ART-Type Artificial Neural Networks Applications for ... - Springer Link

6 downloads 0 Views 219KB Size Report
ART-Type Artificial Neural Networks. Applications for Classification of Operational. States in Wind Turbines. *. Tomasz Barszcz1, Andrzej Bielecki2, and Mateusz ...
ART-Type Artificial Neural Networks Applications for Classification of Operational States in Wind Turbines Tomasz Barszcz1 , Andrzej Bielecki2 , and Mateusz W´ojcik3 1 Chair of Robotics and Mechatronics AGH University of Science and Technology, Al. Mickiewicza 30, 30-059 Krak´ ow, Poland 2 Institute of Computer Science, Jagiellonian University, L  ojasiewicza 6, 30-348 Krak´ ow, Poland 3 Department of Computer Design and Graphics, Jagiellonian University, Reymonta 4, 30-059 Krak´ ow, Poland [email protected], [email protected], [email protected]

Abstract. In recent years wind energy is the fastest growing branch of the power generation industry. The largest cost for the wind turbine is its maintenance. A common technique to decrease this cost is a remote monitoring based on vibration analysis. Growing number of monitored turbines requires an automated way of support for diagnostic experts. As full fault detection and identification is still a very challenging task, it is necessary to prepare an ”early warning” tool, which would focus the attention on cases which are potentially dangerous. There were several attempts to develop such tools, in most cases based on various classification methods (predominantly neural networks). Due to very common lack of sufficient data to perform training of a method, the important problem is the need for creation of new states when there are data different from all known states. As the ART neural networks are capable to perform efficient classification and to recognize new states when necessary, they seems to be a proper tool for classification of operational states in wind turbines. The verification of ART and fuzzy-ART networks efficiency in this task is the topic of this paper.

1

Introduction

In recent years wind energy is the fastest growing branch of the power generation industry. The average yearly growth in the years 1997-2003 achieved 32% in the United States and 22% in the European Union [2] and these figures will hold for at least the next decade. The distribution of costs during the life cycle of 

The paper was supported by the Polish Ministry of Science and Higher Education under Grant No. N504 147838.

L. Rutkowski et al. (Eds.): ICAISC 2010, Part II, LNAI 6114, pp. 11–18, 2010. c Springer-Verlag Berlin Heidelberg 2010 

12

T. Barszcz, A. Bielecki, and M. W´ ojcik

the unit for wind energy is significantly different from that of traditional, fossil fired units [2]. First of all, initial investment costs are relatively higher, whereas in traditional units cost of fuel plays important role (usually it is the second largest cost). After commissioning, the largest cost for the wind turbine (WT for abbreviation) is maintenance. With proper maintenance policies, wind turbines can achieve the highest level of availability in the power generation sector even up to 98%. The basis of proper maintenance is continuous monitoring of the transmission of the wind turbine. There were several attempts to develop various classification tools, in most cases based on various classification methods (predominantly neural networks). Due to very common lack of sufficient data to perform training of a method, the important problem is the need for creation of new states when there are data different from all known states. In most types of artificial neural networks (ANNs for abbreviation) the learning process is unsuitable for cases of continuous machinery intelligence monitoring. This means, among others, that adding a new patterns as inputs requires repetition of the learning process. In ART networks, introduced by Carpenter and Grossberg [3,4], the learning process is not separated from its operation. Furthermore, ART neural networks are capable to add new states when necessary [8,9,12]. Therefore, this sort of ANNs were tested as a tool for classification of operational states in continuous monitoring of wind turbines. The paper is organized in the following way. In section 2 the problem of classification of wind turbine states is discussed. The used ART networks - classical ART-2 network and fuzzy ART network are briefly described in section 3. The obtained results are presented in section 4.

2

Classification Problem of Wind Turbine States

In recent years large development of monitoring and diagnostic technologies for WTs has taken place. The growing number of installed systems created the need for analysis of gigabytes of data created every day by these systems. Apart from the development of several advanced diagnostic methods for this type of machinery there is a need for a group of methods, which will act as an ”early warning”. The idea of this approach could be based on a data driven algorithm, which would decide on a similarity of current data to the data, which are already known. In other words, the data from the turbine should be accounted for one of known states. If this is a state describing a failure, the human expert should be alarmed. If this is an unknown state, the expert should be informed about the situation and asked for a definition of such a new state. This approach could be called ”the blunt expert”, which maybe sounds strange, but gives the most important feature of the proposed method. This approach may brake the biggest barrier of application of artificial neural networks (ANNs) in diagnostics, which is availability of significant amount of training data. As in real cases it is not possible to acquire it, it is only possible to train ANNs for a few cases covered by available data.

