Dual Channel Network for Multivariate Time Series Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional time series retrieval methods fail to account for static system statuses, which are essential in determining system behavior, leading to inaccurate results as they only consider dynamic time series data.
Innovation Solution
A dual-channel network is introduced, combining a Multi-Layer Perceptron (MLP) based static encoder with a Recurrent Neural Network (RNN) based temporal encoder, jointly trained using metric learning loss to encode time series segments with static statuses into compact binary codes, enabling effective multivariate time series retrieval.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional time series retrieval methods are used that only consider dynamic time series data, then the retrieval process is simple, but the retrieval accuracy deteriorates because static system statuses are not accounted for
Solution Approach 1:
The system segments the encoding process into two distinct channels: a static status encoder that processes static system statuses, and a temporal encoder that processes dynamic time series data. These two encoders work in parallel and their outputs are combined to form a comprehensive representation, allowing the system to capture both static and dynamic characteristics without overwhelming complexity
Solution Approach 2:
The patent merges the outputs of the static status encoder and temporal encoder by concatenating their respective feature vectors. This combined representation is then passed through a binary code extractor to generate binary codes that encode both static and dynamic information, achieving accurate retrieval while managing complexity through efficient feature fusion
2Productivity
If binary codes are generated to preserve relative similarity relations, then retrieval efficiency is improved, but information loss may occur during the encoding process
Solution Approach 1:
The system transforms continuous feature vectors from the encoders into discrete binary codes through a binary code extractor. This parameter transformation enables efficient similarity search using Hamming distance while the metric learning loss function ensures that the binary codes preserve relative similarity relationships, minimizing information loss during the discretization process
Solution Approach 2:
The patent replaces traditional continuous-space similarity computation with discrete binary code comparison using Hamming distance. This substitution dramatically improves retrieval efficiency by enabling bitwise operations instead of computationally intensive continuous vector comparisons, while metric learning ensures the binary representation maintains similarity relationships
3Measurement precision
If metric learning loss is used for joint training, then the preservation of relative similarity is improved, but training complexity increases
Solution Approach 1:
The metric learning loss function provides feedback during training by computing the difference between predicted similarity (based on binary code Hamming distance) and actual similarity (based on ground truth labels). This feedback gradient guides the optimization of both static status encoder and temporal encoder parameters, ensuring they learn to produce binary codes that preserve relative similarity relationships
Solution Approach 2:
The metric learning loss function serves multiple purposes: it trains both encoders simultaneously, ensures binary codes preserve similarity relationships, and enables end-to-end optimization of the entire retrieval system. This multi-functional loss formulation manages training complexity by unifying the optimization objective across different components
Data Source
AI summary
A computer implemented method is provided. The method includes jointly encoding, by a dual-channel feature extractor, a current time series segment with corresponding static statuses into a compact feature. The method further includes converting, by a binary code extractor, the compact feature into a binary code. The method also includes computing distances between the binary code and all binary codes stored in a binary code database. The method additionally includes retrieving the top relevant multivariate time series segments based on the distances.