Intelligent Computing Proceedings of the 2021 Computing Conference Volume 2 Lecture Notes in Networks and Systems 284 Kohei Arai (Editor) instant download
Intelligent Computing Proceedings of the 2021 Computing Conference Volume 2 Lecture Notes in Networks and Systems 284 Kohei Arai (Editor) instant download
https://ebookmeta.com/product/intelligent-computing-proceedings-
of-the-2021-computing-conference-volume-3-lecture-notes-in-
networks-and-systems-285-kohei-arai-editor/
https://ebookmeta.com/product/intelligent-computing-proceedings-
of-the-2021-computing-conference-volume-1-1st-edition-kohei-arai/
https://ebookmeta.com/product/wolf-pack-the-u-boats-at-war-
hitler-s-war-machine-bob-carruthers/
The Gaslighting Recovery Workbook Healing From
Emotional Abuse 1st Edition Amy Marlow Macoy
https://ebookmeta.com/product/the-gaslighting-recovery-workbook-
healing-from-emotional-abuse-1st-edition-amy-marlow-macoy/
https://ebookmeta.com/product/throttle-the-driven-world-1st-
edition-p-sawyer/
https://ebookmeta.com/product/dirty-queen-covington-high-3-1st-
edition-amelia-winters/
https://ebookmeta.com/product/political-survival-politicians-and-
public-policy-in-latin-america-barry-ames/
Marketing Management Fifteenth edition Philip Kotler
https://ebookmeta.com/product/marketing-management-fifteenth-
edition-philip-kotler/
Lecture Notes in Networks and Systems 284
Intelligent
Computing
Proceedings of the 2021 Computing
Conference, Volume 2
Lecture Notes in Networks and Systems
Volume 284
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Fernando Gomide, Department of Computer Engineering and Automation—DCA,
School of Electrical and Computer Engineering—FEEC, University of Campinas—
UNICAMP, São Paulo, Brazil
Okyay Kaynak, Department of Electrical and Electronic Engineering,
Bogazici University, Istanbul, Turkey
Derong Liu, Department of Electrical and Computer Engineering, University
of Illinois at Chicago, Chicago, USA; Institute of Automation, Chinese Academy
of Sciences, Beijing, China
Witold Pedrycz, Department of Electrical and Computer Engineering,
University of Alberta, Alberta, Canada; Systems Research Institute,
Polish Academy of Sciences, Warsaw, Poland
Marios M. Polycarpou, Department of Electrical and Computer Engineering,
KIOS Research Center for Intelligent Systems and Networks, University of Cyprus,
Nicosia, Cyprus
Imre J. Rudas, Óbuda University, Budapest, Hungary
Jun Wang, Department of Computer Science, City University of Hong Kong,
Kowloon, Hong Kong
The series “Lecture Notes in Networks and Systems” publishes the latest
developments in Networks and Systems—quickly, informally and with high quality.
Original research reported in proceedings and post-proceedings represents the core
of LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as new
challenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks,
spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor
Networks, Control Systems, Energy Systems, Automotive Systems, Biological
Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems,
Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems,
Robotics, Social Systems, Economic Systems and other. Of particular value to both
the contributors and the readership are the short publication timeframe and the
world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.
The series covers the theory, applications, and perspectives on the state of the art
and future developments relevant to systems and networks, decision making, control,
complex processes and related areas, as embedded in the fields of interdisciplinary
and applied sciences, engineering, computer science, physics, economics, social, and
life sciences, as well as the paradigms and methodologies behind them.
Indexed by SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago.
All books published in the series are submitted for consideration in Web of
Science.
Intelligent Computing
Proceedings of the 2021 Computing
Conference, Volume 2
123
Editor
Kohei Arai
Faculty of Science and Engineering
Saga University
Saga, Japan
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Editor’s Preface
v
vi Editor’s Preface
Hope to see you in 2022, in our next Computing Conference, with the same
amplitude, focus and determination.
Kohei Arai
Contents
vii
viii Contents
Matin Pirouz(B)
1 Introduction
Networks are groups of nodes inter-connected with edges. For example in social
media networks, nodes represent people and edges represent the relationship
among people [10]. There exist some subgroups with more internal edges than
external edges, commonly identified as communities. One of the main tasks in
studying complex networks is to find these structures. Finding communities can
result in finding meaningful information, and is particularly important in social
networks [15] and recommender systems [3, 19].
Several community detection methods exist. Modularity, a quality-function
introduced by Newman, has become one of the major algorithms for which there
have been many optimization methods. On the other hand, the resolution lim-
its [12] problem found by Fortunato and Barthélemy in [1] is a key limitation
of modularity. A metric called resolution value has been used to solving the
problem, where the higher the value, the more communities are found [7].
Another widely known approach is Label Propagation developed by Ragh-
van et al. [12]. Label Propagation is a near linear approach, which is known as
the coloring method as it is based on the flow of uniquely assigned labels in
the network. The problem with this method is that the core method is based
on random distribution; therefore, many different scenarios of answer could be
resulted. Several enhancements have been introduced over the years to solve the
problem and are further studied in the related work section.
In this paper, a new node influence algorithm based on Label Propagation
is proposed to solve the randomness and resolution limit problem in community
detection. Balanced Weighted Label Propagation (BWLP) algorithm has a time
complexity linear to the number of edges. Our approach uses the influence of
nodes on one another to find an edge value for any pair of nodes in any com-
plex graph. The neighboring relationship between two nodes is used as a metric
to define the closeness of the pair. We develop a convergence function which
defines when to stop exchanging labels. Finally, when the distribution of labels
concludes, all the nodes that share the same label find them selves in the same
community.
1.1 Contributions
2 Related Work
size as the first phase. After the maximum cliques have been found, each clique
is enumerated and a clique adjacency matrix is constructed. A parameter K is
set and any clique with size of at least K is considered to be a “part” of a larger
module. Girvan and Newman method [2] creates communities by removing the
highest edge betweenness and re-calculating influenced edges value again. At the
end, it provides a dendrogram with nodes as the leaves. The resolution of com-
munities in ll the optimization method with a fitness function is dependent on
the resolution value chosen for the given function. Lancichinetti et al. [4] choose
a random node and expands a community from there. Similar to modularity, res-
olution value defines the size and quality of the detected communities depends.
