Brain-actuated Control of Wheelchair Using Fuzzy Neural Networks

Transcription

1 Int'l Conf. Artificial Intelligence ICAI'6 67 Brain-actated Control of Wheelchair Using Fzzy Neral Networks Rahib H.Abiyev, Nrllah Akkaya, Ersin Aytac, Irfan Günsel, Ahmet Ça man, Sanan Abizade Near East University, Applied Artificial Intelligence Research Centre, Lefkosa, Mersin-0, North Cyprs Abstract - In this paper, a brain-actated control of the wheelchair for physically disabled people is presented. The design of the system is focsed on receiving, processing and classification of the brain signals and then performing control of the wheelchair. The nmber of experimental measrements of brain activity has been done sing hman control commands of the wheelchair. Using obtained data inclding brain signals and control commands the design of classification system based on fzzy neral networks (FNN) is performed. The strctre and learning algorithm of FNN sed for brain-actated control are presented. The training data is sed to design the system and then test data is applied to measre the performance of the control nder real conditions. The approach sed in the paper allows redcing the probability of misclassification and improving the control accracy of the wheelchair. Keywords Brain-compter interface, wheelchair, fzzy neral networks, electroencephalogram signal. Introdction The measring hman brain signal and converting it into control signals needs the development of the interface between the brain and compter and then implementing the control of devices. A brain compter interface (BCI) provides commnication between compter and mind of ppils. This interface can be based on brain activity dring msclar movements or the changes of the rhythms of brain signals. These brain activities can be detected sing electroencephalographic (EEG) signals. BCI transforms the EEG signals prodced by brain activity into control signals which can then lasted be sed for controlling the wheelchair withot sing mscles. Since the brain signals are very weak we need to apply some spatial and spectral filters and amplifiers to the EEG signals in order to extract characteristic featres of these signals. Several EEG signals can be detected, reslting in different types of BCI. These signals are based on the change of freqencies, change of amplitdes. For example dring volntary thoghts the freqencies of signals are modified, dring movement a synchronisation/ desynchronisation of brain activity which involves rhythm amplitde change. This relevant characteristic makes rhythm based BCI sitable to be sed. Recently some research works have been done to develop many applications of BCI for wheelchairs. BCI is a control interface that translates hman intentions into appropriate motion commands for the wheelchairs, robots, devices, etc. [] considers the application of BCI and control of wheelchair in an experimental sitation. The research considers the driving of a simlated wheelchair in a virtal environment (VE) before sing BCI in a real sitation. [2] describe a BCI system which control the wheelchair that moves in only one direction- move forward. In [3] a simlated robot is designed that performs two actions- trn left then move forward, or trn right then move forward. [4,5] ses three possible commands trn left, trn right and move forward. In [6] BCI is designed sing EEG signal captred by eight electrodes. Wavelet transform was sed for featre extraction and the radial basis networks were sed to classify the predefined movements. In [7] controller based on the brain-emotionallearning algorithm is sed to control the omnidirectional robot. [8] presents the design of an asynchronos BCI based control system for hmanoid robot navigation sing an EEG. [0] considers a non-invasive EEG-based Brain Compter Interface (BCI) system to achieve stable control of a low speed nmanned aerial vehicle for indoor target searching. [0-5] consider the design of brain controlled wheelchair. The constraction of viable brain actated wheelchair that combines brain compter interface with a commercial wheelchair, via a control layer, is considered. BCI allows improving the qality of life of disabled patients and letting them interact with their environment. The processes of featre extraction and classification