Retrospective Self-Training for Time-Series Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Real-time classification of time-series data in machine learning applications, such as assistive devices, faces challenges when the application environment differs from the training data, leading to poor performance due to individual human behavior and environmental variations.

Innovation Solution

A method that adapts the real-time classification model by generating pseudo-labeled data through semi-supervised learning and iterative self-training, allowing the model to be retrained based on both initial labeled data and environment-specific data, combining offline and online classification approaches.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a real-time classification model is pre-trained on large amounts of labeled data, then the model can perform real-time classification, but the performance deteriorates when the application environment differs from the training data due to individual human behavior and environmental variations

Engineering Contradiction:
Improvereal-time classification capabilityVSAvoidclassification accuracy in new environment
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system transitions from a static pre-trained model to a dynamic model that continuously adapts to new environments through iterative self-training. The model updates its parameters using pseudo-labeled data from the specific application environment, enabling it to maintain high accuracy across different users and settings while preserving real-time classification capability.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system performs self-training by automatically generating pseudo-labeled data from unlabeled time-series data in the target environment and using this data to retrain the classification model. This self-service mechanism eliminates the need for manual annotation of environment-specific data, allowing the model to adapt autonomously to new application contexts.

Inventive Principle:
Principle #25Self-service

2Adaptability or versatility

If the classification model is retrained using only initial labeled data, then the model maintains its original training, but it cannot adapt to the specific application environment

Engineering Contradiction:
Improveenvironment adaptation capabilityVSAvoidamount of training data
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The system creates pseudo-labeled copies of unlabeled time-series data by using the pre-trained model to predict labels for data in the target environment. These pseudo-labeled copies are then used to train the model, effectively multiplying the amount of usable training data without requiring manual annotation of each example.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system performs preliminary classification of time-series data in the target environment using the pre-trained model before actual use. This preliminary action generates pseudo-labeled data that can be used for subsequent model refinement, allowing the system to prepare adaptation material in advance without delaying actual classification operations.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If manual labeling of time-series data is performed, then the training data is accurate, but the process is time-consuming and expensive

Engineering Contradiction:
Improvelabel accuracyVSAvoiddata preparation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system automatically generates pseudo-labeled copies of time-series data by using the pre-trained model to predict labels. This copying process replaces manual labeling, producing accurate enough labels for model refinement without requiring human annotators, thereby significantly reducing data preparation time and cost while maintaining sufficient label quality for effective model adaptation.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4195109A1Online time-series classification with retrospective self-training
Publication Date: 2023.06.14 HONDA MOTOR CO LTD
  • EP4195109A1 patent drawingFigure 1
  • EP4195109A1 patent drawingFigure 2
  • EP4195109A1 patent drawingFigure 3

AI summary

The present invention relates to a method for classifying time-series data comprising classifying (S2) the time-series data by a real-time classification using a real-time classification model pre-trained based on labeled data, buffering (S4) the time-series data that has already been classified by the real-time classification and optimizing (S5) the real-time classification model during the real-time classification by generating pseudo-labeled data by semi-supervised learning based on the labeled data and the buffered time-series data and retraining the real-time classification model based on the labeled data and the pseudo-labeled data.