In EAGLE [14] and GCE [6] algorithms, maximal cliques in the network replaces
the random node used in [4].
Zhang et al.’s spectral method [22] is used to lower the dimensions of the
graph. The fuzzy c-mean algorithm is then used for clustering. Psorakis et al.
[11] introduced a novel method utilizing Bayesian non-negative matrix factor-
ization (NMF). The limitations of this method include the need to know the
number of communities and the use of matrices which results in an inefficient
time complexity and increases the space complexity. Zhang et al. [21] used a
stochastic block model to find an unequal size community structure where the
number of nodes inside community should be equal when using stochastic block
models.
Label propagation methods [18] propagate the label of the most popular
nodes throughout the graph to find the structure of communities. Particle Swarm
Optimization (PSO) [23] introduced a new label propagation method mixed
with modularity density. The detected communities have good resolution but the
algorithm has a higher time complexity than other label propagation algorithms.
SLPA introduced a dynamic algorithm to find overlapping community structure.
Li et al. [17] introduced an optimized method called “Stepping LPA-S”, where
similarity for propagating labels is used. In addition, a stepping framework is
used to divide networks. Then, an evaluation function is used to select the final
unique partition.
3 Proposed Approach
This section develops the main idea behind the proposed multi-stage edge weigh-
ing algorithm. In the initialization layer, all the links get their own labels. Assume
a pair of nodes with unique sets of common and exclusive neighbors. Suppose X
and Y are neighbors, then C is the set of their common neighbors and E is the
set of exclusive neighbors between the pair. In this study, we use sets C and E
to find the dissimilarity between nodes X and Y . Function 2 finds a dissimilarity
value between any pair of nodes in the graph based on their structural formats.
The calculated dissimilarity value is assigned to the edge between two nodes as
their edge weight. Algorithm 1 shows the next step which is sorting all nodes
based on their dissimilarity ratio [9].
First, every node in the graph gets their node number as their label. This
way, every node is assigned to its own community. Later, swapping labels
4 M. Pirouz
Lemma 1. If the given algorithm converges on the ring of cliques network where
each clique has the size k>2, then no node will be clustered with nodes outside
of its clique.
Proof. This lemma is proven if we can show that no node has the incentive to
cluster with a node outside of its clique. To this end, let x ∈ V and define C(x)
to be the set of nodes that share a clique with x. Let y be a node in the set
6 M. Pirouz
Algorithm 3: Convergence
1 Input: G network graph, n, Chosen
2 for i ∈ {N eighbors of n} do
3 Chosen = label of neighbor with most similarity
4 m = neighbor with lowest edge value
5 for j ∈ {N eighbors of m} do
6 HAL= Find highest occurrence label
7 end
8 if HAL == Chosen then
9 F lag n as converged
10 end
11 else
12 call Convergence(m, G);
13 end
14 end
N (x) − C(x) and notice that by the construction of the ring of cliques we have
|N (x) ∩ N (y)| = 0. (4)
depending on how we construct the ring of cliques and the choice of x and y
F (x, y)’s value will vary; however
2
min F (x, y) = . (5)
k+1
Let z be any node in C(x). Since z belongs to the same clique as x we have
|N (x) ∩ N (z)| = k − 2. (6)
Again, depending on how we construct the ring of cliques and the choice of x
and z F (x, z)’s value will vary; however
−2(k − 2) 1 1
max F (x, z) = + + . (7)
(k + 2)(k) k + 2 k
√
Finally, we note that max F (x, z) < min F (x, y) as long as k > 12 (1 + 13) since
we assumed that k is an integer that is greater x will never cluster with a node
outside of its clique.
4 Experiment
100000 1.02
Infomap BWLP Infomap BWLP
10000 1
Time in Seconds
0.98
1000
NMI
0.96
100
0.94
10
0.92
1 0.9
5 10 15 20 25 5 10 15 20 25
Average Degree Average Degree
Fig. 1. (a) Run time (seconds) to detect the communities in LFR benchmark datasets
based on varying node degrees. (b) Accuracy of the detected communities using NMI
1.2 1.2
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
1 2 3 4 5 1 2 3 4 5
AƩractor Infomap BWLP Modularity louvain AƩractor Infomap BWLP Modularity louvain
(a) NMI for Small Communities (b) NMI for Large Communities
1.2 1.2
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
1 2 3 4 5 1 2 3 4 5
AƩractor Infomap BWLP Modularity louvain AƩractor Infomap BWLP Modularity louvain
(c) ARI for Small Communities (d) ARI for Large Communities
1000 1000000
100000
100 10000
seconds
seconds
1000
10 100
10
1 1
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Fig. 3. Community detection run time (seconds) in LFR benchmark datasets with
varying mixing parameter µ (horizontal axis)
and from there every node chooses its label (community group) based on its
edge weight and neighboring labels. With this method, we avoid recalculating
and regrouping of the nodes.
In terms of accuracy, Fig. 2.a, 2.b presents BWLP results over two networks
of 1000 vertices with a mixing parameter µ in range of (1–5) on horizontal axis.
Vertical axes shows the correctness measure of ARI [13]. BWLP outperforms
all other algorithms except for the µ = 1 in which Infomap is slightly better in
Fig. 2.a µ = 1. In the Fig. 2.b Louvain is slightly better only at µ = 1.
In addition, Fig. 2.a and 2.b depict the correctness measure of Normalized
Mutual Information (NMI) [16]) on vertical axes and µ in range of (1–5) on
the horizontal axis over two networks of 1000 vertices. The results from Fig. 2
confirm that the communities are detected with a higher accuracy by BWLP as
compared to the other algorithms.
5 Conclusion
In this paper, a new multi-stage community detection algorithm is proposed.
