Marys Medicine

Microsoft word - neuro-fuzzy_measurement_2014_v5_revision.doc

Particle Swarm Optimization (PSO) to improve the model's learning capability. Via this solution, they showed that PSO can be effectively employed for developing industrial model-based control schemes. 2.3. Nowadays Nowadays the emphasis of the research of Neuro-Fuzzy systems is concentrating mainly on their applications on wide practical fields. For instance Guzinski et al. presented a diagnostic system for shaft misalignment detection based on the ANFIS model [13]. They used this system in an adjustable speed sensorless induction motor drive where the Fig. 2. A sketch of a bearing assembly [10]
model is based on the analysis of the stator current, motor speed and load torque processing. The results Fig. 2. shows components of a bearing assembly showed that it can effectively indicate the shaft which can be damaged leading to malfunction. For misalignment. Karimi and Salahshoor also used the the diagnosis of these components the authors ANFIS model and combined it with Principal created a Neuro-Fuzzy algorithm which consists of Component Analysis (PCA) for fault detection and multiple modules. First, an initial set of fuzzy rules diagnosis of distillation column [14]. They used are determined, where the initial large number of PCA to extract the most informative features and at rules is reduced with a heuristic solution based on the same time reduce dimensionality of the the firing strength of each rule. Then the forward measured data then fed the reduced data to ANFIS algorithm calculates the relative strengths of the to discriminate the occurred fault. They rules and the next module uses these values for demonstrated the effectiveness of the proposed creating new rules if necessary. The optimisation method via extensive conducted tests in a module tunes the parameters of the member distillation column benchmark. functions and finally, a pruning is applied to reduce Wali et al. compared intelligent controllers such as the size of the rule set. After the initial set of rules Fuzzy and Neuro-Fuzzy in the task of monitoring has been established, the algorithm repeats itself and control of novel advanced microwave biodiesel iteratively until the desired accuracy is reached. Chen, Roberts and Weston used Neuro-Fuzzy Systems for fault detection and diagnostics of railway track circuits [11]. Fig. 3. Scheme of neuro-fuzzy for fault detection
and diagnosis [11] Fig. 4. The basic block diagram for fuzzy controller
Fig. 3. Shows the scheme of the implemented model where the measurement data associated with different operating conditions is mapped to each Fig. 4. shows the block diagram of the fuzzy failure mode. In their solution, they use a controller which was applied for temperature generalized version of the ANFIS to support control inside the reactor. The authors found that multiple outputs. Each one of the eleven outputs the ANFIS controller is more robust to parameter corresponds to a condition (1 healthy and 10 faulty variations while a pure fuzzy solution is able to condition) while eight current and voltage regulate in minimal overshoot. ANFIS model was measurements are used as the input variables. used by Eristi for fault diagnosis of series Another research was aiming at Neuro-Fuzzy based compensated transmission lines, too [16]. state modelling of a flexible robotic arm using real sensor data [12]. The authors also utilized improved

Fig. 5. Main structure of the proposed algorithm for fault diagnosis, classification and location isolation [16]
Fig. 5. shows the structure of the algorithm where the clustering based fuzzy inference system and the Vabc denotes the Voltage signal and Iabc denotes the ANN method and concluded that it is due to the fact Current signal. Eristi utilized Wavelet that ANFIS inherits the advantages of both of the Transformation (WT) and Norm entropy to achieve effective feature extraction of the fault signals. To One of the major factors in catalytic performance is test the proposed method an extensive data set of the size of the catalyst. In the next application the 23 436 fault cases were used and the results showed authors used response surface methodology and the that the algorithm is effective and robust to ANFIS model to quantify the effects of physical parameter variations. characteristics of magnetite on Fenton-like Concentration estimation of volatile organic oxidation efficiency of methylene blue [18]. For compounds was the goal of Jha et al. [17]. They petroleum products monitoring Roshani et al. built multiple models to measure and compare their applied the ANFIS model to predict fluid density performance on the problem. They found that the for a gamma ray densitometer [19]. Neuro-Fuzzy system (ANFIS) outperformed both Fig. 6. The proposed ANFIS structure to predict fluid density for a gamma ray densitometer [19]
Fig. 6. shows the proposed model where A1-A9 the output values of the corresponding nodes. The denotes the membership functions of the pipe authors simulated the operation of the densitometer diameter variable and B1-B9 denotes the device to obtain the testing dataset and after membership functions of the number of counts (per application of the model they concluded that it can photon) variable. W1-W9, W1-W9 and f1-f9 denotes estimate the fluid density with high accuracy.

