Time-Series Classification With Quickest Change Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing techniques struggle with unclear event boundaries in series data, making it difficult to appropriately assign class labels.

Innovation Solution

A classification apparatus and method that applies quickest change detection repeatedly to series data to assign class labels to each piece of element data, using methods like the dynamic cumulative sum test (D-CUSUM) to determine class transitions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional temporal localization techniques are used, then labeling can be performed on clear events, but labeling accuracy deteriorates when event boundaries are unclear

Engineering Contradiction:

Improvelabeling accuracyVSAvoidperformance on unclear boundaries

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The patent segments the event detection process into multiple stages: initial event candidate extraction, boundary refinement through change detection, and final label assignment. This multi-stage segmentation allows the system to handle unclear boundaries by progressively refining event segments rather than attempting single-step classification.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary actions by first identifying event candidates and establishing initial time ranges before final classification. Change detection algorithms are applied in advance to pre-process the series data, marking potential boundary points before the actual labeling decision is made, which improves accuracy for unclear events.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If repeated quickest change detection is applied to all series data, then labeling accuracy for unclear boundaries improves, but computational complexity increases

Engineering Contradiction:

Improveboundary detection accuracyVSAvoidcomputational complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies quickest change detection selectively rather than uniformly across all data. Change detection is focused on regions where event boundaries are likely to occur or where uncertainty is high, rather than processing every data point with the same computational intensity, thus reducing overall complexity while maintaining accuracy.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent applies change detection algorithms with varying degrees of intensity or repetition based on local data characteristics. For segments with clear boundaries, minimal detection is performed, while segments with unclear boundaries receive repeated or more intensive change detection analysis, optimizing the balance between accuracy and computational load.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS12423331B2Classification apparatus, classification method, and recording medium

Publication Date: 2025.09.23 NEC CORP

AI summary

In order to suitably perform labeling even in a case where there is an unclear event boundary, a classification apparatus (1) includes: an acquiring means (11) for acquiring series data; and a class label assigning means (12) for assigning a class label to each piece of element data contained in the series data, by repeatedly applying quickest change detection to the series data.