ART-Type ANNs Applications for Classification of Operational States in WT

13

The problem of classification was investigated by several authors. One of the first works was research by Shuhui et al. [11], who compared classification techniques for the wind curve estimation. This work was often referenced by others, but from the ANN point of view it only multilayered feed-forward types of networks. Another important contribution was given by Kim [6], who compared performance of several classification methods. His experiments showed that unless the number of independent variables in the system is low, ANNs perform better than other methods. Again, the investigated network was the multi-layer feed-forward network trained by the back-propagation algorithm. There are no works, known to us, which would consider application of ART networks for the classification of WT states. There were also works applying wind turbines for wind and power generation prediction, but this issue is outside the scope of this paper. As the ART networks are capable to perform efficient classification and to recognize new states when necessary [3,4], we performed research of initial classification task. The goal of the experiment was verification of ART classification capabilities with comparison to the human expert. This type of data is acquired in the majority of cases and the successful classifier should create a reasonable number of classes, similar to these by a human expert. This task is the main goal of the following paper. As such a classification was shown, it is thus possible to filter out states, which are known to be correct. The expert can then focus only on ”suspicious” states returned by the algorithm.

3 3.1

Characteristics of the Applied Neural Networks ART-2 Network

Let us briefly recall ART-2 neural network properties tracing [10]. The ART-2 is an unsupervised neural network with based on adaptive resonance theory (ART). A typical ART-2 architecture, introduced in Carpenter and Grossberg (1987), is presented in Fig.1 (only one unit of each type is shown here). In the attentional sub-system, an input pattern s is first presented to the F1 layer, which consists of six kinds of units - the W, X, U, V, P and Q cells. It then undergoes a process of activation, including normalization, noise suppression and updating. This results in an output pattern p from the F1 layer. Responding to this output pattern, an activation is produced across F2 layer through bottom-up weights bij . As the F2 layer is a competitive layer with a winner-takes-all mode, only one stored pattern is a winner. It also represents the best matching pattern for the input pattern at the F1 layer. Furthermore, the pattern of activation on the F2 layer brings about an output pattern that is sent back to the F1 layer via top-down weights tji . For the orienting sub-system, it contains a reset mechanism R and a vigilance parameter q to check for the similarity between the output pattern from the F2 layer and the original input pattern from the F1 layer. If both patterns are concordant, the neural network enters a resonant state where the adaptation of the stored pattern is conducted.

14

T. Barszcz, A. Bielecki, and M. W´ ojcik

Otherwise, the neural network will assign an uncommitted (inhibitory) node on the F2 layer for this input pattern, and thereafter, learn and transform it into a new stored pattern.

Fig. 1. ART-2 architecture

3.2

Fuzzy ART Network

The organization of a fuzzy ART network, introduced by Carpenter et all. [5], is presented in Fig.2. In comparison with a classical ART-2 network, the fuzzy ART network has an additional layer F0 which transforms input vectors using so called complement coding. The fuzzy ART network has a single weights matrix Z, which processes signals being sent both from F1 to F2 layer and vice versa. Operations done by the network are based on the fuzzy logic. The operator fuzzy AND is used for two vectors comparison - see [7]. Signal processing in fuzzy ART network is similar to processing in ART-2 network. The input signal vector, say X, is transformed by F0 and F1 layers producing signals Tj which are put to inputs of the F2 layer. For the neuron which is excited most strongly a vigilance parameter q is used to check the similarity between the output pattern from the F2 layer and the input pattern. If p < ρ then the winner neuron is inhibited and other neuron in F2 layer is searched. Otherwise, the weights matrix Z is modified in order to store the recognized vector features. The learning process is continued until the values of the matrix Z are stabilized.

ART-Type ANNs Applications for Classification of Operational States in WT

15

Fig. 2. Fuzzy ART architecture

4

Results

The practical case study was performed on data from one of wind turbines. The data covering the period from 11.09.2009 till 30.09.2009 were recorded every 10 minutes by the online monitoring system. The recorded data were current values and were not averaged. The data set included 2869 measurements. As the main goal of the work was to test applicability of ART-type ANNs for classification, in the first step we tried to use the network to achieve results similar to a human expert. Other training was not possible, as this type of networks performs only unsupervised learning. The data set contained the most fundamental values, deciding about the operational state of the machine. These were: wind speed, rotational speed of the generator and the generated power. They are related, but only to some extent Table 1. A number of classes recognized by ART-2 and fuzzy ART networks in dependence on the vigilance parameter. Used abbreviations: PW - Power and Wind (2D model), SW - Speed and Wind (2D model), PSW - Power, Speed and Wind (3D model). Vigilance parameter 0.4 0.5 0.6 0.65 0.7 0.75 0.8 0.9 0.9497 0.95 0.97 0.98 0,99

ART-2 network Fuzzy ART network PW SW PSW PW SW PSW 1 1 1 1 2 2 1 1 1 2 4 2 1 1 1 2 4 4 1 1 1 2 5 5 1 1 1 3 7 5 1 1 1 4 8 9 1 1 1 6 10 11 1 2 2 14 24 16 2 2 4 2 2 5 3 2 8 3 5 9 4 8 10 -

16

T. Barszcz, A. Bielecki, and M. W´ ojcik

Fig. 3. Results for fuzzy ART, 3D model, vigilance parameter is equal to 0.55. According to a human expert’s opinion 20 measurements were classified incorrectly which 20 = 0.70%. makes an error equal to 2869