This algorithm automatically identifies the structures in networks based on dis-
similarity between nodes and the structure of the network. There is no pre-
processing or prior knowledge needed of the dataset/network for the proposed
algorithm. We noticed that on average, nodes agree on the same label in less
than three iterations.
The major phases of the proposed algorithm are threefold: The first is to
identify the dissimilarity value and structural form of networks and to sort data
nodes based on their dissimilarity values. The second layer is to find the flow of
the network from each vertex, then re-label it to vertices which belong to that
particular flow. Finally, the last layer converges all the nodes and construct the
final community structures.
Extensive experiments were conducted to demonstrate the correctness and
quality and expenses of the method introduced versus benchmark algorithms.
Results shown in Sect. 4.2 demonstrate that BWLP requires lower resources as
compared to benchmark algorithms. In addition, better accuracy is achieved
particularly over the datasets with higher edge density and complexity. BWLP
Balanced Weighted Label Propagation 11
successfully finds the hidden communities, a known problem for other bench-
mark algorithms. BWLP is a better choice for big data graphs considering it
outperforms the state-of-the-art algorithms, being 550 to 3500 time faster for
given data-sets as shown in Fig. 3(a) and Fig. 3(b) We presented the advantages
of BWLP against several benchmark methods.
Future research plans are threefold: first, to extend the algorithm for weighted
graphs; second, to find overlapping community structure with identical time and
space complexity; and third, to improve the accuracy of the existing algorithms
through introducing more conditions for special graph structures.
References
1. Fortunato, S., Barthélemy, M.: Resolution limit in community detection. Proc.
Natl. Acad. Sci. 104(1), 36–41 (2007). Resolution Limit
2. Girvan, M., Newman, M.E.J.: Community structure in social and biological net-
works. Proc. Natl. Acad. Sci. 99(12), 7821–7826 (2002). Betweenness
3. Ioannidis, V.N., Zamzam, A.S., Giannakis, G.B., Sidiropoulos, N.D.: Coupled
graph and tensor factorization for recommender systems and community detec-
tion. IEEE Trans. Knowl. Data Eng. (2019)
4. Lancichinetti, A., Fortunato, S., Kertész, J.: Detecting the overlapping and hier-
archical community structure in complex networks. New J. Phys. 11(3), 033015
(2009). LFM
5. Lancichinetti, A., Fortunato, S., Radicchi, F.: Benchmark graphs for testing com-
munity detection algorithms. Phys. Rev. E 78, 046110 (2008)
6. Lee, C., Reid, F., McDaid, A., Hurley, N.: Detecting highly overlapping community
structure by greedy clique expansion. arXiv preprint arXiv:1002.1827 (2010). GCE
7. Mukunda, A., Pirouz, M.: Influence-based community detection and ranking. In:
2019 International Conference on Computational Science and Computational Intel-
ligence (CSCI), pp. 1341–1346. IEEE (2019)
8. Palla, G., Derényi, I., Farkas, I., Vicsek, T.: Uncovering the overlapping com-
munity structure of complex networks in nature and society. arXiv preprint
physics/0506133 (2005). CPM
9. Pirouz, M., Zhan, J.: Optimized label propagation community detection on big
data networks. In: Proceedings of the 2018 International Conference on Big Data
and Education, pp. 57–62 (2018)
10. Pirouz, M., Zhan, J., Tayeb, S.: An optimized approach for community detection
and ranking. J. Big Data 3(1), 1–12 (2016). https://doi.org/10.1186/s40537-016-
0058-z
11. Psorakis, I., Roberts, S., Ebden, M., Sheldon, B.: Overlapping community detec-
tion using Bayesian non-negative matrix factorization. Phys. Rev. E 83(6), 066114
(2011). NMF
12. Raghavan, U.N., Albert, R., Kumara, S.: Near linear time algorithm to detect
community structures in large-scale networks. Phys. Rev. E 76(3), 036106 (2007)
13. Rand, W.M.: Objective criteria for the evaluation of clustering methods. J. Am.
Stat. Assoc. 66(336), 846–850 (1971)
14. Shen, H., Cheng, X., Cai, K., Hu, M.-B.: Detect overlapping and hierarchical com-
munity structure in networks. Physica A: Stat. Mech. Appl. 388(8), 1706–1712
(2009). EAGLE
12 M. Pirouz
15. Singh, S.S., Kumar, A., Singh, K., Biswas, B.: C2IM: community based context-
aware influence maximization in social networks. Physica A: Stat. Mech. Appl.
514, 796–818 (2019)
16. Strehl, A., Ghosh, J.: Cluster ensembles–a knowledge reuse framework for combin-
ing multiple partitions. J. Mach. Learn. Res. 3(Dec), 583–617 (2002). NMI
17. Li, W., Huang, C., Wang, M., Chen, X.: Stepping community detection algorithm
based on label propagation and similarity. Physica A: Stat. Mech. Appl. 472, 145–
155 (2017)
18. Xie, J., Szymanski, B.K.: Community detection using a neighborhood strength
driven label propagation algorithm. In: 2011 IEEE Network Science Workshop
(NSW), pp. 188–195. IEEE (2011)
19. Xue, C., Wu, S., Zhang, Q., Shao, F.: An incremental group-specific framework
based on community detection for cold start recommendation. IEEE Access 7,
112363–112374 (2019)
20. Zachary, W.W.: An information flow model for conflict and fission in small
groups. J. Anthropol. Res. 33(4), 452–473 (1977). http://www-personal.umich.
edu/∼mejn/netdata/
21. Zhang, P., Moore, C., Newman, M.E.J.: Community detection in networks with
unequal groups. Phys. Rev. E 93(1), 012303 (2016)
22. Zhang, S., Wang, R.-S., Zhang, X.-S.: Identification of overlapping community
structure in complex networks using fuzzy c-means clustering. Physica A: Stat.
