Acoustic Feature Extractor Training for Unsupervised Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In unsupervised anomaly detection, it is challenging to design an acoustic feature extractor for anomalous sound detection without training data for anomalous sounds, leading to low detection accuracy, as existing methods rely on manual design or deep training which are difficult to implement effectively.
Innovation Solution
A training apparatus that includes a first acoustic feature extraction unit for normal sounds, a normal sound model updating unit, a second acoustic feature extraction unit for both normal and simulated anomalous sounds, and an acoustic feature extractor updating unit, which repeatedly processes to optimize the acoustic feature extractor for anomalous sound detection regardless of available training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual design of acoustic feature extractor is used for unsupervised anomaly detection, then the system can operate without anomalous sound training data, but the detection accuracy is low
Solution Approach 1:
The patent applies preliminary action by generating simulated anomalous sound data before the actual detection process. The sound generation unit creates artificial anomalous sounds based on normal sound data, which are then used to train the acoustic feature extractor in advance. This pre-training with simulated data enables the system to achieve high detection accuracy without requiring actual anomalous sound training data, thus resolving the contradiction between detection accuracy and the difficulty of implementing deep training.
2Measurement precision
If deep training is used to automatically design acoustic feature extractor, then detection accuracy improves, but it becomes difficult to implement without anomalous sound training data
Solution Approach 1:
The patent applies copying by creating simulated anomalous sound data that replicates the characteristics of real anomalous sounds. The sound generation unit generates artificial anomalous sound signals that mimic the spectral and temporal features of actual anomalies. These copied/simulated sounds are then used for training the acoustic feature extractor, enabling deep training to work in unsupervised scenarios where real anomalous data is unavailable, thus improving both detection accuracy and adaptability.
3Adaptability or versatility
If general-purpose acoustic features like MFCC are used, then the system can detect various types of anomalous sounds, but the detection accuracy is lower compared to supervised training
Solution Approach 1:
The patent applies dynamics by making the acoustic feature extractor adaptive and learnable rather than static and fixed. The trained acoustic feature extractor dynamically adjusts its parameters based on the simulated anomalous sound data generated during training. This dynamic adaptation allows the system to optimize feature extraction for specific application scenarios while maintaining the ability to detect various types of anomalies, thus achieving both high detection accuracy and general-purpose capability.
Data Source
AI summary
An anomalous sound detection training apparatus includes: a first acoustic feature extraction unit that extracts an acoustic feature of normal sound based on training data for normal sound by using an acoustic feature extractor; a normal sound model updating unit that updates a normal sound model by using the acoustic feature extracted; a second acoustic feature extraction unit that extracts an acoustic feature of anomalous sound based on simulated anomalous sound and extracts the acoustic feature of normal sound based on the training data for normal sound by using the acoustic feature extractor; and an acoustic feature extractor updating unit that updates the acoustic feature extractor by using the acoustic feature of anomalous sound and the acoustic feature of normal sound that have been extracted, in which processing by the units is repeatedly performed.