Anomaly Score Learning Model for Biased IoT Detection Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing abnormality detection systems for IoT devices face increased processing times due to biased data distributions, particularly when learning with imbalanced data sets, leading to erroneous detections.

Innovation Solution

A training device that selectively learns and generates models from unlearned data with anomaly scores above a threshold, iteratively refining the model to reduce biased data and shorten learning time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If learning is performed using all learning data in a single stage, then learning time is short, but learning accuracy deteriorates due to biased data distribution

Engineering Contradiction:
Improvelearning speedVSAvoidlearning accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The learning process is segmented into multiple stages: first learning common patterns from all data, then selectively learning rare patterns from unlearned data with high anomaly scores. This segmentation resolves the contradiction by allowing different learning objectives in different stages, achieving both speed and accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary learning on common data patterns before addressing rare patterns. By first establishing a baseline model from all data, then iteratively refining with selected unlearned data, the system prepares the learning process in advance to handle both common and rare cases effectively.

Inventive Principle:
Principle #10Preliminary action

2Manufacturing precision

If two-stage learning is used to handle biased data distribution, then learning accuracy improves, but processing time increases

Engineering Contradiction:
Improvelearning accuracyVSAvoidprocessing time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The system extracts only the necessary unlearned data (those with high anomaly scores) for the second learning stage, rather than relearning all data. This extraction approach maintains learning accuracy by focusing on rare patterns while significantly reducing processing time compared to complete two-stage learning.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of performing complete two-stage learning on all data, the system applies partial action by selectively processing only unlearned data with high anomaly scores. This partial approach achieves the necessary learning accuracy improvement without the full time cost of traditional two-stage methods.

Inventive Principle:
Principle #16Partial or excessive action

3Adaptability or versatility

If all normal data is used for learning, then model coverage is comprehensive, but detection accuracy for rare patterns deteriorates

Engineering Contradiction:
Improvemodel coverageVSAvoiddetection accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The learning process is segmented into common pattern learning (from all data) and rare pattern learning (from selected unlearned data). This segmentation allows the model to maintain comprehensive coverage while achieving high detection accuracy for rare patterns through focused iterative learning.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary learning on all data to establish comprehensive coverage, then iteratively refines the model by focusing on unlearned data with high anomaly scores. This preliminary comprehensive learning followed by targeted refinement achieves both coverage and accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4202800B1Learning device, learning method, and learning program
Publication Date: 2026.03.11 NT T INC
  • EP4202800B1 patent drawingFigure 1
  • EP4202800B1 patent drawingFigure 2
  • EP4202800B1 patent drawingFigure 3

AI summary

A generation unit (131) learns data selected as unlearned data among learning data and generates a model calculating an anomaly score. A selection unit (133) selects, as unlearned data, at least some of data in which an anomaly score calculated by the model generated by the generation unit (131) is equal to or greater than a threshold among the learning data.