Mech. Appl. 374(1), 483–490 (2007). Spectral, Euclidean Space, Fuzzy C-mean
Algorithm
23. Zhou, D., Wang, X.: A neighborhood-impact based community detection algorithm
via discrete PSO. Math. Prob. Eng. 2016 (2016)
Analysis of the Vegetation Index Dynamics
on the Base of Fuzzy Time Series Model
Abstract. Based on the fuzzy analysis of satellite monitoring data, the annual
dynamics of the vegetation index for the selected crop area is investigated by means
of MODIS images (LPDAAC – the Land Processes Distributed Active Archive
Center). To reconstruct and predict the weakly structured vegetation index time
series, the fuzzy models are proposed, compiled taking into account the analysis
of internal connections of the first and second orders, which are presented in the
form of fuzzy relations. The proposed models were investigated for adequacy and
suitability from the point of view of the analysis of the peculiarities of the intra-
annual mean annual dynamics of the index, typical for the cultivated area. On the
basis of the proposed approach, the results of the study of long-term dynamics of
vegetation indices can be used for a complex analysis of the dynamics of vegeta-
tion cover, including modeling and forecasting the efficiency and productivity of
agricultural crops.
1 Introduction
Modern technologies for satellite monitoring of the Earth’s surface provide agricultural
producers with useful information about the health of crops. The remote sensor’s ability
to detect minor differences in vegetation makes it a useful tool for quantifying variability
within a given field, assessing crop growth, and managing land based on current condi-
tions. Remote sensing data, collected on a regular basis, allows growers and agronomists
to make a current map of the state and strength of crops, analyze the dynamics of changes
in the health of crops, and predict it. For the interpretation of remote sensing data, the
most effective means are all kinds of vegetation indices, in particular, Normalized Dif-
ference Vegetation Index (NDVI), which are calculated empirically, i.e. by operations
with different spectral ranges of satellite monitoring data.
Most agricultural crops are characterized by changes in the phases of development,
which is reflected in the dynamics of the spectral-reflective properties of plants. The
study of seasonal and long-term changes in the spectral and brightness characteristics of
crops is possible through the analysis and modeling of the dynamic series of vegetation
indices, which makes it possible to quantitatively evaluate the features of the vegetation
cover and the patterns of its temporal dynamics. At the same time, standard algorithms
for solving problems of predicting the dynamics of the spectral-reflective properties
of plants work, as a rule, with “crisp” or structured data of satellite earth sensing, i.e.
with data presented as averaged numbers. Therefore, the averaging of the results of
measurements of vegetation indices is one of the most common empirical operations in
systems for collecting data from satellite monitoring and crop management. In particular,
the achievement of the required accuracy in the NDVI averaging process is achieved
by multiple measurements, where the results of individual measurements are partially
compensated by positive and negative deviations from the exact value. The accuracy of
their mutual compensation improves with an increase in the number of measurements,
since the absolute value of the mean of negative deviations approaches to the mean of
positive deviations.
More generally, satellite monitoring data, for example, NDVI values should be con-
sidered as weakly structured, i.e. those about which only their belonging to a certain
type is known [1]. In particular, the interval [NDVImin , NDVImax ], which includes all
its measurements, can serve as an adequate reflection of the poorly structured NDVI
values. Another more adequate reflection of the weakly structured value of NDVI can
be a statement of the form “high”, which, in fact, is one of the terms (values) of the lin-
guistic variable “value of the vegetation index”, which can be formally described by the
appropriate fuzzy set [2]. Therefore, based on this premise, it becomes obvious the impor-
tance and relevance of studying methods for studying seasonal and long-term changes
in the spectral-brightness characteristics of crops by means of time series of satellite
monitoring indicators relative to NDVI, which we will consider as weakly structured
values.
2 Problem Definition
On the basis of reliable information obtained by remote sensing of the Earth, it is nec-
essary to assess the regularities in the change in the average long-term values of NDVI
on the specifically selected area of cropland. More generally, the study of the NDVI
dynamics should include three stages: 1) calculation and analysis of the annual dynam-
ics of the NDVI values averaged for each analyzed date, typical for the selected region;
2) analysis of the time series of long-term NDVI values; 3) estimation of the dynamics
of the average annual NDVI values for cultivated areas for a certain period. The purpose
of this study is to provide a fuzzy analysis of seasonal and long-term characteristics in
the NDVI dynamics for the cultivated areas created by the MODIS platform (LPDAAC)
[3].
Step 1. Sorting the time series data {x t } (t = 1 ÷ n) into an ascending sequence {x p(i) },
where p is a permutation that sorts the data values in ascending order, i.e. x p(t) ≤ x p(t+1) .
Step 2. Calculation of the mean value over the set of all pairwise distances d i = |x p(i) −
x p(i+1) | between any two consecutive values x p(i) and x p(i+1) by the formula:
1 n−1
AD(d1 , d2 , ..., dn ) = xp(i) − xp(i+1) , (1)
n−1 i=1
Step 4. Recalculate the mean distance between any two consecutive values from the set
of values remaining after sorting for outliers.
Step 5. Establishing the universe U in the form of U = [Dmin − AD, Dmax + AD] = [D1 ,
D2 ], where Dmin and Dmax are the minimum and maximum values, respectively, on the
entire data set {x t } (t = 1 ÷ n).
There are various ways to identify membership functions that restore the fuzzy
subsets of the given universe. In particular, one of such methods is symmetric trapezoidal
membership functions of the following form (see Fig. 1):
⎧
⎪
⎪ 0, x < ak1
⎪
⎪
⎪ x−ak1
⎨ ak2 −ak1 , ak1 ≤ x ≤ ak2 ,
μAk (x) = 1, ak2 ≤ x ≤ ak3 , (4)
⎪
⎪ ak4 −x
⎪
⎪ , a ≤ x ≤ a ,
⎪
⎩ ak4 −ak3
k3 k4
0, x > ak4 ,
with parameters satisfying the conditions: ak2 − ak1 = ak3 − ak2 = ak4 − ak3 , where k =
1 ÷ m, m is the total number of fuzzy sets Ak describing the time series data. According
to [4], this number is calculated by the formula:
As an example, it was chosen the time series reflecting the annual dynamics of the
NDVI index (see Table 1) based on MODIS (LPDAAC) images (Fig. 2, [5]) of the
cultivated area in Jonesboro (USA) with geographic coordinates (−90.1614583252562,
35.8135416634583).
