Contents of Volume 31 (2021)

1/2021 2/2021 3/2021

4/2021

  • [1] Malinovský V., (CZ)
    Comparative analysis of freight transport prognoses results provided by transport system model and neural network, pp. 239-259

      Full text         https://doi.org/10.14311/NNW.2021.31.013

    Abstract: This paper deals with problems of processing freight statistic data into the form of time series and analysing consequent results by means of two completely different methods. The first method for calculating chosen transport trends uses the transport model Trans-Tools based on conventional mathematical and statistical functions while the second one uses the Scikit Learn software providing users with development environment including algorithms of neural networks. The obtained results are similar to a certain extent which shows new possibilities of progressive use of neural networks in future and enables modern approach to analysing time series not only in transportation sector. Comparative analysis of results obtained from the same transport data processed by “standard” mathematical (Trans-Tool) method and neuron-network (Scikit Learn) method as well as a research focused on some trends development within the scope of freight transport in EU represent goals of this work.

  • [2] Švorc D., Tichý T., Růžička M. (CZ)
    An infrared video detection and categorization system based on machine learning, pp. 261-277

      Full text         https://doi.org/10.14311/NNW.2021.31.014

    Abstract: The main aim of this paper is to present a new possibility for detection and recognition of different categories of electric and conventional (equipped with combustion engine) vehicles. These possibilities are provided by use of thermal and visual video cameras and two methods of machine learning. The used methods are Haar cascade classifier and convolutional neural network (CNN). The thermal images, obtained through an infrared thermography camera, were used for the training database. The thermal cameras can complement or substitute visible spectrum of video cameras and other conventional sensors and provide detailed recognition and classification data needed for vehicle type recognition. The first listed method was used as an object detector and serves for the localization of the vehicle on the road without any further classification. The second method was trained for vehicle recognition on the thermal image database and classifies a localized object according to one of the defined categories. The results confirmed that it is possible to use infrared thermography for vehicle drive categorization according to the thermal features of vehicle exteriors together with methods of machine learning for vehicle type recognition.

  • [3] Weiß N., Pozzobon E., Mottok J., Matoušek V., (Germany, CZ) ,
    Automated Reverse Engineering of CAN Protocols, pp. 279-295

      Full text         https://doi.org/10.14311/NNW.2021.31.015

    Abstract: Car manufacturers define proprietary protocols to be used inside their vehicular networks, which are kept an industrial secret, therefore impeding independent researchers from extracting information from these networks. This article describes a statistical and a neural network approach that allows reverse engineering proprietary controller area network (CAN)-protocols assuming they were designed using the data base CAN (DBC) file format. The proposed algorithms are tested with CAN traces taken from a real car. We show that our approaches can correctly reverse engineer CAN messages in an automated manner.

  • [4] Liu J., Ni F., Du M., Zhang X., Que Z., Song S., (U.S., China) ,
    Upper bounds on the node numbers of hidden layers in MLPs, pp. 297-309

      Full text         https://doi.org/10.14311/NNW.2021.31.016

    Abstract: It is one of the fundamental and challenging problems to determine the node numbers of hidden layers in neural networks. Various efforts have been made to study the relations between the approximation ability and the number of hidden nodes of some specific neural networks, such as single-hidden-layer and two-hiddenlayer feedforward neural networks with specific or conditional activation functions. However, for arbitrary feedforward neural networks, there are few theoretical results on such issues. This paper gives an upper bound on the node number of each hidden layer for the most general feedforward neural networks called multilayer perceptrons (MLP), from an algebraic point of view. First, we put forward the method of expansion linear spaces to investigate the algebraic structure and properties of the outputs of MLPs. Then it is proved that given k distinct training samples, for any MLP with k nodes in each hidden layer, if a certain optimization problem has solutions, the approximation error keeps invariant with adding nodes to hidden layers. Furthermore, it is shown that for any MLP whose activation function for the output layer is bounded on R, at most k hidden nodes in each hidden layer are needed to learn k training samples.