In many cases very different application fields are Fig. 8. shows a typical artificial accelerogram from targeted by Neuro-Fuzzy solutions as in the case of which the 20 parameters were extracted. The another ANFIS model which was used to detect proposed models were able to produce 98% alterations in sleep EEG activity during hypopnoea recognition rate and thus their application can be episodes by Übeyli et al. [20]. The authors used the tested in real circumstances. ANFIS for classification and they performed Benyamin Khoshnevisan et al. adopted an ANFIS feature extraction by computing of wavelet and an ANN based system to predict potato yield coefficients. In their case four models was used: from energy inputs [23]. The purpose of the study three were fed directly by measured data on the was to create model which helps farmers to electrodes and the fourth had the purpose of estimate the level of production in advance and improving diagnostic accuracy by gaining its inputs make an appropriate plan for the future. The authors from the outputs of the other three systems. Lee and evaluated various topologies to find the optimal Lim also used hybrid fuzzy neural methods to one; moreover they compared the ANN and ANFIS compare Deep Brain Stimulation (DBS) and models and found that the latter one is more capable levodopa as two treatments of Parkinsonian resting of overcoming the problem of inconsistent data because of its rule based architecture. Another application aimed at prediction of vertical stress transmission in real soil profile using ANFIS [24]. The authors built multiple models with different membership function types and for this specific dataset the gauss membership function proved to be the most efficient. Moharana et al. also used the ANFIS model to estimate the roughness coefficients of a meandering channel [25]. This is a complex problem because the coefficients depend Fig. 7. Velocity laser recording of resting tremor
on many hydraulic, geometric and roughness parameters. They compared the result with earlier studies and concluded that the Neuro-Fuzzy model Fig. 7. shows a method for recording resting is more effective than others in terms of estimation tremors which was applied to 16 subjects to build a dataset. The study not only showed that DBS is There is a wide variety of other applications where more effective than levodopa but it did it in a way this kind of systems was successfully implemented that is less time-consuming and less expensive than from the fields of biology and environment to fault MRI and other medical expert dependant solutions. detection and diagnostics as by Kar et al. [26]. In another case ANN and Neuro-Fuzzy based models were applied to classify earthquake Neuro-Fuzzy applications are widely used for damages in buildings which can be utilized to help technical diagnosis and measurement purposes; engineers decide whether some structures are however, neural and fuzzy methods are often used remained safe or not [22]. The presented individually, too. For instance Bilski used an classification models used 20 seismic parameters of artificial intelligence-based model for diagnostics of accelerograms to estimate 4 damage categories. analog systems [27]. He preprocessed the training and testing data sets using statistical methods to minimize the amount of information to be measured in order to optimize the performance of the Artificial Neural Network (ANN) diagnostic modules. In another application Catelani and Ciani analysed the problem of disturbance induced by high energy particles on electronic devices and developed a model to determine whether a system respond to specific requirements [28]. 2.4. Overview of Neuro-Fuzzy System applications to Technical Diagnostics and Measurement Table 1 gives a comprehensive structure to all overviewed applications appointing the main Fig. 8. Artificial seismic accelerogram processed to
functionalities and application fields and the related classify earthquake damages in buildings [22]
Table 1
Applications functionalities, with their fields and related publications
Application field
Authors and Reference
Turbocharger state, air pressure of vehicle wheels Monitoring and
Fenton-like oxidation efficiency Pouran et al. [18] supervision
fluid density for a gamma ray densitometer Roshani et al. [19] Analogue systems Continuous stirred tank reactor process Zhang and Morris [2] Nuclear reactors Evsukoff and Gentil [6] Induction motors Guzinski et al. [13] Fault diagnosis
Rotating machinery Zio and Gole [10] Railway track circuits Chen et al. [11] Distillation column Karimi and Salahshoor [14] Series compensated transmission lines Disturbance induced by high energy particles Catelani and Ciani [28] Adaptive filtering
Oscillatory signal Sakuntala Mahapatra et al. [3] Rotary hammer drill Microwave biodiesel reactors Wali et al. [15] Pattern identification
Amaral et al. [9] System state modelling
Flexible robotic arm Amitava et al. [12] Volatile organic compounds Khoshnevisan et al. [23] Variable estimation
Hamid Taghavifar et al. [24] Meandering channel Moharana et al. [25] Variable change
Sleep EEG activity during hypopnoea episodes Übeyli et al. [20] detection
Parkinsonian resting tremors Lee and Lim [21] Damage classification
