| With the rapid development of Internet of things, a lot of practical applications are required to deploy a large number of sensing devices to collect and process data, and thus it can monitor high-quality of the physical world. However, due to the inherent limitations of these hardware devices or the influence of factors such as environment, single-modal data often fail to complete comprehensive monitoring to the characteristics of the physical world. So in order to avoid the occurrence of this problem, this thesis introduces the concept of multi-source multi-modal data, using different attributes of information (that is, multi-source multi-modal data), obtaining more accurate and objective information on the real physical world. In addition, most of the sensing data collections are based on the equal frequency sampling methods, but for the continuous variation of the physical world, in the energy-constrained conditions, equal frequency sampling method tends to ignore some key points lead to the lack of the critical information, and thus the curve may be distorted. Based on this, this thesis proposes a study for multi-source multi-modal sensory data variable-frequency sampling technology.Initially, this thesis proposes a variable-frequency sampling technique based on multi-source multi-modal sensing data. This approach takes into account the relevant characteristics between each dimension attribute data, using principal component analysis model to find an index which can explain all modal information without redundant information. According to the change of the index, we can determine the next sampling time. The method of sampling frequency is changing along with the comprehensive index. It could guarantee the acquisition of the critical information point, so that the final output of the perceptual curve are O(?) similar to the real perception curve. On this basis, this thesis also defines the accuracy of the data quality for each modality, and put forward the corresponding data improved method according to this definition. Therefore, each modal data quality in the end is also able to meet the accuracy requirements. Then, we test the frequency sampling method while using real data and simulation data to observe methods, such as the performance. Finally we can demonstrate the superiority of the variable frequency sampling algorithm.Secondly, considering the interference of external factors in the sensor sensing environmental information, the unstable environment and the energy problems, the data will be mixed in some abnormal data. Therefore, in order to avoid negative effects on the final result, this thesis also proposes a pretreatment method for multi-source multi-modal data. In this thesis, we use inherent advantages of multi-source multi-modal data. The correlations between the multi-modals data are applied to the mean absolute method. We firstly use the comprehensive index to judge the abnormal points. After that we can find the dimension data in one or more of the existence of abnormal points at that moment. Then we remove and add the corresponding estimates to get more accurate conclusions. And at last, the experimental results are derived by a lot of experiments. It shows that our algorithm is efficient and scalable. Experiments also show that this method can substantially remove the outliers in the data.Finally, this thesis is proposed based on multi-source multi-modal sensing data to fusion technology. The collected data will be output in the form of classified conclusions or decisions, in order to achieve different types of requirements. Meanwhile, this thesis also defines the accuracy of the data quality for classification results. Through experimental verification, you can compare the quality of different fusion methods.In summary, this thesis use inherent characteristics of multi-source multi-modal sensory data. We research the variable frequency sampling method and fusion technology to ensure a better quality of the final data. |
PDF Full Text Request