For the entire set of time series data, after recalculating the average distance and
standard deviation according to formulas (1) and (2) at the 5th step their final values are
16 E. Aliyev et al.
established under fulfilling condition (3): AD = 0.0296 and σ AD = 0.0007. In this case,
the desired universe is defined as the following segment U = [0.2479 – 0.0296, 0.7135 +
Analysis of the Vegetation Index Dynamics 17
0.0296] = [0.2183, 0.7431], where 0.2479 and 0.7135 are the minimum and maximum
values of the NDVI, respectively. To describe the qualitative criteria for evaluating the
NDVI the sufficient number of fuzzy subsets of the U is established from equality
(4) as follows: m = [0.7431 – 0.2183 – 0.0296]/[2 · 0.0296] = 8.3649 ≈ 8. Then
the corresponding 8 trapezoidal membership functions are identified (see Fig. 3), the
parameters of which are summarized in Table 2.
0.8
0.6
0.4
0.2
0
0.2183
0.2479
0.2775
0.3071
0.3367
0.3663
0.3959
0.4255
0.4551
0.4847
0.5143
0.5439
0.5735
0.6031
0.6327
0.6623
0.7063
0.7431
Fig. 3. Trapezoidal membership functions describing time series data.
framework of the NDVI fuzzy time series, the internal relationships of the 1st and 2nd
orders were chosen. These relationships demonstrate fuzzy relations in the implicative
form of the form “If <…>, then <…>” [6]. For example, internal relationships of the
1st order are grouped according to the principle: if the fuzzy set A3 is related to the sets
A4 and A2 , then the group of the 1st order is localized relative to it: A3 ⇒ A2 , A4 (see
Table 4, group G3). The groups of the 1st and 2nd orders are summarized in Table 4 and
Table 5, respectively.
If the NDVI value on the i-th day is denoted by x i , and the NDVI value on the next (i
+ 1)-th day is denoted by x i+1 , then the 1st order internal relationship, for example, A4
Analysis of the Vegetation Index Dynamics 19
where Aα = {u|μA (u) ≥ α, u ∈ U} are the α-level sets (α ∈ [0, 1]); M (Aα ) =
m
k=1 uk /m (uk ∈ Aα ) are the cardinalities of the corresponding α-level sets, for the
fuzzy set A3 = {0/0.3367, 1/0.3663, 1/0.3959, 0/0.4255} (see Table 2), which is a
forecast in the fuzzy relation A4 ⇒ A3 , we have:
For 0 < α < 1, α = 1, A3α = {0.3663, 0.3959}, M(A3α ) = (0.3663 +
0.3959)/2 ≈ 0.3811.
Then, according to (5), the crisp (defuzzified) output of the model is calculated as
1 1
F(A3 ) = M (A3α )d α ≈ M (A3α ) · α = 0.3811.
1 0
Principle 2 [7]. In the case of the fuzzy relation of the form Ai ⇒ Aj , At , Ap , where Ai is
the fuzzy analogue of NDVI on the i-th day, the crisp forecast for the next (i + 1)-th day
20 E. Aliyev et al.
is calculated as the arithmetic mean of the abscissa of the midpoints of the upper bases
of trapeziums corresponding to the fuzzy sets Aj , At and Ap . In particular, according
to the group of internal relationships A8 ⇒ A1 , A6 , A8 the forecast F for 11.07.2000,
27.07.2000 and 28.08.2000 dates is calculated as follows:
The forecasts obtained on the basis of the 1st order predictive model are summarized
in Table 6, and the geometric interpretation of this model is shown in Fig. 5.
The set of internal connections of the 2nd and higher orders coincides with the set
of NDVI fuzzy analogs (see Table 3). Nevertheless, we considered it necessary to show
the defuzzified outputs of the 2nd order fuzzy model (see Table 7) and its geometric
interpretation (see Fig. 5).
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
3/3/2000
4/7/2000
5/5/2000
6/2/2000
6/9/2000
7/7/2000
8/4/2000
9/1/2000
9/8/2000
10/13/2000
2/18/2000
2/25/2000
3/10/2000
3/17/2000
3/24/2000
3/31/2000
4/14/2000
4/21/2000
4/28/2000
5/12/2000
5/19/2000
5/26/2000
6/16/2000
6/23/2000
6/30/2000
7/14/2000
7/21/2000
7/28/2000
8/11/2000
8/18/2000
8/25/2000
9/15/2000
9/22/2000
9/29/2000
10/6/2000
Time series 1-st order fuzzy model 2-nd order fuzzy model
To assess the adequacy of fuzzy time series models the following criteria are used:
1 m |Fj − Aj | 1 m
MAPE = × 100, MSE = (Fj − Aj )2 ,
m j=1 Aj m j=1
where Aj and F j denote the actual value and predict at the t j date, respectively. The
results of forecasting and assessments of their reliability for both models (see Fig. 5) are
summarized in Table 8.
4 Conclusion
The results of the study of NDVI long-term dynamics can be used for modeling and
forecasting the effectivity and productivity of agricultural crops, for the complex analysis
22 E. Aliyev et al.
of the vegetation dynamics. Certainly, the obtained fuzzy models cannot be a reliable
source of information for conclusions about the prospects for the state of vegetation
on a given plot of agricultural land, because, firstly, they rely on a small amount of
annual NDVI data, which is extremely insufficient for analyzing internal cause-effect
relations of higher orders, and, secondly, they describe one annual cycle. Nevertheless,
this approach can be extrapolated to all long-term NDVI data obtained using MODIS
technology. As a result, it is possible to model and accordingly predict, for example, the
dynamics of long-term seasonal NDVI values averaged for each analyzed date, as well
as the dynamics of the average annual, annual minimum and annual maximum NDVI
values.
