Efficient Top-k DTW-Based Sensor Data Similarity Search Using Perceptually Important Points and Dual-Bound Filtering

The proliferation of Internet-of-Things (IoT) devices generates vast amounts of sensor data, making sensor data mining an increasingly important research topic in the IoT domain. The Top-k similarity search plays a crucial role in sensor data mining, forming the foundation for advanced techniques such as sensor data classification, clustering, and anomaly detection. The need for efficient algorithms to enhance search speed drives our research. This article introduces an efficient method named PIPDTW for Top-k similarity search of sensor data using dynamic time warping (DTW) distance, a prevalent but computationally intensive similarity measure in time-series data mining. To address the high computational complexity, PIPDTW first uses perceptually important points (PIPs) to transform lengthy sensor data into concise vector representations. We then introduce a novel distance measure, dynamic time warping based on index and value (DTWIV), based on PIP representations to quickly estimate DTW distances between the query q and all sensor data. Finally, during the online search, we apply two levels of filtering based on approximate ranking and the lower bound technique to efficiently prune dissimilar data, reducing the need for exact DTW distance computations and thereby achieving efficient querying. Experiments conducted on synthetic and real-world publicly available datasets, compared with relevant competitors, demonstrate that PIPDTW slightly sacrifices retrieval accuracy but significantly reduces runtime.