Acoustic Anomaly Recognition via Normal State Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing acoustic anomaly detection systems face challenges in recognizing unknown sound anomalies, are sensitive to changes in acoustic conditions, and raise data protection concerns due to potential voice signal recognition capabilities.
Innovation Solution
A method that involves obtaining long-term recordings to learn the 'acoustic normal state' independently, analyzing audio segments to generate characteristic vectors, and matching these with vectors from new recordings to identify anomalies, without requiring annotated training data, thus enhancing robustness and privacy through a machine learning-based approach.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with annotated training data is used to train classification models, then recognition precision for known anomalies is improved, but adaptability to unknown anomalies and changing acoustic conditions deteriorates
Solution Approach 1:
Instead of training the model to recognize specific anomaly types using supervised learning, the patent inverts the approach by training the model to recognize normal acoustic patterns. The anomaly detection then occurs by identifying deviations from these learned normal patterns, allowing the system to detect unknown anomalies without requiring annotated training data for specific anomaly types.
Solution Approach 2:
The system performs self-adaptation to changing acoustic conditions by continuously learning the normal acoustic environment without requiring external retraining or manual annotation. The anomaly detection mechanism automatically adjusts to new acoustic conditions by comparing against the learned normal patterns, enabling the system to serve itself in adapting to environmental changes.
2Adaptability or versatility
If domain adaptation is performed to improve robustness to changing acoustic conditions, then adaptability is improved, but logistical complexity and costs increase due to requirement of recording and annotating audio data at application locations
Solution Approach 1:
The patent extracts the essential function of domain adaptation by removing the complex logistical requirements of data collection and annotation at application locations. Instead, it uses a general acoustic pattern recognition model that can be deployed without location-specific training data, separating the core anomaly detection capability from the burdensome data collection process.
Solution Approach 2:
The system employs a universal acoustic pattern recognition approach that functions across diverse acoustic environments without requiring environment-specific customization. The single model can detect anomalies in various settings (industrial, environmental, security) without needing to be retrained or adapted to each specific location, achieving multi-functionality across different application domains.
3Difficulty of detecting and measuring
If classification methods are used to detect acoustic anomalies, then anomaly detection capability is improved, but data protection concerns arise due to potential voice signal recognition
Solution Approach 1:
The patent applies local quality by focusing the acoustic analysis on specific frequency ranges and temporal patterns characteristic of anomaly sounds rather than analyzing the complete audio spectrum. This selective analysis of particular acoustic features enables effective anomaly detection while excluding voice signal components, thereby reducing data protection risks.
Solution Approach 2:
The system converts the potential harm of voice recognition capability into a benefit by deliberately designing the acoustic analysis to focus on non-voice characteristics. The anomaly detection mechanism uses acoustic features that are distinct from human speech patterns, thereby preventing unintended voice recognition while maintaining effective anomaly detection capability.
Data Source
AI summary
A method for detecting anomalies has the following steps:Obtaining a long-term recording having a plurality of first audio segments associated to respective first time windows; analyzing the plurality of the first audio segments to obtain, for each of the plurality of the first audio segments, a first characteristic vector describing the respective first audio segment; obtaining a further recording having one or more second audio segments associated to respective second time windows; analyzing the one or more second audio segments to obtain one or more characteristic vectors describing the one or more second audio segments ABCD; matching the one or more second characteristic vectors with the plurality of the first characteristic vectors to recognize at least one anomaly, like a temporal, sound or spatial anomaly.