3/2021

  • [1] Sajanraj T.D., Mulerikkal J., Raghavendra S., Vinith R., Fabera V. (India, CZ) ,
    Passenger flow prediction from AFC data using station memorizing LSTM for metro rail systems, pp. 173-189

      Full text         https://doi.org/10.14311/NNW.2021.31.009

    Abstract: Metro rail systems are increasingly becoming relevant and inevitable in the context of rising demand for sustainable transportation methods. Metros are therefore going to have a consistently expanding user-base and hence user satisfaction will require meticulous planning. Usage forecast is clearly an integral component of metro planning as it enables forward looking and efficient allocation of resources leading to greater commuter satisfaction. An observation from studying the usage of Kochi Metro Rail Ltd. is that there is a consistently occurring temporal pattern in usage for every station. But the patterns differ from station to station. This hinders the search for a global model representing all stations. We propose a way to overcome this by using station memorizing Long Short-Term Memory (LSTM) which takes in stations in encoded form as input along with usage sequence of stations. This is observed to significantly improve the performance of the model. The proposed architecture with station parameter is compared with algorithms like SVR (support vector regression) and neural network implementation with the best architecture to testify the claim. The proposed model can predict the future flow with an error rate of 0.00127 MSE (mean squared error), which is better than the other models tested.

  • [2] Sultana S., Hussain S.S., Hashmani M., Ahmad J., Zubair M. (Pakistan, Malaysia) ,
    A deep learning hybrid ensemble fusion for chest radiograph classification, pp. 191-209

      Full text         https://doi.org/10.14311/NNW.2021.31.010

    Abstract: Biomedical imaging, archiving, and classification is the recent challenge of computer-aided medical imaging. The popular and influential Deep Learning methods predict and congregate distinct markable features of ambiguity in radiographs precisely and accurately. This study submits a new topology of a deep learning network for chest radiograph classification. In this approach, a hybrid ensemble fusion of neural network topology can better diagnose ambiguities with high precision. The proposed topology also compares statistical findings with three optimizers and the most possible varying essential attributes of dropout probabilities and learning rates. The performance as a function of the AUCROC of this model is measured on the Chest Xpert dataset.

  • [3] Zheng S., Jiang A.N., Yang X.R. (China)
    Tunnel displacement prediction under spatial effect based on Gaussian process regression optimized by differential evolution, pp. 211-226

      Full text         https://doi.org/10.14311/NNW.2021.31.011

    Abstract: The prediction and analysis of surrounding rock deformation is a primary risk assessment method in tunnel engineering. However, the accurate prediction result is not easy to achieve due to the influence of multiple factors such as rock mass properties, support structure, and the spatial effect of tunnel construction. In this paper, a multivariate time-series model (MTSM) for tunnel displacement prediction is studied based on Gaussian process regression (GPR) optimized by differential evolutionary (DE) strategy, where the spatial effect is intuitively expressed through an extended time-series model. First, building learning samples for GPR, in which the inputs is the displacement data of the previous n days and the output is the data of the day (n + 1). Then, for each sample, an input item is added successively to form an expanded learning sample, which is the “distance between the construction face and monitoring section” on the day (n + 1). Taking the root mean square error between the regression and measured data as the control index, the GPR model is trained to express the nonlinear mapping relationship between input and output, and the optimal parameters of this model are searched by DE. The displacement multivariate time-series model represented by DE-GPR is known as MTSM. On this basis, the applicability of GPR for tunnel displacement prediction and the necessity of DE optimization are illustrated by comparing the prediction results of several commonly used machine learning models. At the same time, the influence of GPR and DE parameters on the regression result and the computational efficiency of the MTSM model is analyzed, the recommendation for parameter values are given considering both calculation efficiency and accuracy. This method is successfully applied to the Leshanting tunnel of Puyan expressway in Fujian province, China, and the results show that the MTSM based on DE-GPR has a good ability in the deformation prediction of the surrounding rock, which provides a new method for tunnel engineering safety control. Key words: displacement prediction, Gaussian process, differential evolution, spatial effect, parameter analysis

  • [4] Jozová Š., Tobiška J., Nagy I. (CZ)
    On-line recognition of critical driving situations , pp. 227-238

      Full text         https://doi.org/10.14311/NNW.2021.31.012

    Abstract: According to the statistics about vehicle accidents, there are many causes such as traffic violations, reduced concentration, micro sleep, hasty aggression, but the most frequent cause of accidents at highways is a carelessness of the driver and violation of keeping a safe distance. Producers of vehicles try to take into account this situation by development of assistance systems which are able to avoid accidents or at least to mitigate its consequences. This urgent situation leaded to the described project of investigation of behavior of drivers in dangerous situations occurring in vehicle driving. The research is to help in solution of the present unsatisfactory situation in driving accidents. The developed decisionmaking algorithm of detection serious driving situations that can lead to accidents was tested in the laboratory of driving simulators in FTS CTU, Prague. The data for its testing resembled highway traffic.


