Research On Vortex Detection And Feature-Based Flow Visualization

To visualize terascale and petascale data rapidly, feature detection operates as a data compression technique by reducing the amount of data that needs to be analyzed to a set of feature descriptors or a feature catalog. There are two distinct paradigms that can be employed to identify a feature:local and global. The local approach, or point classification, operates on a small neighborhood of the data and performs a binary classification as to whether a discrete point belongs to a feature. The collection of identified data points can then be aggregated to form the feature. In contrast, a global approach identifies a feature by an aggregate classification strategy and requires information from nonlocal regions of the dataset. For certain feature types, the global approach can be more discriminating; however, this increased discrimination comes with an increased cost and, for exascale data, the resulting cost is prohibitive.In many fluid dynamics applications, vortices are the feature of interest. There are many algorithms that have been created to identify vortices. Unfortunately, they may encounter situations in which the detector incorrectly indicates the presence of feature (false positive) or fails to locate an existing feature (false negative).These occur because the detector is not based on a formal, rigorous definition.We conceptualize this as the problem of robustness in feature detection and aim to judiciously combine different detection algorithms by using meachine learning methodologies. With the supervise of expert labels, This compound classifier combines the advantages of each individual local classifier and increasingly approximate the ideal detection results. Since there is no literature about implementing machine learning on flow visualization, the apporach we propose will be guidance on this field. Besides, the apporach could be extended to multi-resolution flow visualization to get more accurate results without expensive cost.The main contributions of the thesis are as follows:1. We propose the algorithm to use experts to identify the regions of the domain that contain vortices area in fluid dynamics datasets relying on streamline.These expert labels will be the training data for machine learning and the criteria to measure the performance of some detectors. Although the streamlines would be affected by some unexpected factors, they have been being the common tools because they present both the local and global feature. Regards of the other detectors, experts in the domain label the flow feature according to the streamlines and other physical feature. The method combines the accuracy of mannual detectors and the flexibility of auto detectors to provide the best base of machine learning.2. We present the algorithm to determine the best threshold which is crucial for detectors. In the algorithm, ROC space is the major tool we use to measure the performance of detectors and to find the best threshold according to the given precision. We define the measure as the distance between coordination of ideal detectors and coordination of a detector. First, meanly abstract temporary thresholds from a range as large as including the best threshold in terms of the distribution of the samples.Compute the perfomanc for each temporary threshold to get the best one. According to the best temporary threshold, change the range to abstract new temporary threshold. Repeat this process under the given precision. The method we described above combines some measure of statistic and is easy to implement on other relevant field.3. We present the algorithm to enhance vortex detection via boosting technology which is classical method of machine learning. The algorithm, based on natural feature of flow data, judiciously combine different detection algorithms to a new strong vortex detector. Ideally, the compound classifier would combine the best of all local vortex detection algorithms as they respond to the underlying physical signal. Because of the inadequation of flow feature and the simliarity of existing vortex detection algorithm, we propose the strategic way to generate branch detectors to avoid that few vortex detectors dominate the final strong detecto. The experiments show that this algorithm outperforms previous method for vortex detection.4. We present the algorithm to enhance vortex detection via CAVIAR which is based on the feature distance between samples. The significant problem of this algorithm is to define feature distance in flow data as well as the samples’ neighbors. Beside the natural physical feature, we introduce the relative position to classify samples to enhance the vortex detection results. Some parameters in the algorithm are determined in the process of cross validation. Our feature-based approach to use CAVIAR in vortex detection can learn selectively and provides diverse enhanced detectors for samples with different feature. Therefore, the final enhanced detectors have better performance than previous vortex detection algorithm.All those techniques can improve the performance to detect vortex in flow. They can be widely used to flow visualization of large dataset to analyze features in flow and get a high effective performance.Since machine learning methodologies have not invade the visualization domain, our research could be a reference of this field.