Earthquake damages in buildings Alvanitopoulos et al. [22] inspired by the behaviour of biological neural 3. Prelude of Neuro-Fuzzy Systems
networks inside the human brain. An ANN implements the functionality of the biological This section discusses the techniques that provided neural networks by building up a network of the theoretical basis and allowed the concept of autonomous computational units (neurons) and Neuro-Fuzzy system to be formed. These connecting them via weighted links defined by the techniques are the Neural Networks and the Fuzzy first pioneers W. S. McCulloch and W. Pitts [29]. Systems which will be presented in the following Such a computational unit e.g. an artificial neuron is shown on Fig. 9. It consists of input links (special neurons sometimes don't have any input links), a 3.1. Artificial Neural Networks (ANNs) transfer function, an activation function and an optional memory component while the weighted links between them are represented by real numbers. In Fig. 9. x1-xn denotes the input values of the neuron, oj denotes the output value and w1j-wnj denotes the weights of the input links. The transfer function aggregate the outputs of the other neurons connected via the input links weighted with the strength of the links. The activation function produces the output value based on the output of the transfer function and the memory serves as a container to store previous states of the neuron (in Fig. 9. An artificial neuron
many cases it is not used or only a part of the state is stored). When a neuron fires the signal The concept of the ANN was established seven propagates through the output links to the decades ago and, as the name suggests, it was connected neurons and the weight of a link (represented by a real number) determines the strength of the connection and weakens the signal Yet another class of ANNs is the Radial Basis Function Network (RBFN) which uses radial basis In 1957 Frank Rosenblatt created the perceptron functions as the activation function of the neurons algorithm for supervised classification of an input [33]. Their training algorithms are extended to into one of two possible outputs [30]. This is a type adjust not only the network's weights but the of linear classifier and at the time, only the single- activation function parameters, too. They have layer perceptron could be trained. many uses such as function approximation, time 18 years later Werbos created the backpropagation series prediction, classification, etc. algorithm for training the MLP [31]. Over the decades ANNs proved to be powerful computational models for solving complex estimation and classification problems as they are robust and are capable of high level generalization, moreover they can already handle incomplete data, too [34]. However no information can be extracted from a trained ANN about the connections between the parameters, e.g. a generic ANN model can only approximate the output parameters but cannot tell what kind of connections exist between the input and output parameters. This is a key disadvantage of the Neural Network model which led to the Fig. 10. The MuliLayer Perceptron model
creation of Neuro-Fuzzy Systems. Fig. 10. shows an MLP model where the neurons 3.2. Fuzzy Systems are organized into layers and each layer is fully connected with the next one. Supervised training of Real life problems often have the tendency to be not an MLP means repeated adjustment of the weight discrete but continuous in nature. A somewhat of each link to receive more and more favourable special case of this phenomenon is to categorize output on specific neurons (output neurons) while objects or theoretical entities because in many stimulating other neurons (input neurons). The cases, categories don't have precisely defined backpropagation algorithm achieves this by criteria of membership. To solve this problem Lofti calculating the derivatives of the network's error Zadeh [35] introduced fuzzy set theory, where the with respect to all of its weights and adjusting the membership of an element is no longer a binary weights to a position where, based on the state but a continuous value e.g. instead of saying derivatives, the error is smaller e.g. moving the that a is an element of A set and not element of B, weights in the direction of the descent of the we can say that a is an element of A fuzzy set by derivatives where the error is a measure of the 0.67 degree and element of B by 0.23 degree. difference between the network's output and the Fuzzy logic is a type of logic that uses fuzzy sets to target values for the same input. represent truth values and consequently it provides After 1975 the MLP became more popular and an effective way to represent human knowledge in a widespread and during the years many other, but mathematical language. Fuzzy logic uses fuzzy not so highly popular ANN model types were inference rules which are able to process the defined, moreover, currently the basic research in continuous truth values and produce an also the field of neural networks is emphasising on the continuous output. Each rule has the form of study of biological neural systems and define new learning algorithms and architectures that are maps if <premise> then <consequent>, of the biological brain systems. One of these new types is the Self-Organizing Map (SOM) which can that uses linguistics variables with