2/2021

  • [1] Jonáková L., Nagy I. (CZ)
    Power purchase strategy of retail customers utilizing advanced classification methods , pp. 89-107

      Full text         https://doi.org/10.14311/NNW.2021.31.005

    Abstract: This study reflects a unique task with significant business potential, on the edge of the wholesale and retail power market, i.e., optimization of power derivatives purchase strategy of retail customers. Even though the definition of the task as well as initial assumptions may be highly complex, essentially, the purpose of this study can be narrowed down to the estimation of buying signals. The price signals are estimated with the use of machine learning techniques, i.e., one-, two- and three-layer perceptron with supervised learning as well as long short-term memory network, which allow modelling of highly complex functional relationships, and with the use of relative strength index, i.e., momentum technical indicator, which on the contrary allows higher flexibility in terms of parameters adjustment as well as easier results interpretation. Thereafter, performance of these methods is compared and evaluated against the established benchmark.

  • [2] Hlaváč V. (CZ)
    Neural Network for the identification of a functional dependence using data preselection , pp. 109-124

      Full text         https://doi.org/10.14311/NNW.2021.31.006

    Abstract: A neural network can be used in the identification of a given functional dependency. An undetermined problem (with more degrees of freedom) has to be converted to a determined one by adding other conditions. This is easy for a well-defined problem, described by a theoretical functional dependency; in this case, no identification (using a neural network) is necessary. The article describes how to apply a fitness (or a penalty) function directly to the data, before a neural network is trained. As a result, the trained neural network is near to the best possible solution according to the selected fitness function. In comparison to implementing the fitness function during the training of the neural network, the method described here is simpler and more reliable. The new method is demonstrated on the kinematics control of a redundant 2D manipulator.

  • [3] Fuangkhon P. (Thailand)
    Normalized data barrier amplifier for feed-forward neural network, pp. 125-157

      Full text         https://doi.org/10.14311/NNW.2021.31.007

    Abstract: A boundary vector generator is a data barrier amplifier that improves the distribution model of the samples to increase the classification accuracy of the feed-forward neural network. It generates new forms of samples, one for amplifying the barrier of their class (fundamental multi-class outpost vectors) and the other for increasing the barrier of the nearest class (additional multi-class outpost vectors). However, these sets of boundary vectors are enormous. The reduced boundary vector generators proposed three boundary vector reduction techniques that scale down fundamental multi-class outpost vectors and additional multi-class outpost vectors. Nevertheless, these techniques do not consider the interval of the attributes, causing some attributes to suppress over the other attributes on the Euclidean distance calculation. The motivation of this study is to explore whether six normalization techniques; min-max, Z-score, mean and mean absolute deviation, median and median absolute deviation, modified hyperbolic tangent, and hyperbolic tangent estimator, can improve the classification performance of the boundary vector generator and the reduced boundary vector generators for maximizing class boundary. Each normalization technique pre-processes the original training set before the boundary vector generator or each of the three reduced boundary vector generators will begin. The experimental results on the real-world datasets generally confirmed that (1) the final training set having only FF-AA reduced boundary vectors can be integrated with one of the normalization techniques effectively when the accuracy and precision are prioritized, (2) the final training set having only the boundary vectors can be integrated with one of the normalization techniques effectively when the recall and F1-score are prioritized, (3) the Z-score normalization can generally improve the accuracy and precision of all types of training sets, (4) the modified hyperbolic tangent normalization can generally improve the recall of all types of training sets, (5) the min-max normalization can generally improve the accuracy and F1-score of all types of training sets, and (6) the selection of the normalization techniques and the training set types depends on the key performance measure for the dataset.

  • [4] Zhang M., Xu P. (China)
    AGAN: Attribute generative adversarial network, pp. 159-172

      Full text         https://doi.org/10.14311/NNW.2021.31.008

    Abstract: Graph generative adversarial network has achieved remarkable effectiveness, such as link prediction, node classification, user recommendation and node visualization in recent years. Most existing methods mainly focus on how to represent the proximity between nodes according to the structure of the graph. However, the graph nodes also have rich attribute information in social networks, the traditional methods mainly consider the node attributes as auxiliary information incorporate into the embedding representation of the graph to improve the accuracy of node classification and link prediction. In fact, in social networks, these node attributes are often sparse. Due to privacy and other reasons, the attributes of many nodes are difficult to obtain. Inspired by the application of generative adversarial network in image field, we propose an innovative framework to discover node latent attribute. Through experiments, we demonstrate the effectiveness of our proposed methods.


