Loading...
Search
Search in this resource
sort by
Send to Friend
 Digital Object List
 Bookmark
 Review List
  1. MIMO radar signal design to improve the MIMO ambiguity function via maximizing its peak
منطبق کردن سری‌ های زمانی با شبکه‌ های عمیق
7822 viewed

منطبق کردن سری‌ های زمانی با شبکه‌ های عمیق

نوربخش، علیرضا Nourbakhsh, Alireza

Deep Learning-Based Multiple Time Series Alignment

Nourbakhsh, Alireza | 2025

9 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 58848 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Mohammadzadeh, Naejesolhoda
  7. Abstract:
  8. The analysis of time series data within machine learning frameworks necessitates robust alignment techniques to address inherent misalignments caused by temporal warping and amplitude variations. Such distortions, prevalent in signals representing human activity, physiological data, and economic indicators, can severely degrade the performance of subsequent classification and analytical models. While the traditional Dynamic Time Warping (DTW) algorithm effectively addresses the pairwise alignment of time series, its quadratic computational complexity renders its direct application to the Multiple Time Series Alignment (MTSA) problem computationally intractable for large-scale datasets. Although Multiple Sequence Alignment (MSA) is a well-established field in bioinformatics, a conspicuous paucity of efficient methodologies persists for the alignment of numerical time series. Here we propose a novel method for MTSA, called MSAN: Multiple Series Alignment Network, leveraging the representational power of deep learning. Diverging from conventional pairwise strategies, MSAN performs a global, simultaneous alignment of the entire set of time series, thereby yielding substantial gains in computational efficiency. We conceptualize the alignment path as a series of piece-wise linear transformations, constrained by boundary, monotonicity and continuity conditions. This modeling enables the formulation of a custom loss function for a Deep Convolutional Neural Network, designed explicitly to eliminate the inherent limitations of DTW. Empirical evaluation on benchmark datasets from the 2018 UCR Time Series Archive demonstrates that our approach affords significant improvements in downstream classification accuracy (3 to 8 percents), warped averaging tasks, and runtime efficiency (more than 4 folds) compared to state-of-the-art methods across a majority of the datasets. Also, MSAN has been applied to important tasks such as Local Alignment and Video Alignment with acceptable results
  9. Keywords:
  10. Time Series ; Dynamic Time Wrapping ; Deep Learning ; Classification ; Time Series Alignment ; Multiple Time Series Alignment (MTSA)

 Digital Object List

 Bookmark

...see more