symbolic terms. be trained by unsupervised learning to produce a Each term represents a fuzzy set. The terms of the low dimensional representation of the input space input space (typically 5-7 for each linguistic [32]. They are mainly used for visualizing high variable) compose the fuzzy partition. The fuzzy dimensional data in low dimensional views. inference mechanism consists of three stages: in the Another type is the Recurrent Neural Network first stage, the values of the numerical inputs are (RNN), they are different from the common mapped by a function according to a degree of feedforward networks as they allow circles in their compatibility of the respective fuzzy sets; this structure e.g. some links propagate the signal back operation can be called fuzzification. In the second to such neurons that sent the original signal. This stage, the fuzzy system processes the rules in feature allows them to establish an internal memory accordance with the firing strengths of the inputs. which can be used to process arbitrary sequences of In the third stage, the resultant fuzzy values are

transformed again into numerical values; this architecture is called ANFIS and it uses the Takagi- operation can be called defuzzification. Essentially, Sugeno-Kang inference system. this procedure makes possible the use fuzzy categories in representation of words and abstracts ideas of the human beings in the description of the decision taking procedure [36]. Fuzzy inference systems have two main type based on the mathematical calculation of the inference. These are the Mamdani type inference [37] and the Takagi-Sugeno-Kang (TSK) type inference [38]. A Mamdani type fuzzy rule can be described as if A is X1 and B is X2 then C is X3,
where A, B, C are variables and X Fig. 11. The ANFIS architecture [39] where x and y
fuzzy sets. In contrast to the Mamdani type, a TSK denote the input variables and z denotes the output rule has the form of if A is X
Fig. 11. shows the ANFIS architecture consisting of 1 and B is X2 then C = aA + bB + c,
six layers. The first layer contains two nodes for where a, b and c are constants. As a result of the input x and y, the second layer is responsible for form of the rules the Mamdani type inference mapping input values to the membership functions. systems are more interpretable because both the The nodes of the third layer correspond to the fuzzy premises and consequents of the rules are fuzzy sets rules in the form of production functions; their while the Takagi-Sugeno-Kang types are more output values are the firing strengths of each rule accurate and computationally efficient, e.g. they while the nodes in the fourth layer calculate the build up more accurate models, however, here, only ratio to the sum of all rules' firing strengths. the premises of the rules are fuzzy sets. Defuzzification happens in the fifth layer and the All in all, Fuzzy Systems have the advantage that sixth layer's output nodes sum their input values. the fuzzy rules, which store the information, are Iterative learning of ANFIS is composed of two easily interpretable. Furthermore they provide a stages. In the first stage the parameters of the simple interface for extending the system with new consequent functions (in the fifth layer) are tuned information (by adding new rules) or manipulating via a least mean square method. During the second the existing rules. The problem with Fuzzy Systems stage the parameters of the premise functions (in lies in the fact that they completely depend on the the second layer) are adjusted by a backpropagation experts who design them. It only uses the algorithm. These two stages are repeated iteratively information which were encoded in the system and for training of the system. It is also worth to cannot learn on its own and it is incapable of mention that this model has the best estimation generalization. The described nature of Fuzzy accuracy based on various benchmarking and Systems indicates that a fusion with ANNs may application results. possibly lead to a new powerful computational 4.2. FALCON Architecture 4. Neuro-Fuzzy System Architectures
Approximately in the same time as ANFIS, the Fuzzy Adaptive Learning Control Network The previous section briefly described the concept (FALCON) architecture was introduced, which is a of the two main components building up a Neuro- system with five layers and uses Mamdani type Fuzzy system individually, so in this section the different architectures can be discussed to show Fig. 12. shows the FALCON architecture. Input how different approaches managed to combine nodes are located in the first layer; second layer has ANNs with Fuzzy Systems. At the end of this term nodes which represent the membership section a table is also presented summarizing the functions for the input values. Each node of the advantages and drawbacks of each presented third layer acts as a fuzzy rule. The fourth layer also consists of term nodes; these represent the membership functions for the outputs. Finally the 4.1. ANFIS Architecture fifth layer is the output layer; here for every output there are two nodes: one is for training data which One of the first Neuro-Fuzzy Systems was is the desired output and the other is for decision introduced by Jang in year 1993 [39][40]. This signal which is the actual output.