1/2021

  • [1] Dai J., He Y.H., Li J.Y. (China)
    An approach for heuristic parallel LDTW distance optimization method with bio-inspired strategy , pp. 1-28

      Full text         https://doi.org/10.14311/NNW.2021.31.001

    Abstract: Dynamic time warping (DTW) is a classical similarity measure for arbitrary length time series. As an effective improvement of DTW, dynamic time warping under limited warping path length (LDTW) oppresses the long-term pathological alignment problem and allows better flexibility. However, since LDTW increases path lengths, it directly leads to higher time-consuming. In this paper, a new method of similarity sequence measurement (ACO LDTW) is proposed by the parallel computing characteristics of ant colony optimization (ACO) algorithm with bio-inspired strategy and the idea of LDTW path restriction. This algorithm searches the optimal path on the restricted distance matrix by simulating the behavior of ant colony parallel foraging. Firstly, the distance matrix is mapped to the 0 - 1 matrix of grid method, and the search range of ants is limited by the warping path in LDTW. Secondly, the state transition probability function, pheromone initialization and update mechanism of ACO algorithm are adapted. On the basis of ensuring the accurate results, the convergence rate can be effectively improved. The validity of ACO LDTW is verified by cases. In the 22 data sets of 1NN classification experiment, ACO LDTW has the lowest classification error rate in 16 data sets, and it is shorter than the calculation time of LDTW. At the same time, it is applied to the field of mechanical fault diagnosis and has the ability to solve practical engineering applications. Experiments show that ACO LDTW is more effective in terms of accuracy and computation time.

  • [2] Samson A.B.P., Chandra S.R.A., Manikant M. (India)
    A deep neural network approach for the prediction of protein subcellular localization , pp. 29-45

      Full text         https://doi.org/10.14311/NNW.2021.31.002

    Abstract: The subcellular localization of proteins is an essential characteristic of human cells, which plays a vital part in understanding distinct functions and cells’ biological processes. The abnormal protein subcellular localization affects protein functionality and may cause many human diseases ranging from metabolic disorders to cancer. Therefore, the prediction of subcellular locations of the proteins is an important task. Artificial neural network has become a popular research topic in machine learning that can achieve remarkable results in learning highlevel latent traits. This paper proposes a deep neural network (DNN) model to predict the human protein subcellular locations. The DNN automatically learns high-level representations of abstract features and proteins by examining nonlinear relationships between different subcellular locations. The experimental results have shown that the proposed method gave better results compared with the classical machine learning techniques such as support vector machine and random forest. This model also outperformed the similar model, which uses stacked auto-encoder (SAE) with a softmax classifier.

  • [3] Garlík B. (CZ)
    Modelling and optimization of an intelligent environmental energy system in an intelligent area, pp. 47-76

      Full text         https://doi.org/10.14311/NNW.2021.31.003

    Abstract: The article deals with the current state of energy consumption, the development of distribution networks in the context of its decentralization and integrated community energy systems. The article focuses on the issue and optimization of the operation of EnergyHubs (EH) – energy centres in terms of solving environmental aspects using a mathematical model in the GAMS environment. The acquired knowledge and results of simulations were then applied to a specific urban area to find the optimal variant of EH. The aim of the research is to present its results at the level of cleaner production, improvement of the environment, significant reduction of CO2 and sustainability of society. My experience proves that the achievement of sustainable development goals represents fundamental gaps in research and practical applications, especially at the level of specific projects. It is mainly the application of insufficient indicators and work methodologies in the design of building projects with almost zero energy consumption. Another shortcoming is the coordination of design procedures and applications of optimization and simulation methods necessary to address the energy performance of buildings or clusters of buildings. In addition, the results show growing expectations about the added value of applying artificial intelligence in meeting sustainable development goals, through new data sources that inevitably enter the energy sustainability design process.

  • [4] Kovanda J., Rulc V. (CZ)
    Pre-crash control strategy of driver assistance system, pp. 77-88

      Full text         https://doi.org/10.14311/NNW.2021.31.004

    Abstract: The aim of the article is the optimisation process of the ADAS (Advanced Driver Assistance Systems) control. The methodology is based on the classification of ADAS systems according to the situations of unavoidable accidents. The evaluation of expected consequences uses injury biomechanics, which represents the extended definition of HMI (Human-Machine Interaction). The evaluation of injury mechanism and the machine intervention enables to control this process with the target to minimise the consequent injuries. Then the decision making takes new inputs to the control process and it enriches the multiparametric control of the system with the target to minimise the consequences.