is very important in BCI design and has a great affect to the performance of the BCI system. Set of research have been done for improvement of the featre extraction and classification algorithms [6-9]. [6,7] considers featre extraction algorithms for Brain-Compter Interfaces. Recently different clstering algorithms based on spport vector machine, linear discriminant analysis, neral networks are applied for classification of brain signals [8]. [9] sed featres, optimised in the sense of statistically significant and potentially discriminative coherences at a specific freqency and applied linear discriminant for classification prpose. SVM based classification [20] and linear discriminant analysis (LDA) [2] are sed for classification prpose of brain signals. [22] ses fzzy logic and [23] ses neral networks with fzzy particle swarm optimisation for BGI design. In [24] continos wavelet transform is sed to extract highly representative featres and an Adaptive Neron-Fzzy Inference System (ANFIS) is sed for classification. Fzzy

2 68 Int'l Conf. Artificial Intelligence ICAI'6 logic provides a simple way for determining a conclsion based pon vage, ambigos, imprecise, noisy, or missing inpt information. Fzzy Logic s approach to control problems mimics how a person wold make faster decisions. As shown featre extraction and classification plays an important role in designing brain based control for obtaining of high classification accracy. In BCI design, a classification error (a wrong command) can case dangeros sitations, so it is crcial to garantee a minimm error rate to keep the sers safe. Different clstering algorithms based on spport vector machine, linear discriminant analysis, neral networks are applied for classification of brain signals [8]. Fzzy classification represents knowledge more natrally to the way of hman thinking and is more robst in tolerating imprecision, conflict, and missing information. In this paper, fzzy neral network is sed for the design of BCI in order to achieve efficient brain based control of wheelchair. Signal processing, featre extraction and classification algorithms are designed for brain-actated control of wheelchair. 2 BCI system architectre Fig. depicts BCI based control of the wheelchair. BCI system consists of an Emotiv headset connected to a compter where classification algorithms are rn which is connected to a micro-controller that controls the movement of motors. A BCI based control system is sally composed of six main nits signal acqisition nit, signal preprocessing nit, featre extraction nit, classification nit, control action nit and wheelchair motors nit. The brain signals are captred sing an emotive headset tilizing 4 channels. These inpt signals are sent to the signal processing nit. The signals in preprocessing block after filtering and scaling are entered to the featre extraction block. The basic featres are extracted and send to the classification system. The otpt signals of the classification block are motor signals (clsters) that are sent to the wheelchair. Preprocessing Signal Signal Featre Extraction Classification Control actions Wheelchair motors Fig.. The BCI based control of the wheelchair In signal acqisition block the EEG signals are captred sing the Emotiv headset. Emotiv EPOC is an EEG Headset which spplies 4 channel EEG data (Fig.2) and 2 gyros for 2- dimensional controls. Its featres are adeqate for a sefl BCI in case of resoltion and bandwidth. Or system ses pper face gestres for actation commands since most Emotiv sensors are located in the frontal cortex they are the most reliable signals to detect. Fig.2. Emotiv's sensor Layot compared to standart 72 sensors layot. The distribtion of EEG electrodes. Forteen channels are marked for data acqisition. The measred signals are sent to the system inpt. The signals are very long for processing. Therefore, the featre extraction techniqe is applied in order to decrease the signal size and extract more important featres for classification. In the paper, two different approaches are sed for processing of the inpt sensor signals With Fast Forier Transform (FFT) and withot FFT. The inpt signal received from the headset is divided into windows having 2 sec time interval with 50% overlap (Fig.3). The se of overlapping windows