Fig. 13. shows the ASN component. The ASN is a five layer network which is responsible for selecting an action based on the current state of the system using fuzzy inference. Input nodes are in the first layer and the second one holds the membership functions. Each node in the third layer represents a fuzzy rule and nodes of the forth layer correspond to consequent labels, e.g. if a consequent label is in a rule then there is a link between the label's node and the rule's node. The fifth layer's nodes calculate the real output values based on the rules' firing strength and the forth layer's outputs. The AEN component of this architecture is a simple feedforward network which predicts reinforcements based on the state variables of the system. And the SAM component stochastically generates an action from based on the recommendation of the AEN GARIC uses gradient descending and reinforcement learning to adjust its internal parameters. 4.4. NEFCON Architecture In parallel with the models mentioned earlier, the Fig. 12. The FALCON architecture
Neural Fuzzy Controller (NEFCON) architecture had been created, which has three layers and Training is done by a two-phase-algorithm. The implements a Mamdani type inference system [43]. first phase is responsible for finding the initial membership functions by a self-organized learning scheme. In the second phase the parameters of the membership functions are adjusted using supervised learning. During the training nodes and links can be deleted or merged reforming the structure of the network. 4.3. GARIC Architecture Another early Neuro-Fuzzy model is the Generalized Approximate Reasoning based Intelligence Control (GARIC) system, which is composed of three components: the Action Selection Network (ASN), the Action Evaluation Fig. 14. The NEFCON architecture where x
Network (AEN) and the Stochastic Action Modifier denote the input variables and c denotes the output Fig. 14. Shows the NEFCON architecture where the circles indicate the nodes which are forming the layers and the rectangles indicate the shared weights of the network. The first layer consists of the input nodes, in the second layer the nodes represent the fuzzy rules and the third layer holds the output nodes. In this architecture the links connecting the nodes are weighted with fuzzy sets. The learning procedure uses reinforcement learning with backpropagation algorithm to either learn the Fig. 13. The ASN component of the GARIC
rule base from the beginning or to optimise an architecture [42] initially defined rule base. Two other systems were developed based on NEFCON which are

specialized versions of the original architecture. These systems are the NEFCLASS [44] which is specialized in classification problems and the NEFPROX [45] which was created for function approximation. 4.5. SONFIN Architecture Self-Constructing Neural Fuzzy Inference Network (SONFIN) is a Takagi-Sugeno-Kang-type fuzzy rule-based model which consists of six layers [46]. Fig. 16. The dmEfuNN architecture
Fig. 16. shows the dmEfuNN architecture. The first layer contains the input nodes and membership functions are in the second layer. Fuzzy rules are represented by the nodes in the third layer. Fourth layer selects a number of rules from the third layer which are the closest to the fuzzy inputs and the fifth layer does the defuzzification and produces Fig. 15. The SONFIN architecture [46] where x1
The dmEfuNN can optimize global generalization and x2 denote the input variables, y1 denotes the error and local generalization error in contrast to output variable, R1-R3 denotes the rule nodes and x MLP and ANFIS which can only optimize global represents the input vector error. As the name suggests the number of nodes and links in the structure can dynamically increase Fig. 15. shows the SONFIN architecture which, in or decrease during the on-line learning while off- fact, is similar to the ANFIS. Layer 1-4 and 6 are line training uses a given structure and optimizes functioning as they are in the ANFIS architecture. internal parameters. The fifth, consequent layer can hold two types of nodes. The first type represents the fuzzy sets by membership functions while the second type is optional and gains its inputs from the first and fourth layer. Constructing of SONFIN happens concurrently by a structure and a parameter learning method. The structure learning identifies both the precondition and consequent parts of the rules by minimizing the number of rules and membership functions for the input and by optimally generating new membership functions for the output variables. Parameter learning uses LMS or RLS algorithms to adjust consequent parameters and backpropagation for precondition parameters. 4.6. dmEfuNN Architecture The Dynamic Evolving Fuzzy Neural Network (dmEfuNN) is a system with five layers which uses the Takagi-Sugeno fuzzy inference mechanism [47]. The predecessors of this model are FuNN [48] and EFuNN [49] which both uses the Mamdani 4.6. Comparison of the Neuro-Fuzzy architectures overviewed Neuro-Fuzzy system types appointing the main advantages and drawbacks of each Table 2 gives a comprehensive comparison to all
Table 2
Advantages and drawbacks of the presented architectures.