References
1. Rzayev, R.R.: Analytical Support for Decision-Making in Organizational Systems. Palmerium
Academic Publishing, Saarbruchen (2016). (in Russian)
2. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning.
Inf. Sci. 8(3), 199–249 (1965)
3. Vegetation Indices 16-Day L3 Global 250 m MOD13Q1. LPDAAC. https://lpdaac.usgs.gov/
dataset_discovery/modis/modis_products_table/mod13q1. Accessed 12 Oct 2020
4. Ortiz-Arroyo, D., Poulsen, J.R.: A weighted fuzzy time series forecasting model. Indian J. Sci.
Technol. 11(27), 1–11 (2018)
5. Vegetation Indices 16-Day L3 Global 250 m MOD13Q1. LPDAAC: https://goo.gl/maps/YAd
domuoXsD4QQN36. Accessed 12 Oct 2020
6. Rzayev, R.R., et al.: Time series modeling based on fuzzy analysis of position-binary compo-
nents of historical data. Nechetkie Sistemy i Myagkie Vychisleniya [Fuzzy Systems and Soft
Computing] 10(1), 35–73 (2015). (in Russian)
7. Chen, S.M.: Forecasting enrollments based on high-order fuzzy time series. Cybern. Syst. Int.
J. 33, 1–16 (2002)
Application of Data–Driven Fault
Diagnosis Design Techniques to a Wind
Turbine Test–Rig
1 Introduction
As the power required worldwide is increasing, and at the same time in order to
meet low carbon requirements, one of the possible solutions consists of increas-
ing the exploitation of wind–generated energy. However, this need implies to
improve the levels of availability and reliability, which represent the fundamen-
tal ‘sustainability’ feature, extremely good for these renewable energy conversion
systems. In fact, wind turbine processes should produce the required amount of
electrical energy continuously and in an effective way, trying to maximise the
wind power capture, on the basis of the grid’s demand and despite malfunctions.
To this end, possible faults affecting this energy conversion system must be prop-
erly diagnosed and managed, in their earlier occurrence, before they degrade into
failures and become critical issues.
Another important aspect regards wind turbines with very large rotors, which
allow to reach megawatt size, possibly maintaining light load carrying structures.
On one hand, they become very expensive and complex plants, which require
advanced control techniques to reduce the effects of induced torques and mechan-
ical stress on the structures. On the other hand, they need an extremely high
level of availability and reliability, in order to maximise the generated energy, to
minimise the production cost, and to reduce the requirements of Operation and
Maintenance (O & M) services. In fact, the final cost of the produced energy
depends first on the installation expenses of the plant (fixed cost). Second,
unplanned O&M costs may increase it up to about 30%, especially if offshore
wind turbine installations are considered, as they work in harsh environments.
These considerations motivate the implementation of condition monitoring
and fault diagnosis techniques that can be integrated into fault tolerant control
strategies, i.e. the ‘sustainable’ solutions. Unfortunately, several wind turbine
manufacturers do not implement ‘active’ approaches against faults, but rather
adopt conservative solutions. For example, this means to resort to the shutdown
of the plant and wait for O&M service. Therefore, more effective tools for man-
aging faults in an active way must be considered, in order to improve the wind
turbine working conditions, and not only during faulty behaviour.
Note also that, in principle, this strategy will be able to prevent critical faults
that might affect other components of the wind turbine system, and may thus
avoid to require unplanned replacement of its functional parts. Moreover, it will
lead to decrease possible O&M costs, while improving the energy production and
reducing the energy cost. On the other hand, the implementation of advanced
control systems, with the synergism with big data tools and artificial intelligence
techniques will lead to the development of real–time condition monitoring, fault
diagnosis and fault tolerant control solutions for these safe–critical systems that
can be enabled also only with on–demand features.
In the last decades several papers have investigated the problem of fault
diagnosis for wind turbine systems, as addressed e.g. in [6]. Some of them have
investigated the diagnosis of particular faults, i.e. those affecting the drive–
train part of the wind turbine nacelle. In fact, sometimes the detection of these
faults can be enhanced if the wind turbine subsystems are compared to other
Data–Driven Fault Diagnosis Application 25
modules of the whole plant [9]. Moreover, more challenging topics regarding the
fault tolerant control of wind turbines have been considered e.g. in [11], also
by proposing international co–operations on common problems, as analysed e.g.
in [7]. Therefore, the key point is represented by the fault diagnosis task when
exploited to achieve the sustainability feature for safety–critical systems, such
as wind turbines. In fact, it has been shown to represent a challenging topic [20],
thus justifying the research issues investigated in this paper.
With reference to this paper, the topic of the fault diagnosis of a wind turbine
system is analysed. In particular, the design of practical and reliable solutions
to Fault Detection and Isolation (FDI) is considered. However, differently from
other works by the same authors, the Fault Tolerant Control (FTC) topic is
not investigated here, even if it can rely on the same tools exploited in this
paper. In fact, the fault diagnosis module provides the reconstruction of the fault
signal affecting the process, which could be actively compensated by means of
a controller accommodation mechanism. Moreover, the fault diagnosis design is
enhanced by the derived fault reconstructors that are estimated via data–driven
approaches, as they also allow to accomplish the fault isolation task.
The first data–driven strategy proposed in this work exploits Takagi–Sugeno
(TS) fuzzy prototypes [1, 4], which are estimated via a clustering algorithm and
exploiting the data–driven algorithm developed in [15]. For comparison purpose,
a further approach is designed, which exploits Neural Networks (NNs) to derive
the nonlinear dynamic relations between the input and output measurements
acquired for the process under diagnosis and the faults affecting the plant. The
selected structures belong to the feed–forward Multi–Layer Perceptron (MLP)
neural network class that include also Auto–Regressive with eXogenous (ARX)
inputs in order to model nonlinear dynamic links among the data. In this way,
the training of these Nonlinear ARX (NARX) prototypes for fault estimation
can exploit standard back–propagation training algorithm, as recalled e.g. in [5].