allows s to increase the accracy of the classification. Each two second window corresponds to 256 samples of data. Each second headset retrns 28 data samples. The obtained signals from the channels, stored as windows, are then sent to normalisation block. Each channel is normalised in order to center each channel on zero by calclating the mean vale of each channel for the window, then sbtracting it from each of the data points in the channel. After normalisation, Hamming window is applied to each channel in the window. EEG signals do not generally repeat exactly, over any given time interval, bt the math of the Forier transform assmes that the signal is periodic over the time interval. This mismatch leads to errors in the transform called spectral leakage. Hamming window is sed to mitigate this problem. Then fast Forier transform (FFT) is applied to each channel in the window to find ot the freqency components of the signal. Each freqency component is sed as a featre, which reslts in 64x4 featres. In order to increase the performance of the classification, the featres are ranked by evalating the worth of a freqency by measring the information gain with respect to the class. The expected information gain is the change in information entropy (H) from a prior state to a state that takes some information as given InfoGain(Class,Freqency) = H(Class) - H(Class Freqency)

3 Int'l Conf. Artificial Intelligence ICAI'6 69 Information Gain, selects a sbset of the original representation attribtes according to Information Theory qality metric, Information Gain. This method comptes the vale of the metric for each attribte, and rank the attribtes. Then simply decide a threshold in the metric and keep the attribtes with a vale over it. Inpt Data Split Data into Windows Having 2 sec. time interval Normalize each window Applying FFT to the window data Combine all channels Redce the nmber of featres Fig.3. Signal Preprocessing nit. After freqency representation, all channels in the window are combined in to a single nit so as to apply classification on all channels at once. The filtering operation is applied in order to select important featres of the brain signals. These featres are sed for classification prpose. The whole signal preprocessing stages are shown in Fig. 3. In the second approach the acqired brain signal after windowing, normalisation and combining operations are sed for classification prpose. These signals are inpt for the classification. After the classification the signals the otpt of classification system is sed to activate the wheelchair. Even thogh dring training system reports 00% sccess rate in real world conditions it does misclassify, a state machine is sed to frther increase safety and redce misclassification. As an example, the system won t transition from forward motion to backward motion withot stopping in netral. The otpt of the state machine drives the microcontroller which controls the motors on the wheelchair. The nmber of classes is eqal to the nmber of control actions. 3 FNN Based Classification The featres extracted from the EEG signals are sed for classification and determining control action. In this paper, we propose a novel approach for the classification of brain signals sing FNN based classifier. The extracted featres are inpt signals of the FNN based classifier. The classifier based on the above featres classifies the signals into the six classes move forward, move backward, switch on, stop, trn left and trn right. The fzzy neral system combines the learning capabilities of neral networks with the lingistic rle interpretation of fzzy inference systems. The design of FNN incldes the generation of IF-THEN rles [25-28]. Here, the problem consists in the optimal definition of the premise and conseqent part of fzzy IF-THEN rles for the classification system throgh the training capability of neral networks, evalating the error response of the system. There are two basic types of IF-THEN rles sed in fzzy systems. These are Mamdani and Takagi-Sgeno-Kang (TSK) type fzzy rles. The first one consists of rles, whose antecedents and conseqents parts tilize fzzy vales. The second type fzzy system ses the rle base that has fzzy antecedent and crisp conseqent parts. The second type of fzzy system approximates