Anomalous Sound Detection Using Multi-Head Embeddings Under Domain Shift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomalous sound detection systems fail when presented with domain shift due to changes in acoustic characteristics, such as different background noise or operating voltages, as they struggle to distinguish between unexpected signal changes caused by anomalous sounds and expected signal changes.
Innovation Solution
A multi-head neural network is used to extract embedding vectors for different attribute types, where one type is independent of the recording environment and the other is dependent on it, allowing for separate classification and improved resilience to domain shift by comparing these vectors with pre-learned normal vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional unsupervised anomalous sound detection methods are used, then the system can detect specific types of anomalies such as abrupt transient disturbances, but the system fails when presented with domain shift due to changes in acoustic characteristics like background noise or operating voltages
Solution Approach 1:
The patent segments the acoustic signal analysis into multiple independent attribute dimensions (e.g., temporal characteristics, spectral characteristics, modulation characteristics). Each dimension is analyzed separately by dedicated neural network branches, allowing the system to identify which specific attribute exhibits anomalous behavior while remaining robust to domain shifts in other attributes. This segmentation enables the system to distinguish between expected domain-specific variations and true anomalies.
Solution Approach 2:
The patent transforms the anomaly detection problem from a single-dimensional approach into a multi-dimensional attribute space. By representing sounds in terms of multiple independent attributes (temporal, spectral, modulation) rather than raw waveforms alone, the system creates additional analytical dimensions that help disentanglement domain-specific variations from true anomalies, improving both reliability and adaptability.
2Measurement precision
If the system uses algorithms that analyze acoustic characteristics changes, then it can detect unexpected signal changes caused by anomalous sounds, but it cannot distinguish between unexpected changes caused by anomalies and expected changes caused by domain shift
Solution Approach 1:
The patent segments the detection task into multiple attribute-specific analysis branches, each monitoring a distinct acoustic characteristic (temporal, spectral, modulation). By analyzing each attribute independently and comparing its anomaly score against thresholds, the system can precisely identify which specific attribute exhibits unexpected changes, thereby distinguishing true anomalies from domain shifts that affect different attributes differently.
Solution Approach 2:
The patent applies local quality by making each neural network branch specialized for a specific attribute type (temporal, spectral, or modulation characteristics). Each branch develops expertise in detecting anomalies within its specific domain, allowing the system to apply appropriate detection sensitivity and thresholds tailored to the characteristics of each attribute, thereby improving measurement precision while reducing false positives from domain shifts.
Data Source
AI summary
A system and a method for detecting anomalous sound are disclosed. The method includes receiving an audio signal from a sound source in a recording environment. The sound source and the recording environment are characterized by a set of attributes including a first attribute pertaining to a first attribute type and a second attribute pertaining to a second attribute type. A multi-head neural network is trained to extract from the received audio signal a first embedding vector indicative of the first attribute type and a second embedding vector indicative of the second attribute type. The first embedding vector is compared with a first set of embedding vectors to classify attributes of the first attribute type and the second embedding vector is compared with a second set of embedding vectors to classify attributes of the second attribute type, to determine a result of anomaly detection.