Machine Learning Model Training With Multi-Interval Normal Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional methods for training machine learning models struggle to effectively detect abnormal system conditions, particularly in devices with intermittent operations, as they require selecting only one time period for training data and fail to combine datasets across multiple intervals or include predicted future values.

Innovation Solution

An information processing apparatus and method that acquires normal data from multiple intervals, combines it to generate combined data, and uses this data to train a machine learning model that outputs a score indicating the system's normal condition, allowing for the detection of abnormal conditions by incorporating both past and predicted data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If only one time period is selected for training data, then the training process is simple, but the detection accuracy of abnormal conditions is insufficient

Engineering Contradiction:
Improvedetection accuracyVSAvoiddata processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent combines multiple time periods of normal operation data into a unified training dataset. The data combining unit integrates data from different time periods (e.g., morning operation, afternoon operation) to create a comprehensive training set, enabling the machine learning model to learn from diverse operational patterns and improve detection accuracy across various conditions.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent segments the operational data into multiple distinct time periods and processes each segment separately before combining them. The data acquiring unit divides the dataset into multiple time periods, allowing selective acquisition of normal data from each segment, and the combining unit then integrates these segmented datasets systematically.

Inventive Principle:
Principle #1Segmentation

2Quantity of substance

If data from multiple intervals are combined, then the training data quantity increases, but the data processing time increases

Engineering Contradiction:
Improvetraining data quantityVSAvoiddata processing time
Core Design Contradiction:
Quantity of substanceVSLoss of time

Solution Approach 1:

The patent performs preliminary processing of data from multiple time periods before combining them. The data acquiring unit pre-acquires and prepares normal data from each time period separately, organizing the data in advance so that the combining unit can efficiently integrate pre-processed datasets without incurring excessive processing delays during model training.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If normal data from multiple time periods are used, then the model generalization improves, but the risk of including abnormal data increases

Engineering Contradiction:
Improvemodel generalizationVSAvoiddata quality
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent extracts only the normal data portions from each time period while excluding abnormal data. The data acquiring unit is configured to acquire specifically normal data from multiple time periods, filtering out abnormal conditions. This extraction approach ensures that the combined training dataset contains only representative normal operational patterns, improving model generalization without compromising reliability.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4332707A1Apparatus, method and programme for processing information
Publication Date: 2024.03.06 YOKOGAWA ELECTRIC CORP
  • EP4332707A1 patent drawingFigure 1
  • EP4332707A1 patent drawingFigure 2
  • EP4332707A1 patent drawingFigure 3~4

AI summary

An information processing apparatus (10) acquires pieces of normal data belonging to a plurality of intervals of measurement data measured by a device, combines the acquired pieces of normal data to generate combined data, and performs machine learning using the combined data to train a machine learning model (14c) that outputs a score indicating a normal condition of a system in which the device is installed, in response to an input of measurement data.