nonlinear system with linear systems and has the following form. If x is A j and x 2 is A 2j and and x m is A mj Then m y b a x = + () j j ij i i= Here x i and y j are inpt and otpt signals of the system, respectively, i=,...,m is the nmber of inpt signals, j= r is the nmber of rles. A ij are inpt fzzy sets, b j and a ij are coefficients. The strctre of fzzy neral networks sed for the classification of EEG signals is based on TSK type fzzy rles and is given in Fig. 4. The FNN incldes six layers. In the first layer, the x i (i=,,m) inpt signals are distribted. The second layer incldes membership fnctions. Here each node corresponds to one lingistic term. Here for each inpt signal entering the system, the membership degree to which inpt vale belongs to a fzzy set is calclated. To describe lingistic terms, the Gassian membership fnction is sed. 2 ( xi cij) 2 σij μ ( x ) = e, i =,..., m, j =,..., r (2) j i where m is a nmber of inpt signals, r is a nmber of fzzy rles (hidden nerons in the third layer). c ij and σ ij are centre and width of the Gassian membership fnctions, respectively. j (x i ) is membership fnction of i-th inpt variable for j-th term. The third layer is a rle layer. Here nmber of nodes is eqal to the nmber of rles. Here R, R 2,,R r represents the rles. The otpt signals of this layer are calclated sing t- norm min (AND) operation., i=,..,m, j=,...,r (3)

4 70 Int'l Conf. Artificial Intelligence ICAI'6 where is the min operation. These j (x) signals are inpt signals for the fifth layer. Forth layer is a conseqent layer. It incldes n linear systems. Here the vales of rles otpt are determined. m y b a x = + (4) j j ij i i= In the fifth layer, the otpt signals of the third layer are mltiplied by the otpt signals of the forth layer. The otpt of j-th node is calclated as y = μ ( x) y j j j In the sixth layer, the otpt signals of FNN are determined as k = r j= r j= w y jk j μ ( x) j Here k are the otpt signals of FNN, (k=,..,n). After calclating the otpt signal, the training of the network starts. (5) and the gradient algorithm is applied to design the conseqent parts of the fzzy rles. Fzzy c-means clstering is applied in order to partition inpt space and constrct antecedent part of fzzy if-then rles. In the reslts of partitioning the determined clster centers will correspond to centers of the membership fnctions sed in inpt layer of FNN. The width of the membership fnction is determined sing distance between clster centers. After the design of the antecedents parts by fzzy clstering, the gradient descent algorithm is applied to design the conseqent parts of the fzzy rles. At the beginning, the parameters of the FNN are generated randomly. To generate a proper FNN model, the training of the parameters has been carried ot. For generality we have given the learning procedre of all parameters of FNN sing gradient descent algorithm. The parameters are the membership fnction of lingistic vales in the second layer of the network and the parameters of the forth and fifth layers. Training incldes the adjsting of the parameter vales. In this paper, we applied gradient learning with adaptive learning rate. The adaptive learning rate garantees the convergence and speeds p the learning of the network. In addition, the momentm is sed to speed-p the learning processes. At first, on the otpt of the network, the vale of cost (x ) R (x) y / x R 2 2(x) y / x 2 x m r(x m) R n r (x) y r / Layer Layer 2 Layer 3 Layer 4 Layer 5 Layer 6 Fig. 4. FNN based identifier 4 Parameter Learning The design of FNN (Fig. 4) incldes determination of the nknown parameters that are the parameters of the antecedent and the conseqent parts of the fzzy if-then rles (). In the antecedent parts, the inpt space is divided into a set of fzzy regions, and in the conseqent parts the system behavior in those regions is described [25-28]. In this paper, the fzzy clstering is applied to design the antecedent (premise) parts, fnction is calclated. E = 2 n d 2 ( k k) (6) k= Here n is the nmber of otpt signals of the network, are desired and crrent otpt vales of the network (k=,..,n), respectively. The parameters w jk, a ij, b j, (i=,..,m, j=,..,r, k=,..,n) in conseqent part of network and

