this flexiblee platform

Transcription

1 COST Office Avenue Louise Brussels, Belgium t: +32 (0) f: +32 (0) Improvement of atmospheric microsensor prototypes and software development for managin ng their operation Scientific report Siamak Aram, Poltecnico di COST STSM TD Host Institute: Prof. Kostas Karatzas, Aristotle University of Thessaloniki, Informatics Systems and Applications Group, Dept. of Mechanical Engineering, Thessaloniki, Greece From to Purpose of the STSM; The main purpose of my short term scientificc mission was to improve a previously proposed prototype for the acquisition of environmental informationn in order to increase its stability in application level with Computational Intelligence methodologies for the analysis and forecasting of air pollutant data. Last but not least to discusss and consult with researchers colleagues about their methods and results that were very important for my PhD thesis and my future science career. and its effectiveness concerning the sensor power consumption. Also, to became familiar 2. Description of the work carried out during the STSM Firstly, the sensor prototypes [1,2], their functionalities and related application are presented to the AUTh scientists. In several working sessions (with K. Karatzas, V. Epitropou, A. Bassoukos and M. Riga at AUTh), possible software and operational refinements of the sensor prototypes and cross testing with existing sensors at AUTh were discussed and studied for designing, developing and improving on current works. In a working session, with student and researchers, working on the development of software applications for Participatory Environmental Sensing (PES) at AUTh, new software design and development methods weree presented. In the frame of working sessions the STSM beneficiary was received training in PES for furtherr developing the software mapping the sensorss and controlling data acquisition. Using different data formats, parsing, and communication channels were found to be key issues for general purpose WSNs applications. Also communicationn of a mobile phone with several different nodes types can be burdensome task to handle. Such communication problems were discussed and related work were studied (literature review),to find which possible approaches can be considered to have pervasivee or flexible platform to simplify the interfacee between a variety of external sensors and Android devices users. Besides, applying this flexiblee platform of reading sensors dataa by smartphones to decrease the power consumption of the network is studied and worked. Moreover, the development and comparison of computational intelligence methods for the investigation and forecasting of the quality of the atmospheric environment was studied by the beneficiary and compared with methods that he had already used in the past. First, the undersampling method in [3,4] were applied in orderr to gain further insight of ANN based data modeling. As a next step, it was agreed to have a detailed scientificc collaboration between AUTh, Neuronica Lab group (Torino) and the Finnish Meteorological Institute (FMI), in orderr to involve data and methods used in [5], to new computations to be performed in Torino after the STSM. In doing so, comparing forecasting algorithms and combining the Torino s group approach with

2 AUTh s group approach were discussed and studied to follow up with a predicting that will be followed by the estimation of the uncertainty of the forecasting. Also, the risk of falsely forecasting an AQ episode was indicated. At the end, in an internal workshop, micro sensors for atmospheric observations were presented and current developments were analyzed for completing and future work. 3. Description of the main results obtained In the first step, the general idea of a pervasive platform for parsing part of acquisition system s dataa collection was discussed and studied. The beneficiary received valuable information and help from experienced AUTh researchers, and as it shown in figure 1, he developed a first and primarily version as a prototype. The AUTh approach was in favor of a platform that generalizes access issues to accept different stream types of dataa from different sensors with different data types. These generalizations make easy to have flexibility of environmental application to focus on other sensor hardware issue enabling WSNs more general. It is tested using Bluetooth sensors to gather environmental information and sampling rates to work on the platform s effectiveness and correctness. A report is written to describe related works and some technical aspects until recent steps. The AUTh s group has been working on this area for years and is now in the process of further developing the software platform for sensor handling, where the beneficiary and Torino group will also collaborate. Figure 1. The interface of the new prototype application that can communicate and reading dataa from two sensors with two different data types. Secondly, in the other main step, data gathering part of the application is studied and worked separately by considering the power consumption and due to the new approach on buffer management and time controlling to read data, as it shown in figure 2., it is possible to have lower power consumption. Reducing the time of reading and controlling buffer that consequently decrease the number of network s communications will be allowed sensors to have more no communication time. In this work, power awaree buffering in battery powered sensor networks for both fixed size and fixed interval buffering schemes is studied. Preliminary results obtained by reading environmental information of AUTh s lab using Bluetooth based sensors shows that it is possible to have more than 50% reduction on time of reading. A report is written to describe related works and technical aspectss until these steps. 2

3 A) B) Time decreasing Figure 2. Power measurement of one sensor during reading and not reading states. A) The power consumption of the Bluetooth based temperature and humidity acquisition system is measured previously and explained in [3]. B) The primary result shows thatt the time is needed to read with the same system of acquisition by handling buffer is 50% reduced. In the other main step, the proposed method of prediction that is used in [4], was applied on Agsofias station data on its four air pollution parameters (Temperature, Humidity, PM10 and O3) as a test. It will be followed by a comparison with the forecasting methods used in [5]. The uncertainty of data measurements is predicted (using a Multi Layer Perceptron MLP). A further measurement is necessary when the uncertainty of the prediction underwent a threshold. Each available measurement is considered together with its uncertainty. The MLP applied a forward prediction on 100 realizations of stochasticc inputs that are mined from a Uniform probability distribution with both mean and range values obtained by the available data and their uncertainty, respectively. The prediction is computed as the mean of the obtained 100 estimations. Then, the present and past data of the obtaining value are restructured by interpolation method from the acquired measurement ts and their uncertainty. Also, an extra interpolation is used for preparing data beforee applying the algorithm. In Figures 3 and 4, a representative example of application of the algorithm to Agsofias data is shown as testing and training set data. The method is adapted to the investigated data, in the sense thatt a larger or lower number of sampless are required from the data monitoring a chaotic or a deterministic system, respectively. The algorithm is adoptedd according to this data by analyzing on the number of iteration and neurons in hidden layer to reduce the percentage of errors in fitting test data after undersampling and errors in fitting for the training data that was high before any modification. As it shown in figures 5 and 6, 70 hidden layer neurons have about 49% of training error and 12% after undersampling. Moreover, ncreasing the number of iteration between 500 and 600 will be decreased error on training and after undersampling to about 50 48% and 12% respectively (as it shown in figures 7 and 8). 3

4 Figure 3. Example of application of the algorithm to the one station of the Thessaloniki s data. The first 60% of samples are used to train the prediction algorithm (based on a Multilayer Perceptron) and the subsequent 40% is used for validation (20%) and testt (20%) of data. Uncertainty of the measure was obtained assuming a uniform variability around the measure. Figure 4. Example of test of application by the algorithm to the one station of the Thessaloniki s data. 4

5 Error % Number of Hidden Layer s Neurons Figure 5. Analyzing the number of neurons in hidden layer in order the percentage of the error of training data. Error % Number of Hidden Layer Neurons Figure 6. Analyzing the number of neurons in hidden layer in order the percentage of the error of fitting after undersampling. 5

6 Error % Number of Iterations Figure 7. Analyzing the number of iteration in order the percentage of the error of training data. Error % Number of Iterations Figure 8. Analyzing the number of iteration in order the percentage of the error of fitting after undersampling. 6

7 4. Future collaboration with host institution (if applicable) Developing and completing the software for a flexible parsing platform for android application and holding a common event to introduce the platform. 5. Foreseen publications/articles resulting or to result from the STSM (if applicable) A Pervasive Parsing platform for Android Based Environmental Sensing Applications Improved Sensor s lifetime Using Smartphone by applying a flexible data gathering 6. Other comments (if any) Siamak Aram, 23 rd October Annex: Confirmation by the host institution of the successful execution of the STSM The letter is attached at the end of report. 7

8