Architecture
Advantages Drawbacks
 As it implements a Takagi-Sugeno-  Cannot handle multiple output systems Kang inference mechanism, it is a  Only fully defined structures can be trained very accurate model (most accurate  There is no dynamic rule creation or reduction among the presented architectures and more accurate than Multi-Layer Perceptron and pure Fuzzy systems)  It has a learning phase for building up  Its Mamdani type inference makes it less its initial structure (number of rules accurate especially for parameter estimation and membership functions can be determined via training)  Its Mamdani type inference makes it more interpretable  It is one of the earliest presented  Isn't a single model but consists of multiple Neuro-Fuzzy system  It is able to reduce the number of  Its Mamdani type inference makes it less rules during its training accurate especially for parameter estimation  Its Mamdani type inference makes it more interpretable  There are no initial rules but rules are  Not as accurate as the ANFIS (which has the created and adapted as on-line most similar structure to this) learning proceeds via simultaneous structure and parameter identification  The number of generated rules and membership functions is small even for modelling of a sophisticated system  During its on-line training it can  Not as accurate as the ANFIS increase or decrease the number of rules in the system
5. Conclusions
work differently in different parts of the parameter space providing a more detailed and somehow Different applications of Neuro-Fuzzy Systems distributed model instead of a general solution. were discussed to show their high potential in The paper briefly reviewed the concept of Artificial technical diagnostics and measurement. This survey Neural Networks and Fuzzy Systems as summarizes in a comprehensive overview the computational models and how they inspired the Neuro-Fuzzy applications in technical diagnostics creation of Neuro-Fuzzy Systems. As it was and measurement with appointing the generalized, discussed this fusion can unite the generalization main, typical functionalities and with highlighting capabilities of Neural Networks with the easy their great variety of application areas. interpretability and high expressive power of fuzzy These systems are successful because of their rules in an effective way. nature that they reveal the nature of the important Six different architectures were presented and it can interdependence between the parameters of the be concluded that these are the most important ones modelled system while they are, in fact, powerful although there are other structure variations, too. approximators. Their ability to discover Usually each architecture organizes its nodes a connections between parameter intervals can be slightly different way and consequently they use extremely useful when applying the model for specific learning algorithms which are adapted to diagnostic and control tasks because entirely the different structures. The different Neuro-Fuzzy different rules can be used to specific subsets of the models were also compared and a table is also problem, e.g. these systems have the capability to presented summarizing the advantages and drawbacks of each presented architecture. ANFIS combination with GAs", Mechanical All in all it can be said that ANFIS architecture is Systems and Signal Processing, vol. 21: (5), the most popular and widespread among the Neuro- pp. 2280-2294, 2007. Fuzzy systems for various applications in the field [9] T. G. Amaral, V. F. Pires, J. F. Martins, A. J. of diagnostics, control or for medical research, civil Pires, M. M. Crisóstomo, "Image processing engineering, etc. This is mainly because the ANFIS to a neuro-fuzzy classifier for detection and model has higher accuracy than the other Neuro- diagnosis of induction motor status fault", 33rd Fuzzy model types which compensates its less Annual Conference of the IEEE Industrial interpretable structure. Electronics Society, Taipei, Taiwan, 2007. [10] E. Zio, G. Gola, "A neuro-fuzzy technique for fault diagnosis and its application to rotating machinery", Reliability Engineering and The authors acknowledge the support of grants of System Safety, vol. 94, pp. 78-88, 2009. the Fraunhofer Project Center for Production [11] J. Chen, C. Roberts, P. Weston, "Fault