Time-Series Feature Comparison for Accurate Abnormal State Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting abnormalities in time-series data face challenges due to the difficulty in balancing the duration of feature waveforms, which affects the variation in feature data and leads to increased false detections.
Innovation Solution
A time-series data processing method that compares feature amounts of partial data segments across different positions in the first time-series data with corresponding segments in the second time-series data, allowing for the detection of specific states based on these comparisons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the period of partial time-series data is set to a long duration, then the variation in feature data is improved, but it becomes difficult to prepare learning data covering features of various types of partial time-series data
Solution Approach 1:
The patent divides the time-series data into multiple periods with different durations (first period, second period, third period) to extract feature amounts. This segmentation allows the system to capture both short-term variations and long-term trends, resolving the contradiction between maintaining high detection accuracy and covering diverse feature types.
Solution Approach 2:
The patent changes the parameter of period duration by using multiple periods with different lengths. The first period has a longer duration for capturing overall trends, while the second and third periods have shorter durations for capturing specific variations. This parameter variation enables the system to maintain both reliability and adaptability in feature extraction.
2Adaptability or versatility
If the period of partial time-series data is set to a short duration, then it is easier to prepare diverse learning data, but the variation in feature of the partial time-series data is reduced, leading to increased false detections
Solution Approach 1:
The patent segments the feature extraction process into multiple periods with different durations. The shorter second and third periods provide diverse learning data, while the longer first period ensures sufficient variation for accurate detection. This multi-scale segmentation resolves the contradiction between data diversity and detection reliability.
Solution Approach 2:
The patent merges the results from multiple periods with different durations to make the final determination. By combining feature amounts from the first period (long duration) and the second/third periods (short duration), the system achieves both diverse learning data coverage and high detection accuracy, eliminating false detections.
3Device complexity
If a single period duration is used for feature extraction, then the processing is simplified, but it is difficult to appropriately detect specific situations such as abnormalities
Solution Approach 1:
The patent divides the feature extraction process into multiple periods (first period, second period, third period) with different durations. This segmentation allows the system to detect specific situations like abnormalities by comparing features across different time scales, while maintaining manageable processing complexity through systematic comparison of feature amounts between periods.
Data Source
AI summary
A time-series data processing system 100 according to the present invention includes a comparison unit 121 configured to compare a feature amount of each of pieces of partial time-series data in a plurality of specific periods at different positions on a time axis in first time-series data and a feature amount of each of pieces of partial time-series data in a plurality of periods located in correspondence with a positional relationship among the plurality of specific periods in second time-series data, and a detection unit 122 configured to detect a specific state in the second time-series data based on a result of the comparison.