Acoustic Feature Extractor Training for Unsupervised Anomaly Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Unsupervised anomaly detection in industrial machines faces challenges in designing acoustic feature extractors without training data for anomalous sounds, leading to low detection accuracy, especially when dealing with diverse anomalous sound characteristics.

Innovation Solution

A method that simultaneously optimizes the normal sound model and acoustic feature extractor using a hierarchical structure, where the acoustic feature extractor is represented as a matrix product or multilayer perceptron, and updates are performed through repeated processing involving acoustic feature extraction units and model updating units, allowing for detection irrespective of available training data for anomalous sounds.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If unsupervised anomalous sound detection is used to avoid collecting anomalous sound training data, then personnel and traveling costs are reduced, but detection accuracy deteriorates due to difficulty in designing acoustic feature extractors

Engineering Contradiction:
Improvecost reductionVSAvoiddetection accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system performs preliminary action by collecting normal sound training data in advance and using it to train a normal sound model before actual anomaly detection. This preliminary training enables the system to establish a baseline of normal operations, which is then used to detect deviations during runtime without requiring pre-collected anomaly data.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary mechanism - a normal sound model trained from normal sound data - that mediates between the available training data and the detection task. This intermediary model serves as a reference framework that enables anomaly detection without direct exposure to anomaly examples, bridging the gap between limited training data and comprehensive detection capability.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If manual design of acoustic feature extractors is performed for supervised detection, then detection accuracy improves for specific sound types, but device complexity and setup costs increase for every machine type

Engineering Contradiction:
Improvedetection accuracyVSAvoidsetup complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements universality by creating a general-purpose anomaly detection framework that can be applied across different machine types without requiring custom feature extractor design for each case. The normal sound model and detection algorithm serve multiple functions across various industrial equipment, eliminating the need for specialized configurations while maintaining effective detection capability.

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

Solution Approach 2:

The system enables self-service by automatically training the normal sound model from collected normal sound data without requiring manual feature extractor design or expert intervention. The automated training process adapts to specific machine characteristics independently, reducing setup complexity and enabling deployment by non-experts while maintaining high detection accuracy.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4113076B1Anomalous sound detection training apparatus, and methods and program for the same
Publication Date: 2024.07.24 NIPPON TELEGRAPH & TELEPHONE CORP
  • EP4113076B1 patent drawingFigure 1
  • EP4113076B1 patent drawingFigure 2
  • EP4113076B1 patent drawingFigure 3

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.