Management and Informatics at SZTAKI, detection and diagnosis for railway track Budapest, Hungary and the Highly industrialised circuits using neuro-fuzzy systems", Control region on the west part of Hungary with limited Engineering Practice, vol. 16: (5), pp. 585- R&D capacity: Research and development programs related to strengthening the strategic [12] Amitava Chatterjee, Ranajit Chatterjee, future-oriented industries manufacturing Fumitoshi Matsuno, Takahiro Endo, "Neuro- technologies and products of regional competences fuzzy state modelling of flexible robotic arm carried out in comprehensive collaboration, employing dynamically varying cognitive and VKSZ_12-1-2013-0038, with iKOMP acronym. social component based PSO", Measurement, vol. 40, pp. 628-643, 2007. References
[13] J. Guzinski, H. Abu-Rub, A. Iqbal, S. M. Ahmed, "Shaft misalignment detection using [1] M. Ayoubi, "Fuzzy systems design based on a ANFIS for speed sensorless AC drive with hybrid neural structure and application to the inverter output filter", IEEE International fault diagnosis of technical processes", Symposium on Industrial Electronics, Gdansk, Control Engineering Practice, vol. 4: (1), pp. [14] I. Karimi, K. Salahshoor, "A new fault [2] J. Zhang, J. Morris, "Process modelling and detection and diagnosis approach for a fault diagnosis using fuzzy neural networks", distillation column based on a combined PCA Fuzzy Sets and Systems, vol. 79, pp. 127-140, and ANFIS scheme", 24th Chinese Control and Decision Conference, Taiyuan, China, 2012. [3] Sakuntala Mahapatra, Santanu K. Nayak, [15] W. A. Wali, K. H. Hassan, J. D. Cullen, A. Samrat L. Sabat, "Neuro fuzzy model for Shaw, A. I. Al-Shamma's, "Real time adaptive filtering of oscillatory signals", monitoring and intelligent control for novel Measurement, vol. 30, pp. 231-239, 2001. advanced microwave biodiesel reactor", [4] C. W. Frey, A. Jacubasch, H.-B. Kuntze, Measurement, vol. 46, pp. 823-839, 2013. „Smart neuro-fuzzy based control of a rotary [16] Huseyin Eristi, "Fault diagnosis system for hammer drill", IEEE International Conference series compensated transmission line based on on Robotics and Automation, vol. 2, pp. 2634- wavelet transform and adaptive neuro-fuzzy inference system", Measurement, vol. 46, pp. [5] W. Wang, F. Ismail, A. F. Golnaraghi, "A neuro-fuzzy approach to gear system [17] Sunil K. Jha, Kenshi Hayashi, R. D. S. monitoring", IEEE Transactions on Fuzzy Yadava, "Neural, fuzzy and neuro-fuzzy Systems, vol. 12: (5), pp. 710-723, 2004. approach for concentration estimation of [6] A. Evsukoff, S. Gentil, "Recurrent neuro- volatile organic compounds by surface fuzzy system for fault detection and isolation acoustic wave sensor array", Measurement, in nuclear reactors", Advanced Engineering vol. 55, pp. 186-195, 2014. Informatics, vol. 19, pp. 55-66, 2005. [18] Shima Rahim Pouran, A. R. Abdul Aziz, Wan [7] Z. Ye, A. Sadeghian, B. Wu, "Mechanical Mohd Ashri Wan Daud, Shahaboddin fault diagnostics for induction motor with Shamshirband, "Estimation of the effect of variable speed drives using Adaptive Neuro- catalyst physical characteristics on Fenton-like fuzzy Inference System", Electric power oxidation efficiency using adaptive neuro- System Research, vol. 76, pp. 742-752, 2006. fuzzy computing technique", Measurement, [8] Y. Lei, Z. He, Y. Zi, Q. Hu, "Fault diagnosis vol. 59, pp. 314-328, 2015. of rotating machinery based on multiple [19] G. H. Roshani, S. A. H. Feghhi, A. Adineh- Vand, M. Khorsandi, "Application of adaptive for Prediction and Analysis in the Behaviour neuro-fuzzy inference system in prediction of Sciences, Ph. D Thesis, Harward University, fluid density for a gamma ray densitometer in Cambridge, 1974. petroleum products monitoring", [32] T. Kohonen, Self-organized Formation of Measurement, vol. 46, pp. 3276-3281, 2013. Topologically Correct Feature Maps, [20] E. D. Übeyli, D. Cvetkovic, G. Holland, I. Biological Cybernetics, 43, pp. 59-69, 1982. Cosic, "Adaptive neuro-fuzzy inference [33] D. S. Broomhead, D. lowe, Radial basis system employing wavelet coefficients for functions, multi-variable functional detection of alterations in sleep EEG activity interpolation and adaptive networks, RSRE during hypopnoea episodes", Digital Signal memorandum No. 4148, 1988. Processing, vol. 20, pp. 