Machine Learning Sub-Models for Personalized Accident Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning algorithms for accident event assessment lack personalization, failing to effectively differentiate between various physical and behavioral characteristics of individuals, which affects their accuracy and speed in detecting falls and near-falls across different scenarios.

Innovation Solution

A computer-implemented method for training machine learning models that involves preparing training data by acquiring sensor data from subjects, dividing them into sub-groups based on unique features, and training both initial and sub-models to cater to specific groups or individuals, using sensors like motion sensors, 3D sensors, and cameras, and incorporating features such as accident type, environmental conditions, age, and medical conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If a universal machine learning model is used for all subjects, then device complexity is reduced and ease of operation is improved, but measurement precision and reliability of accident event detection deteriorate due to inability to differentiate individual physical and behavioral characteristics

Engineering Contradiction:
Improveease of operationVSAvoidmeasurement precision
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent segments the training data by dividing subjects into different groups based on physical and behavioral characteristics. Multiple machine learning models are trained, each specialized for specific subject groups, rather than using a single universal model. This segmentation allows each model to achieve higher measurement precision for its target group while maintaining ease of operation through automated model selection based on subject features.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If multiple customized machine learning models are trained for different sub-groups and individuals, then measurement precision and reliability of accident event detection are improved, but device complexity and training time increase

Engineering Contradiction:
Improvemeasurement precisionVSAvoiddevice complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing subjects into manageable sub-groups based on key characteristics, creating specialized models for each group. This reduces the overall complexity compared to creating fully individualized models for every subject, while still achieving high measurement precision within each segment.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates a universal framework that can handle multiple subject types through a family of models. The system universally processes different subject groups using appropriate specialized models, achieving multi-functionality without requiring completely separate systems for each subject type.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Adaptability or versatility

If training data is collected and processed for individual subjects, then adaptability of the machine learning model to individual characteristics is improved, but loss of time for data collection and model training increases

Engineering Contradiction:
ImproveadaptabilityVSAvoidloss of time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent segments subjects into groups with similar characteristics, allowing adaptability to be achieved at the group level rather than requiring full individualization. This reduces the time needed for data collection and model training compared to completely individualized approaches, while still maintaining high adaptability within each segment.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements partial personalization by creating models adapted to subject groups rather than complete individual customization. This partial action approach achieves sufficient adaptability for accurate accident detection without the excessive time investment required for fully individualized models for every subject.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4414959A1Computer-implemented method for training of a machine learning model in accident event assessment
Publication Date: 2024.08.14 AUTOLIV DEV AB
  • EP4414959A1 patent drawingFigure 1~2
  • EP4414959A1 patent drawingFigure 3~5
  • EP4414959A1 patent drawingFigure 6

AI summary

The present disclosure relates to a computer-implemented method for training of machine learning models in person accident event assessment. The method comprises preparing (S200) training data, by obtaining and automatically acquiring (S210) sensor data generated from sensors (2, 3; 6, 7) at an initial group (200) of subjects (201a-201c), and dividing (S220) the group (200) of subjects (201a-201c) into sub-groups (200a-200h) associated with certain corresponding features that are unique for the subjects (201a-201c) in that sub-group (200a-200h). The method further comprises training (S400) an initial machine learning model for the subjects in the initial group (200) and training (S500) a plurality of machine learning sub-models for the subjects (201a-201c) in the corresponding sub-group (200a-200h), such that one machine learning sub-model for each sub-group (200a-200h) is obtained.