Vehicle Machine Learning Training Using Localized Sensor Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine learning algorithms used in vehicles for safety systems, trained with global data, often fail to accurately identify local environment items, leading to safety concerns for drivers and surroundings due to inadequate training data specific to the vehicle's environment.

Innovation Solution

A method for training machine learning algorithms using localized sensor data collected by vehicle sensors, pre-processed and labeled by users, with the option for automatic triggering based on location, time, or user input, to improve item identification accuracy in specific environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If machine learning algorithms are trained using global data from different parts of the world, then the algorithm can process diverse environments, but the algorithm fails to accurately identify local area items

Engineering Contradiction:
Improveenvironmental adaptabilityVSAvoiditem identification accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by collecting and training on localized sensor data specific to the vehicle's operating environment. The system triggers sensor data collection based on location information (such as frequently visited locations or areas with identification uncertainty) and uses this localized data to re-train the machine learning algorithm, enabling accurate identification of local items while maintaining global knowledge

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent implements dynamics by making the training data collection process adaptive and responsive. The system dynamically triggers sensor data collection based on real-time location information, user input, or identification uncertainty thresholds, and continuously re-trains the algorithm with new localized data to improve performance over time in specific environments

Inventive Principle:
Principle #15Dynamics

2Productivity

If machine learning algorithms are automatically trained by computing devices without human interaction, then the training process is efficient and scalable, but the algorithm produces wrong identification in local environments

Engineering Contradiction:
Improvetraining efficiencyVSAvoididentification reliability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies feedback by incorporating user interaction into the training process. The system transmits pre-processed sensor data and information to a user interface, receives user feedback or corrections, and uses this feedback to re-train the machine learning algorithm. This feedback mechanism ensures the algorithm learns from actual local conditions and user observations, improving identification reliability while maintaining automated training efficiency

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent implements preliminary action by pre-processing sensor data before it reaches the user interface. The machine learning unit performs preliminary processing on sensor data (such as image processing, feature extraction, or data formatting) to prepare it for user review and labeling, making the feedback process more efficient and scalable

Inventive Principle:
Principle #10Preliminary action

3Reliability

If sensor data collection is triggered automatically based on location or time, then the system reduces human error and increases consistency, but the system cannot adapt to user-specific conditions

Engineering Contradiction:
Improvedata collection consistencyVSAvoiduser-specific adaptation
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by designing a multi-functional data collection trigger system that can operate in multiple modes: automatic triggering based on location information (such as frequently visited locations or areas with identification uncertainty), time-based scheduling, user input triggering, or combinations thereof. This allows the system to maintain consistent automated operation while adapting to user-specific conditions through configurable trigger parameters

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

Data Source

PatentEP4231196A1Methods and apparatus for training a machine learning algorithm
Publication Date: 2023.08.23 VOLVO CAR CORP

AI summary

The invention relates to a method for training a machine learning algorithm, such as a deep learning algorithm, for correct item identification performed by a computing unit of a vehicle, wherein the computing unit comprises a processor, a machine learning unit, a communication unit, and a memory and wherein the vehicle comprises a sensor. The method comprises the steps of: triggering sensor data collection and controlling the sensor to acquire sensor data, controlling the sensor to transmit the sensor data to the machine learning unit, optionally controlling the machine learning unit to transmit the data and information related to the data to a user interface and controlling the user interface unit to display, on the display of the user interface, the data and instructions to label the data, based on the information related to the data; assigning labelling information to the data; and controlling the machine learning unit to re-train a machine learning algorithm stored in the memory based on the labelling information.