are used as the original odors. The neural network used here is a conventional layered neural network. The classification is acceptable although the perfect classification could not been achieved.

Bolt tightening control using neural networks

We propose a method of controlling the tightening operation of bolts using a device called an impact wrench. A neural network is used for classifying the material and shape of the work to which the bolts are being tightened. Then another neural network is used for estimating the relationship between the clamping force of the bolt and its incremental angle. The input to the actuator of the impact wrench is determined based on the estimated value of the clamping force. A simulation study shows satisfactory results in comparison to those achieved by a skilled factory worker.

Automatic Inspection of Transmission Devices Using Acoustic Data

Most factories depend on skilled workers to test the quality of transmission devices by listening to the sound. In this paper, an intelligent inspection system is proposed to evaluate the quality of transmission devices in place of experts. Since the causes of faults of transmission devices are complex and a defective product might simultaneously have many types of faults, the discrimination process between defective and nondefective products and the classification process of defective products are treated separately in the proposed system. From the acoustic data of operating transmission devices, we extract the feature vectors based on time-frequency analysis and train a neuroclassifier by using the learning vector quantization (LVQ). Furthermore, the genetic algorithm (GA) with floating point (FP) is utilized to select some significant frequencies from the spectra of acoustic data of defective and nondefective products and to make a quality evaluation rule automatically. The defective products are picked up from the automatic production line according to the evaluation rule and the trained neuroclassifier. At last, the self-organizing feature map (SOM) algorithm is used to identify the kinds of defective products. The experimental results show that the proposed intelligent system is able to perform the quality evaluation of transmission devices successfully. This paper was motivated by the problem of developing an intelligent evaluation system in place of skilled workers to evaluate the quality of transmission devices automatically based on acoustic data. Most of the prior works in quality evaluation of transmission devices are based on processing vibrometer signals for system vibrations. Such vibrometers must be installed on the surface of vibrating part of transmission devices, which alter the physical integrity. In this paper, the acoustic data of operating transmission devices are recorded. We first compute the ASFTS and FVAVT, where ASFTS denotes the average of a serial of spectra calculated from time segments of an acoustic data and FVAVT denotes the feature vector of amplitude variation with time in a certain band. By using the different characteristics of the ASFTS and the FVAVT, we apply the LVQ and the FGA to the quality evaluation of transmission devices, respectively. Utilizing the advantages of the SOM, we classify the defective products successfully. In the industry production, the quality of each batch of products will change according to the situation of equipments. Similarly, the quality evaluation rule will be adjusted according to the yield and the request of customers. The proposed system can evaluate the quality of transmission devices correctly as demanded so long as we change the nondefective and defective samples.