Cosmin Stirbu

Weights set selection method for feed forward neural networks

In this paper is described a method for weights set selection for a feed forward neural network, based on the fault tolerance analysis of the network. For a certain neural network used in a specific problem, one can obtain many weight sets, as a result of backpropagation training algorithm, due to the fact that this algorithm initializes the weights with random numbers. Each time we repeat the training, we obtain a new set of weights. We propose a method to select one of the available training sets of weights, taking into account the fault tolerance of the network. We considered as a typical fault, the fault of the neurons from the hidden layer. We developed a Java application to illustrate the proposed method.

A Java application for the failure rate prediction using feed forward neural networks

In this paper we show the possibility to use feed forward neural networks for failure rate prediction, and this can be used for improving predictive maintenance. We use a series of real values that represent the failure rate of a radio-reception system, and describe a Java application that we developed, that simulates a feed forward neural network which is trained to predict, based on the available series of failure rate values, the next failure rate value. We use a one hidden layer neural network that has a single output, the predicted value. The inputs of network are analog, continuous values between 0 and 1, and represent the previous known values of failure rate. Using the software, we used different numbers of input (2, 3, and 5) and the accuracy of prediction is compared for these different numbers.

Traffic management system

This papers purpose is describing and highlighting of how our system is intended to work and how it can help us in everyday life, as well as making the Earth a safer and cleaner place, by not even noticing its presence. Our system is always monitoring the traffic of every major city, analyzing it and making decisions on how people should drive, in a manner that the roads are safer, the fuel consumption is reduced, as well as the pollution. The system also takes in consideration the traffic jams and gives hints on how we can avoid the jammed traffic so we can be on time to our destination. On the other hand, we have to consider the pollution that appears when everybody drives his own car instead of using the public transportation systems. Our system will receive information from strategic placed sensors that measure the pollution level, and after interpreting the information, takes decisions on how we can reduce the level of pollution.

System for monitoring and controlling renewable energy sources

This article shows the possibility of using renewable energy sources in order to improve energy efficiency, reduce greenhouse gas emissions and therefore prevent climate change. This article proposes a remote monitoring and control system with interfaces and data collectors. In addition to research, development, testing and use of renewable energies, it is also necessary to measure and track online, from a distance all the phenomena that occur. Due to the evolution of information and communications technologies, monitoring and control techniques, they need to be tested in practice before marketing. This ensures both scalability and interoperability, demonstrating the possibility of collecting, centralizing, analyzing energy consumption and CO2 data. In this way, measurable character, data transparency, company acceptability, planning and visibility of renewable energy usage can be improved. And of course, its impact on the environment must be monitored.

Using the Phone’s Light Sensor to Detect the TV Video Stream

Current smart devices (phones, tablets, etc.) have integrated light sensors to adjust the screen’s brightness to the ambient light. The light sensors have become more sensitive and are even able to read the RGB light components. In Android, this information can be accessed without special access rights for the application. An application can use the information from the light sensor to detect the ambient light variations and relay this information to a server where it can be used to determine the video information being displayed. This paper details the data flow and tests the implementation for a single video flow on multiple light sensors.

A Java simulation software for the study of the effects of the short-circuit faults in a feed forward neural network

Feed forward neural networks have an intrinsic fault tolerance to the faults of neurons from the hidden layer. In this paper is presented a simulation program that analyses the behaviour of a feed forward neural network, with a single hidden layer, in the presence of faults. The neural network is used to classify a binary image in four classes. The fault that is analysed is the short circuit of the output of one ore more neurons from the hidden layer. It is also analysed the influence of the dimension of the hidden layer (number of neurons) on the behaviour of the neural network in the presence of faults. For the training of the neural network it is used the backpropagation algorithm. The simulation program is written in the Java language.

A Java application for the selection of a weights file for a two hidden layers feed forward neural network

In this paper we describe the way in which we select the weights set for a two hidden layers feed forward neural network. The weights are selected based on the fault tolerance of the neural network. We developed a Java application for training the network that generates more valid weights files and then the application selects the file that offers the maximum fault tolerance to the faults of the hidden neurons.

Cellular automata based algorithm for image density classification task

In this paper is presented a solution based on a bi-dimensional cellular automata (CA) for image density classification task (DCT). The two necessary properties: density preserving and translation are combined together to obtain the DCT solution. These two properties are achieved using a combination of nine fundamental 2D-CA rules and the proposed solution for DCT has two phases: preprocessing phase and decision phase. The project has been implemented in software using C# programming language and experimental results are presented for images with different sizes. This work is totally related to the idea of understanding the impact of the inherently local information processing of CA on their ability to perform a coordinated computation at the global level, as achieved by an evolutionary process.