The designed fault diagnosis schemes are tested via a high–fidelity simu-
lator of a wind turbine process, which describes its behaviour in healthy and
faulty conditions. This simulator, which represents a benchmark [8], includes
the presence of uncertainty and disturbance effects, thus allowing to verify the
reliability and robustness characteristics of the proposed fault diagnosis method-
ologies. Moreover, this work proposes to validate the efficacy of the designed fault
diagnosis techniques by exploiting a more realistic scenario, which consists of a
Hardware–In–the–Loop (HIL) tool.
It is worth noting the main contributions of this paper with respect to pre-
vious works by the authors. For example, this study analyses the solutions
addressed e.g. in [19] but taking into account a more realistic and real–time
system illustrated in Sect. 4. On the other hand, the fault diagnosis scheme
developed in this paper was designed for a wind turbine system also in [17], but
without considering the HIL environment.
The fuzzy methodology was also proposed in by the authors in [14], which
considered the development of recursive algorithms for the implementation of
adaptive laws relying on Linear Parameter Varying (LPV) systems. The app-
26 S. Simani et al.
roach proposed in this paper estimates the fault diagnosis models by means
of off–line procedures. Moreover, this paper further develops the achievements
obtained e.g. in [18], but concerning the fault diagnosis a wind farm. The paper
[16] proposed the design of a fault tolerant controller using the input–output
data achieved from a single wind turbine, by exploiting the results achieved in
[17]. On the other hand, this work considers the verification and the validation
of the proposed fault diagnosis methodologies by exploiting an original HIL tool,
proposed considered in a preliminary paper by the same authors [13].
The work follows the structure sketched in the following. Section 2 briefly
summarises the wind turbine simulator, as it represents a well-established bench-
mark available in literature [10]. Section 3 describes the fault diagnosis strate-
gies based on Fuzzy Systems (FSs) and NN structures. Section 4 summarises the
obtained results via the simulations and the HIL tool describing the behaviour
of the wind turbine process. Finally, Sect. 5 concludes the work by reporting
the main points of the paper and suggesting some interesting issues for further
research and future investigations.
Rotor
Pitch Pitch
Blade
actuator sensor
controller
controller controller
Switch Switch
This wind turbine benchmark is able to generate different typical fault cases
affecting the sensors, the actuators and the process components. This scenario
comprising 9 fault situations is illustrated by means of Table 1, which reports
the input and output measurements acquired form the WT process signals and
mainly affected by these faults.
In this way, Table 1 reports the most sensitive measurements uj (k) and yl (k)
acquired from the WT system with respect to the fault conditions implemented
in the WT benchmark. In practice, the fault signals of Table 1 were injected
into the WT simulator, assuming that only a single fault may occur. Then, by
checking the Relative Mean Square Errors (RMSEs) between all the fault–free
and faulty measurements from the WT plant, the most sensitive signal uj (k)
and yl (k) was selected and reported in Table 1.
For fault diagnosis purpose, the complete model of the WT benchmark can
be described as a nonlinear continuous–time dynamic model represented by the
function fwt of Eq. (1) including the overall behaviour of the WT process reported
in Fig. 1 with state vector xwt and fed by the driving input vector u:
ẋwt (t) = fwt (xwt , u(t))
(1)
y(t) = xwt (t)
Equation (1) highlights that the simulator allows to measure all the state vector
signals, i.e. the rotor speed, the generator speed and the generated power of the
WT process:
that represent the acquired measurements of the pitch angles from the three WT
blades and the measured generator/converter torque. These signals are acquired
with sample time T in order to obtain N data indicated as u(k) and y(k) with
index k = 1, . . . , N that are exploited to design the fault diagnosis strategies
addressed in this work.
This section recalls the fault diagnosis strategy proposed in this paper that
relies on FS and NN tools, as summarised in Sect. 3.1. These architectures are
able to represent NARX models exploited for estimating the nonlinear dynamic
relations between the input and output measurements of the WT process and
the fault signals. In this sense, these NARX prototypes will be employed as fault
estimators for solving the problem of the fault diagnosis of the WT system.
Under these assumptions, the fault estimators derived by means of a data–
driven approach represent the residual generators r(k), which provide the on–
line reconstruction f̂ (k) of the fault signals summarised in Table tab:faults, as
represented by Eq. (2):
r(k) = f̂ (k) (2)
where the
term f̂ (k) represents the general fault vector of Table tab:faults, i.e.
f̂ (k) = fˆ1 (k), . . . , fˆ9 (k) .
The fault diagnosis scheme exploiting the proposed fault estimators as resid-
ual generator is sketched in Fig. 2. Note that, as already highlighted, this scheme
is also able to solve the fault isolation task [2].
process
Inputs Selector
r2(k) = f2(k)
Output Selector
Fault
sensors y (k) estimator 2
1
…
y2(k)
Selector u j(k)
ym(k) ri (k) = fi (k)
Selector Fault
estimator i
Bus
y (k)
…
Selector
Outputs rr+m(k) = fr+m(k)
Bus Selector Fault
estimatorr+m
Figure 2 shows that the general residual generator exploits the input and
output measurements acquired from the process under diagnosis, u(k) and y(k),
properly selected according to the analysis shown in Table 1. The fault detection
problem can be easily achieved by means of a simple threshold logic applied to
the residuals themselves, as described in [2]. This issue will not be considered in
this paper.
Once the fault detection phase is solved, the fault isolation stage is directly
obtained via the bank of estimators of Fig. 2. In this case, the number of esti-
mators of Fig. 2 is equal to the faults to be detected, i.e. 9, which is lower than
the number of input and output measurements, r + m, acquired from the WT
process.
This condition provides several degrees of freedom, as the i–th reconstructor
of the fault fˆ(k) = ri (k) is a function of the input and output signals u(k)
and y(k). These signals are thus selected in order to be affected sensitive to the
specific fault fi (k), as highlighted in Table 1. This procedure enhances also the
design of the fault reconstructors, as it reduces the number of possible input
and output measurements, uj (k) and yl (k), which have to be considered for the
identification procedure reported in Sect. 3.1.
