Audio Event Classifier Sample Selection for Low-Label Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Obtaining labeled audio data for audio event detection is slow and expensive due to manual annotation, and applications with changing classes of interest or rare events require inefficient relabeling, especially in real-world scenarios.

Innovation Solution

An active machine learning method using a deep reinforcement learning system with a reinforcement learning agent to select informative samples for annotation, combined with a few-shot adaptation module for adapting to new classes, reducing the need for extensive retraining.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual annotation by human experts is used to obtain labeled audio data, then the quality and accuracy of labeled data is improved, but the time consumption and cost increase significantly

Engineering Contradiction:
Improvelabeled data qualityVSAvoidannotation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system enables self-service by using the trained audio event classifier to automatically pre-label unlabeled audio samples. The reinforcement learning agent then selects which samples need human annotation based on uncertainty metrics, allowing the system to serve itself by identifying its own knowledge gaps rather than requiring manual annotation of all samples.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system implements feedback loops where the classifier's performance on selected samples provides feedback to the reinforcement learning agent. The agent uses this feedback to adjust its sampling strategy, continuously improving which samples are selected for annotation to maximize learning efficiency while minimizing annotation costs.

Inventive Principle:
Principle #23Feedback

2Adaptability or versatility

If manual annotation is performed continuously to accommodate changing classes of interest, then the model remains up-to-date with current requirements, but the cost becomes prohibitively expensive

Engineering Contradiction:
Improvemodel adaptability to changing classesVSAvoidannotation cost
Core Design Contradiction:
Adaptability or versatilityVSLoss of energy

Solution Approach 1:

The system dynamically adapts to changing classes by using the reinforcement learning agent to identify which new classes or modified classes require annotation based on current performance metrics. Rather than continuously annotating all samples, the system dynamically adjusts the annotation strategy to focus only on samples relevant to current class changes, making the adaptation process cost-effective.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system performs preliminary classification on unlabeled samples before human annotation. The reinforcement learning agent pre-selects samples that are most likely to be relevant to changing classes based on uncertainty and diversity metrics, so that human annotators only need to label a small subset of pre-screened samples rather than all samples.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If extensive labeled data is collected to train the audio event classifier, then the detection accuracy is improved, but the efficiency of finding and labeling rare events decreases

Engineering Contradiction:
Improvedetection accuracyVSAvoidlabeling efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system applies local quality by focusing annotation efforts on specific regions of the data space where rare events are most likely to occur. The reinforcement learning agent identifies samples with high uncertainty or those that differ significantly from the majority class distribution, concentrating annotation resources on local areas of the feature space where rare events are expected rather than uniformly annotating all samples.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system uses partial action by selecting only the most informative samples for annotation rather than annotating all samples. The reinforcement learning agent determines the optimal subset of samples that will provide the maximum improvement in detection accuracy for rare events, performing annotation on just enough samples to achieve the desired productivity-accuracy balance.

Inventive Principle:
Principle #16Partial or excessive action

4Productivity

If a reinforcement learning agent selects samples for annotation based on environment states, then the sample selection efficiency is improved, but the system complexity increases

Engineering Contradiction:
Improvesample selection efficiencyVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system achieves universality by using the reinforcement learning agent to perform multiple functions: selecting samples for annotation, evaluating classifier performance, and guiding the active learning process. The same agent that selects samples also learns from the results and adjusts its strategy, consolidating multiple functions into a single component to manage complexity while maintaining high selection efficiency.

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

Data Source

PatentEP4679425A1Deep reinforcement active machine learning system for audio event detection and classification
Publication Date: 2026.01.14 ROBERT BOSCH GMBH
  • EP4679425A1 patent drawingFigure 1
  • EP4679425A1 patent drawingFigure 2
  • EP4679425A1 patent drawingFigure 3

AI summary

Active machine learning systems for anomalous event detection and classification. Initial samples from an industrial environment may be received and labeled. Initially, a training pool of audio samples may be labeled. These labeled samples may be used to train an audio event classifier to detect and categorize sounds. Environment states may be calculated using outputs from the classifier. A batch of audio samples may then selected from an unlabeled pool for annotation, guided by a reinforcement learning agent. These selected samples may be annotated and added to the labeled training pool. The classifier may be retrained with this updated pool. Rewards may be calculated for each of the annotated samples based on their annotations. The environment states may be updated using the retrained classifier, and the exploration-exploitation parameter of the reinforcement learning agent may be adjusted. The reinforcement learning agent may be retrained using the updated environment states and rewards.