| 학회 | 한국재료학회 |
| 학술대회 | 2018년 봄 (05/16 ~ 05/18, 삼척 쏠비치 호텔&리조트) |
| 권호 | 24권 1호 |
| 발표분야 | 5. 화학센서용 기능성 소재(Chemical sensors)-오거나이저: 장호원 교수(서울대) |
| 제목 | Novel Signal Processing Techniques based on Adaptive Radial Basis Function Networks for Chemical Sensors Array |
| 초록 | Using a chemical sensors array, it is desirable to discriminate between chemicals and compare one sample with another. The ability to classify pattern characteristics from relatively small pieces of information has led to growing interest in methods of sensor recognition. A variety of pattern recognition algorithms including adaptive radial basis function network (RBFN) may be applicable to gas and/or odour classification. In presentation, we proposed different types of identification techniques based on RBFN as well as drift compensation techniques caused by sensor poisoning and aging. To use the RBFN as a classifier, the parameters such as centers, widths and weights have to be optimized in the training (learning) stage before the classification (testing) stage, in which classification is carried out using the RBFN with the optimized and fixed parameters. Along with how well sensor mechanism is used, how accurately (close to optimum) parameters are acquired is one of the most important factors in classification performances. The centers and widths of the RBFN are calculated by the fuzzy c-means algorithm and the distribution of input patterns. The weights of the network are calculated by singular value decomposition (SVD) method in a single shot process. This RBFN-SVD is considered superior to other learning algorithms, particular in terms of processing speed and solavbility of non-linear pattern responses in gas and/or odour analysis. However, since the centers and widths are fixed after they are chsen, this method often results in an unsatisfying performance when input patterns are not particularly clusterd. So, the fine-tuning of cenetrs and widths is needed and the stochastic gradient (SG) method is successfully applied for this purpose. The adaptive RBFN-SVD-SG algorithm has shown good claasification performance for complex chemical patterns. Fine tuning of all the network parameters, centers, widths and weights, can be carried out by SG (RBFN-SG) with the initial weight values of zero. The learning speed of algorithm emplying the stochastic-gradient descent method is dependent on convergence coeffients. To avoid long training time, we applied Taylor’s expansion similar to that of the normalized least means square (LMS) to obtaining the optium convergence coeeficients as soon as possible. Due to sensors drift, sometime trained (learned) RBFN can no longer be effective in gas and/or odour classification. We propsed enhanced signal processing techniques for readjusting the RBFN weights in the test stage using probability distribution function (PDF) construction and Cross-Correntropy Concept based on Information Theorical Learning (ITL). The RBFN weights are readjusting in the test stage by proposed techniques for sensor drift compensation. |
| 저자 | 변형기, 김남용 |
| 소속 | 강원대 |
| 키워드 | Chemical Sensor Array; Signal Processing; Adaptive Radial Basis Function Networks |