Fig. 4. Results for ART-2, 3D model, vigilance parameter is equal to 0.9497. According to a human expert’s opinion 27 measurements were classified incorrectly which makes 27 = 0.94%. an error equal to 2869

and in fact they are all independent variables. The selection of variables is the same as the human expert would use. Typically the operation of the WT can be divided in a few distinct states: stopped, transient between 0 and 1000 rpm, idle load (rpm about 1000 rpm, no load), low power, high power. Sometimes it is not necessary to distinguish all of them and the first two pairs are sometimes regarded as only two states (i.e. ”stopped or transient” and ”low power” including also idle mode). Very important advantage of chosen set is that it has only 3 variables and can be presented in a graphical way. Thus, it can be easily understood and compared with a human expert. Results create the basis and give some intuition for more advanced research. The main idea of the research was to apply recorded data to the ART network and to investigate what is its behavior (i.e. how many states will be created) and how does this depends on the network parameters. The research was performed

ART-Type ANNs Applications for Classification of Operational States in WT

17

for two types of networks: ART-2 and fuzzy ART. It was found that the single most important parameter is the vigilance. Table 1 presents how the number of created states depends on this parameter. It can be seen that for the ART-2 network reasonable results are obtained for vigilance in the vicinity of 0.95. For fuzzy ART the optimum vigilance is around 0.6. Apart from the number of states it was very important to check how the networks actually divide the data into states. It can be seen that both networks gave very interesting results, creating regions very similar to ”natural” ones. In a few cases created regions stretched across a larger area than a human would create. This was most probably result of normalization and should be further investigated. There are sight differences between the networks, but at this stage one can say that both are performing well and could be the basis for further research.

5

Concluding Remarks

Presented results belong to a broader research activity, aimed at automatic monitoring of rotating machinery. We are interested in investigation of several approaches, which can be applied in the engineering practice. Thus, one has to assume that learning sets are not available or cover only a part of machine states.The problem becomes much more the classification of the current state to one of previously known states or detection of a new state. Ideally, such a new state should be included for further classifications. The initial attempt to the problem was investigation of PNN networks [1]. The study revealed classification capabilities much better than MLP, but handling of new states required retraining of the network. The approach presented in this paper is much more interesting. To start with simple problems, we showed that ART-2 networks are capable to classify typical states of a wind turbine with efficacy very close to the human expert. Such a result is, in our opinion, very interesting and encouraging for further research. Further works will be conducted in two directions. Firstly, the dimensionality of the problem will be increased (i.e. more measurement channels will be taken into account). Secondly, the data with faults will be used to train the network.

References 1. Barszcz, T., Bielecki, A., Romaniuk, T.: Application of probabilistic neural networks for detection of mechanical faults in electric motors. Electrical Review 8/2009, 37–41 (2009) 2. Barszcz, T., Randall, R.B.: Application of spectral kurtosis for detection of a tooth crack in the planetary gear of a wind turbine. Mechanical Systems and Signal Processing 23, 1352–1365 (2009) 3. Carpenter, G.A., Grossberg, S.: A massively parallel architecture for a selforganizing neural pattern recognition machine. Computer Vision, Graphics, and Image Processing 37, 54–115 (1987)

18

T. Barszcz, A. Bielecki, and M. W´ ojcik

4. Carpenter, G.A., Grossberg, S.: ART2: self-organization of stable category recognition codes for analog input pattern. Applied Optics 26, 4919–4930 (1987) 5. Carpenter, G.A., Grossberg, S., Rosen, D.B.: Fuzzy ART: Fast stable learning and categorization of analog patterns by an adaptive resonance system. Neural Networks 4, 759–771 (1991) 6. Kim, Y.S.: Performance evaluation for classification methods: A comparative simulation study. Expert Systems with Applications 37, 2292–2306 (2010) 7. Knosala, R.: Artificial Intelligence Methods Applications in Manufacturing Engineering. WNT, Warsaw (2002) (in Polish) 8. Korbicz, J., Obuchowicz, A., Uci´ nski, D.: Artificial Neural Networks. Foundations and Applications. Academic Press PLJ, Warsaw (1994) (in Polish) 9. Rutkowski, L.: Neural networks and Neurocomputers. Technical University in Cz¸estochowa Press, Cz¸estochowa (1996) (in Polish) 10. Shieh, M.D., Yan, W., Chen, C.H.: Soliciting customer requirements for product redesign based on picture sorts and ART2 neural network. Expert Systems with Applications 34, 194–204 (2008) 11. Shuhui, L., Wunsch, D.C., O’Hair, E., Giesselmann, M.G.: Comparative analysis of regression and artificial neural network models for wind turbine power curve estimation. Journal of Solar Energy Engineering 123, 327–332 (2001) 12. Tadeusiewicz, R.: Neural Networks. Academic Press, Warsaw (1993) (in Polish)