5 Int'l Conf. Artificial Intelligence ICAI'6 7 the parameters of membership fnctions c ij and σ ij (i=,..,m, j=,..,r) of in the premise part of FNN strctre are adjsted sing the following formlas. wjk ( t+ ) = wjk ( t) γ + λ( wjk ( t) wjk ( t )); w jk aij ( t+ ) = aij ( t) γ + λ(a ij ( t) aij ( t )); (7) aij bj( t+ ) = bj( t) γ + λ(b j( t) bj( t )); bj cij ( t+ ) = cij ( t) γ + λ( cij ( t) cij ( t )); cij σij ( t+ ) = σij ( t) γ + λσ ( ij ( t) σ ij ( t )); (8) σ ij i=,..., m; j =,..., r; k =,..., n. Here is the learning rate, is the momentm, m is the nmber of inpt signals of the network (inpt nerons) and r is the nmber of fzzy rles (hidden nerons) ). Using eqations (7) and (8) the correction of the parameters of FNN is carried ot. Convergence is very important problem in learning of FNN model. The convergence of the learning algorithm sing gradient descent depends on the selection of the initial vales of the learning rate. The derivation of the convergence is given in [33, 34]. 5 Experiments and Reslts The BCI system is simlated and sed in real life application. The EEG signals are measred with Signal acqisition nitthe Emotiv EPOC headset. In the experiments, we have tilized 4 channels for measring EEG signals. The measred EEG signals have different rhythms within the freqency band. The experiments show that measring brain signals is difficlt so we have tested or system sing brain mscle signals. The signals obtained from 5 sample channels are shown in Fig.5. Fig.5(a) depicts a netral pose, patient relax not doing anything. Fig.5(b) depicts a positive gestre. As shown in figres, the EEG signals with positive gestre pose are changing more freqently than a netral pose. In the paper, the FFT is applied to extract important featres of the signal. After preprocessing stage, given in section 2, the important featres of these signals are extracted and sed for classification prpose. The nmber of extracted featres was determined as 00. These signal are inpts for FNN system. Otpts of FNN model are clsters. Six clsters are sed in the experiment Move Backward, Move Forward, Switch on, Stop, Trn Left, and Trn Right. For each clster, the system recorded 0 seconds of data. The classification of the EEG signals is performed sing FNN model. To synthesis classification model the FNN strctre with hndred inpt- and six otpt nerons is generated first.. Fig.5. EEG signals for five channels a) netral pose, b) positive gestre pose Fzzy classification is applied in order to partition inpt space and select the parameters of the premise parts, that is the parameters of Gassian membership fnctions sed in the second layer of FNN. Fzzy c-means clstering is sed for the inpt space with 5 clsters for each inpt. 5 fzzy rles are constrcted sing different combination of these clsters for 00 inpts. After clstering inpt space gradient decent algorithm is sed for learning of conseqent parts of the fzzy rles, that is parameters of the 4-th layer of FNN. In learning of FNN 0 fold cross validation is sed for separation the data into training and testing set. The initial vales of the parameters FNN are randomly generated in the interval [-, ] and, sing the gradient algorithm derived above, they are pdated for the given inpt- criterion, RMSE is otpt training pairs. As a performance sed. The training is carried ot for 000 epochs. The vales of the parameters of the FNN system were determined at the conclsion of training. Once the FNN has been sccessflly trained, it is then sed for the classification of the EEG signals. Dring learning, the vale of RMSE was obtained as for training data, and for evalation. After learning, for the test data the vale of RMSE was obtained as with 00% accracy of classification. Fig. 6 depicts RMSE vales obtained dring training. The design of FNN model is performed sing a different nmber of rles. Table incldes reslts of simlations wit 5, 6, 9 and 6 rles respectively. As shown accracy of FNN classification model are 00%. For comparison prpose, we testt the system sing different classification techniqes. In the reslt of classification, the following reslts are obtained (Table 2). As shown the simlation reslts demonstrate the efficiency of application FNN model in the classification of EEG signals. These clsters activate the corresponding control signal which is then sed to actate the motors of the wheelchair..