The sensitivity analysis already represented in Table 1 has to be performed
before the estimation of the fault estimators. Therefore, once the input–output
signals are selected, according to Table 1, the FSs and the NNs used as fault
reconstructors can be developed, as summarised in Sect. 3.1.
Using this approach, in general, the fault signal fˆ(k) is reconstructed by using
suitable data taken from the WT process under diagnosis. In this case, the fault
function fˆ(k) is represented as a weighted average of affine parametric relations
aTi x(k) + bi (consequents) depending on the input and output measurements
collected in x(k). These weights are the fuzzy membership degrees λi (x) of the
system inputs.
The parametric relations of the consequents depend on the unknown variables
ai and bi , which are estimated by means on an identification approach. The rule
30 S. Simani et al.
where ul (·) and yj (·) represent the l–th and j–th components of the actual
WT input and output vectors u(k) and y(k). These components are selected
according to the results reported in Table 1.
The consequent affine parameters of the i–th model of the Eq. (3) are usually
represented with a vector:
T
(i) (i)
ai = α1 , . . . , αo(i) , δ1 , . . . , δo(i) (5)
(i)
where usually the coefficients αj are associated to the delayed output samples,
(i)
whilst δj to the input ones.
The approach proposed in this paper for the derivation of the generic i–th
fault approximator (FIS) starts with the fuzzy clustering of the data u(k) and
y(k) from the WT process. This paper exploits the well-established Gustafson–
Kessel (GK) algorithm [1]. Moreover, the estimation of the FIS parameters is
addressed as a system identification problem from the noisy data of the WT
process. Once the data are clustered, the identification strategy proposed in this
work exploits the methodology developed by the authors in [3].
Another key point not addressed in this work concerns the selection of the
optimal clusters number nC . This issue was investigated and developed by the
authors, which leads to the estimation of the membership degrees λi (x(k))
required in Eq. (3) and solved as a curve fitting problem [1].
This paper considers an alternative data–driven approach, which exploits
neural networks used as fault approximators in the scheme of Fig. 2. Therefore,
in the same way of the fuzzy scheme, the bank of NNs is exploited to reconstruct
the faults affecting the WT system under diagnosis using a proper selection of
the input and the output measurements. The exploited NN structure consists
of a feed–forward Multi–Layer Percepron (MLP) architecture with 3 layers of
neurons [5].
However, as MPL networks represent static relations, the paper suggests to
implement the MLP structure with a tapped delay line. Therefore, this quasi–
static NN represents a powerful way for estimating nonlinear dynamic regres-
sions between the input and output measurements from the WT process and its
fault functions. This solution allows to obtain another Nonlinear ARX (NARX)
Random documents with unrelated
content Scribd suggests to you:
XVII
KUKISTUNUT SUURUUS
"Niin on."
"Kuolemaan asti."
"Kuinka niin?"
"Siksi, että hovi saa tietää, millaisia hommia meillä on, ja lähettää
Bezenvalin sveitsiläiset ja Lambesqin rakuunat, jolloin joudumme
kahden tulen väliin."
Ja neuvottelijat läksivät.
Majuri luki:
de Flesselles.'
Hän tiesi, mistä linnan aukoista murhaavin tuli tuiski; hän erotti
kanuunan huomaamattoman tärinän juuri ennen laukaisemista. Hän
oli lopulta päässyt selville siitä, minä hetkenä vallituksella olevat
ampujat aikoivat laukaista nostosillan läpi.
Tai:
"Herra Billot, rakas herra Billot, vaimonne jää leskeksi, jos saatte
surmanne."
Pitou tuumi, että hyvä asia kävisi paremmaksi, kun sitä olisi kaksin
verroin. Hän seurasi Billotia kiljuen: "Kahdet rattaat! Kahdet rattaat!"
Niin kauan kuin Billot ajatteli vain itseään, vähät hän siitä, jos
Bastilji räjähtikin ilmaan ja hän sen mukana. Mutta tohtori Gilbertin
täytyi kaikin mokomin saada elää.
"Aatelismiehen sanan!"
Kun siis kuultiin, että peloittava Bastilji aikoi antautua, että sen
kuvernööri oli luvannut luovuttaa linnan, ei kukaan ottanut sitä
uskoakseen.
TOHTORI GILBERT
Pitou tuki Billotia. Häneen ei ollut mikään luoti osunut; hän ei ollut
saanut vammaa, mutta putoaminen oli kuitenkin hiukan
pökerryttänyt kunnon vuokratilallista.
Gonchon salli sen tapahtua. Ehkä hän ajatteli, että kansan viha on
kuin tulviva virta, joka saa enemmän vahinkoa aikaan, jos sen
vuolautta koettaa estää, kuin jos sen antaa rauhassa kiitää
eteenpäin.
Bastilji on valloitettu!
"Kolmas Bertaudière."
Vapiseva vanginvartija osui hänen tielleen.
"Missä ne ovat?"
"On. Se on tämä."
"En tiedä."
"Tekö, Billot."
"Niin, niin, hän, hän, me, me!" huusi parikymmentä miestä, jotka
olivat seisahtuneet porrassillakkeelle nähdessään Billotin iskevän
kirveellä oveen.
"Ketkä te?"
Vaikka hän oli ollut vankilassa viisi tai kuusi päivää, oli hän pitänyt
ulkoasustaan samaa huolta kuin muulloinkin. Parta, joka vankilassa
oli kasvanut, tehosti vain ihon kalpeutta, ja se yksinään osoitti
puuttuvaa huolehtimista ulkomuodosta, mutta tämä ei suinkaan
johtunut vangin omasta tahdosta, vaan siitä, että hän ei pyytäessään
ollut saanut partaveistä tai tilaisuutta parturin kutsumiseen.
"Niin on."
"Olen kyllä."
"Lipas on varastettu."
"Niin."