678-691, 2010. [34] Zs. J. Viharos, K. B. Kis, "Diagnostics of [21] Sang-Hong Lee, Joon S. Lim, "Comparison of wind turbines based on incomplete sensor DBS and levopoda on resting tremor using a data", IMEKO World Congress Technical fuzzy neural network system", Measurement, Diagnostics, Republic of Korea, 2012. vol. 46, pp. 1995-2002, 2013. [35] L. A. Zadeh, "Fuzzy Sets", Information and [22] P. F. Alvanitopoulos, I. Andreadis, A. Elenas, Control, vol. 8, pp. 338-353, 1965. "Neuro-fuzzy techniques for the classification [36] J. Vieira, F. M. Dias, A. Mota, "Neuro-Fuzzy of earthquake damages in buildings", Systems: A Survey", 5th WSEAS NNA Measurement, vol. 43, pp. 797-809, 2010. International Conference, 2004. [23] Benyamin Khoshnevisan, Shahin Rafiee, [37] E. H. Mamdani, S. Assilian, "An experiment Mahmoud Omid, Hossein Mousazadeh, in linguistic synthesis with a fuzzy logic "Prediction of potato yield based on energy controller", International Journal of Man- inputs using multi-layer adaptive neuro-fuzzy Machine Studies, vol. 7: (1), pp. 1-13, 1975. inference system", Measurement, vol. 47, pp. [38] T. Takagi, M. Sugeno, "Fuzzy identification of systems and its applications to modeling [24] Hamid Taghavifar, Aref Mardani, and control", IEEE Trans, on Systems, Man "Prognostication of vertical stress and Cybernetics, vol. 15, pp. 116-132, 1985. transmission in soil profile by adaptive neuro- [39] R. Jang, "Fuzzy Modeling Using Generalized fuzzy inference system based modelling Neural Networks and Kalman Filter approach", Measurement, vol. 50, pp. 152- Algorithm", Proc. Ninth Nat. Conf. Artificial Intell., pp. 762-767, 1991. [25] S. Moharana, K. K. Khatua, "Prediction of [40] R. Jang, "ANFIS: Adaptive-Network-Based roughness coefficient of a meandering open Fuzzy Inference System", IEEE Transactions channel flow using Neuro-Fuzzy Inference on systems, man, and cybernetics, vol. 23: (3), System", Measurement, vol. 51, pp. 112-123, pp. 665-685, 1993. [41] T. C. Lin, C. S. Lee, "Neural Network Based [26] S. Kar, S. Das, P. K. Ghosh, "Applications of Fuzzy Logic Control and Decision System", neuro fuzzy systems: A brief review and IEEE Transactions on Computers, vol. 40: future outline", Applied Soft Computing, vol. (12), pp. 1320-1336, 1991. 15, pp. 243-259, 2014. [42] H. R. Berenji, P. Khedkar, "Learning and [27] P. Bilski, "Data set processing for the Tuning Fuzzy Logic Controllers Through optimization of the artificial intelligence-based Reinforcements", IEEE Transactions on diagnostic methods", IMEKO TC10 Neural Networks, vol. 3, pp. 724-740, 1992. Workshop on Technical Diagnostics, [43] D. Nauck, R. Kruse, "NEFCON-I: An X- Florence, Italy, 2013. Window Based Simulator for Neural Fuzzy [28] M. Catelani, L. Ciani, "Diagnostic and Error Controllers", Proc. IEEE Int. Conf. Neural Correction System for Avionics Devices in Networks, 1994. Presence of Single Event Upset (SEU)", [44] D. Nauck, R. Kruse, "NEFCLASS – A Neuro- IMEKO TC10 Workshop on Technical Fuzzy Approach For The Classification Of Diagnostics, Florence, Italy, 2013. Data", Applied Computing Proc. of the 1995 [29] W. S. McCulloch, W. Pitts, "A logical ACM Symposium on Applied Computing, pp. calculus of the ideas immanent in nervous activity", Bulletin of Mathematical Biophysics, [45] D. Nuck, R. Kurse, "Neuro-fuzzy systems for vol. 5, pp. 115-133, 1943. function approximation", Fuzzy Sets and [30] F. Rosenblatt, The Perceptron -- a perceiving Systems, vol. 101, pp. 261-271, 1999. and recognizing automaton, Cornell [46] F. C. Juang, T. C. Lin, "An On-Line Self Aeronautical Laboratory, Report Number: 85- Constructing Neural Fuzzy Inference Network and its applications", IEEE Transactions on [31] P. J. Werbos, Beyond Regression: New Tools Fuzzy Systems, vol. 6, pp. 12-32, 1998. [47] N. Kasabov, Q. Song, Dynamic Evolving Fuzzy Neural Networks with ‘m-out-of-n' Activation Nodes for On-Line Adaptive Systems, Technical Report TR99/04, 1999. [48] N. Kasabov, "Adaptable Connectionist Production Systems", Neurocomputing, vol. 13: (2-4), pp. 95-117, 1996. [49] N. Kasabov, "Evolving Fuzzy Neural Networks – Algorithms, Applications and Biological Motivation", Methodologies for the Conception, Design and Application of Soft Computing, pp. 271-274, 1998.

Source: http://eprints.sztaki.hu/8231/1/Viharos_126_2885145_ny.pdf