6 72 Int'l Conf. Artificial Intelligence ICAI'6 6 Conclsions The paper presents BCI based on FNN for the wheelchair. The emotional and msclar states of the ser are evalated for control prpose. The design of BCI has been done to actate a brain controlled wheelchair sing six mental activities of the ser Move Backward, Move Forward, Switch on, Stop, Trn Left and Trn Right. For classification of EEG signals FNN with 0 fold cross validation data set is sed. The design of the FNN system is implemented sing fzzy c means classification and gradient descent algorithm. The obtained 00% classification reslts prove that the sed techniqes are a potential candidate for the classification of the EEG signals in the design of brain based control system. In the ftre, we are going to improve the nmber of commands for control of wheelchair and decrease detection time of the EEG signal sed for measring brain activities and design efficient brain controlled wheelchair. Table. Classification reslts. Nmber of Rles Correctly Fig. 6. Training of FNN Incorrectly Training RMSE Evalation RMSE 5 92% % % % Table 2. Classification reslts Method Correctly Incorrectl y Mean absolte error SVM 96% 4% MLP (NN) (6 hidden nerons) 00% Bayesian 94% 6% Random tree 74% 26% 0.04 FNN 00% Test RMSE Root mean sqared error References [] Galán, F., Nttin, M., Vanhooydonck, D., Lew, E., Ferrez, P.W., Philips, J., de Millán,J.R. Continos brainwheelchair by hman EEG. actated control of an intelligent In Proceedings of the 4th International Brain-Compter Interface Workshop and Training Corse, TU Graz/Büroservice, Graz, pp (2008) [2] Leeb, R., Friedman, D., Müller-Ptz, G.R., Scherer, R., Slater, M., Pfrtscheller, G. Self-Paced (Asynchronos) BCI Control of a Wheelchair in Virtal Environments A Case Stdy with a Tetraplegic. Comptational Intelligence and Neroscience. Article ID (2007) [3] Tsi, C.S.L., Gan, J.Q. Asynchronos BCI Control of a Robot Simlator with Spervised Online Training. In Yin, H., Tino, P., Corchado, E., Byrne, W., Yao, X. (eds.) IDEAL LNCS, vol. 488, pp Springer, Heidelberg (2007) [4] Scherer, R., Lee, F., Schlögl, A., Leeb, R., Bischof, H., Pfrtscheller, G. Towards elfpaced Brain-Compter Commnication Navigation throgh virtal worlds. IEEE Transactions on Biomedical Engineering 55(2), (2008) [5] Anas Fattoh, Odile Horn, Gy Borhis. Emotional BCI Control of a Smart Wheelchair. IJCSI International Jornal of Compter Science Isses, Vol. 0, Isse 3, No, May 203 [6] Vijay Khare, Jayashree Santhosh, Sneh Anand, Manvir Bhatia. Brain Compter Interface Based Real Time Control of Wheelchair Using Electroencephalogram. International Jornal of Soft Compting and Engineering (IJSCE) SSN , Volme-, Isse-5, November 20. [7] Maziar A. Sharbafi, Caro Lcas, and Roozbeh Daneshvar. Motion Control of Omni-Directional Three-Wheel Robots by

7 Int'l Conf. Artificial Intelligence ICAI'6 73 Brain-Emotional-Learning-Based Intelligent Controller. IEEE Trans. on Systems, Man, and Cybernetics Part C Applications and Reviews, Vol. 40, No. 6, 200 [8] Yongwook Chae, Jaeseng Jeong, and Sngho Jo. Toward Brain-Actated Hmanoid Robots Asynchronos Direct Control Using an EEG-Based BCI. IEEE Trans. On Robotics, Vol. 28, No. 5, 202 [9] Tianwei Shi, Hong Wang, Chi Zhang. Brain Compter Interface system based on indoor semi-atonomos navigation and motor imagery for Unmanned Aerial Vehicle control. Expert Systems with Applications 42 (205) [0] Tom Carlson, Robert Leeb, Ricardo Chavarriaga and Jos e del R. Mill an. The Birth of the Brain Controlled Wheelchair. IEEE/RSJ International Conference on Intelligent Robots and Systems October 7-2, 202. Vilamora, Algarve, Portgal, pp [] B. Rebsamen, C. Gan, H. Zhang, C. Wang, C. Teo, M. H. Ang, Jr., and E. Brdet, A brain controlled wheelchair to navigate in familiar environments, IEEE Trans. Neral Syst. Rehabil. Eng., vol. 8, no. 6, pp , Dec. 200 [2] Itrrate, J. M. Antelis, A. Kbler, and J. Mingez, A noninvasive brain-actated wheelchair based on a p300 nerophysiological protocol and atomated navigation, IEEE Trans. Robot., vol. 25, no. 3, pp , Jn [3] G. Vanacker, J. del R. Mill an, E. Lew, P. W. Ferrez, F. G. Moles, J. Philips, H. V. Brssel, and M. Nttin, Contextbased filtering for brain-actated wheelchair driving, Compt. Intell. Nerosci., vol. 2007, pp. 2, May 2007 [4] Rahib H.Abiyev, Nrllah Akkaya, Ersin Aytac, Irfan Günsel, Ahmet Ça man. Development of Brain Compter Interface for Wheelchair. The International Biomedical Engineering Congress 205 (IBMEC-205), 2-4 March 205, Girne, North Cyprs [5] Lei Cao, Jie Li, Hongfei Ji, Changjn Jiang. A hybrid brain compter interface system based on the nerophysiological protocol and brain-actated switch for wheelchair control. Jornal of Neroscience Methods. Volme 229, 30 May 204, Pages [6] F. Lotte, C.T. Gan, "Reglarizing Common Spatial Patterns to Improve BCI Designs Unified Theory and New Algorithms", IEEE Transactions on Biomedical Engineering, vol. 58, no. 2, pp , 20 [7] Xiaom Song, Sk-Chng Yoon. Improving brain compter interface classification sing adaptive common spatial patterns. Compters in Biology and Medicine 6 (205) [8] Lsheng Bi, Xin-an Fan, Yili Li. EEG-based brain controlled mobile robots a srvey. IEEE Tran. on Hman- Machine Systems, V43,No2,203. [9] Rocio Salazar-Varas, David Gtiérrez. An optimized featre selection and classification method for sing electroencephalographic coherence in brain compter interfaces. Biomedical Signal Processing and Control. 8 (205) 8 [20] E. Hortal,, D. Planelles, A. Costa, E. Iáñez, A. Úbeda, J.M. Azorín, E. Fernández. SVM-based Brain Machine Interface for controlling a robot arm throgh for mental tasks. Nerocompting, Volme 5, Part, 3 March 205, Pages 6 2 [2] Yonghi Fang, Minyo Chen, Xfei Zheng. Extracting featres from phase space of EEG signals in brain compter interfaces. Nerocompting 5 (205) [22] Mandeep Kar & Poonam Tanwar. Developing brain compter interface sing fzzy logic. International Jornal of Information Technology and Knowledge Management Jly- December 200, Volme 2, No. 2, pp [23] Chai R, Ling SH, Hnter GP, Tran Y, Ngyen HT.Braincompter interface classifier for wheelchair commands sing neral network with fzzy particle swarm optimization. IEEE J Biomed Health Inform. 204 Sep;8(5) doi 0.09/JBHI [24] Darvishi S, Al-Ani A. Brain-compter interface analysis sing continos wavelet transform and adaptive nero-fzzy classifier. Conf Proc IEEE Eng Med Biol Soc. 2007; [25] Rahib H.Abiyev. Fzzy Wavelet Neral Network Based on Fzzy Clstering and Gradient Techniqes for Time Series Prediction. Neral Compting & Applications, Vol. 20, No. 2, pp , 20 [26] Rahib Hidayat Abiyev, Controller based on Fzzy Wavelet Neral Network for Control of Technological Processes. In proceeding of IEEE International Conference on Comptational Intelligence for Measrement Systems and Applications, IEEE CIMSA 2005, pp.25-29, Giardini Naxos - Taormina, Sicily, ITALY, Jly [27] Rahib H. Abiyev, Time Series Prediction Using Fzzy Wavelet Neral Network Model. ICANN Lectre Notes in Compter Sciences, Springer-Verlag, Berlin Heidelberg, pp [28] Abiyev R.H, Abiyev V, Ardil C. Electricity Consmption Prediction Model sing Nero-Fzzy System. Proceedings of the World Academy of Science Engineering and Technology, Vol.8,pp.28-3, Oct 26-28, 2005, Bdapest, Hngary.