• TECHNOLOGICAL BASIS OF “INDUSTRY 4.0”

    Guidelines for the application of artificial intelligence in the study of the influence of climate change on transport infrastructure

    Industry 4.0, Vol. 8 (2023), Issue 3, pg(s) 75-78

    We are witnessing the massive and impressive penetration of artificial intelligence (AI) into many areas of human activity. This process is expected to intensify in the next few decades. In most technical fields, there will be a preponderance of the so-called narrow artificial intelligence with clearly defined tasks and functions. It is usually a coherent set of neural networks trained to solve specific problems. The advantage of narrow AI is that it is entirely controllable and, at the same time, has excellent capabilities. This publication aims to outline guidelines for applying narrow artificial intelligence in investigating the impact of climate change on transport infrastructure. After a brief introduction to narrow artificial intelligence and climate change, various possible areas suitable for AI modeling are explored. Directions and preparatory tasks for collecting climate-sensitive local data on the condition and changes in the transport infrastructure in Bulgaria necessary for AI training are identified.

  • DOMINANT TECHNOLOGIES IN “INDUSTRY 4.0”

    Solver parameter influence on the results of multilayer perceptron for estimating power output of a combined cycle power plant

    Industry 4.0, Vol. 5 (2020), Issue 3, pg(s) 114-117

    Previous work has determined the ability of using the Multilayer Perceptron (MLP) type of Artificial Neural Network (ANN) to estimate the power output of a Combined Cycle Power Plant (CCPP) in which optimization did not focus on the solver parameter optimization. In previous work, the solvers used the default parameters. Possibility exists that optimizing solver parameters will net better results. Two solver algorithm’s parameters are optimized: Stochastic Gradient Descent (SGD) and Adam, with 140 and 720 parameter combinations respectively. Solutions are estimated through the use of Root Mean Square Error (RMSE). Lowest RMSE achieved is 4.275 [MW] for SGD and 4.259 [MW] for Adam, achieved with parameters: = 0.05, = 0.02, and nesterov=True for SGD and with parameters = 0.001, 1 = 0.95, 2 = 0.99, and amsgrad=False for Adam. Only a slight improvement is shown in comparison to previous results (RMSE=4.305 [MW]) which points towards the fact that solver parameter optimization with the goal of improving results does not justify the extra time taken for training.

  • BUSINESS

    Оn predictability of precious metals towards robust trading

    Science. Business. Society., Vol. 5 (2020), Issue 1, pg(s) 14-16

    Large amounts of liquidity flow into several precious metals every day. Investment decisions are mainly based on predicting the future movements of the instrument(s) in question. However, high frequency financial data are somewhat hard to model or predict as stochastic processes and many other random factors are involved. It would be valuable information for the investor if he or she knew which precious metals were quantitatively more predictable, that would also be a good basis for more robust trading decisions. The objective in this study is to build predictive models on high frequency precious metal data and compare predictabilities of different metals using only past price and volume values which should be a basis for robust trading decisions. The data used consist of various frequencies from 1-minute to 4-hour covering a period of almost 20 years for each instrument and frequency. Artificial Neural Network (ANN) and Gradient Boosted Decision Tree (XGB) methods are applied. Comparable results are achieved.

  • BUSINESS & “INDUSTRY 4.0”

    Оn predictability of precious metals towards robust trading

    Industry 4.0, Vol. 5 (2020), Issue 2, pg(s) 87-89

    Large amounts of liquidity flow into several precious metals every day. Investment decisions are mainly based on predicting the future movements of the instrument(s) in question. However, high frequency financial data are somewhat hard to model or predict as stochastic processes and many other random factors are involved. It would be valuable information for the investor if he or she knew which precious metals were quantitatively more predictable, that would also be a good basis for more robust trading decisions. The objective in this study is to build predictive models on high frequency precious metal data and compare predictabilities of different metals using only past price and volume values which should be a basis for robust trading decisions. The data used consist of various frequencies from 1-minute to 4-hour covering a period of almost 20 years for each instrument and frequency. Artificial Neural Network (ANN) and Gradient Boosted Decision Tree (XGB) methods are applied. Comparable results are achieved

  • Finding anomalies with artificial neural network

    Industry 4.0, Vol. 4 (2019), Issue 3, pg(s) 128-129

    Nowadays all companies and corporations have their own external and internal servers with information that require specialized software for their support and configuration. Sometimes when data is exchanged with other external or internal sources for unwanted reasons data traffic may be different than expected. In this case, artificial neural networks may be used to monitor the traction. Thanks to their ability to learn the artificial neural networks can detect an anomaly in communication traffic.

  • SCIENTIFIC BASES OF CREATION OF HIGHLY EFFECTIVE BIOACTIVE COATINGS FOR BONE IMPLANTS

    Machines. Technologies. Materials., Vol. 8 (2014), Issue 9, pg(s) 32-35

    In this paper, we present an object detection system and its application to plasma sprayed coatings implants with the classifier based on the Principal Component Analysis (PCA). In order to improve performance of the classifier, we used combinations of halftone images and gradient images generated by the Sobel operator. To improve the quality of plasma coatings we apply intelligent control methods based on artificial neural networks and Bayesian network for optimization the weights