پیش بینی رویداد نادر مبتنی بر شبکه های بیزی با داده های سنسور

A Bayesian network is a powerful graphical model. It is advantageous for real-world data analysis and finding relations among variables. Knowledge presentation and rule generation, based on a Bayesian approach, have been studied and reported in many research papers across various fields. Since a Bayesian network has both causal and probabilistic semantics, it is regarded as an ideal representation to combine background knowledge and real data. Rare event predictions have been performed using several methods, but remain a challenge. We design and implement a Bayesian network model to forecast daily ozone states. We evaluate the proposed Bayesian network model, comparing it to traditional decision tree models, to examine its utility.

Scientists have reported the adverse effects of high ozone (O3) concentration [4] and [20]. Kelsall et al. [17] researched the cause and effect of air pollution and mortality using 15 years data from Philadelphia. Schwartz et al. [26] found associations between acute respiratory symptoms in children and summer air pollution. In the light of these health effects, many models have been suggested and developed during the past two decades. Previous models are mainly based on empirical models and/or expert opinions, statistical models, causal models, or combinations of these [5], [15], [16] and [23]. Despite the various models developed, many scientists still seek more accurate and reliable models, because O3 changes its reaction mechanism depending on altitude, location, and other factors. In this paper, Bayesian approaches are adopted to combine real measured data and expert knowledge to overcome complexity and nonlinearity of O3 reactions. The proposed approach has the following advantages compared to traditional models: • It is an almost ideal combination of sensor data and knowledge. • It is simple to handle missing or abnormal sensor data. • It is easy to understand a forecast result and its cause and effect. • It is easy to analyze high concentration O3 episodes. • It is easily applied and customized to other areas. In this paper, we especially consider the rare event state prediction field with a Bayesian network in daily maximum O3 forecasts in Seoul, Korea. The construction of the proposed Bayesian network is shown in Fig. 1. This Bayesian approach starts from historic data and expert knowledge. The historic data contains weather and pollution data when high concentration O3 events have occurred. A variety of data analysis methods have been applied to extract features and create episodes from historic data. Conversely, expert knowledge included O3 related chemical equations and hypotheses make a skeletal Bayesian network. Data analysis and the skeletal Bayesian network are completed as a Bayesian network model by the learning algorithm and belief propagation. Full-size image (27 K) Fig. 1. Construction of the proposed Bayesian network model. Figure options The Bayesian network related to forecasting O3 is discussed in Section 2. Section 3 includes the design procedures for the Bayesian network model, methods of representing the expert knowledge, and atmospheric chemistry hypotheses. The results of an evaluation of the proposed model and typical decision tree models are compared and discussed in Section 4. Conclusions will be addressed in Section 5.

In this paper, we propose a Bayesian network-based on rare event prediction methods for high concentrations of O3. Rare data prediction is a particularly challenging application for data mining, used for such things as fraud detection in finance, diagnosis in industry, and affect analysis in chemistry. It is challenging, not only because building cases of training sets is difficult, but also because the cases may have many forms, causes, and unknown relationships. We describe heuristic and statistical data analysis approaches and whole Bayesian network model design procedures using chemical reaction equations and expert knowledge. Then, we evaluate the performance of the suggested models, comparing them to decision tree models. In summary, the Bayesian model is superior to the decision tree model as evaluated by performance indices. the other merits of the Bayesian network were not described enough, because historic data was used in the experiments. However, when the belief propagation property is applied to the model, it can also be used for high concentration O3 cause analysis and to determine plans to reduce O3 in specific local areas. In the future, the suggested model will be applied to on-line O3 state prediction systems and its predictive performance evaluated. This model should also have other advantages of the Bayesian methods including robustness and reliability against disturbance and noisy data.